1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGObjCRuntime.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/ASTContext.h" 19 #include "clang/AST/Decl.h" 20 #include "clang/Basic/TargetBuiltins.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/CodeGen/CGFunctionInfo.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/IR/Intrinsics.h" 25 26 using namespace clang; 27 using namespace CodeGen; 28 using namespace llvm; 29 30 /// getBuiltinLibFunction - Given a builtin id for a function like 31 /// "__builtin_fabsf", return a Function* for "fabsf". 32 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 33 unsigned BuiltinID) { 34 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 35 36 // Get the name, skip over the __builtin_ prefix (if necessary). 37 StringRef Name; 38 GlobalDecl D(FD); 39 40 // If the builtin has been declared explicitly with an assembler label, 41 // use the mangled name. This differs from the plain label on platforms 42 // that prefix labels. 43 if (FD->hasAttr<AsmLabelAttr>()) 44 Name = getMangledName(D); 45 else 46 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 47 48 llvm::FunctionType *Ty = 49 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 50 51 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 52 } 53 54 /// Emit the conversions required to turn the given value into an 55 /// integer of the given size. 56 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 57 QualType T, llvm::IntegerType *IntType) { 58 V = CGF.EmitToMemory(V, T); 59 60 if (V->getType()->isPointerTy()) 61 return CGF.Builder.CreatePtrToInt(V, IntType); 62 63 assert(V->getType() == IntType); 64 return V; 65 } 66 67 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 68 QualType T, llvm::Type *ResultType) { 69 V = CGF.EmitFromMemory(V, T); 70 71 if (ResultType->isPointerTy()) 72 return CGF.Builder.CreateIntToPtr(V, ResultType); 73 74 assert(V->getType() == ResultType); 75 return V; 76 } 77 78 /// Utility to insert an atomic instruction based on Instrinsic::ID 79 /// and the expression node. 80 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 81 llvm::AtomicRMWInst::BinOp Kind, 82 const CallExpr *E) { 83 QualType T = E->getType(); 84 assert(E->getArg(0)->getType()->isPointerType()); 85 assert(CGF.getContext().hasSameUnqualifiedType(T, 86 E->getArg(0)->getType()->getPointeeType())); 87 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 88 89 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 90 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 91 92 llvm::IntegerType *IntType = 93 llvm::IntegerType::get(CGF.getLLVMContext(), 94 CGF.getContext().getTypeSize(T)); 95 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 96 97 llvm::Value *Args[2]; 98 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 99 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 100 llvm::Type *ValueType = Args[1]->getType(); 101 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 102 103 llvm::Value *Result = 104 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 105 llvm::SequentiallyConsistent); 106 Result = EmitFromInt(CGF, Result, T, ValueType); 107 return RValue::get(Result); 108 } 109 110 /// Utility to insert an atomic instruction based Instrinsic::ID and 111 /// the expression node, where the return value is the result of the 112 /// operation. 113 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 114 llvm::AtomicRMWInst::BinOp Kind, 115 const CallExpr *E, 116 Instruction::BinaryOps Op) { 117 QualType T = E->getType(); 118 assert(E->getArg(0)->getType()->isPointerType()); 119 assert(CGF.getContext().hasSameUnqualifiedType(T, 120 E->getArg(0)->getType()->getPointeeType())); 121 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 122 123 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 124 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 125 126 llvm::IntegerType *IntType = 127 llvm::IntegerType::get(CGF.getLLVMContext(), 128 CGF.getContext().getTypeSize(T)); 129 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 130 131 llvm::Value *Args[2]; 132 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 133 llvm::Type *ValueType = Args[1]->getType(); 134 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 135 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 136 137 llvm::Value *Result = 138 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 139 llvm::SequentiallyConsistent); 140 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 141 Result = EmitFromInt(CGF, Result, T, ValueType); 142 return RValue::get(Result); 143 } 144 145 /// EmitFAbs - Emit a call to @llvm.fabs(). 146 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 147 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 148 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 149 Call->setDoesNotAccessMemory(); 150 return Call; 151 } 152 153 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 154 const CallExpr *E, llvm::Value *calleeValue) { 155 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 156 ReturnValueSlot(), Fn); 157 } 158 159 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 160 /// depending on IntrinsicID. 161 /// 162 /// \arg CGF The current codegen function. 163 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 164 /// \arg X The first argument to the llvm.*.with.overflow.*. 165 /// \arg Y The second argument to the llvm.*.with.overflow.*. 166 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 167 /// \returns The result (i.e. sum/product) returned by the intrinsic. 168 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 169 const llvm::Intrinsic::ID IntrinsicID, 170 llvm::Value *X, llvm::Value *Y, 171 llvm::Value *&Carry) { 172 // Make sure we have integers of the same width. 173 assert(X->getType() == Y->getType() && 174 "Arguments must be the same type. (Did you forget to make sure both " 175 "arguments have the same integer width?)"); 176 177 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 178 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 179 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 180 return CGF.Builder.CreateExtractValue(Tmp, 0); 181 } 182 183 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 184 unsigned BuiltinID, const CallExpr *E) { 185 // See if we can constant fold this builtin. If so, don't emit it at all. 186 Expr::EvalResult Result; 187 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 188 !Result.hasSideEffects()) { 189 if (Result.Val.isInt()) 190 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 191 Result.Val.getInt())); 192 if (Result.Val.isFloat()) 193 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 194 Result.Val.getFloat())); 195 } 196 197 switch (BuiltinID) { 198 default: break; // Handle intrinsics and libm functions below. 199 case Builtin::BI__builtin___CFStringMakeConstantString: 200 case Builtin::BI__builtin___NSStringMakeConstantString: 201 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 202 case Builtin::BI__builtin_stdarg_start: 203 case Builtin::BI__builtin_va_start: 204 case Builtin::BI__va_start: 205 case Builtin::BI__builtin_va_end: { 206 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 207 ? EmitScalarExpr(E->getArg(0)) 208 : EmitVAListRef(E->getArg(0)); 209 llvm::Type *DestType = Int8PtrTy; 210 if (ArgValue->getType() != DestType) 211 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 212 ArgValue->getName().data()); 213 214 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 215 Intrinsic::vaend : Intrinsic::vastart; 216 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 217 } 218 case Builtin::BI__builtin_va_copy: { 219 Value *DstPtr = EmitVAListRef(E->getArg(0)); 220 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 221 222 llvm::Type *Type = Int8PtrTy; 223 224 DstPtr = Builder.CreateBitCast(DstPtr, Type); 225 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 226 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 227 DstPtr, SrcPtr)); 228 } 229 case Builtin::BI__builtin_abs: 230 case Builtin::BI__builtin_labs: 231 case Builtin::BI__builtin_llabs: { 232 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 233 234 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 235 Value *CmpResult = 236 Builder.CreateICmpSGE(ArgValue, 237 llvm::Constant::getNullValue(ArgValue->getType()), 238 "abscond"); 239 Value *Result = 240 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 241 242 return RValue::get(Result); 243 } 244 case Builtin::BI__builtin_fmod: 245 case Builtin::BI__builtin_fmodf: 246 case Builtin::BI__builtin_fmodl: { 247 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 248 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 249 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 250 return RValue::get(Result); 251 } 252 253 case Builtin::BI__builtin_conj: 254 case Builtin::BI__builtin_conjf: 255 case Builtin::BI__builtin_conjl: { 256 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 257 Value *Real = ComplexVal.first; 258 Value *Imag = ComplexVal.second; 259 Value *Zero = 260 Imag->getType()->isFPOrFPVectorTy() 261 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 262 : llvm::Constant::getNullValue(Imag->getType()); 263 264 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 265 return RValue::getComplex(std::make_pair(Real, Imag)); 266 } 267 case Builtin::BI__builtin_creal: 268 case Builtin::BI__builtin_crealf: 269 case Builtin::BI__builtin_creall: 270 case Builtin::BIcreal: 271 case Builtin::BIcrealf: 272 case Builtin::BIcreall: { 273 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 274 return RValue::get(ComplexVal.first); 275 } 276 277 case Builtin::BI__builtin_cimag: 278 case Builtin::BI__builtin_cimagf: 279 case Builtin::BI__builtin_cimagl: 280 case Builtin::BIcimag: 281 case Builtin::BIcimagf: 282 case Builtin::BIcimagl: { 283 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 284 return RValue::get(ComplexVal.second); 285 } 286 287 case Builtin::BI__builtin_ctzs: 288 case Builtin::BI__builtin_ctz: 289 case Builtin::BI__builtin_ctzl: 290 case Builtin::BI__builtin_ctzll: { 291 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 292 293 llvm::Type *ArgType = ArgValue->getType(); 294 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 295 296 llvm::Type *ResultType = ConvertType(E->getType()); 297 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 298 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 299 if (Result->getType() != ResultType) 300 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 301 "cast"); 302 return RValue::get(Result); 303 } 304 case Builtin::BI__builtin_clzs: 305 case Builtin::BI__builtin_clz: 306 case Builtin::BI__builtin_clzl: 307 case Builtin::BI__builtin_clzll: { 308 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 309 310 llvm::Type *ArgType = ArgValue->getType(); 311 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 312 313 llvm::Type *ResultType = ConvertType(E->getType()); 314 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 315 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 316 if (Result->getType() != ResultType) 317 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 318 "cast"); 319 return RValue::get(Result); 320 } 321 case Builtin::BI__builtin_ffs: 322 case Builtin::BI__builtin_ffsl: 323 case Builtin::BI__builtin_ffsll: { 324 // ffs(x) -> x ? cttz(x) + 1 : 0 325 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 326 327 llvm::Type *ArgType = ArgValue->getType(); 328 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 329 330 llvm::Type *ResultType = ConvertType(E->getType()); 331 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 332 Builder.getTrue()), 333 llvm::ConstantInt::get(ArgType, 1)); 334 Value *Zero = llvm::Constant::getNullValue(ArgType); 335 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 336 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 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_parity: 343 case Builtin::BI__builtin_parityl: 344 case Builtin::BI__builtin_parityll: { 345 // parity(x) -> ctpop(x) & 1 346 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 347 348 llvm::Type *ArgType = ArgValue->getType(); 349 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 350 351 llvm::Type *ResultType = ConvertType(E->getType()); 352 Value *Tmp = Builder.CreateCall(F, ArgValue); 353 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 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_popcount: 360 case Builtin::BI__builtin_popcountl: 361 case Builtin::BI__builtin_popcountll: { 362 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 363 364 llvm::Type *ArgType = ArgValue->getType(); 365 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 366 367 llvm::Type *ResultType = ConvertType(E->getType()); 368 Value *Result = Builder.CreateCall(F, ArgValue); 369 if (Result->getType() != ResultType) 370 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 371 "cast"); 372 return RValue::get(Result); 373 } 374 case Builtin::BI__builtin_expect: { 375 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 376 llvm::Type *ArgType = ArgValue->getType(); 377 378 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 379 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 380 381 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 382 "expval"); 383 return RValue::get(Result); 384 } 385 case Builtin::BI__builtin_assume_aligned: { 386 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 387 Value *OffsetValue = 388 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 389 390 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 391 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 392 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 393 394 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 395 return RValue::get(PtrValue); 396 } 397 case Builtin::BI__assume: 398 case Builtin::BI__builtin_assume: { 399 if (E->getArg(0)->HasSideEffects(getContext())) 400 return RValue::get(nullptr); 401 402 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 403 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 404 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 405 } 406 case Builtin::BI__builtin_bswap16: 407 case Builtin::BI__builtin_bswap32: 408 case Builtin::BI__builtin_bswap64: { 409 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 410 llvm::Type *ArgType = ArgValue->getType(); 411 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 412 return RValue::get(Builder.CreateCall(F, ArgValue)); 413 } 414 case Builtin::BI__builtin_object_size: { 415 // We rely on constant folding to deal with expressions with side effects. 416 assert(!E->getArg(0)->HasSideEffects(getContext()) && 417 "should have been constant folded"); 418 419 // We pass this builtin onto the optimizer so that it can 420 // figure out the object size in more complex cases. 421 llvm::Type *ResType = ConvertType(E->getType()); 422 423 // LLVM only supports 0 and 2, make sure that we pass along that 424 // as a boolean. 425 Value *Ty = EmitScalarExpr(E->getArg(1)); 426 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 427 assert(CI); 428 uint64_t val = CI->getZExtValue(); 429 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 430 // FIXME: Get right address space. 431 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 432 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 433 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 434 } 435 case Builtin::BI__builtin_prefetch: { 436 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 437 // FIXME: Technically these constants should of type 'int', yes? 438 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 439 llvm::ConstantInt::get(Int32Ty, 0); 440 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 441 llvm::ConstantInt::get(Int32Ty, 3); 442 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 443 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 444 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 445 } 446 case Builtin::BI__builtin_readcyclecounter: { 447 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 448 return RValue::get(Builder.CreateCall(F)); 449 } 450 case Builtin::BI__builtin___clear_cache: { 451 Value *Begin = EmitScalarExpr(E->getArg(0)); 452 Value *End = EmitScalarExpr(E->getArg(1)); 453 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 454 return RValue::get(Builder.CreateCall2(F, Begin, End)); 455 } 456 case Builtin::BI__builtin_trap: { 457 Value *F = CGM.getIntrinsic(Intrinsic::trap); 458 return RValue::get(Builder.CreateCall(F)); 459 } 460 case Builtin::BI__debugbreak: { 461 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 462 return RValue::get(Builder.CreateCall(F)); 463 } 464 case Builtin::BI__builtin_unreachable: { 465 if (SanOpts->Unreachable) { 466 SanitizerScope SanScope(this); 467 EmitCheck(Builder.getFalse(), "builtin_unreachable", 468 EmitCheckSourceLocation(E->getExprLoc()), 469 None, CRK_Unrecoverable); 470 } else 471 Builder.CreateUnreachable(); 472 473 // We do need to preserve an insertion point. 474 EmitBlock(createBasicBlock("unreachable.cont")); 475 476 return RValue::get(nullptr); 477 } 478 479 case Builtin::BI__builtin_powi: 480 case Builtin::BI__builtin_powif: 481 case Builtin::BI__builtin_powil: { 482 Value *Base = EmitScalarExpr(E->getArg(0)); 483 Value *Exponent = EmitScalarExpr(E->getArg(1)); 484 llvm::Type *ArgType = Base->getType(); 485 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 486 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 487 } 488 489 case Builtin::BI__builtin_isgreater: 490 case Builtin::BI__builtin_isgreaterequal: 491 case Builtin::BI__builtin_isless: 492 case Builtin::BI__builtin_islessequal: 493 case Builtin::BI__builtin_islessgreater: 494 case Builtin::BI__builtin_isunordered: { 495 // Ordered comparisons: we know the arguments to these are matching scalar 496 // floating point values. 497 Value *LHS = EmitScalarExpr(E->getArg(0)); 498 Value *RHS = EmitScalarExpr(E->getArg(1)); 499 500 switch (BuiltinID) { 501 default: llvm_unreachable("Unknown ordered comparison"); 502 case Builtin::BI__builtin_isgreater: 503 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 504 break; 505 case Builtin::BI__builtin_isgreaterequal: 506 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 507 break; 508 case Builtin::BI__builtin_isless: 509 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 510 break; 511 case Builtin::BI__builtin_islessequal: 512 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 513 break; 514 case Builtin::BI__builtin_islessgreater: 515 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 516 break; 517 case Builtin::BI__builtin_isunordered: 518 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 519 break; 520 } 521 // ZExt bool to int type. 522 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 523 } 524 case Builtin::BI__builtin_isnan: { 525 Value *V = EmitScalarExpr(E->getArg(0)); 526 V = Builder.CreateFCmpUNO(V, V, "cmp"); 527 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 528 } 529 530 case Builtin::BI__builtin_isinf: { 531 // isinf(x) --> fabs(x) == infinity 532 Value *V = EmitScalarExpr(E->getArg(0)); 533 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 534 535 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 536 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 537 } 538 539 // TODO: BI__builtin_isinf_sign 540 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 541 542 case Builtin::BI__builtin_isnormal: { 543 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 544 Value *V = EmitScalarExpr(E->getArg(0)); 545 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 546 547 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 548 Value *IsLessThanInf = 549 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 550 APFloat Smallest = APFloat::getSmallestNormalized( 551 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 552 Value *IsNormal = 553 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 554 "isnormal"); 555 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 556 V = Builder.CreateAnd(V, IsNormal, "and"); 557 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 558 } 559 560 case Builtin::BI__builtin_isfinite: { 561 // isfinite(x) --> x == x && fabs(x) != infinity; 562 Value *V = EmitScalarExpr(E->getArg(0)); 563 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 564 565 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 566 Value *IsNotInf = 567 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 568 569 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 570 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 571 } 572 573 case Builtin::BI__builtin_fpclassify: { 574 Value *V = EmitScalarExpr(E->getArg(5)); 575 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 576 577 // Create Result 578 BasicBlock *Begin = Builder.GetInsertBlock(); 579 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 580 Builder.SetInsertPoint(End); 581 PHINode *Result = 582 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 583 "fpclassify_result"); 584 585 // if (V==0) return FP_ZERO 586 Builder.SetInsertPoint(Begin); 587 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 588 "iszero"); 589 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 590 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 591 Builder.CreateCondBr(IsZero, End, NotZero); 592 Result->addIncoming(ZeroLiteral, Begin); 593 594 // if (V != V) return FP_NAN 595 Builder.SetInsertPoint(NotZero); 596 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 597 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 598 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 599 Builder.CreateCondBr(IsNan, End, NotNan); 600 Result->addIncoming(NanLiteral, NotZero); 601 602 // if (fabs(V) == infinity) return FP_INFINITY 603 Builder.SetInsertPoint(NotNan); 604 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 605 Value *IsInf = 606 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 607 "isinf"); 608 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 609 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 610 Builder.CreateCondBr(IsInf, End, NotInf); 611 Result->addIncoming(InfLiteral, NotNan); 612 613 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 614 Builder.SetInsertPoint(NotInf); 615 APFloat Smallest = APFloat::getSmallestNormalized( 616 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 617 Value *IsNormal = 618 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 619 "isnormal"); 620 Value *NormalResult = 621 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 622 EmitScalarExpr(E->getArg(3))); 623 Builder.CreateBr(End); 624 Result->addIncoming(NormalResult, NotInf); 625 626 // return Result 627 Builder.SetInsertPoint(End); 628 return RValue::get(Result); 629 } 630 631 case Builtin::BIalloca: 632 case Builtin::BI_alloca: 633 case Builtin::BI__builtin_alloca: { 634 Value *Size = EmitScalarExpr(E->getArg(0)); 635 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 636 } 637 case Builtin::BIbzero: 638 case Builtin::BI__builtin_bzero: { 639 std::pair<llvm::Value*, unsigned> Dest = 640 EmitPointerWithAlignment(E->getArg(0)); 641 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 642 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 643 Dest.second, false); 644 return RValue::get(Dest.first); 645 } 646 case Builtin::BImemcpy: 647 case Builtin::BI__builtin_memcpy: { 648 std::pair<llvm::Value*, unsigned> Dest = 649 EmitPointerWithAlignment(E->getArg(0)); 650 std::pair<llvm::Value*, unsigned> Src = 651 EmitPointerWithAlignment(E->getArg(1)); 652 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 653 unsigned Align = std::min(Dest.second, Src.second); 654 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 655 return RValue::get(Dest.first); 656 } 657 658 case Builtin::BI__builtin___memcpy_chk: { 659 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 660 llvm::APSInt Size, DstSize; 661 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 662 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 663 break; 664 if (Size.ugt(DstSize)) 665 break; 666 std::pair<llvm::Value*, unsigned> Dest = 667 EmitPointerWithAlignment(E->getArg(0)); 668 std::pair<llvm::Value*, unsigned> Src = 669 EmitPointerWithAlignment(E->getArg(1)); 670 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 671 unsigned Align = std::min(Dest.second, Src.second); 672 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 673 return RValue::get(Dest.first); 674 } 675 676 case Builtin::BI__builtin_objc_memmove_collectable: { 677 Value *Address = EmitScalarExpr(E->getArg(0)); 678 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 679 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 680 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 681 Address, SrcAddr, SizeVal); 682 return RValue::get(Address); 683 } 684 685 case Builtin::BI__builtin___memmove_chk: { 686 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 687 llvm::APSInt Size, DstSize; 688 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 689 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 690 break; 691 if (Size.ugt(DstSize)) 692 break; 693 std::pair<llvm::Value*, unsigned> Dest = 694 EmitPointerWithAlignment(E->getArg(0)); 695 std::pair<llvm::Value*, unsigned> Src = 696 EmitPointerWithAlignment(E->getArg(1)); 697 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 698 unsigned Align = std::min(Dest.second, Src.second); 699 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 700 return RValue::get(Dest.first); 701 } 702 703 case Builtin::BImemmove: 704 case Builtin::BI__builtin_memmove: { 705 std::pair<llvm::Value*, unsigned> Dest = 706 EmitPointerWithAlignment(E->getArg(0)); 707 std::pair<llvm::Value*, unsigned> Src = 708 EmitPointerWithAlignment(E->getArg(1)); 709 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 710 unsigned Align = std::min(Dest.second, Src.second); 711 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 712 return RValue::get(Dest.first); 713 } 714 case Builtin::BImemset: 715 case Builtin::BI__builtin_memset: { 716 std::pair<llvm::Value*, unsigned> Dest = 717 EmitPointerWithAlignment(E->getArg(0)); 718 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 719 Builder.getInt8Ty()); 720 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 721 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 722 return RValue::get(Dest.first); 723 } 724 case Builtin::BI__builtin___memset_chk: { 725 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 726 llvm::APSInt Size, DstSize; 727 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 728 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 729 break; 730 if (Size.ugt(DstSize)) 731 break; 732 std::pair<llvm::Value*, unsigned> Dest = 733 EmitPointerWithAlignment(E->getArg(0)); 734 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 735 Builder.getInt8Ty()); 736 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 737 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 738 return RValue::get(Dest.first); 739 } 740 case Builtin::BI__builtin_dwarf_cfa: { 741 // The offset in bytes from the first argument to the CFA. 742 // 743 // Why on earth is this in the frontend? Is there any reason at 744 // all that the backend can't reasonably determine this while 745 // lowering llvm.eh.dwarf.cfa()? 746 // 747 // TODO: If there's a satisfactory reason, add a target hook for 748 // this instead of hard-coding 0, which is correct for most targets. 749 int32_t Offset = 0; 750 751 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 752 return RValue::get(Builder.CreateCall(F, 753 llvm::ConstantInt::get(Int32Ty, Offset))); 754 } 755 case Builtin::BI__builtin_return_address: { 756 Value *Depth = EmitScalarExpr(E->getArg(0)); 757 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 758 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 759 return RValue::get(Builder.CreateCall(F, Depth)); 760 } 761 case Builtin::BI__builtin_frame_address: { 762 Value *Depth = EmitScalarExpr(E->getArg(0)); 763 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 764 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 765 return RValue::get(Builder.CreateCall(F, Depth)); 766 } 767 case Builtin::BI__builtin_extract_return_addr: { 768 Value *Address = EmitScalarExpr(E->getArg(0)); 769 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 770 return RValue::get(Result); 771 } 772 case Builtin::BI__builtin_frob_return_addr: { 773 Value *Address = EmitScalarExpr(E->getArg(0)); 774 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 775 return RValue::get(Result); 776 } 777 case Builtin::BI__builtin_dwarf_sp_column: { 778 llvm::IntegerType *Ty 779 = cast<llvm::IntegerType>(ConvertType(E->getType())); 780 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 781 if (Column == -1) { 782 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 783 return RValue::get(llvm::UndefValue::get(Ty)); 784 } 785 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 786 } 787 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 788 Value *Address = EmitScalarExpr(E->getArg(0)); 789 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 790 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 791 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 792 } 793 case Builtin::BI__builtin_eh_return: { 794 Value *Int = EmitScalarExpr(E->getArg(0)); 795 Value *Ptr = EmitScalarExpr(E->getArg(1)); 796 797 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 798 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 799 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 800 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 801 ? Intrinsic::eh_return_i32 802 : Intrinsic::eh_return_i64); 803 Builder.CreateCall2(F, Int, Ptr); 804 Builder.CreateUnreachable(); 805 806 // We do need to preserve an insertion point. 807 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 808 809 return RValue::get(nullptr); 810 } 811 case Builtin::BI__builtin_unwind_init: { 812 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 813 return RValue::get(Builder.CreateCall(F)); 814 } 815 case Builtin::BI__builtin_extend_pointer: { 816 // Extends a pointer to the size of an _Unwind_Word, which is 817 // uint64_t on all platforms. Generally this gets poked into a 818 // register and eventually used as an address, so if the 819 // addressing registers are wider than pointers and the platform 820 // doesn't implicitly ignore high-order bits when doing 821 // addressing, we need to make sure we zext / sext based on 822 // the platform's expectations. 823 // 824 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 825 826 // Cast the pointer to intptr_t. 827 Value *Ptr = EmitScalarExpr(E->getArg(0)); 828 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 829 830 // If that's 64 bits, we're done. 831 if (IntPtrTy->getBitWidth() == 64) 832 return RValue::get(Result); 833 834 // Otherwise, ask the codegen data what to do. 835 if (getTargetHooks().extendPointerWithSExt()) 836 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 837 else 838 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 839 } 840 case Builtin::BI__builtin_setjmp: { 841 // Buffer is a void**. 842 Value *Buf = EmitScalarExpr(E->getArg(0)); 843 844 // Store the frame pointer to the setjmp buffer. 845 Value *FrameAddr = 846 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 847 ConstantInt::get(Int32Ty, 0)); 848 Builder.CreateStore(FrameAddr, Buf); 849 850 // Store the stack pointer to the setjmp buffer. 851 Value *StackAddr = 852 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 853 Value *StackSaveSlot = 854 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 855 Builder.CreateStore(StackAddr, StackSaveSlot); 856 857 // Call LLVM's EH setjmp, which is lightweight. 858 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 859 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 860 return RValue::get(Builder.CreateCall(F, Buf)); 861 } 862 case Builtin::BI__builtin_longjmp: { 863 Value *Buf = EmitScalarExpr(E->getArg(0)); 864 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 865 866 // Call LLVM's EH longjmp, which is lightweight. 867 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 868 869 // longjmp doesn't return; mark this as unreachable. 870 Builder.CreateUnreachable(); 871 872 // We do need to preserve an insertion point. 873 EmitBlock(createBasicBlock("longjmp.cont")); 874 875 return RValue::get(nullptr); 876 } 877 case Builtin::BI__sync_fetch_and_add: 878 case Builtin::BI__sync_fetch_and_sub: 879 case Builtin::BI__sync_fetch_and_or: 880 case Builtin::BI__sync_fetch_and_and: 881 case Builtin::BI__sync_fetch_and_xor: 882 case Builtin::BI__sync_add_and_fetch: 883 case Builtin::BI__sync_sub_and_fetch: 884 case Builtin::BI__sync_and_and_fetch: 885 case Builtin::BI__sync_or_and_fetch: 886 case Builtin::BI__sync_xor_and_fetch: 887 case Builtin::BI__sync_val_compare_and_swap: 888 case Builtin::BI__sync_bool_compare_and_swap: 889 case Builtin::BI__sync_lock_test_and_set: 890 case Builtin::BI__sync_lock_release: 891 case Builtin::BI__sync_swap: 892 llvm_unreachable("Shouldn't make it through sema"); 893 case Builtin::BI__sync_fetch_and_add_1: 894 case Builtin::BI__sync_fetch_and_add_2: 895 case Builtin::BI__sync_fetch_and_add_4: 896 case Builtin::BI__sync_fetch_and_add_8: 897 case Builtin::BI__sync_fetch_and_add_16: 898 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 899 case Builtin::BI__sync_fetch_and_sub_1: 900 case Builtin::BI__sync_fetch_and_sub_2: 901 case Builtin::BI__sync_fetch_and_sub_4: 902 case Builtin::BI__sync_fetch_and_sub_8: 903 case Builtin::BI__sync_fetch_and_sub_16: 904 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 905 case Builtin::BI__sync_fetch_and_or_1: 906 case Builtin::BI__sync_fetch_and_or_2: 907 case Builtin::BI__sync_fetch_and_or_4: 908 case Builtin::BI__sync_fetch_and_or_8: 909 case Builtin::BI__sync_fetch_and_or_16: 910 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 911 case Builtin::BI__sync_fetch_and_and_1: 912 case Builtin::BI__sync_fetch_and_and_2: 913 case Builtin::BI__sync_fetch_and_and_4: 914 case Builtin::BI__sync_fetch_and_and_8: 915 case Builtin::BI__sync_fetch_and_and_16: 916 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 917 case Builtin::BI__sync_fetch_and_xor_1: 918 case Builtin::BI__sync_fetch_and_xor_2: 919 case Builtin::BI__sync_fetch_and_xor_4: 920 case Builtin::BI__sync_fetch_and_xor_8: 921 case Builtin::BI__sync_fetch_and_xor_16: 922 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 923 924 // Clang extensions: not overloaded yet. 925 case Builtin::BI__sync_fetch_and_min: 926 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 927 case Builtin::BI__sync_fetch_and_max: 928 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 929 case Builtin::BI__sync_fetch_and_umin: 930 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 931 case Builtin::BI__sync_fetch_and_umax: 932 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 933 934 case Builtin::BI__sync_add_and_fetch_1: 935 case Builtin::BI__sync_add_and_fetch_2: 936 case Builtin::BI__sync_add_and_fetch_4: 937 case Builtin::BI__sync_add_and_fetch_8: 938 case Builtin::BI__sync_add_and_fetch_16: 939 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 940 llvm::Instruction::Add); 941 case Builtin::BI__sync_sub_and_fetch_1: 942 case Builtin::BI__sync_sub_and_fetch_2: 943 case Builtin::BI__sync_sub_and_fetch_4: 944 case Builtin::BI__sync_sub_and_fetch_8: 945 case Builtin::BI__sync_sub_and_fetch_16: 946 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 947 llvm::Instruction::Sub); 948 case Builtin::BI__sync_and_and_fetch_1: 949 case Builtin::BI__sync_and_and_fetch_2: 950 case Builtin::BI__sync_and_and_fetch_4: 951 case Builtin::BI__sync_and_and_fetch_8: 952 case Builtin::BI__sync_and_and_fetch_16: 953 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 954 llvm::Instruction::And); 955 case Builtin::BI__sync_or_and_fetch_1: 956 case Builtin::BI__sync_or_and_fetch_2: 957 case Builtin::BI__sync_or_and_fetch_4: 958 case Builtin::BI__sync_or_and_fetch_8: 959 case Builtin::BI__sync_or_and_fetch_16: 960 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 961 llvm::Instruction::Or); 962 case Builtin::BI__sync_xor_and_fetch_1: 963 case Builtin::BI__sync_xor_and_fetch_2: 964 case Builtin::BI__sync_xor_and_fetch_4: 965 case Builtin::BI__sync_xor_and_fetch_8: 966 case Builtin::BI__sync_xor_and_fetch_16: 967 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 968 llvm::Instruction::Xor); 969 970 case Builtin::BI__sync_val_compare_and_swap_1: 971 case Builtin::BI__sync_val_compare_and_swap_2: 972 case Builtin::BI__sync_val_compare_and_swap_4: 973 case Builtin::BI__sync_val_compare_and_swap_8: 974 case Builtin::BI__sync_val_compare_and_swap_16: { 975 QualType T = E->getType(); 976 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 977 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 978 979 llvm::IntegerType *IntType = 980 llvm::IntegerType::get(getLLVMContext(), 981 getContext().getTypeSize(T)); 982 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 983 984 Value *Args[3]; 985 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 986 Args[1] = EmitScalarExpr(E->getArg(1)); 987 llvm::Type *ValueType = Args[1]->getType(); 988 Args[1] = EmitToInt(*this, Args[1], T, IntType); 989 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 990 991 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 992 llvm::SequentiallyConsistent, 993 llvm::SequentiallyConsistent); 994 Result = Builder.CreateExtractValue(Result, 0); 995 Result = EmitFromInt(*this, Result, T, ValueType); 996 return RValue::get(Result); 997 } 998 999 case Builtin::BI__sync_bool_compare_and_swap_1: 1000 case Builtin::BI__sync_bool_compare_and_swap_2: 1001 case Builtin::BI__sync_bool_compare_and_swap_4: 1002 case Builtin::BI__sync_bool_compare_and_swap_8: 1003 case Builtin::BI__sync_bool_compare_and_swap_16: { 1004 QualType T = E->getArg(1)->getType(); 1005 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1006 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1007 1008 llvm::IntegerType *IntType = 1009 llvm::IntegerType::get(getLLVMContext(), 1010 getContext().getTypeSize(T)); 1011 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1012 1013 Value *Args[3]; 1014 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1015 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1016 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1017 1018 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1019 llvm::SequentiallyConsistent, 1020 llvm::SequentiallyConsistent); 1021 Value *Result = Builder.CreateExtractValue(Pair, 1); 1022 // zext bool to int. 1023 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1024 return RValue::get(Result); 1025 } 1026 1027 case Builtin::BI__sync_swap_1: 1028 case Builtin::BI__sync_swap_2: 1029 case Builtin::BI__sync_swap_4: 1030 case Builtin::BI__sync_swap_8: 1031 case Builtin::BI__sync_swap_16: 1032 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1033 1034 case Builtin::BI__sync_lock_test_and_set_1: 1035 case Builtin::BI__sync_lock_test_and_set_2: 1036 case Builtin::BI__sync_lock_test_and_set_4: 1037 case Builtin::BI__sync_lock_test_and_set_8: 1038 case Builtin::BI__sync_lock_test_and_set_16: 1039 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1040 1041 case Builtin::BI__sync_lock_release_1: 1042 case Builtin::BI__sync_lock_release_2: 1043 case Builtin::BI__sync_lock_release_4: 1044 case Builtin::BI__sync_lock_release_8: 1045 case Builtin::BI__sync_lock_release_16: { 1046 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1047 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1048 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1049 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1050 StoreSize.getQuantity() * 8); 1051 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1052 llvm::StoreInst *Store = 1053 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1054 Store->setAlignment(StoreSize.getQuantity()); 1055 Store->setAtomic(llvm::Release); 1056 return RValue::get(nullptr); 1057 } 1058 1059 case Builtin::BI__sync_synchronize: { 1060 // We assume this is supposed to correspond to a C++0x-style 1061 // sequentially-consistent fence (i.e. this is only usable for 1062 // synchonization, not device I/O or anything like that). This intrinsic 1063 // is really badly designed in the sense that in theory, there isn't 1064 // any way to safely use it... but in practice, it mostly works 1065 // to use it with non-atomic loads and stores to get acquire/release 1066 // semantics. 1067 Builder.CreateFence(llvm::SequentiallyConsistent); 1068 return RValue::get(nullptr); 1069 } 1070 1071 case Builtin::BI__c11_atomic_is_lock_free: 1072 case Builtin::BI__atomic_is_lock_free: { 1073 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1074 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1075 // _Atomic(T) is always properly-aligned. 1076 const char *LibCallName = "__atomic_is_lock_free"; 1077 CallArgList Args; 1078 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1079 getContext().getSizeType()); 1080 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1081 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1082 getContext().VoidPtrTy); 1083 else 1084 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1085 getContext().VoidPtrTy); 1086 const CGFunctionInfo &FuncInfo = 1087 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1088 FunctionType::ExtInfo(), 1089 RequiredArgs::All); 1090 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1091 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1092 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1093 } 1094 1095 case Builtin::BI__atomic_test_and_set: { 1096 // Look at the argument type to determine whether this is a volatile 1097 // operation. The parameter type is always volatile. 1098 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1099 bool Volatile = 1100 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1101 1102 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1103 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1104 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1105 Value *NewVal = Builder.getInt8(1); 1106 Value *Order = EmitScalarExpr(E->getArg(1)); 1107 if (isa<llvm::ConstantInt>(Order)) { 1108 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1109 AtomicRMWInst *Result = nullptr; 1110 switch (ord) { 1111 case 0: // memory_order_relaxed 1112 default: // invalid order 1113 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1114 Ptr, NewVal, 1115 llvm::Monotonic); 1116 break; 1117 case 1: // memory_order_consume 1118 case 2: // memory_order_acquire 1119 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1120 Ptr, NewVal, 1121 llvm::Acquire); 1122 break; 1123 case 3: // memory_order_release 1124 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1125 Ptr, NewVal, 1126 llvm::Release); 1127 break; 1128 case 4: // memory_order_acq_rel 1129 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1130 Ptr, NewVal, 1131 llvm::AcquireRelease); 1132 break; 1133 case 5: // memory_order_seq_cst 1134 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1135 Ptr, NewVal, 1136 llvm::SequentiallyConsistent); 1137 break; 1138 } 1139 Result->setVolatile(Volatile); 1140 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1141 } 1142 1143 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1144 1145 llvm::BasicBlock *BBs[5] = { 1146 createBasicBlock("monotonic", CurFn), 1147 createBasicBlock("acquire", CurFn), 1148 createBasicBlock("release", CurFn), 1149 createBasicBlock("acqrel", CurFn), 1150 createBasicBlock("seqcst", CurFn) 1151 }; 1152 llvm::AtomicOrdering Orders[5] = { 1153 llvm::Monotonic, llvm::Acquire, llvm::Release, 1154 llvm::AcquireRelease, llvm::SequentiallyConsistent 1155 }; 1156 1157 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1158 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1159 1160 Builder.SetInsertPoint(ContBB); 1161 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1162 1163 for (unsigned i = 0; i < 5; ++i) { 1164 Builder.SetInsertPoint(BBs[i]); 1165 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1166 Ptr, NewVal, Orders[i]); 1167 RMW->setVolatile(Volatile); 1168 Result->addIncoming(RMW, BBs[i]); 1169 Builder.CreateBr(ContBB); 1170 } 1171 1172 SI->addCase(Builder.getInt32(0), BBs[0]); 1173 SI->addCase(Builder.getInt32(1), BBs[1]); 1174 SI->addCase(Builder.getInt32(2), BBs[1]); 1175 SI->addCase(Builder.getInt32(3), BBs[2]); 1176 SI->addCase(Builder.getInt32(4), BBs[3]); 1177 SI->addCase(Builder.getInt32(5), BBs[4]); 1178 1179 Builder.SetInsertPoint(ContBB); 1180 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1181 } 1182 1183 case Builtin::BI__atomic_clear: { 1184 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1185 bool Volatile = 1186 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1187 1188 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1189 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1190 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1191 Value *NewVal = Builder.getInt8(0); 1192 Value *Order = EmitScalarExpr(E->getArg(1)); 1193 if (isa<llvm::ConstantInt>(Order)) { 1194 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1195 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1196 Store->setAlignment(1); 1197 switch (ord) { 1198 case 0: // memory_order_relaxed 1199 default: // invalid order 1200 Store->setOrdering(llvm::Monotonic); 1201 break; 1202 case 3: // memory_order_release 1203 Store->setOrdering(llvm::Release); 1204 break; 1205 case 5: // memory_order_seq_cst 1206 Store->setOrdering(llvm::SequentiallyConsistent); 1207 break; 1208 } 1209 return RValue::get(nullptr); 1210 } 1211 1212 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1213 1214 llvm::BasicBlock *BBs[3] = { 1215 createBasicBlock("monotonic", CurFn), 1216 createBasicBlock("release", CurFn), 1217 createBasicBlock("seqcst", CurFn) 1218 }; 1219 llvm::AtomicOrdering Orders[3] = { 1220 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1221 }; 1222 1223 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1224 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1225 1226 for (unsigned i = 0; i < 3; ++i) { 1227 Builder.SetInsertPoint(BBs[i]); 1228 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1229 Store->setAlignment(1); 1230 Store->setOrdering(Orders[i]); 1231 Builder.CreateBr(ContBB); 1232 } 1233 1234 SI->addCase(Builder.getInt32(0), BBs[0]); 1235 SI->addCase(Builder.getInt32(3), BBs[1]); 1236 SI->addCase(Builder.getInt32(5), BBs[2]); 1237 1238 Builder.SetInsertPoint(ContBB); 1239 return RValue::get(nullptr); 1240 } 1241 1242 case Builtin::BI__atomic_thread_fence: 1243 case Builtin::BI__atomic_signal_fence: 1244 case Builtin::BI__c11_atomic_thread_fence: 1245 case Builtin::BI__c11_atomic_signal_fence: { 1246 llvm::SynchronizationScope Scope; 1247 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1248 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1249 Scope = llvm::SingleThread; 1250 else 1251 Scope = llvm::CrossThread; 1252 Value *Order = EmitScalarExpr(E->getArg(0)); 1253 if (isa<llvm::ConstantInt>(Order)) { 1254 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1255 switch (ord) { 1256 case 0: // memory_order_relaxed 1257 default: // invalid order 1258 break; 1259 case 1: // memory_order_consume 1260 case 2: // memory_order_acquire 1261 Builder.CreateFence(llvm::Acquire, Scope); 1262 break; 1263 case 3: // memory_order_release 1264 Builder.CreateFence(llvm::Release, Scope); 1265 break; 1266 case 4: // memory_order_acq_rel 1267 Builder.CreateFence(llvm::AcquireRelease, Scope); 1268 break; 1269 case 5: // memory_order_seq_cst 1270 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1271 break; 1272 } 1273 return RValue::get(nullptr); 1274 } 1275 1276 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1277 AcquireBB = createBasicBlock("acquire", CurFn); 1278 ReleaseBB = createBasicBlock("release", CurFn); 1279 AcqRelBB = createBasicBlock("acqrel", CurFn); 1280 SeqCstBB = createBasicBlock("seqcst", CurFn); 1281 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1282 1283 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1284 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1285 1286 Builder.SetInsertPoint(AcquireBB); 1287 Builder.CreateFence(llvm::Acquire, Scope); 1288 Builder.CreateBr(ContBB); 1289 SI->addCase(Builder.getInt32(1), AcquireBB); 1290 SI->addCase(Builder.getInt32(2), AcquireBB); 1291 1292 Builder.SetInsertPoint(ReleaseBB); 1293 Builder.CreateFence(llvm::Release, Scope); 1294 Builder.CreateBr(ContBB); 1295 SI->addCase(Builder.getInt32(3), ReleaseBB); 1296 1297 Builder.SetInsertPoint(AcqRelBB); 1298 Builder.CreateFence(llvm::AcquireRelease, Scope); 1299 Builder.CreateBr(ContBB); 1300 SI->addCase(Builder.getInt32(4), AcqRelBB); 1301 1302 Builder.SetInsertPoint(SeqCstBB); 1303 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1304 Builder.CreateBr(ContBB); 1305 SI->addCase(Builder.getInt32(5), SeqCstBB); 1306 1307 Builder.SetInsertPoint(ContBB); 1308 return RValue::get(nullptr); 1309 } 1310 1311 // Library functions with special handling. 1312 case Builtin::BIsqrt: 1313 case Builtin::BIsqrtf: 1314 case Builtin::BIsqrtl: { 1315 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1316 // in finite- or unsafe-math mode (the intrinsic has different semantics 1317 // for handling negative numbers compared to the library function, so 1318 // -fmath-errno=0 is not enough). 1319 if (!FD->hasAttr<ConstAttr>()) 1320 break; 1321 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1322 CGM.getCodeGenOpts().NoNaNsFPMath)) 1323 break; 1324 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1325 llvm::Type *ArgType = Arg0->getType(); 1326 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1327 return RValue::get(Builder.CreateCall(F, Arg0)); 1328 } 1329 1330 case Builtin::BIpow: 1331 case Builtin::BIpowf: 1332 case Builtin::BIpowl: { 1333 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1334 if (!FD->hasAttr<ConstAttr>()) 1335 break; 1336 Value *Base = EmitScalarExpr(E->getArg(0)); 1337 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1338 llvm::Type *ArgType = Base->getType(); 1339 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1340 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1341 } 1342 1343 case Builtin::BIfma: 1344 case Builtin::BIfmaf: 1345 case Builtin::BIfmal: 1346 case Builtin::BI__builtin_fma: 1347 case Builtin::BI__builtin_fmaf: 1348 case Builtin::BI__builtin_fmal: { 1349 // Rewrite fma to intrinsic. 1350 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1351 llvm::Type *ArgType = FirstArg->getType(); 1352 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1353 return RValue::get(Builder.CreateCall3(F, FirstArg, 1354 EmitScalarExpr(E->getArg(1)), 1355 EmitScalarExpr(E->getArg(2)))); 1356 } 1357 1358 case Builtin::BI__builtin_signbit: 1359 case Builtin::BI__builtin_signbitf: 1360 case Builtin::BI__builtin_signbitl: { 1361 LLVMContext &C = CGM.getLLVMContext(); 1362 1363 Value *Arg = EmitScalarExpr(E->getArg(0)); 1364 llvm::Type *ArgTy = Arg->getType(); 1365 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1366 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1367 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1368 if (ArgTy->isPPC_FP128Ty()) { 1369 // The higher-order double comes first, and so we need to truncate the 1370 // pair to extract the overall sign. The order of the pair is the same 1371 // in both little- and big-Endian modes. 1372 ArgWidth >>= 1; 1373 ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1374 BCArg = Builder.CreateTrunc(BCArg, ArgIntTy); 1375 } 1376 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1377 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1378 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1379 } 1380 case Builtin::BI__builtin_annotation: { 1381 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1382 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1383 AnnVal->getType()); 1384 1385 // Get the annotation string, go through casts. Sema requires this to be a 1386 // non-wide string literal, potentially casted, so the cast<> is safe. 1387 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1388 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1389 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1390 } 1391 case Builtin::BI__builtin_addcb: 1392 case Builtin::BI__builtin_addcs: 1393 case Builtin::BI__builtin_addc: 1394 case Builtin::BI__builtin_addcl: 1395 case Builtin::BI__builtin_addcll: 1396 case Builtin::BI__builtin_subcb: 1397 case Builtin::BI__builtin_subcs: 1398 case Builtin::BI__builtin_subc: 1399 case Builtin::BI__builtin_subcl: 1400 case Builtin::BI__builtin_subcll: { 1401 1402 // We translate all of these builtins from expressions of the form: 1403 // int x = ..., y = ..., carryin = ..., carryout, result; 1404 // result = __builtin_addc(x, y, carryin, &carryout); 1405 // 1406 // to LLVM IR of the form: 1407 // 1408 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1409 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1410 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1411 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1412 // i32 %carryin) 1413 // %result = extractvalue {i32, i1} %tmp2, 0 1414 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1415 // %tmp3 = or i1 %carry1, %carry2 1416 // %tmp4 = zext i1 %tmp3 to i32 1417 // store i32 %tmp4, i32* %carryout 1418 1419 // Scalarize our inputs. 1420 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1421 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1422 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1423 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1424 EmitPointerWithAlignment(E->getArg(3)); 1425 1426 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1427 llvm::Intrinsic::ID IntrinsicId; 1428 switch (BuiltinID) { 1429 default: llvm_unreachable("Unknown multiprecision builtin id."); 1430 case Builtin::BI__builtin_addcb: 1431 case Builtin::BI__builtin_addcs: 1432 case Builtin::BI__builtin_addc: 1433 case Builtin::BI__builtin_addcl: 1434 case Builtin::BI__builtin_addcll: 1435 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1436 break; 1437 case Builtin::BI__builtin_subcb: 1438 case Builtin::BI__builtin_subcs: 1439 case Builtin::BI__builtin_subc: 1440 case Builtin::BI__builtin_subcl: 1441 case Builtin::BI__builtin_subcll: 1442 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1443 break; 1444 } 1445 1446 // Construct our resulting LLVM IR expression. 1447 llvm::Value *Carry1; 1448 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1449 X, Y, Carry1); 1450 llvm::Value *Carry2; 1451 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1452 Sum1, Carryin, Carry2); 1453 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1454 X->getType()); 1455 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1456 CarryOutPtr.first); 1457 CarryOutStore->setAlignment(CarryOutPtr.second); 1458 return RValue::get(Sum2); 1459 } 1460 case Builtin::BI__builtin_uadd_overflow: 1461 case Builtin::BI__builtin_uaddl_overflow: 1462 case Builtin::BI__builtin_uaddll_overflow: 1463 case Builtin::BI__builtin_usub_overflow: 1464 case Builtin::BI__builtin_usubl_overflow: 1465 case Builtin::BI__builtin_usubll_overflow: 1466 case Builtin::BI__builtin_umul_overflow: 1467 case Builtin::BI__builtin_umull_overflow: 1468 case Builtin::BI__builtin_umulll_overflow: 1469 case Builtin::BI__builtin_sadd_overflow: 1470 case Builtin::BI__builtin_saddl_overflow: 1471 case Builtin::BI__builtin_saddll_overflow: 1472 case Builtin::BI__builtin_ssub_overflow: 1473 case Builtin::BI__builtin_ssubl_overflow: 1474 case Builtin::BI__builtin_ssubll_overflow: 1475 case Builtin::BI__builtin_smul_overflow: 1476 case Builtin::BI__builtin_smull_overflow: 1477 case Builtin::BI__builtin_smulll_overflow: { 1478 1479 // We translate all of these builtins directly to the relevant llvm IR node. 1480 1481 // Scalarize our inputs. 1482 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1483 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1484 std::pair<llvm::Value *, unsigned> SumOutPtr = 1485 EmitPointerWithAlignment(E->getArg(2)); 1486 1487 // Decide which of the overflow intrinsics we are lowering to: 1488 llvm::Intrinsic::ID IntrinsicId; 1489 switch (BuiltinID) { 1490 default: llvm_unreachable("Unknown security overflow builtin id."); 1491 case Builtin::BI__builtin_uadd_overflow: 1492 case Builtin::BI__builtin_uaddl_overflow: 1493 case Builtin::BI__builtin_uaddll_overflow: 1494 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1495 break; 1496 case Builtin::BI__builtin_usub_overflow: 1497 case Builtin::BI__builtin_usubl_overflow: 1498 case Builtin::BI__builtin_usubll_overflow: 1499 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1500 break; 1501 case Builtin::BI__builtin_umul_overflow: 1502 case Builtin::BI__builtin_umull_overflow: 1503 case Builtin::BI__builtin_umulll_overflow: 1504 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1505 break; 1506 case Builtin::BI__builtin_sadd_overflow: 1507 case Builtin::BI__builtin_saddl_overflow: 1508 case Builtin::BI__builtin_saddll_overflow: 1509 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1510 break; 1511 case Builtin::BI__builtin_ssub_overflow: 1512 case Builtin::BI__builtin_ssubl_overflow: 1513 case Builtin::BI__builtin_ssubll_overflow: 1514 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1515 break; 1516 case Builtin::BI__builtin_smul_overflow: 1517 case Builtin::BI__builtin_smull_overflow: 1518 case Builtin::BI__builtin_smulll_overflow: 1519 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1520 break; 1521 } 1522 1523 1524 llvm::Value *Carry; 1525 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1526 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1527 SumOutStore->setAlignment(SumOutPtr.second); 1528 1529 return RValue::get(Carry); 1530 } 1531 case Builtin::BI__builtin_addressof: 1532 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1533 case Builtin::BI__builtin_operator_new: 1534 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1535 E->getArg(0), false); 1536 case Builtin::BI__builtin_operator_delete: 1537 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1538 E->getArg(0), true); 1539 case Builtin::BI__noop: 1540 // __noop always evaluates to an integer literal zero. 1541 return RValue::get(ConstantInt::get(IntTy, 0)); 1542 case Builtin::BI_InterlockedExchange: 1543 case Builtin::BI_InterlockedExchangePointer: 1544 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1545 case Builtin::BI_InterlockedCompareExchangePointer: { 1546 llvm::Type *RTy; 1547 llvm::IntegerType *IntType = 1548 IntegerType::get(getLLVMContext(), 1549 getContext().getTypeSize(E->getType())); 1550 llvm::Type *IntPtrType = IntType->getPointerTo(); 1551 1552 llvm::Value *Destination = 1553 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1554 1555 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1556 RTy = Exchange->getType(); 1557 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1558 1559 llvm::Value *Comparand = 1560 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1561 1562 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1563 SequentiallyConsistent, 1564 SequentiallyConsistent); 1565 Result->setVolatile(true); 1566 1567 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1568 0), 1569 RTy)); 1570 } 1571 case Builtin::BI_InterlockedCompareExchange: { 1572 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1573 EmitScalarExpr(E->getArg(0)), 1574 EmitScalarExpr(E->getArg(2)), 1575 EmitScalarExpr(E->getArg(1)), 1576 SequentiallyConsistent, 1577 SequentiallyConsistent); 1578 CXI->setVolatile(true); 1579 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1580 } 1581 case Builtin::BI_InterlockedIncrement: { 1582 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1583 AtomicRMWInst::Add, 1584 EmitScalarExpr(E->getArg(0)), 1585 ConstantInt::get(Int32Ty, 1), 1586 llvm::SequentiallyConsistent); 1587 RMWI->setVolatile(true); 1588 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1589 } 1590 case Builtin::BI_InterlockedDecrement: { 1591 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1592 AtomicRMWInst::Sub, 1593 EmitScalarExpr(E->getArg(0)), 1594 ConstantInt::get(Int32Ty, 1), 1595 llvm::SequentiallyConsistent); 1596 RMWI->setVolatile(true); 1597 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1598 } 1599 case Builtin::BI_InterlockedExchangeAdd: { 1600 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1601 AtomicRMWInst::Add, 1602 EmitScalarExpr(E->getArg(0)), 1603 EmitScalarExpr(E->getArg(1)), 1604 llvm::SequentiallyConsistent); 1605 RMWI->setVolatile(true); 1606 return RValue::get(RMWI); 1607 } 1608 } 1609 1610 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1611 // the call using the normal call path, but using the unmangled 1612 // version of the function name. 1613 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1614 return emitLibraryCall(*this, FD, E, 1615 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1616 1617 // If this is a predefined lib function (e.g. malloc), emit the call 1618 // using exactly the normal call path. 1619 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1620 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1621 1622 // See if we have a target specific intrinsic. 1623 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1624 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1625 if (const char *Prefix = 1626 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1627 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1628 // NOTE we dont need to perform a compatibility flag check here since the 1629 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1630 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1631 if (IntrinsicID == Intrinsic::not_intrinsic) 1632 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1633 } 1634 1635 if (IntrinsicID != Intrinsic::not_intrinsic) { 1636 SmallVector<Value*, 16> Args; 1637 1638 // Find out if any arguments are required to be integer constant 1639 // expressions. 1640 unsigned ICEArguments = 0; 1641 ASTContext::GetBuiltinTypeError Error; 1642 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1643 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1644 1645 Function *F = CGM.getIntrinsic(IntrinsicID); 1646 llvm::FunctionType *FTy = F->getFunctionType(); 1647 1648 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1649 Value *ArgValue; 1650 // If this is a normal argument, just emit it as a scalar. 1651 if ((ICEArguments & (1 << i)) == 0) { 1652 ArgValue = EmitScalarExpr(E->getArg(i)); 1653 } else { 1654 // If this is required to be a constant, constant fold it so that we 1655 // know that the generated intrinsic gets a ConstantInt. 1656 llvm::APSInt Result; 1657 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1658 assert(IsConst && "Constant arg isn't actually constant?"); 1659 (void)IsConst; 1660 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1661 } 1662 1663 // If the intrinsic arg type is different from the builtin arg type 1664 // we need to do a bit cast. 1665 llvm::Type *PTy = FTy->getParamType(i); 1666 if (PTy != ArgValue->getType()) { 1667 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1668 "Must be able to losslessly bit cast to param"); 1669 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1670 } 1671 1672 Args.push_back(ArgValue); 1673 } 1674 1675 Value *V = Builder.CreateCall(F, Args); 1676 QualType BuiltinRetType = E->getType(); 1677 1678 llvm::Type *RetTy = VoidTy; 1679 if (!BuiltinRetType->isVoidType()) 1680 RetTy = ConvertType(BuiltinRetType); 1681 1682 if (RetTy != V->getType()) { 1683 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1684 "Must be able to losslessly bit cast result type"); 1685 V = Builder.CreateBitCast(V, RetTy); 1686 } 1687 1688 return RValue::get(V); 1689 } 1690 1691 // See if we have a target specific builtin that needs to be lowered. 1692 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1693 return RValue::get(V); 1694 1695 ErrorUnsupported(E, "builtin function"); 1696 1697 // Unknown builtin, for now just dump it out and return undef. 1698 return GetUndefRValue(E->getType()); 1699 } 1700 1701 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1702 const CallExpr *E) { 1703 switch (getTarget().getTriple().getArch()) { 1704 case llvm::Triple::arm: 1705 case llvm::Triple::armeb: 1706 case llvm::Triple::thumb: 1707 case llvm::Triple::thumbeb: 1708 return EmitARMBuiltinExpr(BuiltinID, E); 1709 case llvm::Triple::aarch64: 1710 case llvm::Triple::aarch64_be: 1711 return EmitAArch64BuiltinExpr(BuiltinID, E); 1712 case llvm::Triple::x86: 1713 case llvm::Triple::x86_64: 1714 return EmitX86BuiltinExpr(BuiltinID, E); 1715 case llvm::Triple::ppc: 1716 case llvm::Triple::ppc64: 1717 case llvm::Triple::ppc64le: 1718 return EmitPPCBuiltinExpr(BuiltinID, E); 1719 case llvm::Triple::r600: 1720 return EmitR600BuiltinExpr(BuiltinID, E); 1721 default: 1722 return nullptr; 1723 } 1724 } 1725 1726 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1727 NeonTypeFlags TypeFlags, 1728 bool V1Ty=false) { 1729 int IsQuad = TypeFlags.isQuad(); 1730 switch (TypeFlags.getEltType()) { 1731 case NeonTypeFlags::Int8: 1732 case NeonTypeFlags::Poly8: 1733 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1734 case NeonTypeFlags::Int16: 1735 case NeonTypeFlags::Poly16: 1736 case NeonTypeFlags::Float16: 1737 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1738 case NeonTypeFlags::Int32: 1739 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1740 case NeonTypeFlags::Int64: 1741 case NeonTypeFlags::Poly64: 1742 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1743 case NeonTypeFlags::Poly128: 1744 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1745 // There is a lot of i128 and f128 API missing. 1746 // so we use v16i8 to represent poly128 and get pattern matched. 1747 return llvm::VectorType::get(CGF->Int8Ty, 16); 1748 case NeonTypeFlags::Float32: 1749 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1750 case NeonTypeFlags::Float64: 1751 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1752 } 1753 llvm_unreachable("Unknown vector element type!"); 1754 } 1755 1756 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1757 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1758 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1759 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1760 } 1761 1762 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1763 const char *name, 1764 unsigned shift, bool rightshift) { 1765 unsigned j = 0; 1766 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1767 ai != ae; ++ai, ++j) 1768 if (shift > 0 && shift == j) 1769 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1770 else 1771 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1772 1773 return Builder.CreateCall(F, Ops, name); 1774 } 1775 1776 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1777 bool neg) { 1778 int SV = cast<ConstantInt>(V)->getSExtValue(); 1779 1780 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1781 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1782 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1783 } 1784 1785 // \brief Right-shift a vector by a constant. 1786 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1787 llvm::Type *Ty, bool usgn, 1788 const char *name) { 1789 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1790 1791 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1792 int EltSize = VTy->getScalarSizeInBits(); 1793 1794 Vec = Builder.CreateBitCast(Vec, Ty); 1795 1796 // lshr/ashr are undefined when the shift amount is equal to the vector 1797 // element size. 1798 if (ShiftAmt == EltSize) { 1799 if (usgn) { 1800 // Right-shifting an unsigned value by its size yields 0. 1801 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1802 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1803 } else { 1804 // Right-shifting a signed value by its size is equivalent 1805 // to a shift of size-1. 1806 --ShiftAmt; 1807 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1808 } 1809 } 1810 1811 Shift = EmitNeonShiftVector(Shift, Ty, false); 1812 if (usgn) 1813 return Builder.CreateLShr(Vec, Shift, name); 1814 else 1815 return Builder.CreateAShr(Vec, Shift, name); 1816 } 1817 1818 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1819 /// alignment of the type referenced by the pointer. Skip over implicit 1820 /// casts. 1821 std::pair<llvm::Value*, unsigned> 1822 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1823 assert(Addr->getType()->isPointerType()); 1824 Addr = Addr->IgnoreParens(); 1825 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1826 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1827 ICE->getSubExpr()->getType()->isPointerType()) { 1828 std::pair<llvm::Value*, unsigned> Ptr = 1829 EmitPointerWithAlignment(ICE->getSubExpr()); 1830 Ptr.first = Builder.CreateBitCast(Ptr.first, 1831 ConvertType(Addr->getType())); 1832 return Ptr; 1833 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1834 LValue LV = EmitLValue(ICE->getSubExpr()); 1835 unsigned Align = LV.getAlignment().getQuantity(); 1836 if (!Align) { 1837 // FIXME: Once LValues are fixed to always set alignment, 1838 // zap this code. 1839 QualType PtTy = ICE->getSubExpr()->getType(); 1840 if (!PtTy->isIncompleteType()) 1841 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1842 else 1843 Align = 1; 1844 } 1845 return std::make_pair(LV.getAddress(), Align); 1846 } 1847 } 1848 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1849 if (UO->getOpcode() == UO_AddrOf) { 1850 LValue LV = EmitLValue(UO->getSubExpr()); 1851 unsigned Align = LV.getAlignment().getQuantity(); 1852 if (!Align) { 1853 // FIXME: Once LValues are fixed to always set alignment, 1854 // zap this code. 1855 QualType PtTy = UO->getSubExpr()->getType(); 1856 if (!PtTy->isIncompleteType()) 1857 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1858 else 1859 Align = 1; 1860 } 1861 return std::make_pair(LV.getAddress(), Align); 1862 } 1863 } 1864 1865 unsigned Align = 1; 1866 QualType PtTy = Addr->getType()->getPointeeType(); 1867 if (!PtTy->isIncompleteType()) 1868 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1869 1870 return std::make_pair(EmitScalarExpr(Addr), Align); 1871 } 1872 1873 enum { 1874 AddRetType = (1 << 0), 1875 Add1ArgType = (1 << 1), 1876 Add2ArgTypes = (1 << 2), 1877 1878 VectorizeRetType = (1 << 3), 1879 VectorizeArgTypes = (1 << 4), 1880 1881 InventFloatType = (1 << 5), 1882 UnsignedAlts = (1 << 6), 1883 1884 Use64BitVectors = (1 << 7), 1885 Use128BitVectors = (1 << 8), 1886 1887 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 1888 VectorRet = AddRetType | VectorizeRetType, 1889 VectorRetGetArgs01 = 1890 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 1891 FpCmpzModifiers = 1892 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 1893 }; 1894 1895 struct NeonIntrinsicInfo { 1896 unsigned BuiltinID; 1897 unsigned LLVMIntrinsic; 1898 unsigned AltLLVMIntrinsic; 1899 const char *NameHint; 1900 unsigned TypeModifier; 1901 1902 bool operator<(unsigned RHSBuiltinID) const { 1903 return BuiltinID < RHSBuiltinID; 1904 } 1905 }; 1906 1907 #define NEONMAP0(NameBase) \ 1908 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 1909 1910 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 1911 { NEON:: BI__builtin_neon_ ## NameBase, \ 1912 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 1913 1914 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 1915 { NEON:: BI__builtin_neon_ ## NameBase, \ 1916 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 1917 #NameBase, TypeModifier } 1918 1919 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 1920 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1921 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1922 NEONMAP1(vabs_v, arm_neon_vabs, 0), 1923 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 1924 NEONMAP0(vaddhn_v), 1925 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 1926 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 1927 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 1928 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 1929 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 1930 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 1931 NEONMAP1(vcage_v, arm_neon_vacge, 0), 1932 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 1933 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 1934 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 1935 NEONMAP1(vcale_v, arm_neon_vacge, 0), 1936 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 1937 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 1938 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 1939 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 1940 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 1941 NEONMAP1(vclz_v, ctlz, Add1ArgType), 1942 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 1943 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 1944 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 1945 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 1946 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 1947 NEONMAP0(vcvt_f32_v), 1948 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1949 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1950 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1951 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1952 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1953 NEONMAP0(vcvt_s32_v), 1954 NEONMAP0(vcvt_s64_v), 1955 NEONMAP0(vcvt_u32_v), 1956 NEONMAP0(vcvt_u64_v), 1957 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 1958 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 1959 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 1960 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 1961 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 1962 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 1963 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 1964 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 1965 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 1966 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 1967 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 1968 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 1969 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 1970 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 1971 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 1972 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 1973 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 1974 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 1975 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 1976 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 1977 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 1978 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 1979 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 1980 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 1981 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 1982 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 1983 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 1984 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 1985 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 1986 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 1987 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 1988 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 1989 NEONMAP0(vcvtq_f32_v), 1990 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1991 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1992 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1993 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1994 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1995 NEONMAP0(vcvtq_s32_v), 1996 NEONMAP0(vcvtq_s64_v), 1997 NEONMAP0(vcvtq_u32_v), 1998 NEONMAP0(vcvtq_u64_v), 1999 NEONMAP0(vext_v), 2000 NEONMAP0(vextq_v), 2001 NEONMAP0(vfma_v), 2002 NEONMAP0(vfmaq_v), 2003 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2004 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2005 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2006 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2007 NEONMAP0(vld1_dup_v), 2008 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2009 NEONMAP0(vld1q_dup_v), 2010 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2011 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2012 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2013 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2014 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2015 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2016 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2017 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2018 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2019 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2020 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2021 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2022 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2023 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2024 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2025 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2026 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2027 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2028 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2029 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2030 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2031 NEONMAP0(vmovl_v), 2032 NEONMAP0(vmovn_v), 2033 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2034 NEONMAP0(vmull_v), 2035 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2036 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2037 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2038 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2039 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2040 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2041 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2042 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2043 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2044 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2045 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2046 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2047 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2048 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2049 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2050 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2051 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2052 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2053 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2054 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2055 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2056 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2057 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2058 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2059 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2060 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2061 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2062 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2063 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2064 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2065 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2066 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2067 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2068 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2069 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2070 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2071 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2072 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2073 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2074 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2075 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2076 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2077 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2078 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2079 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2080 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2081 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2082 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2083 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2084 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2085 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2086 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2087 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2088 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2089 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2090 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2091 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2092 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2093 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2094 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2095 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2096 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2097 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2098 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2099 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2100 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2101 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2102 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2103 NEONMAP0(vshl_n_v), 2104 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2105 NEONMAP0(vshll_n_v), 2106 NEONMAP0(vshlq_n_v), 2107 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2108 NEONMAP0(vshr_n_v), 2109 NEONMAP0(vshrn_n_v), 2110 NEONMAP0(vshrq_n_v), 2111 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2112 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2113 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2114 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2115 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2116 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2117 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2118 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2119 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2120 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2121 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2122 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2123 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2124 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2125 NEONMAP0(vsubhn_v), 2126 NEONMAP0(vtrn_v), 2127 NEONMAP0(vtrnq_v), 2128 NEONMAP0(vtst_v), 2129 NEONMAP0(vtstq_v), 2130 NEONMAP0(vuzp_v), 2131 NEONMAP0(vuzpq_v), 2132 NEONMAP0(vzip_v), 2133 NEONMAP0(vzipq_v) 2134 }; 2135 2136 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2137 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2138 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2139 NEONMAP0(vaddhn_v), 2140 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2141 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2142 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2143 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2144 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2145 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2146 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2147 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2148 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2149 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2150 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2151 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2152 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2153 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2154 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2155 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2156 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2157 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2158 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2159 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2160 NEONMAP0(vcvt_f32_v), 2161 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2162 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2163 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2164 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2165 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2166 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2167 NEONMAP0(vcvtq_f32_v), 2168 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2169 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2170 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2171 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2172 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2173 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2174 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2175 NEONMAP0(vext_v), 2176 NEONMAP0(vextq_v), 2177 NEONMAP0(vfma_v), 2178 NEONMAP0(vfmaq_v), 2179 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2180 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2181 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2182 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2183 NEONMAP0(vmovl_v), 2184 NEONMAP0(vmovn_v), 2185 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2186 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2187 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2188 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2189 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2190 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2191 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2192 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2193 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2194 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2195 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2196 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2197 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2198 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2199 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2200 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2201 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2202 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2203 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2204 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2205 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2206 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2207 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2208 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2209 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2210 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2211 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2212 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2213 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2214 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2215 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2216 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2217 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2218 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2219 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2220 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2221 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2222 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2223 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2224 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2225 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2226 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2227 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2228 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2229 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2230 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2231 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2232 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2233 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2234 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2235 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2236 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2237 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2238 NEONMAP0(vshl_n_v), 2239 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2240 NEONMAP0(vshll_n_v), 2241 NEONMAP0(vshlq_n_v), 2242 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2243 NEONMAP0(vshr_n_v), 2244 NEONMAP0(vshrn_n_v), 2245 NEONMAP0(vshrq_n_v), 2246 NEONMAP0(vsubhn_v), 2247 NEONMAP0(vtst_v), 2248 NEONMAP0(vtstq_v), 2249 }; 2250 2251 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2252 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2253 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2254 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2255 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2256 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2257 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2258 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2259 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2260 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2261 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2262 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2263 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2264 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2265 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2266 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2267 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2268 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2269 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2270 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2271 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2272 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2273 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2274 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2275 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2276 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2277 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2278 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2279 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2280 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2281 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2282 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2283 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2284 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2285 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2286 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2287 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2288 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2289 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2290 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2291 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2292 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2293 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2294 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2295 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2296 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2297 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2298 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2299 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2300 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2301 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2302 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2303 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2304 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2305 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2306 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2307 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2308 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2309 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2310 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2311 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2312 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2313 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2314 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2315 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2316 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2317 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2318 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2319 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2320 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2321 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2322 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2323 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2324 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2325 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2326 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2327 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2328 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2329 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2330 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2331 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2332 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2333 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2334 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2335 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2336 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2337 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2338 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2339 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2340 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2341 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2342 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2343 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2344 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2345 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2346 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2347 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2348 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2349 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2350 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2351 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2352 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2353 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2354 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2355 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2356 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2357 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2358 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2359 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2360 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2361 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2362 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2363 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2364 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2365 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2366 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2367 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2368 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2369 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2370 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2371 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2372 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2373 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2374 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2375 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2376 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2377 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2378 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2379 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2380 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2381 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2382 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2383 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2384 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2385 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2386 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2387 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2388 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2389 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2390 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2391 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2392 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2393 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2394 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2395 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2396 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2397 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2398 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2399 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2400 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2401 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2402 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2403 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2404 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2405 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2406 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2407 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2408 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2409 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2410 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2411 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2412 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2413 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2414 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2415 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2416 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2417 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2418 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2419 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2420 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2421 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2422 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2423 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2424 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2425 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2426 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2427 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2428 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2429 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2430 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2431 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2432 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2433 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2434 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2435 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2436 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2437 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2438 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2439 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2440 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2441 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2442 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2443 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2444 }; 2445 2446 #undef NEONMAP0 2447 #undef NEONMAP1 2448 #undef NEONMAP2 2449 2450 static bool NEONSIMDIntrinsicsProvenSorted = false; 2451 2452 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2453 static bool AArch64SISDIntrinsicsProvenSorted = false; 2454 2455 2456 static const NeonIntrinsicInfo * 2457 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2458 unsigned BuiltinID, bool &MapProvenSorted) { 2459 2460 #ifndef NDEBUG 2461 if (!MapProvenSorted) { 2462 // FIXME: use std::is_sorted once C++11 is allowed 2463 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2464 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2465 MapProvenSorted = true; 2466 } 2467 #endif 2468 2469 const NeonIntrinsicInfo *Builtin = 2470 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2471 2472 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2473 return Builtin; 2474 2475 return nullptr; 2476 } 2477 2478 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2479 unsigned Modifier, 2480 llvm::Type *ArgType, 2481 const CallExpr *E) { 2482 int VectorSize = 0; 2483 if (Modifier & Use64BitVectors) 2484 VectorSize = 64; 2485 else if (Modifier & Use128BitVectors) 2486 VectorSize = 128; 2487 2488 // Return type. 2489 SmallVector<llvm::Type *, 3> Tys; 2490 if (Modifier & AddRetType) { 2491 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2492 if (Modifier & VectorizeRetType) 2493 Ty = llvm::VectorType::get( 2494 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2495 2496 Tys.push_back(Ty); 2497 } 2498 2499 // Arguments. 2500 if (Modifier & VectorizeArgTypes) { 2501 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2502 ArgType = llvm::VectorType::get(ArgType, Elts); 2503 } 2504 2505 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2506 Tys.push_back(ArgType); 2507 2508 if (Modifier & Add2ArgTypes) 2509 Tys.push_back(ArgType); 2510 2511 if (Modifier & InventFloatType) 2512 Tys.push_back(FloatTy); 2513 2514 return CGM.getIntrinsic(IntrinsicID, Tys); 2515 } 2516 2517 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2518 const NeonIntrinsicInfo &SISDInfo, 2519 SmallVectorImpl<Value *> &Ops, 2520 const CallExpr *E) { 2521 unsigned BuiltinID = SISDInfo.BuiltinID; 2522 unsigned int Int = SISDInfo.LLVMIntrinsic; 2523 unsigned Modifier = SISDInfo.TypeModifier; 2524 const char *s = SISDInfo.NameHint; 2525 2526 switch (BuiltinID) { 2527 case NEON::BI__builtin_neon_vcled_s64: 2528 case NEON::BI__builtin_neon_vcled_u64: 2529 case NEON::BI__builtin_neon_vcles_f32: 2530 case NEON::BI__builtin_neon_vcled_f64: 2531 case NEON::BI__builtin_neon_vcltd_s64: 2532 case NEON::BI__builtin_neon_vcltd_u64: 2533 case NEON::BI__builtin_neon_vclts_f32: 2534 case NEON::BI__builtin_neon_vcltd_f64: 2535 case NEON::BI__builtin_neon_vcales_f32: 2536 case NEON::BI__builtin_neon_vcaled_f64: 2537 case NEON::BI__builtin_neon_vcalts_f32: 2538 case NEON::BI__builtin_neon_vcaltd_f64: 2539 // Only one direction of comparisons actually exist, cmle is actually a cmge 2540 // with swapped operands. The table gives us the right intrinsic but we 2541 // still need to do the swap. 2542 std::swap(Ops[0], Ops[1]); 2543 break; 2544 } 2545 2546 assert(Int && "Generic code assumes a valid intrinsic"); 2547 2548 // Determine the type(s) of this overloaded AArch64 intrinsic. 2549 const Expr *Arg = E->getArg(0); 2550 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2551 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2552 2553 int j = 0; 2554 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2555 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2556 ai != ae; ++ai, ++j) { 2557 llvm::Type *ArgTy = ai->getType(); 2558 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2559 ArgTy->getPrimitiveSizeInBits()) 2560 continue; 2561 2562 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2563 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2564 // it before inserting. 2565 Ops[j] = 2566 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2567 Ops[j] = 2568 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2569 } 2570 2571 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2572 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2573 if (ResultType->getPrimitiveSizeInBits() < 2574 Result->getType()->getPrimitiveSizeInBits()) 2575 return CGF.Builder.CreateExtractElement(Result, C0); 2576 2577 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2578 } 2579 2580 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2581 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2582 const char *NameHint, unsigned Modifier, const CallExpr *E, 2583 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2584 // Get the last argument, which specifies the vector type. 2585 llvm::APSInt NeonTypeConst; 2586 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2587 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2588 return nullptr; 2589 2590 // Determine the type of this overloaded NEON intrinsic. 2591 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2592 bool Usgn = Type.isUnsigned(); 2593 bool Quad = Type.isQuad(); 2594 2595 llvm::VectorType *VTy = GetNeonType(this, Type); 2596 llvm::Type *Ty = VTy; 2597 if (!Ty) 2598 return nullptr; 2599 2600 unsigned Int = LLVMIntrinsic; 2601 if ((Modifier & UnsignedAlts) && !Usgn) 2602 Int = AltLLVMIntrinsic; 2603 2604 switch (BuiltinID) { 2605 default: break; 2606 case NEON::BI__builtin_neon_vabs_v: 2607 case NEON::BI__builtin_neon_vabsq_v: 2608 if (VTy->getElementType()->isFloatingPointTy()) 2609 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2610 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2611 case NEON::BI__builtin_neon_vaddhn_v: { 2612 llvm::VectorType *SrcTy = 2613 llvm::VectorType::getExtendedElementVectorType(VTy); 2614 2615 // %sum = add <4 x i32> %lhs, %rhs 2616 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2617 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2618 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2619 2620 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2621 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2622 SrcTy->getScalarSizeInBits() / 2); 2623 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2624 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2625 2626 // %res = trunc <4 x i32> %high to <4 x i16> 2627 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2628 } 2629 case NEON::BI__builtin_neon_vcale_v: 2630 case NEON::BI__builtin_neon_vcaleq_v: 2631 case NEON::BI__builtin_neon_vcalt_v: 2632 case NEON::BI__builtin_neon_vcaltq_v: 2633 std::swap(Ops[0], Ops[1]); 2634 case NEON::BI__builtin_neon_vcage_v: 2635 case NEON::BI__builtin_neon_vcageq_v: 2636 case NEON::BI__builtin_neon_vcagt_v: 2637 case NEON::BI__builtin_neon_vcagtq_v: { 2638 llvm::Type *VecFlt = llvm::VectorType::get( 2639 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2640 VTy->getNumElements()); 2641 llvm::Type *Tys[] = { VTy, VecFlt }; 2642 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2643 return EmitNeonCall(F, Ops, NameHint); 2644 } 2645 case NEON::BI__builtin_neon_vclz_v: 2646 case NEON::BI__builtin_neon_vclzq_v: 2647 // We generate target-independent intrinsic, which needs a second argument 2648 // for whether or not clz of zero is undefined; on ARM it isn't. 2649 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2650 break; 2651 case NEON::BI__builtin_neon_vcvt_f32_v: 2652 case NEON::BI__builtin_neon_vcvtq_f32_v: 2653 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2654 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2655 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2656 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2657 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2658 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2659 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2660 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2661 bool Double = 2662 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2663 llvm::Type *FloatTy = 2664 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2665 : NeonTypeFlags::Float32, 2666 false, Quad)); 2667 llvm::Type *Tys[2] = { FloatTy, Ty }; 2668 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2669 Function *F = CGM.getIntrinsic(Int, Tys); 2670 return EmitNeonCall(F, Ops, "vcvt_n"); 2671 } 2672 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2673 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2674 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2675 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2676 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2677 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2678 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2679 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2680 bool Double = 2681 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2682 llvm::Type *FloatTy = 2683 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2684 : NeonTypeFlags::Float32, 2685 false, Quad)); 2686 llvm::Type *Tys[2] = { Ty, FloatTy }; 2687 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2688 return EmitNeonCall(F, Ops, "vcvt_n"); 2689 } 2690 case NEON::BI__builtin_neon_vcvt_s32_v: 2691 case NEON::BI__builtin_neon_vcvt_u32_v: 2692 case NEON::BI__builtin_neon_vcvt_s64_v: 2693 case NEON::BI__builtin_neon_vcvt_u64_v: 2694 case NEON::BI__builtin_neon_vcvtq_s32_v: 2695 case NEON::BI__builtin_neon_vcvtq_u32_v: 2696 case NEON::BI__builtin_neon_vcvtq_s64_v: 2697 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2698 bool Double = 2699 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2700 llvm::Type *FloatTy = 2701 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2702 : NeonTypeFlags::Float32, 2703 false, Quad)); 2704 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2705 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2706 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2707 } 2708 case NEON::BI__builtin_neon_vcvta_s32_v: 2709 case NEON::BI__builtin_neon_vcvta_s64_v: 2710 case NEON::BI__builtin_neon_vcvta_u32_v: 2711 case NEON::BI__builtin_neon_vcvta_u64_v: 2712 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2713 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2714 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2715 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2716 case NEON::BI__builtin_neon_vcvtn_s32_v: 2717 case NEON::BI__builtin_neon_vcvtn_s64_v: 2718 case NEON::BI__builtin_neon_vcvtn_u32_v: 2719 case NEON::BI__builtin_neon_vcvtn_u64_v: 2720 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2721 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2722 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2723 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2724 case NEON::BI__builtin_neon_vcvtp_s32_v: 2725 case NEON::BI__builtin_neon_vcvtp_s64_v: 2726 case NEON::BI__builtin_neon_vcvtp_u32_v: 2727 case NEON::BI__builtin_neon_vcvtp_u64_v: 2728 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2729 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2730 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2731 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2732 case NEON::BI__builtin_neon_vcvtm_s32_v: 2733 case NEON::BI__builtin_neon_vcvtm_s64_v: 2734 case NEON::BI__builtin_neon_vcvtm_u32_v: 2735 case NEON::BI__builtin_neon_vcvtm_u64_v: 2736 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2737 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2738 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2739 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2740 bool Double = 2741 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2742 llvm::Type *InTy = 2743 GetNeonType(this, 2744 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2745 : NeonTypeFlags::Float32, false, Quad)); 2746 llvm::Type *Tys[2] = { Ty, InTy }; 2747 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2748 } 2749 case NEON::BI__builtin_neon_vext_v: 2750 case NEON::BI__builtin_neon_vextq_v: { 2751 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2752 SmallVector<Constant*, 16> Indices; 2753 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2754 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2755 2756 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2757 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2758 Value *SV = llvm::ConstantVector::get(Indices); 2759 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2760 } 2761 case NEON::BI__builtin_neon_vfma_v: 2762 case NEON::BI__builtin_neon_vfmaq_v: { 2763 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2764 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2765 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2766 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2767 2768 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2769 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2770 } 2771 case NEON::BI__builtin_neon_vld1_v: 2772 case NEON::BI__builtin_neon_vld1q_v: 2773 Ops.push_back(Align); 2774 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2775 case NEON::BI__builtin_neon_vld2_v: 2776 case NEON::BI__builtin_neon_vld2q_v: 2777 case NEON::BI__builtin_neon_vld3_v: 2778 case NEON::BI__builtin_neon_vld3q_v: 2779 case NEON::BI__builtin_neon_vld4_v: 2780 case NEON::BI__builtin_neon_vld4q_v: { 2781 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2782 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2783 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2784 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2785 return Builder.CreateStore(Ops[1], Ops[0]); 2786 } 2787 case NEON::BI__builtin_neon_vld1_dup_v: 2788 case NEON::BI__builtin_neon_vld1q_dup_v: { 2789 Value *V = UndefValue::get(Ty); 2790 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2791 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2792 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2793 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2794 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2795 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2796 return EmitNeonSplat(Ops[0], CI); 2797 } 2798 case NEON::BI__builtin_neon_vld2_lane_v: 2799 case NEON::BI__builtin_neon_vld2q_lane_v: 2800 case NEON::BI__builtin_neon_vld3_lane_v: 2801 case NEON::BI__builtin_neon_vld3q_lane_v: 2802 case NEON::BI__builtin_neon_vld4_lane_v: 2803 case NEON::BI__builtin_neon_vld4q_lane_v: { 2804 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2805 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2806 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2807 Ops.push_back(Align); 2808 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2809 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2810 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2811 return Builder.CreateStore(Ops[1], Ops[0]); 2812 } 2813 case NEON::BI__builtin_neon_vmovl_v: { 2814 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2815 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2816 if (Usgn) 2817 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2818 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2819 } 2820 case NEON::BI__builtin_neon_vmovn_v: { 2821 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2822 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2823 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2824 } 2825 case NEON::BI__builtin_neon_vmull_v: 2826 // FIXME: the integer vmull operations could be emitted in terms of pure 2827 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2828 // hoisting the exts outside loops. Until global ISel comes along that can 2829 // see through such movement this leads to bad CodeGen. So we need an 2830 // intrinsic for now. 2831 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2832 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2833 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2834 case NEON::BI__builtin_neon_vpadal_v: 2835 case NEON::BI__builtin_neon_vpadalq_v: { 2836 // The source operand type has twice as many elements of half the size. 2837 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2838 llvm::Type *EltTy = 2839 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2840 llvm::Type *NarrowTy = 2841 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2842 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2843 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2844 } 2845 case NEON::BI__builtin_neon_vpaddl_v: 2846 case NEON::BI__builtin_neon_vpaddlq_v: { 2847 // The source operand type has twice as many elements of half the size. 2848 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2849 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2850 llvm::Type *NarrowTy = 2851 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2852 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2853 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2854 } 2855 case NEON::BI__builtin_neon_vqdmlal_v: 2856 case NEON::BI__builtin_neon_vqdmlsl_v: { 2857 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2858 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2859 MulOps, "vqdmlal"); 2860 2861 SmallVector<Value *, 2> AccumOps; 2862 AccumOps.push_back(Ops[0]); 2863 AccumOps.push_back(Mul); 2864 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2865 AccumOps, NameHint); 2866 } 2867 case NEON::BI__builtin_neon_vqshl_n_v: 2868 case NEON::BI__builtin_neon_vqshlq_n_v: 2869 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2870 1, false); 2871 case NEON::BI__builtin_neon_vqshlu_n_v: 2872 case NEON::BI__builtin_neon_vqshluq_n_v: 2873 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 2874 1, false); 2875 case NEON::BI__builtin_neon_vrecpe_v: 2876 case NEON::BI__builtin_neon_vrecpeq_v: 2877 case NEON::BI__builtin_neon_vrsqrte_v: 2878 case NEON::BI__builtin_neon_vrsqrteq_v: 2879 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2880 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2881 2882 case NEON::BI__builtin_neon_vrshr_n_v: 2883 case NEON::BI__builtin_neon_vrshrq_n_v: 2884 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 2885 1, true); 2886 case NEON::BI__builtin_neon_vshl_n_v: 2887 case NEON::BI__builtin_neon_vshlq_n_v: 2888 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2889 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2890 "vshl_n"); 2891 case NEON::BI__builtin_neon_vshll_n_v: { 2892 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2893 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2894 if (Usgn) 2895 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2896 else 2897 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2898 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2899 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2900 } 2901 case NEON::BI__builtin_neon_vshrn_n_v: { 2902 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2903 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2904 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2905 if (Usgn) 2906 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2907 else 2908 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2909 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2910 } 2911 case NEON::BI__builtin_neon_vshr_n_v: 2912 case NEON::BI__builtin_neon_vshrq_n_v: 2913 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2914 case NEON::BI__builtin_neon_vst1_v: 2915 case NEON::BI__builtin_neon_vst1q_v: 2916 case NEON::BI__builtin_neon_vst2_v: 2917 case NEON::BI__builtin_neon_vst2q_v: 2918 case NEON::BI__builtin_neon_vst3_v: 2919 case NEON::BI__builtin_neon_vst3q_v: 2920 case NEON::BI__builtin_neon_vst4_v: 2921 case NEON::BI__builtin_neon_vst4q_v: 2922 case NEON::BI__builtin_neon_vst2_lane_v: 2923 case NEON::BI__builtin_neon_vst2q_lane_v: 2924 case NEON::BI__builtin_neon_vst3_lane_v: 2925 case NEON::BI__builtin_neon_vst3q_lane_v: 2926 case NEON::BI__builtin_neon_vst4_lane_v: 2927 case NEON::BI__builtin_neon_vst4q_lane_v: 2928 Ops.push_back(Align); 2929 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 2930 case NEON::BI__builtin_neon_vsubhn_v: { 2931 llvm::VectorType *SrcTy = 2932 llvm::VectorType::getExtendedElementVectorType(VTy); 2933 2934 // %sum = add <4 x i32> %lhs, %rhs 2935 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2936 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2937 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2938 2939 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2940 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2941 SrcTy->getScalarSizeInBits() / 2); 2942 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2943 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2944 2945 // %res = trunc <4 x i32> %high to <4 x i16> 2946 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2947 } 2948 case NEON::BI__builtin_neon_vtrn_v: 2949 case NEON::BI__builtin_neon_vtrnq_v: { 2950 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2951 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2952 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2953 Value *SV = nullptr; 2954 2955 for (unsigned vi = 0; vi != 2; ++vi) { 2956 SmallVector<Constant*, 16> Indices; 2957 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2958 Indices.push_back(Builder.getInt32(i+vi)); 2959 Indices.push_back(Builder.getInt32(i+e+vi)); 2960 } 2961 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2962 SV = llvm::ConstantVector::get(Indices); 2963 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2964 SV = Builder.CreateStore(SV, Addr); 2965 } 2966 return SV; 2967 } 2968 case NEON::BI__builtin_neon_vtst_v: 2969 case NEON::BI__builtin_neon_vtstq_v: { 2970 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2971 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2972 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2973 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2974 ConstantAggregateZero::get(Ty)); 2975 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2976 } 2977 case NEON::BI__builtin_neon_vuzp_v: 2978 case NEON::BI__builtin_neon_vuzpq_v: { 2979 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2980 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2981 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2982 Value *SV = nullptr; 2983 2984 for (unsigned vi = 0; vi != 2; ++vi) { 2985 SmallVector<Constant*, 16> Indices; 2986 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2987 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2988 2989 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2990 SV = llvm::ConstantVector::get(Indices); 2991 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2992 SV = Builder.CreateStore(SV, Addr); 2993 } 2994 return SV; 2995 } 2996 case NEON::BI__builtin_neon_vzip_v: 2997 case NEON::BI__builtin_neon_vzipq_v: { 2998 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2999 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3000 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3001 Value *SV = nullptr; 3002 3003 for (unsigned vi = 0; vi != 2; ++vi) { 3004 SmallVector<Constant*, 16> Indices; 3005 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3006 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3007 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3008 } 3009 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3010 SV = llvm::ConstantVector::get(Indices); 3011 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3012 SV = Builder.CreateStore(SV, Addr); 3013 } 3014 return SV; 3015 } 3016 } 3017 3018 assert(Int && "Expected valid intrinsic number"); 3019 3020 // Determine the type(s) of this overloaded AArch64 intrinsic. 3021 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3022 3023 Value *Result = EmitNeonCall(F, Ops, NameHint); 3024 llvm::Type *ResultType = ConvertType(E->getType()); 3025 // AArch64 intrinsic one-element vector type cast to 3026 // scalar type expected by the builtin 3027 return Builder.CreateBitCast(Result, ResultType, NameHint); 3028 } 3029 3030 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3031 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3032 const CmpInst::Predicate Ip, const Twine &Name) { 3033 llvm::Type *OTy = Op->getType(); 3034 3035 // FIXME: this is utterly horrific. We should not be looking at previous 3036 // codegen context to find out what needs doing. Unfortunately TableGen 3037 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3038 // (etc). 3039 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3040 OTy = BI->getOperand(0)->getType(); 3041 3042 Op = Builder.CreateBitCast(Op, OTy); 3043 if (OTy->getScalarType()->isFloatingPointTy()) { 3044 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3045 } else { 3046 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3047 } 3048 return Builder.CreateSExt(Op, Ty, Name); 3049 } 3050 3051 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3052 Value *ExtOp, Value *IndexOp, 3053 llvm::Type *ResTy, unsigned IntID, 3054 const char *Name) { 3055 SmallVector<Value *, 2> TblOps; 3056 if (ExtOp) 3057 TblOps.push_back(ExtOp); 3058 3059 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3060 SmallVector<Constant*, 16> Indices; 3061 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3062 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3063 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3064 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3065 } 3066 Value *SV = llvm::ConstantVector::get(Indices); 3067 3068 int PairPos = 0, End = Ops.size() - 1; 3069 while (PairPos < End) { 3070 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3071 Ops[PairPos+1], SV, Name)); 3072 PairPos += 2; 3073 } 3074 3075 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3076 // of the 128-bit lookup table with zero. 3077 if (PairPos == End) { 3078 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3079 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3080 ZeroTbl, SV, Name)); 3081 } 3082 3083 Function *TblF; 3084 TblOps.push_back(IndexOp); 3085 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3086 3087 return CGF.EmitNeonCall(TblF, TblOps, Name); 3088 } 3089 3090 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3091 const CallExpr *E) { 3092 unsigned HintID = static_cast<unsigned>(-1); 3093 switch (BuiltinID) { 3094 default: break; 3095 case ARM::BI__builtin_arm_nop: 3096 HintID = 0; 3097 break; 3098 case ARM::BI__builtin_arm_yield: 3099 case ARM::BI__yield: 3100 HintID = 1; 3101 break; 3102 case ARM::BI__builtin_arm_wfe: 3103 case ARM::BI__wfe: 3104 HintID = 2; 3105 break; 3106 case ARM::BI__builtin_arm_wfi: 3107 case ARM::BI__wfi: 3108 HintID = 3; 3109 break; 3110 case ARM::BI__builtin_arm_sev: 3111 case ARM::BI__sev: 3112 HintID = 4; 3113 break; 3114 case ARM::BI__builtin_arm_sevl: 3115 case ARM::BI__sevl: 3116 HintID = 5; 3117 break; 3118 } 3119 3120 if (HintID != static_cast<unsigned>(-1)) { 3121 Function *F = CGM.getIntrinsic(Intrinsic::arm_hint); 3122 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3123 } 3124 3125 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3126 Value *Option = EmitScalarExpr(E->getArg(0)); 3127 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3128 } 3129 3130 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3131 Value *Address = EmitScalarExpr(E->getArg(0)); 3132 Value *RW = EmitScalarExpr(E->getArg(1)); 3133 Value *IsData = EmitScalarExpr(E->getArg(2)); 3134 3135 // Locality is not supported on ARM target 3136 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3137 3138 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3139 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3140 } 3141 3142 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3143 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3144 EmitScalarExpr(E->getArg(0)), 3145 "rbit"); 3146 } 3147 3148 if (BuiltinID == ARM::BI__clear_cache) { 3149 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3150 const FunctionDecl *FD = E->getDirectCallee(); 3151 SmallVector<Value*, 2> Ops; 3152 for (unsigned i = 0; i < 2; i++) 3153 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3154 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3155 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3156 StringRef Name = FD->getName(); 3157 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3158 } 3159 3160 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3161 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3162 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3163 getContext().getTypeSize(E->getType()) == 64) || 3164 BuiltinID == ARM::BI__ldrexd) { 3165 Function *F; 3166 3167 switch (BuiltinID) { 3168 default: llvm_unreachable("unexpected builtin"); 3169 case ARM::BI__builtin_arm_ldaex: 3170 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3171 break; 3172 case ARM::BI__builtin_arm_ldrexd: 3173 case ARM::BI__builtin_arm_ldrex: 3174 case ARM::BI__ldrexd: 3175 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3176 break; 3177 } 3178 3179 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3180 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3181 "ldrexd"); 3182 3183 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3184 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3185 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3186 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3187 3188 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3189 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3190 Val = Builder.CreateOr(Val, Val1); 3191 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3192 } 3193 3194 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3195 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3196 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3197 3198 QualType Ty = E->getType(); 3199 llvm::Type *RealResTy = ConvertType(Ty); 3200 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3201 getContext().getTypeSize(Ty)); 3202 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3203 3204 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3205 ? Intrinsic::arm_ldaex 3206 : Intrinsic::arm_ldrex, 3207 LoadAddr->getType()); 3208 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3209 3210 if (RealResTy->isPointerTy()) 3211 return Builder.CreateIntToPtr(Val, RealResTy); 3212 else { 3213 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3214 return Builder.CreateBitCast(Val, RealResTy); 3215 } 3216 } 3217 3218 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3219 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3220 BuiltinID == ARM::BI__builtin_arm_strex) && 3221 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3222 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3223 ? Intrinsic::arm_stlexd 3224 : Intrinsic::arm_strexd); 3225 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 3226 3227 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3228 Value *Val = EmitScalarExpr(E->getArg(0)); 3229 Builder.CreateStore(Val, Tmp); 3230 3231 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3232 Val = Builder.CreateLoad(LdPtr); 3233 3234 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3235 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3236 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3237 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3238 } 3239 3240 if (BuiltinID == ARM::BI__builtin_arm_strex || 3241 BuiltinID == ARM::BI__builtin_arm_stlex) { 3242 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3243 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3244 3245 QualType Ty = E->getArg(0)->getType(); 3246 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3247 getContext().getTypeSize(Ty)); 3248 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3249 3250 if (StoreVal->getType()->isPointerTy()) 3251 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3252 else { 3253 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3254 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3255 } 3256 3257 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3258 ? Intrinsic::arm_stlex 3259 : Intrinsic::arm_strex, 3260 StoreAddr->getType()); 3261 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3262 } 3263 3264 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3265 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3266 return Builder.CreateCall(F); 3267 } 3268 3269 // CRC32 3270 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3271 switch (BuiltinID) { 3272 case ARM::BI__builtin_arm_crc32b: 3273 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3274 case ARM::BI__builtin_arm_crc32cb: 3275 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3276 case ARM::BI__builtin_arm_crc32h: 3277 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3278 case ARM::BI__builtin_arm_crc32ch: 3279 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3280 case ARM::BI__builtin_arm_crc32w: 3281 case ARM::BI__builtin_arm_crc32d: 3282 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3283 case ARM::BI__builtin_arm_crc32cw: 3284 case ARM::BI__builtin_arm_crc32cd: 3285 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3286 } 3287 3288 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3289 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3290 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3291 3292 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3293 // intrinsics, hence we need different codegen for these cases. 3294 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3295 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3296 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3297 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3298 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3299 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3300 3301 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3302 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3303 return Builder.CreateCall2(F, Res, Arg1b); 3304 } else { 3305 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3306 3307 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3308 return Builder.CreateCall2(F, Arg0, Arg1); 3309 } 3310 } 3311 3312 SmallVector<Value*, 4> Ops; 3313 llvm::Value *Align = nullptr; 3314 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3315 if (i == 0) { 3316 switch (BuiltinID) { 3317 case NEON::BI__builtin_neon_vld1_v: 3318 case NEON::BI__builtin_neon_vld1q_v: 3319 case NEON::BI__builtin_neon_vld1q_lane_v: 3320 case NEON::BI__builtin_neon_vld1_lane_v: 3321 case NEON::BI__builtin_neon_vld1_dup_v: 3322 case NEON::BI__builtin_neon_vld1q_dup_v: 3323 case NEON::BI__builtin_neon_vst1_v: 3324 case NEON::BI__builtin_neon_vst1q_v: 3325 case NEON::BI__builtin_neon_vst1q_lane_v: 3326 case NEON::BI__builtin_neon_vst1_lane_v: 3327 case NEON::BI__builtin_neon_vst2_v: 3328 case NEON::BI__builtin_neon_vst2q_v: 3329 case NEON::BI__builtin_neon_vst2_lane_v: 3330 case NEON::BI__builtin_neon_vst2q_lane_v: 3331 case NEON::BI__builtin_neon_vst3_v: 3332 case NEON::BI__builtin_neon_vst3q_v: 3333 case NEON::BI__builtin_neon_vst3_lane_v: 3334 case NEON::BI__builtin_neon_vst3q_lane_v: 3335 case NEON::BI__builtin_neon_vst4_v: 3336 case NEON::BI__builtin_neon_vst4q_v: 3337 case NEON::BI__builtin_neon_vst4_lane_v: 3338 case NEON::BI__builtin_neon_vst4q_lane_v: 3339 // Get the alignment for the argument in addition to the value; 3340 // we'll use it later. 3341 std::pair<llvm::Value*, unsigned> Src = 3342 EmitPointerWithAlignment(E->getArg(0)); 3343 Ops.push_back(Src.first); 3344 Align = Builder.getInt32(Src.second); 3345 continue; 3346 } 3347 } 3348 if (i == 1) { 3349 switch (BuiltinID) { 3350 case NEON::BI__builtin_neon_vld2_v: 3351 case NEON::BI__builtin_neon_vld2q_v: 3352 case NEON::BI__builtin_neon_vld3_v: 3353 case NEON::BI__builtin_neon_vld3q_v: 3354 case NEON::BI__builtin_neon_vld4_v: 3355 case NEON::BI__builtin_neon_vld4q_v: 3356 case NEON::BI__builtin_neon_vld2_lane_v: 3357 case NEON::BI__builtin_neon_vld2q_lane_v: 3358 case NEON::BI__builtin_neon_vld3_lane_v: 3359 case NEON::BI__builtin_neon_vld3q_lane_v: 3360 case NEON::BI__builtin_neon_vld4_lane_v: 3361 case NEON::BI__builtin_neon_vld4q_lane_v: 3362 case NEON::BI__builtin_neon_vld2_dup_v: 3363 case NEON::BI__builtin_neon_vld3_dup_v: 3364 case NEON::BI__builtin_neon_vld4_dup_v: 3365 // Get the alignment for the argument in addition to the value; 3366 // we'll use it later. 3367 std::pair<llvm::Value*, unsigned> Src = 3368 EmitPointerWithAlignment(E->getArg(1)); 3369 Ops.push_back(Src.first); 3370 Align = Builder.getInt32(Src.second); 3371 continue; 3372 } 3373 } 3374 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3375 } 3376 3377 switch (BuiltinID) { 3378 default: break; 3379 // vget_lane and vset_lane are not overloaded and do not have an extra 3380 // argument that specifies the vector type. 3381 case NEON::BI__builtin_neon_vget_lane_i8: 3382 case NEON::BI__builtin_neon_vget_lane_i16: 3383 case NEON::BI__builtin_neon_vget_lane_i32: 3384 case NEON::BI__builtin_neon_vget_lane_i64: 3385 case NEON::BI__builtin_neon_vget_lane_f32: 3386 case NEON::BI__builtin_neon_vgetq_lane_i8: 3387 case NEON::BI__builtin_neon_vgetq_lane_i16: 3388 case NEON::BI__builtin_neon_vgetq_lane_i32: 3389 case NEON::BI__builtin_neon_vgetq_lane_i64: 3390 case NEON::BI__builtin_neon_vgetq_lane_f32: 3391 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3392 "vget_lane"); 3393 case NEON::BI__builtin_neon_vset_lane_i8: 3394 case NEON::BI__builtin_neon_vset_lane_i16: 3395 case NEON::BI__builtin_neon_vset_lane_i32: 3396 case NEON::BI__builtin_neon_vset_lane_i64: 3397 case NEON::BI__builtin_neon_vset_lane_f32: 3398 case NEON::BI__builtin_neon_vsetq_lane_i8: 3399 case NEON::BI__builtin_neon_vsetq_lane_i16: 3400 case NEON::BI__builtin_neon_vsetq_lane_i32: 3401 case NEON::BI__builtin_neon_vsetq_lane_i64: 3402 case NEON::BI__builtin_neon_vsetq_lane_f32: 3403 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3404 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3405 3406 // Non-polymorphic crypto instructions also not overloaded 3407 case NEON::BI__builtin_neon_vsha1h_u32: 3408 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3409 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3410 "vsha1h"); 3411 case NEON::BI__builtin_neon_vsha1cq_u32: 3412 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3413 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3414 "vsha1h"); 3415 case NEON::BI__builtin_neon_vsha1pq_u32: 3416 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3417 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3418 "vsha1h"); 3419 case NEON::BI__builtin_neon_vsha1mq_u32: 3420 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3421 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3422 "vsha1h"); 3423 } 3424 3425 // Get the last argument, which specifies the vector type. 3426 llvm::APSInt Result; 3427 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3428 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3429 return nullptr; 3430 3431 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3432 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3433 // Determine the overloaded type of this builtin. 3434 llvm::Type *Ty; 3435 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3436 Ty = FloatTy; 3437 else 3438 Ty = DoubleTy; 3439 3440 // Determine whether this is an unsigned conversion or not. 3441 bool usgn = Result.getZExtValue() == 1; 3442 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3443 3444 // Call the appropriate intrinsic. 3445 Function *F = CGM.getIntrinsic(Int, Ty); 3446 return Builder.CreateCall(F, Ops, "vcvtr"); 3447 } 3448 3449 // Determine the type of this overloaded NEON intrinsic. 3450 NeonTypeFlags Type(Result.getZExtValue()); 3451 bool usgn = Type.isUnsigned(); 3452 bool rightShift = false; 3453 3454 llvm::VectorType *VTy = GetNeonType(this, Type); 3455 llvm::Type *Ty = VTy; 3456 if (!Ty) 3457 return nullptr; 3458 3459 // Many NEON builtins have identical semantics and uses in ARM and 3460 // AArch64. Emit these in a single function. 3461 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3462 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3463 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3464 if (Builtin) 3465 return EmitCommonNeonBuiltinExpr( 3466 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3467 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3468 3469 unsigned Int; 3470 switch (BuiltinID) { 3471 default: return nullptr; 3472 case NEON::BI__builtin_neon_vld1q_lane_v: 3473 // Handle 64-bit integer elements as a special case. Use shuffles of 3474 // one-element vectors to avoid poor code for i64 in the backend. 3475 if (VTy->getElementType()->isIntegerTy(64)) { 3476 // Extract the other lane. 3477 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3478 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3479 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3480 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3481 // Load the value as a one-element vector. 3482 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3483 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3484 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3485 // Combine them. 3486 SmallVector<Constant*, 2> Indices; 3487 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3488 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3489 SV = llvm::ConstantVector::get(Indices); 3490 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3491 } 3492 // fall through 3493 case NEON::BI__builtin_neon_vld1_lane_v: { 3494 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3495 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3496 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3497 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3498 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3499 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3500 } 3501 case NEON::BI__builtin_neon_vld2_dup_v: 3502 case NEON::BI__builtin_neon_vld3_dup_v: 3503 case NEON::BI__builtin_neon_vld4_dup_v: { 3504 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3505 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3506 switch (BuiltinID) { 3507 case NEON::BI__builtin_neon_vld2_dup_v: 3508 Int = Intrinsic::arm_neon_vld2; 3509 break; 3510 case NEON::BI__builtin_neon_vld3_dup_v: 3511 Int = Intrinsic::arm_neon_vld3; 3512 break; 3513 case NEON::BI__builtin_neon_vld4_dup_v: 3514 Int = Intrinsic::arm_neon_vld4; 3515 break; 3516 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3517 } 3518 Function *F = CGM.getIntrinsic(Int, Ty); 3519 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3520 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3521 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3522 return Builder.CreateStore(Ops[1], Ops[0]); 3523 } 3524 switch (BuiltinID) { 3525 case NEON::BI__builtin_neon_vld2_dup_v: 3526 Int = Intrinsic::arm_neon_vld2lane; 3527 break; 3528 case NEON::BI__builtin_neon_vld3_dup_v: 3529 Int = Intrinsic::arm_neon_vld3lane; 3530 break; 3531 case NEON::BI__builtin_neon_vld4_dup_v: 3532 Int = Intrinsic::arm_neon_vld4lane; 3533 break; 3534 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3535 } 3536 Function *F = CGM.getIntrinsic(Int, Ty); 3537 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3538 3539 SmallVector<Value*, 6> Args; 3540 Args.push_back(Ops[1]); 3541 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3542 3543 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3544 Args.push_back(CI); 3545 Args.push_back(Align); 3546 3547 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3548 // splat lane 0 to all elts in each vector of the result. 3549 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3550 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3551 Value *Elt = Builder.CreateBitCast(Val, Ty); 3552 Elt = EmitNeonSplat(Elt, CI); 3553 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3554 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3555 } 3556 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3557 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3558 return Builder.CreateStore(Ops[1], Ops[0]); 3559 } 3560 case NEON::BI__builtin_neon_vqrshrn_n_v: 3561 Int = 3562 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3563 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3564 1, true); 3565 case NEON::BI__builtin_neon_vqrshrun_n_v: 3566 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3567 Ops, "vqrshrun_n", 1, true); 3568 case NEON::BI__builtin_neon_vqshrn_n_v: 3569 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3570 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3571 1, true); 3572 case NEON::BI__builtin_neon_vqshrun_n_v: 3573 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3574 Ops, "vqshrun_n", 1, true); 3575 case NEON::BI__builtin_neon_vrecpe_v: 3576 case NEON::BI__builtin_neon_vrecpeq_v: 3577 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3578 Ops, "vrecpe"); 3579 case NEON::BI__builtin_neon_vrshrn_n_v: 3580 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3581 Ops, "vrshrn_n", 1, true); 3582 case NEON::BI__builtin_neon_vrsra_n_v: 3583 case NEON::BI__builtin_neon_vrsraq_n_v: 3584 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3585 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3586 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3587 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3588 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3589 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3590 case NEON::BI__builtin_neon_vsri_n_v: 3591 case NEON::BI__builtin_neon_vsriq_n_v: 3592 rightShift = true; 3593 case NEON::BI__builtin_neon_vsli_n_v: 3594 case NEON::BI__builtin_neon_vsliq_n_v: 3595 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3596 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3597 Ops, "vsli_n"); 3598 case NEON::BI__builtin_neon_vsra_n_v: 3599 case NEON::BI__builtin_neon_vsraq_n_v: 3600 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3601 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3602 return Builder.CreateAdd(Ops[0], Ops[1]); 3603 case NEON::BI__builtin_neon_vst1q_lane_v: 3604 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3605 // a one-element vector and avoid poor code for i64 in the backend. 3606 if (VTy->getElementType()->isIntegerTy(64)) { 3607 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3608 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3609 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3610 Ops[2] = Align; 3611 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3612 Ops[1]->getType()), Ops); 3613 } 3614 // fall through 3615 case NEON::BI__builtin_neon_vst1_lane_v: { 3616 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3617 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3618 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3619 StoreInst *St = Builder.CreateStore(Ops[1], 3620 Builder.CreateBitCast(Ops[0], Ty)); 3621 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3622 return St; 3623 } 3624 case NEON::BI__builtin_neon_vtbl1_v: 3625 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3626 Ops, "vtbl1"); 3627 case NEON::BI__builtin_neon_vtbl2_v: 3628 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3629 Ops, "vtbl2"); 3630 case NEON::BI__builtin_neon_vtbl3_v: 3631 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3632 Ops, "vtbl3"); 3633 case NEON::BI__builtin_neon_vtbl4_v: 3634 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3635 Ops, "vtbl4"); 3636 case NEON::BI__builtin_neon_vtbx1_v: 3637 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3638 Ops, "vtbx1"); 3639 case NEON::BI__builtin_neon_vtbx2_v: 3640 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3641 Ops, "vtbx2"); 3642 case NEON::BI__builtin_neon_vtbx3_v: 3643 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3644 Ops, "vtbx3"); 3645 case NEON::BI__builtin_neon_vtbx4_v: 3646 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3647 Ops, "vtbx4"); 3648 } 3649 } 3650 3651 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3652 const CallExpr *E, 3653 SmallVectorImpl<Value *> &Ops) { 3654 unsigned int Int = 0; 3655 const char *s = nullptr; 3656 3657 switch (BuiltinID) { 3658 default: 3659 return nullptr; 3660 case NEON::BI__builtin_neon_vtbl1_v: 3661 case NEON::BI__builtin_neon_vqtbl1_v: 3662 case NEON::BI__builtin_neon_vqtbl1q_v: 3663 case NEON::BI__builtin_neon_vtbl2_v: 3664 case NEON::BI__builtin_neon_vqtbl2_v: 3665 case NEON::BI__builtin_neon_vqtbl2q_v: 3666 case NEON::BI__builtin_neon_vtbl3_v: 3667 case NEON::BI__builtin_neon_vqtbl3_v: 3668 case NEON::BI__builtin_neon_vqtbl3q_v: 3669 case NEON::BI__builtin_neon_vtbl4_v: 3670 case NEON::BI__builtin_neon_vqtbl4_v: 3671 case NEON::BI__builtin_neon_vqtbl4q_v: 3672 break; 3673 case NEON::BI__builtin_neon_vtbx1_v: 3674 case NEON::BI__builtin_neon_vqtbx1_v: 3675 case NEON::BI__builtin_neon_vqtbx1q_v: 3676 case NEON::BI__builtin_neon_vtbx2_v: 3677 case NEON::BI__builtin_neon_vqtbx2_v: 3678 case NEON::BI__builtin_neon_vqtbx2q_v: 3679 case NEON::BI__builtin_neon_vtbx3_v: 3680 case NEON::BI__builtin_neon_vqtbx3_v: 3681 case NEON::BI__builtin_neon_vqtbx3q_v: 3682 case NEON::BI__builtin_neon_vtbx4_v: 3683 case NEON::BI__builtin_neon_vqtbx4_v: 3684 case NEON::BI__builtin_neon_vqtbx4q_v: 3685 break; 3686 } 3687 3688 assert(E->getNumArgs() >= 3); 3689 3690 // Get the last argument, which specifies the vector type. 3691 llvm::APSInt Result; 3692 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3693 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3694 return nullptr; 3695 3696 // Determine the type of this overloaded NEON intrinsic. 3697 NeonTypeFlags Type(Result.getZExtValue()); 3698 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3699 llvm::Type *Ty = VTy; 3700 if (!Ty) 3701 return nullptr; 3702 3703 unsigned nElts = VTy->getNumElements(); 3704 3705 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3706 3707 // AArch64 scalar builtins are not overloaded, they do not have an extra 3708 // argument that specifies the vector type, need to handle each case. 3709 SmallVector<Value *, 2> TblOps; 3710 switch (BuiltinID) { 3711 case NEON::BI__builtin_neon_vtbl1_v: { 3712 TblOps.push_back(Ops[0]); 3713 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3714 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3715 } 3716 case NEON::BI__builtin_neon_vtbl2_v: { 3717 TblOps.push_back(Ops[0]); 3718 TblOps.push_back(Ops[1]); 3719 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3720 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3721 } 3722 case NEON::BI__builtin_neon_vtbl3_v: { 3723 TblOps.push_back(Ops[0]); 3724 TblOps.push_back(Ops[1]); 3725 TblOps.push_back(Ops[2]); 3726 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3727 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3728 } 3729 case NEON::BI__builtin_neon_vtbl4_v: { 3730 TblOps.push_back(Ops[0]); 3731 TblOps.push_back(Ops[1]); 3732 TblOps.push_back(Ops[2]); 3733 TblOps.push_back(Ops[3]); 3734 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3735 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3736 } 3737 case NEON::BI__builtin_neon_vtbx1_v: { 3738 TblOps.push_back(Ops[1]); 3739 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3740 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3741 3742 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3743 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3744 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3745 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3746 3747 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3748 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3749 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3750 } 3751 case NEON::BI__builtin_neon_vtbx2_v: { 3752 TblOps.push_back(Ops[1]); 3753 TblOps.push_back(Ops[2]); 3754 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3755 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3756 } 3757 case NEON::BI__builtin_neon_vtbx3_v: { 3758 TblOps.push_back(Ops[1]); 3759 TblOps.push_back(Ops[2]); 3760 TblOps.push_back(Ops[3]); 3761 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3762 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3763 3764 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3765 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3766 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3767 TwentyFourV); 3768 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3769 3770 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3771 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3772 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3773 } 3774 case NEON::BI__builtin_neon_vtbx4_v: { 3775 TblOps.push_back(Ops[1]); 3776 TblOps.push_back(Ops[2]); 3777 TblOps.push_back(Ops[3]); 3778 TblOps.push_back(Ops[4]); 3779 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3780 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3781 } 3782 case NEON::BI__builtin_neon_vqtbl1_v: 3783 case NEON::BI__builtin_neon_vqtbl1q_v: 3784 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3785 case NEON::BI__builtin_neon_vqtbl2_v: 3786 case NEON::BI__builtin_neon_vqtbl2q_v: { 3787 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3788 case NEON::BI__builtin_neon_vqtbl3_v: 3789 case NEON::BI__builtin_neon_vqtbl3q_v: 3790 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3791 case NEON::BI__builtin_neon_vqtbl4_v: 3792 case NEON::BI__builtin_neon_vqtbl4q_v: 3793 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3794 case NEON::BI__builtin_neon_vqtbx1_v: 3795 case NEON::BI__builtin_neon_vqtbx1q_v: 3796 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3797 case NEON::BI__builtin_neon_vqtbx2_v: 3798 case NEON::BI__builtin_neon_vqtbx2q_v: 3799 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3800 case NEON::BI__builtin_neon_vqtbx3_v: 3801 case NEON::BI__builtin_neon_vqtbx3q_v: 3802 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3803 case NEON::BI__builtin_neon_vqtbx4_v: 3804 case NEON::BI__builtin_neon_vqtbx4q_v: 3805 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3806 } 3807 } 3808 3809 if (!Int) 3810 return nullptr; 3811 3812 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3813 return CGF.EmitNeonCall(F, Ops, s); 3814 } 3815 3816 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3817 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3818 Op = Builder.CreateBitCast(Op, Int16Ty); 3819 Value *V = UndefValue::get(VTy); 3820 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3821 Op = Builder.CreateInsertElement(V, Op, CI); 3822 return Op; 3823 } 3824 3825 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3826 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3827 Op = Builder.CreateBitCast(Op, Int8Ty); 3828 Value *V = UndefValue::get(VTy); 3829 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3830 Op = Builder.CreateInsertElement(V, Op, CI); 3831 return Op; 3832 } 3833 3834 Value *CodeGenFunction:: 3835 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3836 const char *Name) { 3837 // i8 is not a legal types for AArch64, so we can't just use 3838 // a normal overloaded intrinsic call for these scalar types. Instead 3839 // we'll build 64-bit vectors w/ lane zero being our input values and 3840 // perform the operation on that. The back end can pattern match directly 3841 // to the scalar instruction. 3842 Ops[0] = vectorWrapScalar8(Ops[0]); 3843 Ops[1] = vectorWrapScalar8(Ops[1]); 3844 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3845 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3846 Constant *CI = ConstantInt::get(SizeTy, 0); 3847 return Builder.CreateExtractElement(V, CI, "lane0"); 3848 } 3849 3850 Value *CodeGenFunction:: 3851 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3852 const char *Name) { 3853 // i16 is not a legal types for AArch64, so we can't just use 3854 // a normal overloaded intrinsic call for these scalar types. Instead 3855 // we'll build 64-bit vectors w/ lane zero being our input values and 3856 // perform the operation on that. The back end can pattern match directly 3857 // to the scalar instruction. 3858 Ops[0] = vectorWrapScalar16(Ops[0]); 3859 Ops[1] = vectorWrapScalar16(Ops[1]); 3860 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3861 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3862 Constant *CI = ConstantInt::get(SizeTy, 0); 3863 return Builder.CreateExtractElement(V, CI, "lane0"); 3864 } 3865 3866 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3867 const CallExpr *E) { 3868 unsigned HintID = static_cast<unsigned>(-1); 3869 switch (BuiltinID) { 3870 default: break; 3871 case AArch64::BI__builtin_arm_nop: 3872 HintID = 0; 3873 break; 3874 case AArch64::BI__builtin_arm_yield: 3875 HintID = 1; 3876 break; 3877 case AArch64::BI__builtin_arm_wfe: 3878 HintID = 2; 3879 break; 3880 case AArch64::BI__builtin_arm_wfi: 3881 HintID = 3; 3882 break; 3883 case AArch64::BI__builtin_arm_sev: 3884 HintID = 4; 3885 break; 3886 case AArch64::BI__builtin_arm_sevl: 3887 HintID = 5; 3888 break; 3889 } 3890 3891 if (HintID != static_cast<unsigned>(-1)) { 3892 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 3893 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3894 } 3895 3896 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 3897 Value *Address = EmitScalarExpr(E->getArg(0)); 3898 Value *RW = EmitScalarExpr(E->getArg(1)); 3899 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 3900 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 3901 Value *IsData = EmitScalarExpr(E->getArg(4)); 3902 3903 Value *Locality = nullptr; 3904 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 3905 // Temporal fetch, needs to convert cache level to locality. 3906 Locality = llvm::ConstantInt::get(Int32Ty, 3907 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 3908 } else { 3909 // Streaming fetch. 3910 Locality = llvm::ConstantInt::get(Int32Ty, 0); 3911 } 3912 3913 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 3914 // PLDL3STRM or PLDL2STRM. 3915 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3916 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3917 } 3918 3919 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 3920 assert((getContext().getTypeSize(E->getType()) == 32) && 3921 "rbit of unusual size!"); 3922 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3923 return Builder.CreateCall( 3924 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3925 } 3926 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 3927 assert((getContext().getTypeSize(E->getType()) == 64) && 3928 "rbit of unusual size!"); 3929 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3930 return Builder.CreateCall( 3931 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3932 } 3933 3934 if (BuiltinID == AArch64::BI__clear_cache) { 3935 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3936 const FunctionDecl *FD = E->getDirectCallee(); 3937 SmallVector<Value*, 2> Ops; 3938 for (unsigned i = 0; i < 2; i++) 3939 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3940 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3941 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3942 StringRef Name = FD->getName(); 3943 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3944 } 3945 3946 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 3947 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 3948 getContext().getTypeSize(E->getType()) == 128) { 3949 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 3950 ? Intrinsic::aarch64_ldaxp 3951 : Intrinsic::aarch64_ldxp); 3952 3953 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3954 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3955 "ldxp"); 3956 3957 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3958 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3959 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 3960 Val0 = Builder.CreateZExt(Val0, Int128Ty); 3961 Val1 = Builder.CreateZExt(Val1, Int128Ty); 3962 3963 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 3964 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3965 Val = Builder.CreateOr(Val, Val1); 3966 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3967 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 3968 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 3969 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3970 3971 QualType Ty = E->getType(); 3972 llvm::Type *RealResTy = ConvertType(Ty); 3973 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3974 getContext().getTypeSize(Ty)); 3975 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3976 3977 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 3978 ? Intrinsic::aarch64_ldaxr 3979 : Intrinsic::aarch64_ldxr, 3980 LoadAddr->getType()); 3981 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 3982 3983 if (RealResTy->isPointerTy()) 3984 return Builder.CreateIntToPtr(Val, RealResTy); 3985 3986 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3987 return Builder.CreateBitCast(Val, RealResTy); 3988 } 3989 3990 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 3991 BuiltinID == AArch64::BI__builtin_arm_stlex) && 3992 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 3993 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 3994 ? Intrinsic::aarch64_stlxp 3995 : Intrinsic::aarch64_stxp); 3996 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, NULL); 3997 3998 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 3999 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4000 One); 4001 Value *Val = EmitScalarExpr(E->getArg(0)); 4002 Builder.CreateStore(Val, Tmp); 4003 4004 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4005 Val = Builder.CreateLoad(LdPtr); 4006 4007 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4008 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4009 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4010 Int8PtrTy); 4011 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 4012 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 4013 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4014 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4015 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4016 4017 QualType Ty = E->getArg(0)->getType(); 4018 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4019 getContext().getTypeSize(Ty)); 4020 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4021 4022 if (StoreVal->getType()->isPointerTy()) 4023 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4024 else { 4025 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4026 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4027 } 4028 4029 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4030 ? Intrinsic::aarch64_stlxr 4031 : Intrinsic::aarch64_stxr, 4032 StoreAddr->getType()); 4033 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4034 } 4035 4036 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4037 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4038 return Builder.CreateCall(F); 4039 } 4040 4041 // CRC32 4042 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4043 switch (BuiltinID) { 4044 case AArch64::BI__builtin_arm_crc32b: 4045 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4046 case AArch64::BI__builtin_arm_crc32cb: 4047 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4048 case AArch64::BI__builtin_arm_crc32h: 4049 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4050 case AArch64::BI__builtin_arm_crc32ch: 4051 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4052 case AArch64::BI__builtin_arm_crc32w: 4053 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4054 case AArch64::BI__builtin_arm_crc32cw: 4055 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4056 case AArch64::BI__builtin_arm_crc32d: 4057 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4058 case AArch64::BI__builtin_arm_crc32cd: 4059 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4060 } 4061 4062 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4063 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4064 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4065 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4066 4067 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4068 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4069 4070 return Builder.CreateCall2(F, Arg0, Arg1); 4071 } 4072 4073 llvm::SmallVector<Value*, 4> Ops; 4074 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4075 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4076 4077 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4078 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4079 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4080 4081 if (Builtin) { 4082 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4083 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4084 assert(Result && "SISD intrinsic should have been handled"); 4085 return Result; 4086 } 4087 4088 llvm::APSInt Result; 4089 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4090 NeonTypeFlags Type(0); 4091 if (Arg->isIntegerConstantExpr(Result, getContext())) 4092 // Determine the type of this overloaded NEON intrinsic. 4093 Type = NeonTypeFlags(Result.getZExtValue()); 4094 4095 bool usgn = Type.isUnsigned(); 4096 bool quad = Type.isQuad(); 4097 4098 // Handle non-overloaded intrinsics first. 4099 switch (BuiltinID) { 4100 default: break; 4101 case NEON::BI__builtin_neon_vldrq_p128: { 4102 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4103 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4104 return Builder.CreateLoad(Ptr); 4105 } 4106 case NEON::BI__builtin_neon_vstrq_p128: { 4107 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4108 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4109 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4110 } 4111 case NEON::BI__builtin_neon_vcvts_u32_f32: 4112 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4113 usgn = true; 4114 // FALL THROUGH 4115 case NEON::BI__builtin_neon_vcvts_s32_f32: 4116 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4117 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4118 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4119 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4120 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4121 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4122 if (usgn) 4123 return Builder.CreateFPToUI(Ops[0], InTy); 4124 return Builder.CreateFPToSI(Ops[0], InTy); 4125 } 4126 case NEON::BI__builtin_neon_vcvts_f32_u32: 4127 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4128 usgn = true; 4129 // FALL THROUGH 4130 case NEON::BI__builtin_neon_vcvts_f32_s32: 4131 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4132 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4133 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4134 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4135 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4136 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4137 if (usgn) 4138 return Builder.CreateUIToFP(Ops[0], FTy); 4139 return Builder.CreateSIToFP(Ops[0], FTy); 4140 } 4141 case NEON::BI__builtin_neon_vpaddd_s64: { 4142 llvm::Type *Ty = 4143 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4144 Value *Vec = EmitScalarExpr(E->getArg(0)); 4145 // The vector is v2f64, so make sure it's bitcast to that. 4146 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4147 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4148 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4149 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4150 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4151 // Pairwise addition of a v2f64 into a scalar f64. 4152 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4153 } 4154 case NEON::BI__builtin_neon_vpaddd_f64: { 4155 llvm::Type *Ty = 4156 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4157 Value *Vec = EmitScalarExpr(E->getArg(0)); 4158 // The vector is v2f64, so make sure it's bitcast to that. 4159 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4160 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4161 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4162 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4163 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4164 // Pairwise addition of a v2f64 into a scalar f64. 4165 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4166 } 4167 case NEON::BI__builtin_neon_vpadds_f32: { 4168 llvm::Type *Ty = 4169 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4170 Value *Vec = EmitScalarExpr(E->getArg(0)); 4171 // The vector is v2f32, so make sure it's bitcast to that. 4172 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4173 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4174 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4175 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4176 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4177 // Pairwise addition of a v2f32 into a scalar f32. 4178 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4179 } 4180 case NEON::BI__builtin_neon_vceqzd_s64: 4181 case NEON::BI__builtin_neon_vceqzd_f64: 4182 case NEON::BI__builtin_neon_vceqzs_f32: 4183 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4184 return EmitAArch64CompareBuiltinExpr( 4185 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4186 ICmpInst::ICMP_EQ, "vceqz"); 4187 case NEON::BI__builtin_neon_vcgezd_s64: 4188 case NEON::BI__builtin_neon_vcgezd_f64: 4189 case NEON::BI__builtin_neon_vcgezs_f32: 4190 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4191 return EmitAArch64CompareBuiltinExpr( 4192 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4193 ICmpInst::ICMP_SGE, "vcgez"); 4194 case NEON::BI__builtin_neon_vclezd_s64: 4195 case NEON::BI__builtin_neon_vclezd_f64: 4196 case NEON::BI__builtin_neon_vclezs_f32: 4197 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4198 return EmitAArch64CompareBuiltinExpr( 4199 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4200 ICmpInst::ICMP_SLE, "vclez"); 4201 case NEON::BI__builtin_neon_vcgtzd_s64: 4202 case NEON::BI__builtin_neon_vcgtzd_f64: 4203 case NEON::BI__builtin_neon_vcgtzs_f32: 4204 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4205 return EmitAArch64CompareBuiltinExpr( 4206 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4207 ICmpInst::ICMP_SGT, "vcgtz"); 4208 case NEON::BI__builtin_neon_vcltzd_s64: 4209 case NEON::BI__builtin_neon_vcltzd_f64: 4210 case NEON::BI__builtin_neon_vcltzs_f32: 4211 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4212 return EmitAArch64CompareBuiltinExpr( 4213 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4214 ICmpInst::ICMP_SLT, "vcltz"); 4215 4216 case NEON::BI__builtin_neon_vceqzd_u64: { 4217 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4218 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4219 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4220 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4221 llvm::Constant::getNullValue(Ty)); 4222 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4223 } 4224 case NEON::BI__builtin_neon_vceqd_f64: 4225 case NEON::BI__builtin_neon_vcled_f64: 4226 case NEON::BI__builtin_neon_vcltd_f64: 4227 case NEON::BI__builtin_neon_vcged_f64: 4228 case NEON::BI__builtin_neon_vcgtd_f64: { 4229 llvm::CmpInst::Predicate P; 4230 switch (BuiltinID) { 4231 default: llvm_unreachable("missing builtin ID in switch!"); 4232 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4233 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4234 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4235 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4236 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4237 } 4238 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4239 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4240 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4241 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4242 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4243 } 4244 case NEON::BI__builtin_neon_vceqs_f32: 4245 case NEON::BI__builtin_neon_vcles_f32: 4246 case NEON::BI__builtin_neon_vclts_f32: 4247 case NEON::BI__builtin_neon_vcges_f32: 4248 case NEON::BI__builtin_neon_vcgts_f32: { 4249 llvm::CmpInst::Predicate P; 4250 switch (BuiltinID) { 4251 default: llvm_unreachable("missing builtin ID in switch!"); 4252 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4253 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4254 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4255 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4256 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4257 } 4258 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4259 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4260 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4261 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4262 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4263 } 4264 case NEON::BI__builtin_neon_vceqd_s64: 4265 case NEON::BI__builtin_neon_vceqd_u64: 4266 case NEON::BI__builtin_neon_vcgtd_s64: 4267 case NEON::BI__builtin_neon_vcgtd_u64: 4268 case NEON::BI__builtin_neon_vcltd_s64: 4269 case NEON::BI__builtin_neon_vcltd_u64: 4270 case NEON::BI__builtin_neon_vcged_u64: 4271 case NEON::BI__builtin_neon_vcged_s64: 4272 case NEON::BI__builtin_neon_vcled_u64: 4273 case NEON::BI__builtin_neon_vcled_s64: { 4274 llvm::CmpInst::Predicate P; 4275 switch (BuiltinID) { 4276 default: llvm_unreachable("missing builtin ID in switch!"); 4277 case NEON::BI__builtin_neon_vceqd_s64: 4278 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4279 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4280 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4281 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4282 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4283 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4284 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4285 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4286 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4287 } 4288 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4289 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4290 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4291 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4292 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4293 } 4294 case NEON::BI__builtin_neon_vtstd_s64: 4295 case NEON::BI__builtin_neon_vtstd_u64: { 4296 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4297 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4298 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4299 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4300 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4301 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4302 llvm::Constant::getNullValue(Ty)); 4303 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4304 } 4305 case NEON::BI__builtin_neon_vset_lane_i8: 4306 case NEON::BI__builtin_neon_vset_lane_i16: 4307 case NEON::BI__builtin_neon_vset_lane_i32: 4308 case NEON::BI__builtin_neon_vset_lane_i64: 4309 case NEON::BI__builtin_neon_vset_lane_f32: 4310 case NEON::BI__builtin_neon_vsetq_lane_i8: 4311 case NEON::BI__builtin_neon_vsetq_lane_i16: 4312 case NEON::BI__builtin_neon_vsetq_lane_i32: 4313 case NEON::BI__builtin_neon_vsetq_lane_i64: 4314 case NEON::BI__builtin_neon_vsetq_lane_f32: 4315 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4316 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4317 case NEON::BI__builtin_neon_vset_lane_f64: 4318 // The vector type needs a cast for the v1f64 variant. 4319 Ops[1] = Builder.CreateBitCast(Ops[1], 4320 llvm::VectorType::get(DoubleTy, 1)); 4321 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4322 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4323 case NEON::BI__builtin_neon_vsetq_lane_f64: 4324 // The vector type needs a cast for the v2f64 variant. 4325 Ops[1] = Builder.CreateBitCast(Ops[1], 4326 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4327 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4328 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4329 4330 case NEON::BI__builtin_neon_vget_lane_i8: 4331 case NEON::BI__builtin_neon_vdupb_lane_i8: 4332 Ops[0] = Builder.CreateBitCast(Ops[0], 4333 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4334 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4335 "vget_lane"); 4336 case NEON::BI__builtin_neon_vgetq_lane_i8: 4337 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4338 Ops[0] = Builder.CreateBitCast(Ops[0], 4339 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4340 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4341 "vgetq_lane"); 4342 case NEON::BI__builtin_neon_vget_lane_i16: 4343 case NEON::BI__builtin_neon_vduph_lane_i16: 4344 Ops[0] = Builder.CreateBitCast(Ops[0], 4345 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4346 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4347 "vget_lane"); 4348 case NEON::BI__builtin_neon_vgetq_lane_i16: 4349 case NEON::BI__builtin_neon_vduph_laneq_i16: 4350 Ops[0] = Builder.CreateBitCast(Ops[0], 4351 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4352 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4353 "vgetq_lane"); 4354 case NEON::BI__builtin_neon_vget_lane_i32: 4355 case NEON::BI__builtin_neon_vdups_lane_i32: 4356 Ops[0] = Builder.CreateBitCast( 4357 Ops[0], 4358 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4359 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4360 "vget_lane"); 4361 case NEON::BI__builtin_neon_vdups_lane_f32: 4362 Ops[0] = Builder.CreateBitCast(Ops[0], 4363 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4364 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4365 "vdups_lane"); 4366 case NEON::BI__builtin_neon_vgetq_lane_i32: 4367 case NEON::BI__builtin_neon_vdups_laneq_i32: 4368 Ops[0] = Builder.CreateBitCast(Ops[0], 4369 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4370 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4371 "vgetq_lane"); 4372 case NEON::BI__builtin_neon_vget_lane_i64: 4373 case NEON::BI__builtin_neon_vdupd_lane_i64: 4374 Ops[0] = Builder.CreateBitCast(Ops[0], 4375 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4376 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4377 "vget_lane"); 4378 case NEON::BI__builtin_neon_vdupd_lane_f64: 4379 Ops[0] = Builder.CreateBitCast(Ops[0], 4380 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4381 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4382 "vdupd_lane"); 4383 case NEON::BI__builtin_neon_vgetq_lane_i64: 4384 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4385 Ops[0] = Builder.CreateBitCast(Ops[0], 4386 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4387 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4388 "vgetq_lane"); 4389 case NEON::BI__builtin_neon_vget_lane_f32: 4390 Ops[0] = Builder.CreateBitCast(Ops[0], 4391 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4392 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4393 "vget_lane"); 4394 case NEON::BI__builtin_neon_vget_lane_f64: 4395 Ops[0] = Builder.CreateBitCast(Ops[0], 4396 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4397 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4398 "vget_lane"); 4399 case NEON::BI__builtin_neon_vgetq_lane_f32: 4400 case NEON::BI__builtin_neon_vdups_laneq_f32: 4401 Ops[0] = Builder.CreateBitCast(Ops[0], 4402 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4403 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4404 "vgetq_lane"); 4405 case NEON::BI__builtin_neon_vgetq_lane_f64: 4406 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4407 Ops[0] = Builder.CreateBitCast(Ops[0], 4408 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4409 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4410 "vgetq_lane"); 4411 case NEON::BI__builtin_neon_vaddd_s64: 4412 case NEON::BI__builtin_neon_vaddd_u64: 4413 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4414 case NEON::BI__builtin_neon_vsubd_s64: 4415 case NEON::BI__builtin_neon_vsubd_u64: 4416 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4417 case NEON::BI__builtin_neon_vqdmlalh_s16: 4418 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4419 SmallVector<Value *, 2> ProductOps; 4420 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4421 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4422 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4423 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4424 ProductOps, "vqdmlXl"); 4425 Constant *CI = ConstantInt::get(SizeTy, 0); 4426 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4427 4428 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4429 ? Intrinsic::aarch64_neon_sqadd 4430 : Intrinsic::aarch64_neon_sqsub; 4431 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4432 } 4433 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4434 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4435 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4436 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4437 Ops, "vqshlu_n"); 4438 } 4439 case NEON::BI__builtin_neon_vqshld_n_u64: 4440 case NEON::BI__builtin_neon_vqshld_n_s64: { 4441 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4442 ? Intrinsic::aarch64_neon_uqshl 4443 : Intrinsic::aarch64_neon_sqshl; 4444 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4445 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4446 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4447 } 4448 case NEON::BI__builtin_neon_vrshrd_n_u64: 4449 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4450 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4451 ? Intrinsic::aarch64_neon_urshl 4452 : Intrinsic::aarch64_neon_srshl; 4453 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4454 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4455 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4456 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4457 } 4458 case NEON::BI__builtin_neon_vrsrad_n_u64: 4459 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4460 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4461 ? Intrinsic::aarch64_neon_urshl 4462 : Intrinsic::aarch64_neon_srshl; 4463 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4464 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4465 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4466 Builder.CreateSExt(Ops[2], Int64Ty)); 4467 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4468 } 4469 case NEON::BI__builtin_neon_vshld_n_s64: 4470 case NEON::BI__builtin_neon_vshld_n_u64: { 4471 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4472 return Builder.CreateShl( 4473 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4474 } 4475 case NEON::BI__builtin_neon_vshrd_n_s64: { 4476 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4477 return Builder.CreateAShr( 4478 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4479 Amt->getZExtValue())), 4480 "shrd_n"); 4481 } 4482 case NEON::BI__builtin_neon_vshrd_n_u64: { 4483 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4484 uint64_t ShiftAmt = Amt->getZExtValue(); 4485 // Right-shifting an unsigned value by its size yields 0. 4486 if (ShiftAmt == 64) 4487 return ConstantInt::get(Int64Ty, 0); 4488 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4489 "shrd_n"); 4490 } 4491 case NEON::BI__builtin_neon_vsrad_n_s64: { 4492 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4493 Ops[1] = Builder.CreateAShr( 4494 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4495 Amt->getZExtValue())), 4496 "shrd_n"); 4497 return Builder.CreateAdd(Ops[0], Ops[1]); 4498 } 4499 case NEON::BI__builtin_neon_vsrad_n_u64: { 4500 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4501 uint64_t ShiftAmt = Amt->getZExtValue(); 4502 // Right-shifting an unsigned value by its size yields 0. 4503 // As Op + 0 = Op, return Ops[0] directly. 4504 if (ShiftAmt == 64) 4505 return Ops[0]; 4506 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4507 "shrd_n"); 4508 return Builder.CreateAdd(Ops[0], Ops[1]); 4509 } 4510 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4511 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4512 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4513 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4514 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4515 "lane"); 4516 SmallVector<Value *, 2> ProductOps; 4517 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4518 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4519 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4520 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4521 ProductOps, "vqdmlXl"); 4522 Constant *CI = ConstantInt::get(SizeTy, 0); 4523 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4524 Ops.pop_back(); 4525 4526 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4527 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4528 ? Intrinsic::aarch64_neon_sqadd 4529 : Intrinsic::aarch64_neon_sqsub; 4530 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4531 } 4532 case NEON::BI__builtin_neon_vqdmlals_s32: 4533 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4534 SmallVector<Value *, 2> ProductOps; 4535 ProductOps.push_back(Ops[1]); 4536 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4537 Ops[1] = 4538 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4539 ProductOps, "vqdmlXl"); 4540 4541 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4542 ? Intrinsic::aarch64_neon_sqadd 4543 : Intrinsic::aarch64_neon_sqsub; 4544 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4545 } 4546 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4547 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4548 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4549 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4550 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4551 "lane"); 4552 SmallVector<Value *, 2> ProductOps; 4553 ProductOps.push_back(Ops[1]); 4554 ProductOps.push_back(Ops[2]); 4555 Ops[1] = 4556 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4557 ProductOps, "vqdmlXl"); 4558 Ops.pop_back(); 4559 4560 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4561 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4562 ? Intrinsic::aarch64_neon_sqadd 4563 : Intrinsic::aarch64_neon_sqsub; 4564 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4565 } 4566 } 4567 4568 llvm::VectorType *VTy = GetNeonType(this, Type); 4569 llvm::Type *Ty = VTy; 4570 if (!Ty) 4571 return nullptr; 4572 4573 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4574 // defer to common code if it's been added to our special map. 4575 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4576 AArch64SIMDIntrinsicsProvenSorted); 4577 4578 if (Builtin) 4579 return EmitCommonNeonBuiltinExpr( 4580 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4581 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4582 4583 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4584 return V; 4585 4586 unsigned Int; 4587 switch (BuiltinID) { 4588 default: return nullptr; 4589 case NEON::BI__builtin_neon_vbsl_v: 4590 case NEON::BI__builtin_neon_vbslq_v: { 4591 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4592 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4593 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4594 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4595 4596 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4597 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4598 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4599 return Builder.CreateBitCast(Ops[0], Ty); 4600 } 4601 case NEON::BI__builtin_neon_vfma_lane_v: 4602 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4603 // The ARM builtins (and instructions) have the addend as the first 4604 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4605 Value *Addend = Ops[0]; 4606 Value *Multiplicand = Ops[1]; 4607 Value *LaneSource = Ops[2]; 4608 Ops[0] = Multiplicand; 4609 Ops[1] = LaneSource; 4610 Ops[2] = Addend; 4611 4612 // Now adjust things to handle the lane access. 4613 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4614 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4615 VTy; 4616 llvm::Constant *cst = cast<Constant>(Ops[3]); 4617 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4618 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4619 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4620 4621 Ops.pop_back(); 4622 Int = Intrinsic::fma; 4623 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4624 } 4625 case NEON::BI__builtin_neon_vfma_laneq_v: { 4626 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4627 // v1f64 fma should be mapped to Neon scalar f64 fma 4628 if (VTy && VTy->getElementType() == DoubleTy) { 4629 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4630 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4631 llvm::Type *VTy = GetNeonType(this, 4632 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4633 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4634 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4635 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4636 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4637 return Builder.CreateBitCast(Result, Ty); 4638 } 4639 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4640 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4641 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4642 4643 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4644 VTy->getNumElements() * 2); 4645 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4646 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4647 cast<ConstantInt>(Ops[3])); 4648 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4649 4650 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4651 } 4652 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4653 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4654 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4655 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4656 4657 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4658 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4659 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4660 } 4661 case NEON::BI__builtin_neon_vfmas_lane_f32: 4662 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4663 case NEON::BI__builtin_neon_vfmad_lane_f64: 4664 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4665 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4666 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4667 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4668 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4669 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4670 } 4671 case NEON::BI__builtin_neon_vfms_v: 4672 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4673 // FIXME: probably remove when we no longer support aarch64_simd.h 4674 // (arm_neon.h delegates to vfma). 4675 4676 // The ARM builtins (and instructions) have the addend as the first 4677 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4678 Value *Subtrahend = Ops[0]; 4679 Value *Multiplicand = Ops[2]; 4680 Ops[0] = Multiplicand; 4681 Ops[2] = Subtrahend; 4682 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4683 Ops[1] = Builder.CreateFNeg(Ops[1]); 4684 Int = Intrinsic::fma; 4685 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4686 } 4687 case NEON::BI__builtin_neon_vmull_v: 4688 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4689 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4690 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4691 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4692 case NEON::BI__builtin_neon_vmax_v: 4693 case NEON::BI__builtin_neon_vmaxq_v: 4694 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4695 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4696 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4697 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4698 case NEON::BI__builtin_neon_vmin_v: 4699 case NEON::BI__builtin_neon_vminq_v: 4700 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4701 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4702 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4703 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4704 case NEON::BI__builtin_neon_vabd_v: 4705 case NEON::BI__builtin_neon_vabdq_v: 4706 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4707 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4708 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4709 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4710 case NEON::BI__builtin_neon_vpadal_v: 4711 case NEON::BI__builtin_neon_vpadalq_v: { 4712 unsigned ArgElts = VTy->getNumElements(); 4713 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4714 unsigned BitWidth = EltTy->getBitWidth(); 4715 llvm::Type *ArgTy = llvm::VectorType::get( 4716 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4717 llvm::Type* Tys[2] = { VTy, ArgTy }; 4718 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4719 SmallVector<llvm::Value*, 1> TmpOps; 4720 TmpOps.push_back(Ops[1]); 4721 Function *F = CGM.getIntrinsic(Int, Tys); 4722 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4723 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4724 return Builder.CreateAdd(tmp, addend); 4725 } 4726 case NEON::BI__builtin_neon_vpmin_v: 4727 case NEON::BI__builtin_neon_vpminq_v: 4728 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4729 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4730 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4731 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4732 case NEON::BI__builtin_neon_vpmax_v: 4733 case NEON::BI__builtin_neon_vpmaxq_v: 4734 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4735 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4736 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4737 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4738 case NEON::BI__builtin_neon_vminnm_v: 4739 case NEON::BI__builtin_neon_vminnmq_v: 4740 Int = Intrinsic::aarch64_neon_fminnm; 4741 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4742 case NEON::BI__builtin_neon_vmaxnm_v: 4743 case NEON::BI__builtin_neon_vmaxnmq_v: 4744 Int = Intrinsic::aarch64_neon_fmaxnm; 4745 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4746 case NEON::BI__builtin_neon_vrecpss_f32: { 4747 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4748 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4749 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4750 Ops, "vrecps"); 4751 } 4752 case NEON::BI__builtin_neon_vrecpsd_f64: { 4753 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4754 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4755 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4756 Ops, "vrecps"); 4757 } 4758 case NEON::BI__builtin_neon_vqshrun_n_v: 4759 Int = Intrinsic::aarch64_neon_sqshrun; 4760 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4761 case NEON::BI__builtin_neon_vqrshrun_n_v: 4762 Int = Intrinsic::aarch64_neon_sqrshrun; 4763 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4764 case NEON::BI__builtin_neon_vqshrn_n_v: 4765 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4766 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4767 case NEON::BI__builtin_neon_vrshrn_n_v: 4768 Int = Intrinsic::aarch64_neon_rshrn; 4769 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4770 case NEON::BI__builtin_neon_vqrshrn_n_v: 4771 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4772 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4773 case NEON::BI__builtin_neon_vrnda_v: 4774 case NEON::BI__builtin_neon_vrndaq_v: { 4775 Int = Intrinsic::round; 4776 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4777 } 4778 case NEON::BI__builtin_neon_vrndi_v: 4779 case NEON::BI__builtin_neon_vrndiq_v: { 4780 Int = Intrinsic::nearbyint; 4781 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4782 } 4783 case NEON::BI__builtin_neon_vrndm_v: 4784 case NEON::BI__builtin_neon_vrndmq_v: { 4785 Int = Intrinsic::floor; 4786 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4787 } 4788 case NEON::BI__builtin_neon_vrndn_v: 4789 case NEON::BI__builtin_neon_vrndnq_v: { 4790 Int = Intrinsic::aarch64_neon_frintn; 4791 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4792 } 4793 case NEON::BI__builtin_neon_vrndp_v: 4794 case NEON::BI__builtin_neon_vrndpq_v: { 4795 Int = Intrinsic::ceil; 4796 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4797 } 4798 case NEON::BI__builtin_neon_vrndx_v: 4799 case NEON::BI__builtin_neon_vrndxq_v: { 4800 Int = Intrinsic::rint; 4801 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4802 } 4803 case NEON::BI__builtin_neon_vrnd_v: 4804 case NEON::BI__builtin_neon_vrndq_v: { 4805 Int = Intrinsic::trunc; 4806 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4807 } 4808 case NEON::BI__builtin_neon_vceqz_v: 4809 case NEON::BI__builtin_neon_vceqzq_v: 4810 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4811 ICmpInst::ICMP_EQ, "vceqz"); 4812 case NEON::BI__builtin_neon_vcgez_v: 4813 case NEON::BI__builtin_neon_vcgezq_v: 4814 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4815 ICmpInst::ICMP_SGE, "vcgez"); 4816 case NEON::BI__builtin_neon_vclez_v: 4817 case NEON::BI__builtin_neon_vclezq_v: 4818 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4819 ICmpInst::ICMP_SLE, "vclez"); 4820 case NEON::BI__builtin_neon_vcgtz_v: 4821 case NEON::BI__builtin_neon_vcgtzq_v: 4822 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4823 ICmpInst::ICMP_SGT, "vcgtz"); 4824 case NEON::BI__builtin_neon_vcltz_v: 4825 case NEON::BI__builtin_neon_vcltzq_v: 4826 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4827 ICmpInst::ICMP_SLT, "vcltz"); 4828 case NEON::BI__builtin_neon_vcvt_f64_v: 4829 case NEON::BI__builtin_neon_vcvtq_f64_v: 4830 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4831 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4832 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4833 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4834 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4835 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4836 "unexpected vcvt_f64_f32 builtin"); 4837 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4838 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4839 4840 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4841 } 4842 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4843 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4844 "unexpected vcvt_f32_f64 builtin"); 4845 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4846 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4847 4848 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4849 } 4850 case NEON::BI__builtin_neon_vcvt_s32_v: 4851 case NEON::BI__builtin_neon_vcvt_u32_v: 4852 case NEON::BI__builtin_neon_vcvt_s64_v: 4853 case NEON::BI__builtin_neon_vcvt_u64_v: 4854 case NEON::BI__builtin_neon_vcvtq_s32_v: 4855 case NEON::BI__builtin_neon_vcvtq_u32_v: 4856 case NEON::BI__builtin_neon_vcvtq_s64_v: 4857 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4858 bool Double = 4859 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4860 llvm::Type *InTy = 4861 GetNeonType(this, 4862 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4863 : NeonTypeFlags::Float32, false, quad)); 4864 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4865 if (usgn) 4866 return Builder.CreateFPToUI(Ops[0], Ty); 4867 return Builder.CreateFPToSI(Ops[0], Ty); 4868 } 4869 case NEON::BI__builtin_neon_vcvta_s32_v: 4870 case NEON::BI__builtin_neon_vcvtaq_s32_v: 4871 case NEON::BI__builtin_neon_vcvta_u32_v: 4872 case NEON::BI__builtin_neon_vcvtaq_u32_v: 4873 case NEON::BI__builtin_neon_vcvta_s64_v: 4874 case NEON::BI__builtin_neon_vcvtaq_s64_v: 4875 case NEON::BI__builtin_neon_vcvta_u64_v: 4876 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4877 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 4878 bool Double = 4879 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4880 llvm::Type *InTy = 4881 GetNeonType(this, 4882 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4883 : NeonTypeFlags::Float32, false, quad)); 4884 llvm::Type *Tys[2] = { Ty, InTy }; 4885 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 4886 } 4887 case NEON::BI__builtin_neon_vcvtm_s32_v: 4888 case NEON::BI__builtin_neon_vcvtmq_s32_v: 4889 case NEON::BI__builtin_neon_vcvtm_u32_v: 4890 case NEON::BI__builtin_neon_vcvtmq_u32_v: 4891 case NEON::BI__builtin_neon_vcvtm_s64_v: 4892 case NEON::BI__builtin_neon_vcvtmq_s64_v: 4893 case NEON::BI__builtin_neon_vcvtm_u64_v: 4894 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4895 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 4896 bool Double = 4897 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4898 llvm::Type *InTy = 4899 GetNeonType(this, 4900 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4901 : NeonTypeFlags::Float32, false, quad)); 4902 llvm::Type *Tys[2] = { Ty, InTy }; 4903 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 4904 } 4905 case NEON::BI__builtin_neon_vcvtn_s32_v: 4906 case NEON::BI__builtin_neon_vcvtnq_s32_v: 4907 case NEON::BI__builtin_neon_vcvtn_u32_v: 4908 case NEON::BI__builtin_neon_vcvtnq_u32_v: 4909 case NEON::BI__builtin_neon_vcvtn_s64_v: 4910 case NEON::BI__builtin_neon_vcvtnq_s64_v: 4911 case NEON::BI__builtin_neon_vcvtn_u64_v: 4912 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4913 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 4914 bool Double = 4915 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4916 llvm::Type *InTy = 4917 GetNeonType(this, 4918 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4919 : NeonTypeFlags::Float32, false, quad)); 4920 llvm::Type *Tys[2] = { Ty, InTy }; 4921 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 4922 } 4923 case NEON::BI__builtin_neon_vcvtp_s32_v: 4924 case NEON::BI__builtin_neon_vcvtpq_s32_v: 4925 case NEON::BI__builtin_neon_vcvtp_u32_v: 4926 case NEON::BI__builtin_neon_vcvtpq_u32_v: 4927 case NEON::BI__builtin_neon_vcvtp_s64_v: 4928 case NEON::BI__builtin_neon_vcvtpq_s64_v: 4929 case NEON::BI__builtin_neon_vcvtp_u64_v: 4930 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4931 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 4932 bool Double = 4933 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4934 llvm::Type *InTy = 4935 GetNeonType(this, 4936 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4937 : NeonTypeFlags::Float32, false, quad)); 4938 llvm::Type *Tys[2] = { Ty, InTy }; 4939 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 4940 } 4941 case NEON::BI__builtin_neon_vmulx_v: 4942 case NEON::BI__builtin_neon_vmulxq_v: { 4943 Int = Intrinsic::aarch64_neon_fmulx; 4944 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 4945 } 4946 case NEON::BI__builtin_neon_vmul_lane_v: 4947 case NEON::BI__builtin_neon_vmul_laneq_v: { 4948 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 4949 bool Quad = false; 4950 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 4951 Quad = true; 4952 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4953 llvm::Type *VTy = GetNeonType(this, 4954 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 4955 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4956 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 4957 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 4958 return Builder.CreateBitCast(Result, Ty); 4959 } 4960 case NEON::BI__builtin_neon_vnegd_s64: 4961 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 4962 case NEON::BI__builtin_neon_vpmaxnm_v: 4963 case NEON::BI__builtin_neon_vpmaxnmq_v: { 4964 Int = Intrinsic::aarch64_neon_fmaxnmp; 4965 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 4966 } 4967 case NEON::BI__builtin_neon_vpminnm_v: 4968 case NEON::BI__builtin_neon_vpminnmq_v: { 4969 Int = Intrinsic::aarch64_neon_fminnmp; 4970 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 4971 } 4972 case NEON::BI__builtin_neon_vsqrt_v: 4973 case NEON::BI__builtin_neon_vsqrtq_v: { 4974 Int = Intrinsic::sqrt; 4975 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4976 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 4977 } 4978 case NEON::BI__builtin_neon_vrbit_v: 4979 case NEON::BI__builtin_neon_vrbitq_v: { 4980 Int = Intrinsic::aarch64_neon_rbit; 4981 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 4982 } 4983 case NEON::BI__builtin_neon_vaddv_u8: 4984 // FIXME: These are handled by the AArch64 scalar code. 4985 usgn = true; 4986 // FALLTHROUGH 4987 case NEON::BI__builtin_neon_vaddv_s8: { 4988 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 4989 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 4990 VTy = 4991 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 4992 llvm::Type *Tys[2] = { Ty, VTy }; 4993 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4994 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 4995 return Builder.CreateTrunc(Ops[0], 4996 llvm::IntegerType::get(getLLVMContext(), 8)); 4997 } 4998 case NEON::BI__builtin_neon_vaddv_u16: 4999 usgn = true; 5000 // FALLTHROUGH 5001 case NEON::BI__builtin_neon_vaddv_s16: { 5002 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5003 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5004 VTy = 5005 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5006 llvm::Type *Tys[2] = { Ty, VTy }; 5007 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5008 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5009 return Builder.CreateTrunc(Ops[0], 5010 llvm::IntegerType::get(getLLVMContext(), 16)); 5011 } 5012 case NEON::BI__builtin_neon_vaddvq_u8: 5013 usgn = true; 5014 // FALLTHROUGH 5015 case NEON::BI__builtin_neon_vaddvq_s8: { 5016 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5017 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5018 VTy = 5019 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5020 llvm::Type *Tys[2] = { Ty, VTy }; 5021 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5022 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5023 return Builder.CreateTrunc(Ops[0], 5024 llvm::IntegerType::get(getLLVMContext(), 8)); 5025 } 5026 case NEON::BI__builtin_neon_vaddvq_u16: 5027 usgn = true; 5028 // FALLTHROUGH 5029 case NEON::BI__builtin_neon_vaddvq_s16: { 5030 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5031 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5032 VTy = 5033 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5034 llvm::Type *Tys[2] = { Ty, VTy }; 5035 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5036 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5037 return Builder.CreateTrunc(Ops[0], 5038 llvm::IntegerType::get(getLLVMContext(), 16)); 5039 } 5040 case NEON::BI__builtin_neon_vmaxv_u8: { 5041 Int = Intrinsic::aarch64_neon_umaxv; 5042 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5043 VTy = 5044 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5045 llvm::Type *Tys[2] = { Ty, VTy }; 5046 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5047 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5048 return Builder.CreateTrunc(Ops[0], 5049 llvm::IntegerType::get(getLLVMContext(), 8)); 5050 } 5051 case NEON::BI__builtin_neon_vmaxv_u16: { 5052 Int = Intrinsic::aarch64_neon_umaxv; 5053 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5054 VTy = 5055 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5056 llvm::Type *Tys[2] = { Ty, VTy }; 5057 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5058 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5059 return Builder.CreateTrunc(Ops[0], 5060 llvm::IntegerType::get(getLLVMContext(), 16)); 5061 } 5062 case NEON::BI__builtin_neon_vmaxvq_u8: { 5063 Int = Intrinsic::aarch64_neon_umaxv; 5064 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5065 VTy = 5066 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5067 llvm::Type *Tys[2] = { Ty, VTy }; 5068 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5069 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5070 return Builder.CreateTrunc(Ops[0], 5071 llvm::IntegerType::get(getLLVMContext(), 8)); 5072 } 5073 case NEON::BI__builtin_neon_vmaxvq_u16: { 5074 Int = Intrinsic::aarch64_neon_umaxv; 5075 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5076 VTy = 5077 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5078 llvm::Type *Tys[2] = { Ty, VTy }; 5079 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5080 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5081 return Builder.CreateTrunc(Ops[0], 5082 llvm::IntegerType::get(getLLVMContext(), 16)); 5083 } 5084 case NEON::BI__builtin_neon_vmaxv_s8: { 5085 Int = Intrinsic::aarch64_neon_smaxv; 5086 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5087 VTy = 5088 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5089 llvm::Type *Tys[2] = { Ty, VTy }; 5090 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5091 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5092 return Builder.CreateTrunc(Ops[0], 5093 llvm::IntegerType::get(getLLVMContext(), 8)); 5094 } 5095 case NEON::BI__builtin_neon_vmaxv_s16: { 5096 Int = Intrinsic::aarch64_neon_smaxv; 5097 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5098 VTy = 5099 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5100 llvm::Type *Tys[2] = { Ty, VTy }; 5101 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5102 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5103 return Builder.CreateTrunc(Ops[0], 5104 llvm::IntegerType::get(getLLVMContext(), 16)); 5105 } 5106 case NEON::BI__builtin_neon_vmaxvq_s8: { 5107 Int = Intrinsic::aarch64_neon_smaxv; 5108 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5109 VTy = 5110 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5111 llvm::Type *Tys[2] = { Ty, VTy }; 5112 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5113 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5114 return Builder.CreateTrunc(Ops[0], 5115 llvm::IntegerType::get(getLLVMContext(), 8)); 5116 } 5117 case NEON::BI__builtin_neon_vmaxvq_s16: { 5118 Int = Intrinsic::aarch64_neon_smaxv; 5119 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5120 VTy = 5121 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5122 llvm::Type *Tys[2] = { Ty, VTy }; 5123 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5124 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5125 return Builder.CreateTrunc(Ops[0], 5126 llvm::IntegerType::get(getLLVMContext(), 16)); 5127 } 5128 case NEON::BI__builtin_neon_vminv_u8: { 5129 Int = Intrinsic::aarch64_neon_uminv; 5130 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5131 VTy = 5132 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5133 llvm::Type *Tys[2] = { Ty, VTy }; 5134 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5135 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5136 return Builder.CreateTrunc(Ops[0], 5137 llvm::IntegerType::get(getLLVMContext(), 8)); 5138 } 5139 case NEON::BI__builtin_neon_vminv_u16: { 5140 Int = Intrinsic::aarch64_neon_uminv; 5141 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5142 VTy = 5143 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5144 llvm::Type *Tys[2] = { Ty, VTy }; 5145 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5146 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5147 return Builder.CreateTrunc(Ops[0], 5148 llvm::IntegerType::get(getLLVMContext(), 16)); 5149 } 5150 case NEON::BI__builtin_neon_vminvq_u8: { 5151 Int = Intrinsic::aarch64_neon_uminv; 5152 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5153 VTy = 5154 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5155 llvm::Type *Tys[2] = { Ty, VTy }; 5156 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5157 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5158 return Builder.CreateTrunc(Ops[0], 5159 llvm::IntegerType::get(getLLVMContext(), 8)); 5160 } 5161 case NEON::BI__builtin_neon_vminvq_u16: { 5162 Int = Intrinsic::aarch64_neon_uminv; 5163 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5164 VTy = 5165 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5166 llvm::Type *Tys[2] = { Ty, VTy }; 5167 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5168 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5169 return Builder.CreateTrunc(Ops[0], 5170 llvm::IntegerType::get(getLLVMContext(), 16)); 5171 } 5172 case NEON::BI__builtin_neon_vminv_s8: { 5173 Int = Intrinsic::aarch64_neon_sminv; 5174 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5175 VTy = 5176 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5177 llvm::Type *Tys[2] = { Ty, VTy }; 5178 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5179 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5180 return Builder.CreateTrunc(Ops[0], 5181 llvm::IntegerType::get(getLLVMContext(), 8)); 5182 } 5183 case NEON::BI__builtin_neon_vminv_s16: { 5184 Int = Intrinsic::aarch64_neon_sminv; 5185 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5186 VTy = 5187 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5188 llvm::Type *Tys[2] = { Ty, VTy }; 5189 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5190 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5191 return Builder.CreateTrunc(Ops[0], 5192 llvm::IntegerType::get(getLLVMContext(), 16)); 5193 } 5194 case NEON::BI__builtin_neon_vminvq_s8: { 5195 Int = Intrinsic::aarch64_neon_sminv; 5196 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5197 VTy = 5198 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5199 llvm::Type *Tys[2] = { Ty, VTy }; 5200 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5201 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5202 return Builder.CreateTrunc(Ops[0], 5203 llvm::IntegerType::get(getLLVMContext(), 8)); 5204 } 5205 case NEON::BI__builtin_neon_vminvq_s16: { 5206 Int = Intrinsic::aarch64_neon_sminv; 5207 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5208 VTy = 5209 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5210 llvm::Type *Tys[2] = { Ty, VTy }; 5211 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5212 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5213 return Builder.CreateTrunc(Ops[0], 5214 llvm::IntegerType::get(getLLVMContext(), 16)); 5215 } 5216 case NEON::BI__builtin_neon_vmul_n_f64: { 5217 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5218 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5219 return Builder.CreateFMul(Ops[0], RHS); 5220 } 5221 case NEON::BI__builtin_neon_vaddlv_u8: { 5222 Int = Intrinsic::aarch64_neon_uaddlv; 5223 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5224 VTy = 5225 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5226 llvm::Type *Tys[2] = { Ty, VTy }; 5227 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5228 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5229 return Builder.CreateTrunc(Ops[0], 5230 llvm::IntegerType::get(getLLVMContext(), 16)); 5231 } 5232 case NEON::BI__builtin_neon_vaddlv_u16: { 5233 Int = Intrinsic::aarch64_neon_uaddlv; 5234 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5235 VTy = 5236 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5237 llvm::Type *Tys[2] = { Ty, VTy }; 5238 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5239 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5240 } 5241 case NEON::BI__builtin_neon_vaddlvq_u8: { 5242 Int = Intrinsic::aarch64_neon_uaddlv; 5243 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5244 VTy = 5245 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5246 llvm::Type *Tys[2] = { Ty, VTy }; 5247 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5248 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5249 return Builder.CreateTrunc(Ops[0], 5250 llvm::IntegerType::get(getLLVMContext(), 16)); 5251 } 5252 case NEON::BI__builtin_neon_vaddlvq_u16: { 5253 Int = Intrinsic::aarch64_neon_uaddlv; 5254 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5255 VTy = 5256 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5257 llvm::Type *Tys[2] = { Ty, VTy }; 5258 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5259 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5260 } 5261 case NEON::BI__builtin_neon_vaddlv_s8: { 5262 Int = Intrinsic::aarch64_neon_saddlv; 5263 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5264 VTy = 5265 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5266 llvm::Type *Tys[2] = { Ty, VTy }; 5267 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5268 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5269 return Builder.CreateTrunc(Ops[0], 5270 llvm::IntegerType::get(getLLVMContext(), 16)); 5271 } 5272 case NEON::BI__builtin_neon_vaddlv_s16: { 5273 Int = Intrinsic::aarch64_neon_saddlv; 5274 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5275 VTy = 5276 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5277 llvm::Type *Tys[2] = { Ty, VTy }; 5278 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5279 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5280 } 5281 case NEON::BI__builtin_neon_vaddlvq_s8: { 5282 Int = Intrinsic::aarch64_neon_saddlv; 5283 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5284 VTy = 5285 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5286 llvm::Type *Tys[2] = { Ty, VTy }; 5287 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5288 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5289 return Builder.CreateTrunc(Ops[0], 5290 llvm::IntegerType::get(getLLVMContext(), 16)); 5291 } 5292 case NEON::BI__builtin_neon_vaddlvq_s16: { 5293 Int = Intrinsic::aarch64_neon_saddlv; 5294 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5295 VTy = 5296 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5297 llvm::Type *Tys[2] = { Ty, VTy }; 5298 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5299 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5300 } 5301 case NEON::BI__builtin_neon_vsri_n_v: 5302 case NEON::BI__builtin_neon_vsriq_n_v: { 5303 Int = Intrinsic::aarch64_neon_vsri; 5304 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5305 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5306 } 5307 case NEON::BI__builtin_neon_vsli_n_v: 5308 case NEON::BI__builtin_neon_vsliq_n_v: { 5309 Int = Intrinsic::aarch64_neon_vsli; 5310 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5311 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5312 } 5313 case NEON::BI__builtin_neon_vsra_n_v: 5314 case NEON::BI__builtin_neon_vsraq_n_v: 5315 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5316 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5317 return Builder.CreateAdd(Ops[0], Ops[1]); 5318 case NEON::BI__builtin_neon_vrsra_n_v: 5319 case NEON::BI__builtin_neon_vrsraq_n_v: { 5320 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5321 SmallVector<llvm::Value*,2> TmpOps; 5322 TmpOps.push_back(Ops[1]); 5323 TmpOps.push_back(Ops[2]); 5324 Function* F = CGM.getIntrinsic(Int, Ty); 5325 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5326 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5327 return Builder.CreateAdd(Ops[0], tmp); 5328 } 5329 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5330 // of an Align parameter here. 5331 case NEON::BI__builtin_neon_vld1_x2_v: 5332 case NEON::BI__builtin_neon_vld1q_x2_v: 5333 case NEON::BI__builtin_neon_vld1_x3_v: 5334 case NEON::BI__builtin_neon_vld1q_x3_v: 5335 case NEON::BI__builtin_neon_vld1_x4_v: 5336 case NEON::BI__builtin_neon_vld1q_x4_v: { 5337 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5338 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5339 llvm::Type *Tys[2] = { VTy, PTy }; 5340 unsigned Int; 5341 switch (BuiltinID) { 5342 case NEON::BI__builtin_neon_vld1_x2_v: 5343 case NEON::BI__builtin_neon_vld1q_x2_v: 5344 Int = Intrinsic::aarch64_neon_ld1x2; 5345 break; 5346 case NEON::BI__builtin_neon_vld1_x3_v: 5347 case NEON::BI__builtin_neon_vld1q_x3_v: 5348 Int = Intrinsic::aarch64_neon_ld1x3; 5349 break; 5350 case NEON::BI__builtin_neon_vld1_x4_v: 5351 case NEON::BI__builtin_neon_vld1q_x4_v: 5352 Int = Intrinsic::aarch64_neon_ld1x4; 5353 break; 5354 } 5355 Function *F = CGM.getIntrinsic(Int, Tys); 5356 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5357 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5358 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5359 return Builder.CreateStore(Ops[1], Ops[0]); 5360 } 5361 case NEON::BI__builtin_neon_vst1_x2_v: 5362 case NEON::BI__builtin_neon_vst1q_x2_v: 5363 case NEON::BI__builtin_neon_vst1_x3_v: 5364 case NEON::BI__builtin_neon_vst1q_x3_v: 5365 case NEON::BI__builtin_neon_vst1_x4_v: 5366 case NEON::BI__builtin_neon_vst1q_x4_v: { 5367 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5368 llvm::Type *Tys[2] = { VTy, PTy }; 5369 unsigned Int; 5370 switch (BuiltinID) { 5371 case NEON::BI__builtin_neon_vst1_x2_v: 5372 case NEON::BI__builtin_neon_vst1q_x2_v: 5373 Int = Intrinsic::aarch64_neon_st1x2; 5374 break; 5375 case NEON::BI__builtin_neon_vst1_x3_v: 5376 case NEON::BI__builtin_neon_vst1q_x3_v: 5377 Int = Intrinsic::aarch64_neon_st1x3; 5378 break; 5379 case NEON::BI__builtin_neon_vst1_x4_v: 5380 case NEON::BI__builtin_neon_vst1q_x4_v: 5381 Int = Intrinsic::aarch64_neon_st1x4; 5382 break; 5383 } 5384 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5385 IntOps.push_back(Ops[0]); 5386 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5387 } 5388 case NEON::BI__builtin_neon_vld1_v: 5389 case NEON::BI__builtin_neon_vld1q_v: 5390 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5391 return Builder.CreateLoad(Ops[0]); 5392 case NEON::BI__builtin_neon_vst1_v: 5393 case NEON::BI__builtin_neon_vst1q_v: 5394 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5395 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5396 return Builder.CreateStore(Ops[1], Ops[0]); 5397 case NEON::BI__builtin_neon_vld1_lane_v: 5398 case NEON::BI__builtin_neon_vld1q_lane_v: 5399 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5400 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5401 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5402 Ops[0] = Builder.CreateLoad(Ops[0]); 5403 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5404 case NEON::BI__builtin_neon_vld1_dup_v: 5405 case NEON::BI__builtin_neon_vld1q_dup_v: { 5406 Value *V = UndefValue::get(Ty); 5407 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5408 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5409 Ops[0] = Builder.CreateLoad(Ops[0]); 5410 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5411 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5412 return EmitNeonSplat(Ops[0], CI); 5413 } 5414 case NEON::BI__builtin_neon_vst1_lane_v: 5415 case NEON::BI__builtin_neon_vst1q_lane_v: 5416 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5417 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5418 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5419 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5420 case NEON::BI__builtin_neon_vld2_v: 5421 case NEON::BI__builtin_neon_vld2q_v: { 5422 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5423 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5424 llvm::Type *Tys[2] = { VTy, PTy }; 5425 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5426 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5427 Ops[0] = Builder.CreateBitCast(Ops[0], 5428 llvm::PointerType::getUnqual(Ops[1]->getType())); 5429 return Builder.CreateStore(Ops[1], Ops[0]); 5430 } 5431 case NEON::BI__builtin_neon_vld3_v: 5432 case NEON::BI__builtin_neon_vld3q_v: { 5433 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5434 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5435 llvm::Type *Tys[2] = { VTy, PTy }; 5436 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5437 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5438 Ops[0] = Builder.CreateBitCast(Ops[0], 5439 llvm::PointerType::getUnqual(Ops[1]->getType())); 5440 return Builder.CreateStore(Ops[1], Ops[0]); 5441 } 5442 case NEON::BI__builtin_neon_vld4_v: 5443 case NEON::BI__builtin_neon_vld4q_v: { 5444 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5445 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5446 llvm::Type *Tys[2] = { VTy, PTy }; 5447 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5448 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5449 Ops[0] = Builder.CreateBitCast(Ops[0], 5450 llvm::PointerType::getUnqual(Ops[1]->getType())); 5451 return Builder.CreateStore(Ops[1], Ops[0]); 5452 } 5453 case NEON::BI__builtin_neon_vld2_dup_v: 5454 case NEON::BI__builtin_neon_vld2q_dup_v: { 5455 llvm::Type *PTy = 5456 llvm::PointerType::getUnqual(VTy->getElementType()); 5457 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5458 llvm::Type *Tys[2] = { VTy, PTy }; 5459 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5460 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5461 Ops[0] = Builder.CreateBitCast(Ops[0], 5462 llvm::PointerType::getUnqual(Ops[1]->getType())); 5463 return Builder.CreateStore(Ops[1], Ops[0]); 5464 } 5465 case NEON::BI__builtin_neon_vld3_dup_v: 5466 case NEON::BI__builtin_neon_vld3q_dup_v: { 5467 llvm::Type *PTy = 5468 llvm::PointerType::getUnqual(VTy->getElementType()); 5469 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5470 llvm::Type *Tys[2] = { VTy, PTy }; 5471 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5472 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5473 Ops[0] = Builder.CreateBitCast(Ops[0], 5474 llvm::PointerType::getUnqual(Ops[1]->getType())); 5475 return Builder.CreateStore(Ops[1], Ops[0]); 5476 } 5477 case NEON::BI__builtin_neon_vld4_dup_v: 5478 case NEON::BI__builtin_neon_vld4q_dup_v: { 5479 llvm::Type *PTy = 5480 llvm::PointerType::getUnqual(VTy->getElementType()); 5481 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5482 llvm::Type *Tys[2] = { VTy, PTy }; 5483 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5484 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5485 Ops[0] = Builder.CreateBitCast(Ops[0], 5486 llvm::PointerType::getUnqual(Ops[1]->getType())); 5487 return Builder.CreateStore(Ops[1], Ops[0]); 5488 } 5489 case NEON::BI__builtin_neon_vld2_lane_v: 5490 case NEON::BI__builtin_neon_vld2q_lane_v: { 5491 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5492 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5493 Ops.push_back(Ops[1]); 5494 Ops.erase(Ops.begin()+1); 5495 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5496 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5497 Ops[3] = Builder.CreateZExt(Ops[3], 5498 llvm::IntegerType::get(getLLVMContext(), 64)); 5499 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5500 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5501 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5502 return Builder.CreateStore(Ops[1], Ops[0]); 5503 } 5504 case NEON::BI__builtin_neon_vld3_lane_v: 5505 case NEON::BI__builtin_neon_vld3q_lane_v: { 5506 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5507 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5508 Ops.push_back(Ops[1]); 5509 Ops.erase(Ops.begin()+1); 5510 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5511 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5512 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5513 Ops[4] = Builder.CreateZExt(Ops[4], 5514 llvm::IntegerType::get(getLLVMContext(), 64)); 5515 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5516 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5517 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5518 return Builder.CreateStore(Ops[1], Ops[0]); 5519 } 5520 case NEON::BI__builtin_neon_vld4_lane_v: 5521 case NEON::BI__builtin_neon_vld4q_lane_v: { 5522 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5523 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5524 Ops.push_back(Ops[1]); 5525 Ops.erase(Ops.begin()+1); 5526 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5527 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5528 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5529 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5530 Ops[5] = Builder.CreateZExt(Ops[5], 5531 llvm::IntegerType::get(getLLVMContext(), 64)); 5532 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5533 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5534 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5535 return Builder.CreateStore(Ops[1], Ops[0]); 5536 } 5537 case NEON::BI__builtin_neon_vst2_v: 5538 case NEON::BI__builtin_neon_vst2q_v: { 5539 Ops.push_back(Ops[0]); 5540 Ops.erase(Ops.begin()); 5541 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5542 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5543 Ops, ""); 5544 } 5545 case NEON::BI__builtin_neon_vst2_lane_v: 5546 case NEON::BI__builtin_neon_vst2q_lane_v: { 5547 Ops.push_back(Ops[0]); 5548 Ops.erase(Ops.begin()); 5549 Ops[2] = Builder.CreateZExt(Ops[2], 5550 llvm::IntegerType::get(getLLVMContext(), 64)); 5551 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5552 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5553 Ops, ""); 5554 } 5555 case NEON::BI__builtin_neon_vst3_v: 5556 case NEON::BI__builtin_neon_vst3q_v: { 5557 Ops.push_back(Ops[0]); 5558 Ops.erase(Ops.begin()); 5559 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5560 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5561 Ops, ""); 5562 } 5563 case NEON::BI__builtin_neon_vst3_lane_v: 5564 case NEON::BI__builtin_neon_vst3q_lane_v: { 5565 Ops.push_back(Ops[0]); 5566 Ops.erase(Ops.begin()); 5567 Ops[3] = Builder.CreateZExt(Ops[3], 5568 llvm::IntegerType::get(getLLVMContext(), 64)); 5569 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5570 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5571 Ops, ""); 5572 } 5573 case NEON::BI__builtin_neon_vst4_v: 5574 case NEON::BI__builtin_neon_vst4q_v: { 5575 Ops.push_back(Ops[0]); 5576 Ops.erase(Ops.begin()); 5577 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5578 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5579 Ops, ""); 5580 } 5581 case NEON::BI__builtin_neon_vst4_lane_v: 5582 case NEON::BI__builtin_neon_vst4q_lane_v: { 5583 Ops.push_back(Ops[0]); 5584 Ops.erase(Ops.begin()); 5585 Ops[4] = Builder.CreateZExt(Ops[4], 5586 llvm::IntegerType::get(getLLVMContext(), 64)); 5587 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5588 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5589 Ops, ""); 5590 } 5591 case NEON::BI__builtin_neon_vtrn_v: 5592 case NEON::BI__builtin_neon_vtrnq_v: { 5593 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5594 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5595 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5596 Value *SV = nullptr; 5597 5598 for (unsigned vi = 0; vi != 2; ++vi) { 5599 SmallVector<Constant*, 16> Indices; 5600 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5601 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5602 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5603 } 5604 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5605 SV = llvm::ConstantVector::get(Indices); 5606 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5607 SV = Builder.CreateStore(SV, Addr); 5608 } 5609 return SV; 5610 } 5611 case NEON::BI__builtin_neon_vuzp_v: 5612 case NEON::BI__builtin_neon_vuzpq_v: { 5613 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5614 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5615 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5616 Value *SV = nullptr; 5617 5618 for (unsigned vi = 0; vi != 2; ++vi) { 5619 SmallVector<Constant*, 16> Indices; 5620 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5621 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5622 5623 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5624 SV = llvm::ConstantVector::get(Indices); 5625 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5626 SV = Builder.CreateStore(SV, Addr); 5627 } 5628 return SV; 5629 } 5630 case NEON::BI__builtin_neon_vzip_v: 5631 case NEON::BI__builtin_neon_vzipq_v: { 5632 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5633 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5634 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5635 Value *SV = nullptr; 5636 5637 for (unsigned vi = 0; vi != 2; ++vi) { 5638 SmallVector<Constant*, 16> Indices; 5639 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5640 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5641 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5642 } 5643 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5644 SV = llvm::ConstantVector::get(Indices); 5645 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5646 SV = Builder.CreateStore(SV, Addr); 5647 } 5648 return SV; 5649 } 5650 case NEON::BI__builtin_neon_vqtbl1q_v: { 5651 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5652 Ops, "vtbl1"); 5653 } 5654 case NEON::BI__builtin_neon_vqtbl2q_v: { 5655 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5656 Ops, "vtbl2"); 5657 } 5658 case NEON::BI__builtin_neon_vqtbl3q_v: { 5659 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5660 Ops, "vtbl3"); 5661 } 5662 case NEON::BI__builtin_neon_vqtbl4q_v: { 5663 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5664 Ops, "vtbl4"); 5665 } 5666 case NEON::BI__builtin_neon_vqtbx1q_v: { 5667 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5668 Ops, "vtbx1"); 5669 } 5670 case NEON::BI__builtin_neon_vqtbx2q_v: { 5671 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5672 Ops, "vtbx2"); 5673 } 5674 case NEON::BI__builtin_neon_vqtbx3q_v: { 5675 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5676 Ops, "vtbx3"); 5677 } 5678 case NEON::BI__builtin_neon_vqtbx4q_v: { 5679 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5680 Ops, "vtbx4"); 5681 } 5682 case NEON::BI__builtin_neon_vsqadd_v: 5683 case NEON::BI__builtin_neon_vsqaddq_v: { 5684 Int = Intrinsic::aarch64_neon_usqadd; 5685 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5686 } 5687 case NEON::BI__builtin_neon_vuqadd_v: 5688 case NEON::BI__builtin_neon_vuqaddq_v: { 5689 Int = Intrinsic::aarch64_neon_suqadd; 5690 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5691 } 5692 } 5693 } 5694 5695 llvm::Value *CodeGenFunction:: 5696 BuildVector(ArrayRef<llvm::Value*> Ops) { 5697 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5698 "Not a power-of-two sized vector!"); 5699 bool AllConstants = true; 5700 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5701 AllConstants &= isa<Constant>(Ops[i]); 5702 5703 // If this is a constant vector, create a ConstantVector. 5704 if (AllConstants) { 5705 SmallVector<llvm::Constant*, 16> CstOps; 5706 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5707 CstOps.push_back(cast<Constant>(Ops[i])); 5708 return llvm::ConstantVector::get(CstOps); 5709 } 5710 5711 // Otherwise, insertelement the values to build the vector. 5712 Value *Result = 5713 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5714 5715 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5716 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5717 5718 return Result; 5719 } 5720 5721 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5722 const CallExpr *E) { 5723 SmallVector<Value*, 4> Ops; 5724 5725 // Find out if any arguments are required to be integer constant expressions. 5726 unsigned ICEArguments = 0; 5727 ASTContext::GetBuiltinTypeError Error; 5728 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5729 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5730 5731 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5732 // If this is a normal argument, just emit it as a scalar. 5733 if ((ICEArguments & (1 << i)) == 0) { 5734 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5735 continue; 5736 } 5737 5738 // If this is required to be a constant, constant fold it so that we know 5739 // that the generated intrinsic gets a ConstantInt. 5740 llvm::APSInt Result; 5741 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5742 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5743 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5744 } 5745 5746 switch (BuiltinID) { 5747 default: return nullptr; 5748 case X86::BI_mm_prefetch: { 5749 Value *Address = EmitScalarExpr(E->getArg(0)); 5750 Value *RW = ConstantInt::get(Int32Ty, 0); 5751 Value *Locality = EmitScalarExpr(E->getArg(1)); 5752 Value *Data = ConstantInt::get(Int32Ty, 1); 5753 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5754 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5755 } 5756 case X86::BI__builtin_ia32_vec_init_v8qi: 5757 case X86::BI__builtin_ia32_vec_init_v4hi: 5758 case X86::BI__builtin_ia32_vec_init_v2si: 5759 return Builder.CreateBitCast(BuildVector(Ops), 5760 llvm::Type::getX86_MMXTy(getLLVMContext())); 5761 case X86::BI__builtin_ia32_vec_ext_v2si: 5762 return Builder.CreateExtractElement(Ops[0], 5763 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5764 case X86::BI__builtin_ia32_ldmxcsr: { 5765 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5766 Builder.CreateStore(Ops[0], Tmp); 5767 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5768 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5769 } 5770 case X86::BI__builtin_ia32_stmxcsr: { 5771 Value *Tmp = CreateMemTemp(E->getType()); 5772 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5773 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5774 return Builder.CreateLoad(Tmp, "stmxcsr"); 5775 } 5776 case X86::BI__builtin_ia32_storehps: 5777 case X86::BI__builtin_ia32_storelps: { 5778 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5779 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5780 5781 // cast val v2i64 5782 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5783 5784 // extract (0, 1) 5785 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5786 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5787 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5788 5789 // cast pointer to i64 & store 5790 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5791 return Builder.CreateStore(Ops[1], Ops[0]); 5792 } 5793 case X86::BI__builtin_ia32_palignr: { 5794 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5795 5796 // If palignr is shifting the pair of input vectors less than 9 bytes, 5797 // emit a shuffle instruction. 5798 if (shiftVal <= 8) { 5799 SmallVector<llvm::Constant*, 8> Indices; 5800 for (unsigned i = 0; i != 8; ++i) 5801 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5802 5803 Value* SV = llvm::ConstantVector::get(Indices); 5804 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5805 } 5806 5807 // If palignr is shifting the pair of input vectors more than 8 but less 5808 // than 16 bytes, emit a logical right shift of the destination. 5809 if (shiftVal < 16) { 5810 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 5811 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 5812 5813 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5814 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 5815 5816 // create i32 constant 5817 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 5818 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5819 } 5820 5821 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 5822 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5823 } 5824 case X86::BI__builtin_ia32_palignr128: { 5825 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5826 5827 // If palignr is shifting the pair of input vectors less than 17 bytes, 5828 // emit a shuffle instruction. 5829 if (shiftVal <= 16) { 5830 SmallVector<llvm::Constant*, 16> Indices; 5831 for (unsigned i = 0; i != 16; ++i) 5832 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5833 5834 Value* SV = llvm::ConstantVector::get(Indices); 5835 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5836 } 5837 5838 // If palignr is shifting the pair of input vectors more than 16 but less 5839 // than 32 bytes, emit a logical right shift of the destination. 5840 if (shiftVal < 32) { 5841 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5842 5843 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5844 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5845 5846 // create i32 constant 5847 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 5848 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5849 } 5850 5851 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5852 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5853 } 5854 case X86::BI__builtin_ia32_palignr256: { 5855 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5856 5857 // If palignr is shifting the pair of input vectors less than 17 bytes, 5858 // emit a shuffle instruction. 5859 if (shiftVal <= 16) { 5860 SmallVector<llvm::Constant*, 32> Indices; 5861 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5862 for (unsigned l = 0; l != 2; ++l) { 5863 unsigned LaneStart = l * 16; 5864 unsigned LaneEnd = (l+1) * 16; 5865 for (unsigned i = 0; i != 16; ++i) { 5866 unsigned Idx = shiftVal + i + LaneStart; 5867 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 5868 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 5869 } 5870 } 5871 5872 Value* SV = llvm::ConstantVector::get(Indices); 5873 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5874 } 5875 5876 // If palignr is shifting the pair of input vectors more than 16 but less 5877 // than 32 bytes, emit a logical right shift of the destination. 5878 if (shiftVal < 32) { 5879 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 5880 5881 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5882 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5883 5884 // create i32 constant 5885 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 5886 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5887 } 5888 5889 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5890 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5891 } 5892 case X86::BI__builtin_ia32_movntps: 5893 case X86::BI__builtin_ia32_movntps256: 5894 case X86::BI__builtin_ia32_movntpd: 5895 case X86::BI__builtin_ia32_movntpd256: 5896 case X86::BI__builtin_ia32_movntdq: 5897 case X86::BI__builtin_ia32_movntdq256: 5898 case X86::BI__builtin_ia32_movnti: 5899 case X86::BI__builtin_ia32_movnti64: { 5900 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 5901 Builder.getInt32(1)); 5902 5903 // Convert the type of the pointer to a pointer to the stored type. 5904 Value *BC = Builder.CreateBitCast(Ops[0], 5905 llvm::PointerType::getUnqual(Ops[1]->getType()), 5906 "cast"); 5907 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 5908 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 5909 5910 // If the operand is an integer, we can't assume alignment. Otherwise, 5911 // assume natural alignment. 5912 QualType ArgTy = E->getArg(1)->getType(); 5913 unsigned Align; 5914 if (ArgTy->isIntegerType()) 5915 Align = 1; 5916 else 5917 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 5918 SI->setAlignment(Align); 5919 return SI; 5920 } 5921 // 3DNow! 5922 case X86::BI__builtin_ia32_pswapdsf: 5923 case X86::BI__builtin_ia32_pswapdsi: { 5924 const char *name = nullptr; 5925 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5926 switch(BuiltinID) { 5927 default: llvm_unreachable("Unsupported intrinsic!"); 5928 case X86::BI__builtin_ia32_pswapdsf: 5929 case X86::BI__builtin_ia32_pswapdsi: 5930 name = "pswapd"; 5931 ID = Intrinsic::x86_3dnowa_pswapd; 5932 break; 5933 } 5934 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 5935 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 5936 llvm::Function *F = CGM.getIntrinsic(ID); 5937 return Builder.CreateCall(F, Ops, name); 5938 } 5939 case X86::BI__builtin_ia32_rdrand16_step: 5940 case X86::BI__builtin_ia32_rdrand32_step: 5941 case X86::BI__builtin_ia32_rdrand64_step: 5942 case X86::BI__builtin_ia32_rdseed16_step: 5943 case X86::BI__builtin_ia32_rdseed32_step: 5944 case X86::BI__builtin_ia32_rdseed64_step: { 5945 Intrinsic::ID ID; 5946 switch (BuiltinID) { 5947 default: llvm_unreachable("Unsupported intrinsic!"); 5948 case X86::BI__builtin_ia32_rdrand16_step: 5949 ID = Intrinsic::x86_rdrand_16; 5950 break; 5951 case X86::BI__builtin_ia32_rdrand32_step: 5952 ID = Intrinsic::x86_rdrand_32; 5953 break; 5954 case X86::BI__builtin_ia32_rdrand64_step: 5955 ID = Intrinsic::x86_rdrand_64; 5956 break; 5957 case X86::BI__builtin_ia32_rdseed16_step: 5958 ID = Intrinsic::x86_rdseed_16; 5959 break; 5960 case X86::BI__builtin_ia32_rdseed32_step: 5961 ID = Intrinsic::x86_rdseed_32; 5962 break; 5963 case X86::BI__builtin_ia32_rdseed64_step: 5964 ID = Intrinsic::x86_rdseed_64; 5965 break; 5966 } 5967 5968 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 5969 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 5970 return Builder.CreateExtractValue(Call, 1); 5971 } 5972 // AVX2 broadcast 5973 case X86::BI__builtin_ia32_vbroadcastsi256: { 5974 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 5975 Builder.CreateStore(Ops[0], VecTmp); 5976 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 5977 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 5978 } 5979 } 5980 } 5981 5982 5983 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 5984 const CallExpr *E) { 5985 SmallVector<Value*, 4> Ops; 5986 5987 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 5988 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5989 5990 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5991 5992 switch (BuiltinID) { 5993 default: return nullptr; 5994 5995 // vec_ld, vec_lvsl, vec_lvsr 5996 case PPC::BI__builtin_altivec_lvx: 5997 case PPC::BI__builtin_altivec_lvxl: 5998 case PPC::BI__builtin_altivec_lvebx: 5999 case PPC::BI__builtin_altivec_lvehx: 6000 case PPC::BI__builtin_altivec_lvewx: 6001 case PPC::BI__builtin_altivec_lvsl: 6002 case PPC::BI__builtin_altivec_lvsr: 6003 { 6004 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6005 6006 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6007 Ops.pop_back(); 6008 6009 switch (BuiltinID) { 6010 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6011 case PPC::BI__builtin_altivec_lvx: 6012 ID = Intrinsic::ppc_altivec_lvx; 6013 break; 6014 case PPC::BI__builtin_altivec_lvxl: 6015 ID = Intrinsic::ppc_altivec_lvxl; 6016 break; 6017 case PPC::BI__builtin_altivec_lvebx: 6018 ID = Intrinsic::ppc_altivec_lvebx; 6019 break; 6020 case PPC::BI__builtin_altivec_lvehx: 6021 ID = Intrinsic::ppc_altivec_lvehx; 6022 break; 6023 case PPC::BI__builtin_altivec_lvewx: 6024 ID = Intrinsic::ppc_altivec_lvewx; 6025 break; 6026 case PPC::BI__builtin_altivec_lvsl: 6027 ID = Intrinsic::ppc_altivec_lvsl; 6028 break; 6029 case PPC::BI__builtin_altivec_lvsr: 6030 ID = Intrinsic::ppc_altivec_lvsr; 6031 break; 6032 } 6033 llvm::Function *F = CGM.getIntrinsic(ID); 6034 return Builder.CreateCall(F, Ops, ""); 6035 } 6036 6037 // vec_st 6038 case PPC::BI__builtin_altivec_stvx: 6039 case PPC::BI__builtin_altivec_stvxl: 6040 case PPC::BI__builtin_altivec_stvebx: 6041 case PPC::BI__builtin_altivec_stvehx: 6042 case PPC::BI__builtin_altivec_stvewx: 6043 { 6044 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6045 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6046 Ops.pop_back(); 6047 6048 switch (BuiltinID) { 6049 default: llvm_unreachable("Unsupported st intrinsic!"); 6050 case PPC::BI__builtin_altivec_stvx: 6051 ID = Intrinsic::ppc_altivec_stvx; 6052 break; 6053 case PPC::BI__builtin_altivec_stvxl: 6054 ID = Intrinsic::ppc_altivec_stvxl; 6055 break; 6056 case PPC::BI__builtin_altivec_stvebx: 6057 ID = Intrinsic::ppc_altivec_stvebx; 6058 break; 6059 case PPC::BI__builtin_altivec_stvehx: 6060 ID = Intrinsic::ppc_altivec_stvehx; 6061 break; 6062 case PPC::BI__builtin_altivec_stvewx: 6063 ID = Intrinsic::ppc_altivec_stvewx; 6064 break; 6065 } 6066 llvm::Function *F = CGM.getIntrinsic(ID); 6067 return Builder.CreateCall(F, Ops, ""); 6068 } 6069 } 6070 } 6071 6072 // Emit an intrinsic that has 1 float or double. 6073 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6074 const CallExpr *E, 6075 unsigned IntrinsicID) { 6076 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6077 6078 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6079 return CGF.Builder.CreateCall(F, Src0); 6080 } 6081 6082 // Emit an intrinsic that has 3 float or double operands. 6083 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6084 const CallExpr *E, 6085 unsigned IntrinsicID) { 6086 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6087 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6088 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6089 6090 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6091 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6092 } 6093 6094 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6095 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6096 const CallExpr *E, 6097 unsigned IntrinsicID) { 6098 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6099 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6100 6101 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6102 return CGF.Builder.CreateCall2(F, Src0, Src1); 6103 } 6104 6105 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6106 const CallExpr *E) { 6107 switch (BuiltinID) { 6108 case R600::BI__builtin_amdgpu_div_scale: 6109 case R600::BI__builtin_amdgpu_div_scalef: { 6110 // Translate from the intrinsics's struct return to the builtin's out 6111 // argument. 6112 6113 std::pair<llvm::Value *, unsigned> FlagOutPtr 6114 = EmitPointerWithAlignment(E->getArg(3)); 6115 6116 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6117 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6118 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6119 6120 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6121 X->getType()); 6122 6123 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6124 6125 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6126 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6127 6128 llvm::Type *RealFlagType 6129 = FlagOutPtr.first->getType()->getPointerElementType(); 6130 6131 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6132 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6133 FlagStore->setAlignment(FlagOutPtr.second); 6134 return Result; 6135 } 6136 case R600::BI__builtin_amdgpu_div_fmas: 6137 case R600::BI__builtin_amdgpu_div_fmasf: 6138 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fmas); 6139 case R600::BI__builtin_amdgpu_div_fixup: 6140 case R600::BI__builtin_amdgpu_div_fixupf: 6141 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6142 case R600::BI__builtin_amdgpu_trig_preop: 6143 case R600::BI__builtin_amdgpu_trig_preopf: 6144 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6145 case R600::BI__builtin_amdgpu_rcp: 6146 case R600::BI__builtin_amdgpu_rcpf: 6147 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6148 case R600::BI__builtin_amdgpu_rsq: 6149 case R600::BI__builtin_amdgpu_rsqf: 6150 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6151 case R600::BI__builtin_amdgpu_rsq_clamped: 6152 case R600::BI__builtin_amdgpu_rsq_clampedf: 6153 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6154 case R600::BI__builtin_amdgpu_ldexp: 6155 case R600::BI__builtin_amdgpu_ldexpf: 6156 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6157 default: 6158 return nullptr; 6159 } 6160 } 6161