1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGCXXABI.h" 16 #include "CGObjCRuntime.h" 17 #include "CodeGenModule.h" 18 #include "TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "clang/Basic/TargetInfo.h" 23 #include "clang/CodeGen/CGFunctionInfo.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 30 using namespace clang; 31 using namespace CodeGen; 32 using namespace llvm; 33 34 /// getBuiltinLibFunction - Given a builtin id for a function like 35 /// "__builtin_fabsf", return a Function* for "fabsf". 36 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 37 unsigned BuiltinID) { 38 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 39 40 // Get the name, skip over the __builtin_ prefix (if necessary). 41 StringRef Name; 42 GlobalDecl D(FD); 43 44 // If the builtin has been declared explicitly with an assembler label, 45 // use the mangled name. This differs from the plain label on platforms 46 // that prefix labels. 47 if (FD->hasAttr<AsmLabelAttr>()) 48 Name = getMangledName(D); 49 else 50 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 51 52 llvm::FunctionType *Ty = 53 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 54 55 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 56 } 57 58 /// Emit the conversions required to turn the given value into an 59 /// integer of the given size. 60 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 61 QualType T, llvm::IntegerType *IntType) { 62 V = CGF.EmitToMemory(V, T); 63 64 if (V->getType()->isPointerTy()) 65 return CGF.Builder.CreatePtrToInt(V, IntType); 66 67 assert(V->getType() == IntType); 68 return V; 69 } 70 71 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 72 QualType T, llvm::Type *ResultType) { 73 V = CGF.EmitFromMemory(V, T); 74 75 if (ResultType->isPointerTy()) 76 return CGF.Builder.CreateIntToPtr(V, ResultType); 77 78 assert(V->getType() == ResultType); 79 return V; 80 } 81 82 /// Utility to insert an atomic instruction based on Instrinsic::ID 83 /// and the expression node. 84 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 85 llvm::AtomicRMWInst::BinOp Kind, 86 const CallExpr *E) { 87 QualType T = E->getType(); 88 assert(E->getArg(0)->getType()->isPointerType()); 89 assert(CGF.getContext().hasSameUnqualifiedType(T, 90 E->getArg(0)->getType()->getPointeeType())); 91 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 92 93 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 94 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 95 96 llvm::IntegerType *IntType = 97 llvm::IntegerType::get(CGF.getLLVMContext(), 98 CGF.getContext().getTypeSize(T)); 99 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 100 101 llvm::Value *Args[2]; 102 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 103 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 104 llvm::Type *ValueType = Args[1]->getType(); 105 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 106 107 llvm::Value *Result = 108 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 109 llvm::SequentiallyConsistent); 110 Result = EmitFromInt(CGF, Result, T, ValueType); 111 return RValue::get(Result); 112 } 113 114 /// Utility to insert an atomic instruction based Instrinsic::ID and 115 /// the expression node, where the return value is the result of the 116 /// operation. 117 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 118 llvm::AtomicRMWInst::BinOp Kind, 119 const CallExpr *E, 120 Instruction::BinaryOps Op, 121 bool Invert = false) { 122 QualType T = E->getType(); 123 assert(E->getArg(0)->getType()->isPointerType()); 124 assert(CGF.getContext().hasSameUnqualifiedType(T, 125 E->getArg(0)->getType()->getPointeeType())); 126 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 127 128 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 129 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 130 131 llvm::IntegerType *IntType = 132 llvm::IntegerType::get(CGF.getLLVMContext(), 133 CGF.getContext().getTypeSize(T)); 134 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 135 136 llvm::Value *Args[2]; 137 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 138 llvm::Type *ValueType = Args[1]->getType(); 139 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 140 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 141 142 llvm::Value *Result = 143 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 144 llvm::SequentiallyConsistent); 145 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 146 if (Invert) 147 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 148 llvm::ConstantInt::get(IntType, -1)); 149 Result = EmitFromInt(CGF, Result, T, ValueType); 150 return RValue::get(Result); 151 } 152 153 /// EmitFAbs - Emit a call to @llvm.fabs(). 154 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 155 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 156 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 157 Call->setDoesNotAccessMemory(); 158 return Call; 159 } 160 161 /// Emit the computation of the sign bit for a floating point value. Returns 162 /// the i1 sign bit value. 163 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 164 LLVMContext &C = CGF.CGM.getLLVMContext(); 165 166 llvm::Type *Ty = V->getType(); 167 int Width = Ty->getPrimitiveSizeInBits(); 168 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 169 V = CGF.Builder.CreateBitCast(V, IntTy); 170 if (Ty->isPPC_FP128Ty()) { 171 // The higher-order double comes first, and so we need to truncate the 172 // pair to extract the overall sign. The order of the pair is the same 173 // in both little- and big-Endian modes. 174 Width >>= 1; 175 IntTy = llvm::IntegerType::get(C, Width); 176 V = CGF.Builder.CreateTrunc(V, IntTy); 177 } 178 Value *Zero = llvm::Constant::getNullValue(IntTy); 179 return CGF.Builder.CreateICmpSLT(V, Zero); 180 } 181 182 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 183 const CallExpr *E, llvm::Value *calleeValue) { 184 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 185 ReturnValueSlot(), Fn); 186 } 187 188 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 189 /// depending on IntrinsicID. 190 /// 191 /// \arg CGF The current codegen function. 192 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 193 /// \arg X The first argument to the llvm.*.with.overflow.*. 194 /// \arg Y The second argument to the llvm.*.with.overflow.*. 195 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 196 /// \returns The result (i.e. sum/product) returned by the intrinsic. 197 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 198 const llvm::Intrinsic::ID IntrinsicID, 199 llvm::Value *X, llvm::Value *Y, 200 llvm::Value *&Carry) { 201 // Make sure we have integers of the same width. 202 assert(X->getType() == Y->getType() && 203 "Arguments must be the same type. (Did you forget to make sure both " 204 "arguments have the same integer width?)"); 205 206 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 207 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 208 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 209 return CGF.Builder.CreateExtractValue(Tmp, 0); 210 } 211 212 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 213 unsigned BuiltinID, const CallExpr *E, 214 ReturnValueSlot ReturnValue) { 215 // See if we can constant fold this builtin. If so, don't emit it at all. 216 Expr::EvalResult Result; 217 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 218 !Result.hasSideEffects()) { 219 if (Result.Val.isInt()) 220 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 221 Result.Val.getInt())); 222 if (Result.Val.isFloat()) 223 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 224 Result.Val.getFloat())); 225 } 226 227 switch (BuiltinID) { 228 default: break; // Handle intrinsics and libm functions below. 229 case Builtin::BI__builtin___CFStringMakeConstantString: 230 case Builtin::BI__builtin___NSStringMakeConstantString: 231 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 232 case Builtin::BI__builtin_stdarg_start: 233 case Builtin::BI__builtin_va_start: 234 case Builtin::BI__va_start: 235 case Builtin::BI__builtin_va_end: { 236 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 237 ? EmitScalarExpr(E->getArg(0)) 238 : EmitVAListRef(E->getArg(0)); 239 llvm::Type *DestType = Int8PtrTy; 240 if (ArgValue->getType() != DestType) 241 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 242 ArgValue->getName().data()); 243 244 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 245 Intrinsic::vaend : Intrinsic::vastart; 246 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 247 } 248 case Builtin::BI__builtin_va_copy: { 249 Value *DstPtr = EmitVAListRef(E->getArg(0)); 250 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 251 252 llvm::Type *Type = Int8PtrTy; 253 254 DstPtr = Builder.CreateBitCast(DstPtr, Type); 255 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 256 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 257 DstPtr, SrcPtr)); 258 } 259 case Builtin::BI__builtin_abs: 260 case Builtin::BI__builtin_labs: 261 case Builtin::BI__builtin_llabs: { 262 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 263 264 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 265 Value *CmpResult = 266 Builder.CreateICmpSGE(ArgValue, 267 llvm::Constant::getNullValue(ArgValue->getType()), 268 "abscond"); 269 Value *Result = 270 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 271 272 return RValue::get(Result); 273 } 274 case Builtin::BI__builtin_fabs: 275 case Builtin::BI__builtin_fabsf: 276 case Builtin::BI__builtin_fabsl: { 277 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 278 Value *Result = EmitFAbs(*this, Arg1); 279 return RValue::get(Result); 280 } 281 case Builtin::BI__builtin_fmod: 282 case Builtin::BI__builtin_fmodf: 283 case Builtin::BI__builtin_fmodl: { 284 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 285 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 286 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 287 return RValue::get(Result); 288 } 289 290 case Builtin::BI__builtin_conj: 291 case Builtin::BI__builtin_conjf: 292 case Builtin::BI__builtin_conjl: { 293 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 294 Value *Real = ComplexVal.first; 295 Value *Imag = ComplexVal.second; 296 Value *Zero = 297 Imag->getType()->isFPOrFPVectorTy() 298 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 299 : llvm::Constant::getNullValue(Imag->getType()); 300 301 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 302 return RValue::getComplex(std::make_pair(Real, Imag)); 303 } 304 case Builtin::BI__builtin_creal: 305 case Builtin::BI__builtin_crealf: 306 case Builtin::BI__builtin_creall: 307 case Builtin::BIcreal: 308 case Builtin::BIcrealf: 309 case Builtin::BIcreall: { 310 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 311 return RValue::get(ComplexVal.first); 312 } 313 314 case Builtin::BI__builtin_cimag: 315 case Builtin::BI__builtin_cimagf: 316 case Builtin::BI__builtin_cimagl: 317 case Builtin::BIcimag: 318 case Builtin::BIcimagf: 319 case Builtin::BIcimagl: { 320 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 321 return RValue::get(ComplexVal.second); 322 } 323 324 case Builtin::BI__builtin_ctzs: 325 case Builtin::BI__builtin_ctz: 326 case Builtin::BI__builtin_ctzl: 327 case Builtin::BI__builtin_ctzll: { 328 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 329 330 llvm::Type *ArgType = ArgValue->getType(); 331 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 332 333 llvm::Type *ResultType = ConvertType(E->getType()); 334 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 335 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 336 if (Result->getType() != ResultType) 337 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 338 "cast"); 339 return RValue::get(Result); 340 } 341 case Builtin::BI__builtin_clzs: 342 case Builtin::BI__builtin_clz: 343 case Builtin::BI__builtin_clzl: 344 case Builtin::BI__builtin_clzll: { 345 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 346 347 llvm::Type *ArgType = ArgValue->getType(); 348 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 349 350 llvm::Type *ResultType = ConvertType(E->getType()); 351 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 352 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 353 if (Result->getType() != ResultType) 354 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 355 "cast"); 356 return RValue::get(Result); 357 } 358 case Builtin::BI__builtin_ffs: 359 case Builtin::BI__builtin_ffsl: 360 case Builtin::BI__builtin_ffsll: { 361 // ffs(x) -> x ? cttz(x) + 1 : 0 362 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 363 364 llvm::Type *ArgType = ArgValue->getType(); 365 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 366 367 llvm::Type *ResultType = ConvertType(E->getType()); 368 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 369 Builder.getTrue()), 370 llvm::ConstantInt::get(ArgType, 1)); 371 Value *Zero = llvm::Constant::getNullValue(ArgType); 372 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 373 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 374 if (Result->getType() != ResultType) 375 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 376 "cast"); 377 return RValue::get(Result); 378 } 379 case Builtin::BI__builtin_parity: 380 case Builtin::BI__builtin_parityl: 381 case Builtin::BI__builtin_parityll: { 382 // parity(x) -> ctpop(x) & 1 383 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 384 385 llvm::Type *ArgType = ArgValue->getType(); 386 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 387 388 llvm::Type *ResultType = ConvertType(E->getType()); 389 Value *Tmp = Builder.CreateCall(F, ArgValue); 390 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 391 if (Result->getType() != ResultType) 392 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 393 "cast"); 394 return RValue::get(Result); 395 } 396 case Builtin::BI__builtin_popcount: 397 case Builtin::BI__builtin_popcountl: 398 case Builtin::BI__builtin_popcountll: { 399 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 400 401 llvm::Type *ArgType = ArgValue->getType(); 402 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 403 404 llvm::Type *ResultType = ConvertType(E->getType()); 405 Value *Result = Builder.CreateCall(F, ArgValue); 406 if (Result->getType() != ResultType) 407 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 408 "cast"); 409 return RValue::get(Result); 410 } 411 case Builtin::BI__builtin_expect: { 412 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 413 llvm::Type *ArgType = ArgValue->getType(); 414 415 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 416 // Don't generate llvm.expect on -O0 as the backend won't use it for 417 // anything. 418 // Note, we still IRGen ExpectedValue because it could have side-effects. 419 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 420 return RValue::get(ArgValue); 421 422 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 423 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 424 "expval"); 425 return RValue::get(Result); 426 } 427 case Builtin::BI__builtin_assume_aligned: { 428 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 429 Value *OffsetValue = 430 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 431 432 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 433 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 434 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 435 436 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 437 return RValue::get(PtrValue); 438 } 439 case Builtin::BI__assume: 440 case Builtin::BI__builtin_assume: { 441 if (E->getArg(0)->HasSideEffects(getContext())) 442 return RValue::get(nullptr); 443 444 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 445 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 446 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 447 } 448 case Builtin::BI__builtin_bswap16: 449 case Builtin::BI__builtin_bswap32: 450 case Builtin::BI__builtin_bswap64: { 451 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 452 llvm::Type *ArgType = ArgValue->getType(); 453 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 454 return RValue::get(Builder.CreateCall(F, ArgValue)); 455 } 456 case Builtin::BI__builtin_object_size: { 457 // We rely on constant folding to deal with expressions with side effects. 458 assert(!E->getArg(0)->HasSideEffects(getContext()) && 459 "should have been constant folded"); 460 461 // We pass this builtin onto the optimizer so that it can 462 // figure out the object size in more complex cases. 463 llvm::Type *ResType = ConvertType(E->getType()); 464 465 // LLVM only supports 0 and 2, make sure that we pass along that 466 // as a boolean. 467 Value *Ty = EmitScalarExpr(E->getArg(1)); 468 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 469 assert(CI); 470 uint64_t val = CI->getZExtValue(); 471 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 472 // FIXME: Get right address space. 473 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 474 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 475 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 476 } 477 case Builtin::BI__builtin_prefetch: { 478 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 479 // FIXME: Technically these constants should of type 'int', yes? 480 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 481 llvm::ConstantInt::get(Int32Ty, 0); 482 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 483 llvm::ConstantInt::get(Int32Ty, 3); 484 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 485 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 486 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 487 } 488 case Builtin::BI__builtin_readcyclecounter: { 489 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 490 return RValue::get(Builder.CreateCall(F)); 491 } 492 case Builtin::BI__builtin___clear_cache: { 493 Value *Begin = EmitScalarExpr(E->getArg(0)); 494 Value *End = EmitScalarExpr(E->getArg(1)); 495 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 496 return RValue::get(Builder.CreateCall2(F, Begin, End)); 497 } 498 case Builtin::BI__builtin_trap: { 499 Value *F = CGM.getIntrinsic(Intrinsic::trap); 500 return RValue::get(Builder.CreateCall(F)); 501 } 502 case Builtin::BI__debugbreak: { 503 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 504 return RValue::get(Builder.CreateCall(F)); 505 } 506 case Builtin::BI__builtin_unreachable: { 507 if (SanOpts.has(SanitizerKind::Unreachable)) { 508 SanitizerScope SanScope(this); 509 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 510 SanitizerKind::Unreachable), 511 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 512 None); 513 } else 514 Builder.CreateUnreachable(); 515 516 // We do need to preserve an insertion point. 517 EmitBlock(createBasicBlock("unreachable.cont")); 518 519 return RValue::get(nullptr); 520 } 521 522 case Builtin::BI__builtin_powi: 523 case Builtin::BI__builtin_powif: 524 case Builtin::BI__builtin_powil: { 525 Value *Base = EmitScalarExpr(E->getArg(0)); 526 Value *Exponent = EmitScalarExpr(E->getArg(1)); 527 llvm::Type *ArgType = Base->getType(); 528 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 529 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 530 } 531 532 case Builtin::BI__builtin_isgreater: 533 case Builtin::BI__builtin_isgreaterequal: 534 case Builtin::BI__builtin_isless: 535 case Builtin::BI__builtin_islessequal: 536 case Builtin::BI__builtin_islessgreater: 537 case Builtin::BI__builtin_isunordered: { 538 // Ordered comparisons: we know the arguments to these are matching scalar 539 // floating point values. 540 Value *LHS = EmitScalarExpr(E->getArg(0)); 541 Value *RHS = EmitScalarExpr(E->getArg(1)); 542 543 switch (BuiltinID) { 544 default: llvm_unreachable("Unknown ordered comparison"); 545 case Builtin::BI__builtin_isgreater: 546 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 547 break; 548 case Builtin::BI__builtin_isgreaterequal: 549 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 550 break; 551 case Builtin::BI__builtin_isless: 552 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 553 break; 554 case Builtin::BI__builtin_islessequal: 555 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 556 break; 557 case Builtin::BI__builtin_islessgreater: 558 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 559 break; 560 case Builtin::BI__builtin_isunordered: 561 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 562 break; 563 } 564 // ZExt bool to int type. 565 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 566 } 567 case Builtin::BI__builtin_isnan: { 568 Value *V = EmitScalarExpr(E->getArg(0)); 569 V = Builder.CreateFCmpUNO(V, V, "cmp"); 570 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 571 } 572 573 case Builtin::BI__builtin_isinf: { 574 // isinf(x) --> fabs(x) == infinity 575 Value *V = EmitScalarExpr(E->getArg(0)); 576 V = EmitFAbs(*this, V); 577 578 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 579 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 580 } 581 582 case Builtin::BI__builtin_isinf_sign: { 583 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 584 Value *Arg = EmitScalarExpr(E->getArg(0)); 585 Value *AbsArg = EmitFAbs(*this, Arg); 586 Value *IsInf = Builder.CreateFCmpOEQ( 587 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 588 Value *IsNeg = EmitSignBit(*this, Arg); 589 590 llvm::Type *IntTy = ConvertType(E->getType()); 591 Value *Zero = Constant::getNullValue(IntTy); 592 Value *One = ConstantInt::get(IntTy, 1); 593 Value *NegativeOne = ConstantInt::get(IntTy, -1); 594 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 595 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 596 return RValue::get(Result); 597 } 598 599 case Builtin::BI__builtin_isnormal: { 600 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 601 Value *V = EmitScalarExpr(E->getArg(0)); 602 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 603 604 Value *Abs = EmitFAbs(*this, V); 605 Value *IsLessThanInf = 606 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 607 APFloat Smallest = APFloat::getSmallestNormalized( 608 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 609 Value *IsNormal = 610 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 611 "isnormal"); 612 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 613 V = Builder.CreateAnd(V, IsNormal, "and"); 614 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 615 } 616 617 case Builtin::BI__builtin_isfinite: { 618 // isfinite(x) --> x == x && fabs(x) != infinity; 619 Value *V = EmitScalarExpr(E->getArg(0)); 620 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 621 622 Value *Abs = EmitFAbs(*this, V); 623 Value *IsNotInf = 624 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 625 626 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 627 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 628 } 629 630 case Builtin::BI__builtin_fpclassify: { 631 Value *V = EmitScalarExpr(E->getArg(5)); 632 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 633 634 // Create Result 635 BasicBlock *Begin = Builder.GetInsertBlock(); 636 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 637 Builder.SetInsertPoint(End); 638 PHINode *Result = 639 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 640 "fpclassify_result"); 641 642 // if (V==0) return FP_ZERO 643 Builder.SetInsertPoint(Begin); 644 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 645 "iszero"); 646 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 647 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 648 Builder.CreateCondBr(IsZero, End, NotZero); 649 Result->addIncoming(ZeroLiteral, Begin); 650 651 // if (V != V) return FP_NAN 652 Builder.SetInsertPoint(NotZero); 653 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 654 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 655 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 656 Builder.CreateCondBr(IsNan, End, NotNan); 657 Result->addIncoming(NanLiteral, NotZero); 658 659 // if (fabs(V) == infinity) return FP_INFINITY 660 Builder.SetInsertPoint(NotNan); 661 Value *VAbs = EmitFAbs(*this, V); 662 Value *IsInf = 663 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 664 "isinf"); 665 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 666 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 667 Builder.CreateCondBr(IsInf, End, NotInf); 668 Result->addIncoming(InfLiteral, NotNan); 669 670 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 671 Builder.SetInsertPoint(NotInf); 672 APFloat Smallest = APFloat::getSmallestNormalized( 673 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 674 Value *IsNormal = 675 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 676 "isnormal"); 677 Value *NormalResult = 678 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 679 EmitScalarExpr(E->getArg(3))); 680 Builder.CreateBr(End); 681 Result->addIncoming(NormalResult, NotInf); 682 683 // return Result 684 Builder.SetInsertPoint(End); 685 return RValue::get(Result); 686 } 687 688 case Builtin::BIalloca: 689 case Builtin::BI_alloca: 690 case Builtin::BI__builtin_alloca: { 691 Value *Size = EmitScalarExpr(E->getArg(0)); 692 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 693 } 694 case Builtin::BIbzero: 695 case Builtin::BI__builtin_bzero: { 696 std::pair<llvm::Value*, unsigned> Dest = 697 EmitPointerWithAlignment(E->getArg(0)); 698 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 699 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 700 Dest.second, false); 701 return RValue::get(Dest.first); 702 } 703 case Builtin::BImemcpy: 704 case Builtin::BI__builtin_memcpy: { 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.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 712 return RValue::get(Dest.first); 713 } 714 715 case Builtin::BI__builtin___memcpy_chk: { 716 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 717 llvm::APSInt Size, DstSize; 718 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 719 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 720 break; 721 if (Size.ugt(DstSize)) 722 break; 723 std::pair<llvm::Value*, unsigned> Dest = 724 EmitPointerWithAlignment(E->getArg(0)); 725 std::pair<llvm::Value*, unsigned> Src = 726 EmitPointerWithAlignment(E->getArg(1)); 727 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 728 unsigned Align = std::min(Dest.second, Src.second); 729 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 730 return RValue::get(Dest.first); 731 } 732 733 case Builtin::BI__builtin_objc_memmove_collectable: { 734 Value *Address = EmitScalarExpr(E->getArg(0)); 735 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 736 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 737 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 738 Address, SrcAddr, SizeVal); 739 return RValue::get(Address); 740 } 741 742 case Builtin::BI__builtin___memmove_chk: { 743 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 744 llvm::APSInt Size, DstSize; 745 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 746 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 747 break; 748 if (Size.ugt(DstSize)) 749 break; 750 std::pair<llvm::Value*, unsigned> Dest = 751 EmitPointerWithAlignment(E->getArg(0)); 752 std::pair<llvm::Value*, unsigned> Src = 753 EmitPointerWithAlignment(E->getArg(1)); 754 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 755 unsigned Align = std::min(Dest.second, Src.second); 756 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 757 return RValue::get(Dest.first); 758 } 759 760 case Builtin::BImemmove: 761 case Builtin::BI__builtin_memmove: { 762 std::pair<llvm::Value*, unsigned> Dest = 763 EmitPointerWithAlignment(E->getArg(0)); 764 std::pair<llvm::Value*, unsigned> Src = 765 EmitPointerWithAlignment(E->getArg(1)); 766 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 767 unsigned Align = std::min(Dest.second, Src.second); 768 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 769 return RValue::get(Dest.first); 770 } 771 case Builtin::BImemset: 772 case Builtin::BI__builtin_memset: { 773 std::pair<llvm::Value*, unsigned> Dest = 774 EmitPointerWithAlignment(E->getArg(0)); 775 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 776 Builder.getInt8Ty()); 777 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 778 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 779 return RValue::get(Dest.first); 780 } 781 case Builtin::BI__builtin___memset_chk: { 782 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 783 llvm::APSInt Size, DstSize; 784 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 785 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 786 break; 787 if (Size.ugt(DstSize)) 788 break; 789 std::pair<llvm::Value*, unsigned> Dest = 790 EmitPointerWithAlignment(E->getArg(0)); 791 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 792 Builder.getInt8Ty()); 793 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 794 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 795 return RValue::get(Dest.first); 796 } 797 case Builtin::BI__builtin_dwarf_cfa: { 798 // The offset in bytes from the first argument to the CFA. 799 // 800 // Why on earth is this in the frontend? Is there any reason at 801 // all that the backend can't reasonably determine this while 802 // lowering llvm.eh.dwarf.cfa()? 803 // 804 // TODO: If there's a satisfactory reason, add a target hook for 805 // this instead of hard-coding 0, which is correct for most targets. 806 int32_t Offset = 0; 807 808 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 809 return RValue::get(Builder.CreateCall(F, 810 llvm::ConstantInt::get(Int32Ty, Offset))); 811 } 812 case Builtin::BI__builtin_return_address: { 813 Value *Depth = EmitScalarExpr(E->getArg(0)); 814 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 815 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 816 return RValue::get(Builder.CreateCall(F, Depth)); 817 } 818 case Builtin::BI__builtin_frame_address: { 819 Value *Depth = EmitScalarExpr(E->getArg(0)); 820 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 821 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 822 return RValue::get(Builder.CreateCall(F, Depth)); 823 } 824 case Builtin::BI__builtin_extract_return_addr: { 825 Value *Address = EmitScalarExpr(E->getArg(0)); 826 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 827 return RValue::get(Result); 828 } 829 case Builtin::BI__builtin_frob_return_addr: { 830 Value *Address = EmitScalarExpr(E->getArg(0)); 831 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 832 return RValue::get(Result); 833 } 834 case Builtin::BI__builtin_dwarf_sp_column: { 835 llvm::IntegerType *Ty 836 = cast<llvm::IntegerType>(ConvertType(E->getType())); 837 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 838 if (Column == -1) { 839 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 840 return RValue::get(llvm::UndefValue::get(Ty)); 841 } 842 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 843 } 844 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 845 Value *Address = EmitScalarExpr(E->getArg(0)); 846 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 847 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 848 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 849 } 850 case Builtin::BI__builtin_eh_return: { 851 Value *Int = EmitScalarExpr(E->getArg(0)); 852 Value *Ptr = EmitScalarExpr(E->getArg(1)); 853 854 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 855 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 856 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 857 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 858 ? Intrinsic::eh_return_i32 859 : Intrinsic::eh_return_i64); 860 Builder.CreateCall2(F, Int, Ptr); 861 Builder.CreateUnreachable(); 862 863 // We do need to preserve an insertion point. 864 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 865 866 return RValue::get(nullptr); 867 } 868 case Builtin::BI__builtin_unwind_init: { 869 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 870 return RValue::get(Builder.CreateCall(F)); 871 } 872 case Builtin::BI__builtin_extend_pointer: { 873 // Extends a pointer to the size of an _Unwind_Word, which is 874 // uint64_t on all platforms. Generally this gets poked into a 875 // register and eventually used as an address, so if the 876 // addressing registers are wider than pointers and the platform 877 // doesn't implicitly ignore high-order bits when doing 878 // addressing, we need to make sure we zext / sext based on 879 // the platform's expectations. 880 // 881 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 882 883 // Cast the pointer to intptr_t. 884 Value *Ptr = EmitScalarExpr(E->getArg(0)); 885 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 886 887 // If that's 64 bits, we're done. 888 if (IntPtrTy->getBitWidth() == 64) 889 return RValue::get(Result); 890 891 // Otherwise, ask the codegen data what to do. 892 if (getTargetHooks().extendPointerWithSExt()) 893 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 894 else 895 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 896 } 897 case Builtin::BI__builtin_setjmp: { 898 // Buffer is a void**. 899 Value *Buf = EmitScalarExpr(E->getArg(0)); 900 901 // Store the frame pointer to the setjmp buffer. 902 Value *FrameAddr = 903 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 904 ConstantInt::get(Int32Ty, 0)); 905 Builder.CreateStore(FrameAddr, Buf); 906 907 // Store the stack pointer to the setjmp buffer. 908 Value *StackAddr = 909 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 910 Value *StackSaveSlot = 911 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 912 Builder.CreateStore(StackAddr, StackSaveSlot); 913 914 // Call LLVM's EH setjmp, which is lightweight. 915 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 916 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 917 return RValue::get(Builder.CreateCall(F, Buf)); 918 } 919 case Builtin::BI__builtin_longjmp: { 920 Value *Buf = EmitScalarExpr(E->getArg(0)); 921 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 922 923 // Call LLVM's EH longjmp, which is lightweight. 924 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 925 926 // longjmp doesn't return; mark this as unreachable. 927 Builder.CreateUnreachable(); 928 929 // We do need to preserve an insertion point. 930 EmitBlock(createBasicBlock("longjmp.cont")); 931 932 return RValue::get(nullptr); 933 } 934 case Builtin::BI__sync_fetch_and_add: 935 case Builtin::BI__sync_fetch_and_sub: 936 case Builtin::BI__sync_fetch_and_or: 937 case Builtin::BI__sync_fetch_and_and: 938 case Builtin::BI__sync_fetch_and_xor: 939 case Builtin::BI__sync_fetch_and_nand: 940 case Builtin::BI__sync_add_and_fetch: 941 case Builtin::BI__sync_sub_and_fetch: 942 case Builtin::BI__sync_and_and_fetch: 943 case Builtin::BI__sync_or_and_fetch: 944 case Builtin::BI__sync_xor_and_fetch: 945 case Builtin::BI__sync_nand_and_fetch: 946 case Builtin::BI__sync_val_compare_and_swap: 947 case Builtin::BI__sync_bool_compare_and_swap: 948 case Builtin::BI__sync_lock_test_and_set: 949 case Builtin::BI__sync_lock_release: 950 case Builtin::BI__sync_swap: 951 llvm_unreachable("Shouldn't make it through sema"); 952 case Builtin::BI__sync_fetch_and_add_1: 953 case Builtin::BI__sync_fetch_and_add_2: 954 case Builtin::BI__sync_fetch_and_add_4: 955 case Builtin::BI__sync_fetch_and_add_8: 956 case Builtin::BI__sync_fetch_and_add_16: 957 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 958 case Builtin::BI__sync_fetch_and_sub_1: 959 case Builtin::BI__sync_fetch_and_sub_2: 960 case Builtin::BI__sync_fetch_and_sub_4: 961 case Builtin::BI__sync_fetch_and_sub_8: 962 case Builtin::BI__sync_fetch_and_sub_16: 963 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 964 case Builtin::BI__sync_fetch_and_or_1: 965 case Builtin::BI__sync_fetch_and_or_2: 966 case Builtin::BI__sync_fetch_and_or_4: 967 case Builtin::BI__sync_fetch_and_or_8: 968 case Builtin::BI__sync_fetch_and_or_16: 969 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 970 case Builtin::BI__sync_fetch_and_and_1: 971 case Builtin::BI__sync_fetch_and_and_2: 972 case Builtin::BI__sync_fetch_and_and_4: 973 case Builtin::BI__sync_fetch_and_and_8: 974 case Builtin::BI__sync_fetch_and_and_16: 975 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 976 case Builtin::BI__sync_fetch_and_xor_1: 977 case Builtin::BI__sync_fetch_and_xor_2: 978 case Builtin::BI__sync_fetch_and_xor_4: 979 case Builtin::BI__sync_fetch_and_xor_8: 980 case Builtin::BI__sync_fetch_and_xor_16: 981 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 982 case Builtin::BI__sync_fetch_and_nand_1: 983 case Builtin::BI__sync_fetch_and_nand_2: 984 case Builtin::BI__sync_fetch_and_nand_4: 985 case Builtin::BI__sync_fetch_and_nand_8: 986 case Builtin::BI__sync_fetch_and_nand_16: 987 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 988 989 // Clang extensions: not overloaded yet. 990 case Builtin::BI__sync_fetch_and_min: 991 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 992 case Builtin::BI__sync_fetch_and_max: 993 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 994 case Builtin::BI__sync_fetch_and_umin: 995 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 996 case Builtin::BI__sync_fetch_and_umax: 997 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 998 999 case Builtin::BI__sync_add_and_fetch_1: 1000 case Builtin::BI__sync_add_and_fetch_2: 1001 case Builtin::BI__sync_add_and_fetch_4: 1002 case Builtin::BI__sync_add_and_fetch_8: 1003 case Builtin::BI__sync_add_and_fetch_16: 1004 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1005 llvm::Instruction::Add); 1006 case Builtin::BI__sync_sub_and_fetch_1: 1007 case Builtin::BI__sync_sub_and_fetch_2: 1008 case Builtin::BI__sync_sub_and_fetch_4: 1009 case Builtin::BI__sync_sub_and_fetch_8: 1010 case Builtin::BI__sync_sub_and_fetch_16: 1011 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1012 llvm::Instruction::Sub); 1013 case Builtin::BI__sync_and_and_fetch_1: 1014 case Builtin::BI__sync_and_and_fetch_2: 1015 case Builtin::BI__sync_and_and_fetch_4: 1016 case Builtin::BI__sync_and_and_fetch_8: 1017 case Builtin::BI__sync_and_and_fetch_16: 1018 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1019 llvm::Instruction::And); 1020 case Builtin::BI__sync_or_and_fetch_1: 1021 case Builtin::BI__sync_or_and_fetch_2: 1022 case Builtin::BI__sync_or_and_fetch_4: 1023 case Builtin::BI__sync_or_and_fetch_8: 1024 case Builtin::BI__sync_or_and_fetch_16: 1025 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1026 llvm::Instruction::Or); 1027 case Builtin::BI__sync_xor_and_fetch_1: 1028 case Builtin::BI__sync_xor_and_fetch_2: 1029 case Builtin::BI__sync_xor_and_fetch_4: 1030 case Builtin::BI__sync_xor_and_fetch_8: 1031 case Builtin::BI__sync_xor_and_fetch_16: 1032 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1033 llvm::Instruction::Xor); 1034 case Builtin::BI__sync_nand_and_fetch_1: 1035 case Builtin::BI__sync_nand_and_fetch_2: 1036 case Builtin::BI__sync_nand_and_fetch_4: 1037 case Builtin::BI__sync_nand_and_fetch_8: 1038 case Builtin::BI__sync_nand_and_fetch_16: 1039 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1040 llvm::Instruction::And, true); 1041 1042 case Builtin::BI__sync_val_compare_and_swap_1: 1043 case Builtin::BI__sync_val_compare_and_swap_2: 1044 case Builtin::BI__sync_val_compare_and_swap_4: 1045 case Builtin::BI__sync_val_compare_and_swap_8: 1046 case Builtin::BI__sync_val_compare_and_swap_16: { 1047 QualType T = E->getType(); 1048 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1049 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1050 1051 llvm::IntegerType *IntType = 1052 llvm::IntegerType::get(getLLVMContext(), 1053 getContext().getTypeSize(T)); 1054 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1055 1056 Value *Args[3]; 1057 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1058 Args[1] = EmitScalarExpr(E->getArg(1)); 1059 llvm::Type *ValueType = Args[1]->getType(); 1060 Args[1] = EmitToInt(*this, Args[1], T, IntType); 1061 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1062 1063 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1064 llvm::SequentiallyConsistent, 1065 llvm::SequentiallyConsistent); 1066 Result = Builder.CreateExtractValue(Result, 0); 1067 Result = EmitFromInt(*this, Result, T, ValueType); 1068 return RValue::get(Result); 1069 } 1070 1071 case Builtin::BI__sync_bool_compare_and_swap_1: 1072 case Builtin::BI__sync_bool_compare_and_swap_2: 1073 case Builtin::BI__sync_bool_compare_and_swap_4: 1074 case Builtin::BI__sync_bool_compare_and_swap_8: 1075 case Builtin::BI__sync_bool_compare_and_swap_16: { 1076 QualType T = E->getArg(1)->getType(); 1077 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1078 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1079 1080 llvm::IntegerType *IntType = 1081 llvm::IntegerType::get(getLLVMContext(), 1082 getContext().getTypeSize(T)); 1083 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1084 1085 Value *Args[3]; 1086 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1087 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1088 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1089 1090 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1091 llvm::SequentiallyConsistent, 1092 llvm::SequentiallyConsistent); 1093 Value *Result = Builder.CreateExtractValue(Pair, 1); 1094 // zext bool to int. 1095 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1096 return RValue::get(Result); 1097 } 1098 1099 case Builtin::BI__sync_swap_1: 1100 case Builtin::BI__sync_swap_2: 1101 case Builtin::BI__sync_swap_4: 1102 case Builtin::BI__sync_swap_8: 1103 case Builtin::BI__sync_swap_16: 1104 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1105 1106 case Builtin::BI__sync_lock_test_and_set_1: 1107 case Builtin::BI__sync_lock_test_and_set_2: 1108 case Builtin::BI__sync_lock_test_and_set_4: 1109 case Builtin::BI__sync_lock_test_and_set_8: 1110 case Builtin::BI__sync_lock_test_and_set_16: 1111 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1112 1113 case Builtin::BI__sync_lock_release_1: 1114 case Builtin::BI__sync_lock_release_2: 1115 case Builtin::BI__sync_lock_release_4: 1116 case Builtin::BI__sync_lock_release_8: 1117 case Builtin::BI__sync_lock_release_16: { 1118 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1119 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1120 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1121 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1122 StoreSize.getQuantity() * 8); 1123 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1124 llvm::StoreInst *Store = 1125 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1126 Store->setAlignment(StoreSize.getQuantity()); 1127 Store->setAtomic(llvm::Release); 1128 return RValue::get(nullptr); 1129 } 1130 1131 case Builtin::BI__sync_synchronize: { 1132 // We assume this is supposed to correspond to a C++0x-style 1133 // sequentially-consistent fence (i.e. this is only usable for 1134 // synchonization, not device I/O or anything like that). This intrinsic 1135 // is really badly designed in the sense that in theory, there isn't 1136 // any way to safely use it... but in practice, it mostly works 1137 // to use it with non-atomic loads and stores to get acquire/release 1138 // semantics. 1139 Builder.CreateFence(llvm::SequentiallyConsistent); 1140 return RValue::get(nullptr); 1141 } 1142 1143 case Builtin::BI__c11_atomic_is_lock_free: 1144 case Builtin::BI__atomic_is_lock_free: { 1145 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1146 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1147 // _Atomic(T) is always properly-aligned. 1148 const char *LibCallName = "__atomic_is_lock_free"; 1149 CallArgList Args; 1150 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1151 getContext().getSizeType()); 1152 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1153 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1154 getContext().VoidPtrTy); 1155 else 1156 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1157 getContext().VoidPtrTy); 1158 const CGFunctionInfo &FuncInfo = 1159 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1160 FunctionType::ExtInfo(), 1161 RequiredArgs::All); 1162 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1163 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1164 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1165 } 1166 1167 case Builtin::BI__atomic_test_and_set: { 1168 // Look at the argument type to determine whether this is a volatile 1169 // operation. The parameter type is always volatile. 1170 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1171 bool Volatile = 1172 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1173 1174 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1175 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1176 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1177 Value *NewVal = Builder.getInt8(1); 1178 Value *Order = EmitScalarExpr(E->getArg(1)); 1179 if (isa<llvm::ConstantInt>(Order)) { 1180 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1181 AtomicRMWInst *Result = nullptr; 1182 switch (ord) { 1183 case 0: // memory_order_relaxed 1184 default: // invalid order 1185 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1186 Ptr, NewVal, 1187 llvm::Monotonic); 1188 break; 1189 case 1: // memory_order_consume 1190 case 2: // memory_order_acquire 1191 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1192 Ptr, NewVal, 1193 llvm::Acquire); 1194 break; 1195 case 3: // memory_order_release 1196 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1197 Ptr, NewVal, 1198 llvm::Release); 1199 break; 1200 case 4: // memory_order_acq_rel 1201 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1202 Ptr, NewVal, 1203 llvm::AcquireRelease); 1204 break; 1205 case 5: // memory_order_seq_cst 1206 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1207 Ptr, NewVal, 1208 llvm::SequentiallyConsistent); 1209 break; 1210 } 1211 Result->setVolatile(Volatile); 1212 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1213 } 1214 1215 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1216 1217 llvm::BasicBlock *BBs[5] = { 1218 createBasicBlock("monotonic", CurFn), 1219 createBasicBlock("acquire", CurFn), 1220 createBasicBlock("release", CurFn), 1221 createBasicBlock("acqrel", CurFn), 1222 createBasicBlock("seqcst", CurFn) 1223 }; 1224 llvm::AtomicOrdering Orders[5] = { 1225 llvm::Monotonic, llvm::Acquire, llvm::Release, 1226 llvm::AcquireRelease, llvm::SequentiallyConsistent 1227 }; 1228 1229 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1230 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1231 1232 Builder.SetInsertPoint(ContBB); 1233 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1234 1235 for (unsigned i = 0; i < 5; ++i) { 1236 Builder.SetInsertPoint(BBs[i]); 1237 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1238 Ptr, NewVal, Orders[i]); 1239 RMW->setVolatile(Volatile); 1240 Result->addIncoming(RMW, BBs[i]); 1241 Builder.CreateBr(ContBB); 1242 } 1243 1244 SI->addCase(Builder.getInt32(0), BBs[0]); 1245 SI->addCase(Builder.getInt32(1), BBs[1]); 1246 SI->addCase(Builder.getInt32(2), BBs[1]); 1247 SI->addCase(Builder.getInt32(3), BBs[2]); 1248 SI->addCase(Builder.getInt32(4), BBs[3]); 1249 SI->addCase(Builder.getInt32(5), BBs[4]); 1250 1251 Builder.SetInsertPoint(ContBB); 1252 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1253 } 1254 1255 case Builtin::BI__atomic_clear: { 1256 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1257 bool Volatile = 1258 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1259 1260 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1261 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1262 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1263 Value *NewVal = Builder.getInt8(0); 1264 Value *Order = EmitScalarExpr(E->getArg(1)); 1265 if (isa<llvm::ConstantInt>(Order)) { 1266 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1267 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1268 Store->setAlignment(1); 1269 switch (ord) { 1270 case 0: // memory_order_relaxed 1271 default: // invalid order 1272 Store->setOrdering(llvm::Monotonic); 1273 break; 1274 case 3: // memory_order_release 1275 Store->setOrdering(llvm::Release); 1276 break; 1277 case 5: // memory_order_seq_cst 1278 Store->setOrdering(llvm::SequentiallyConsistent); 1279 break; 1280 } 1281 return RValue::get(nullptr); 1282 } 1283 1284 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1285 1286 llvm::BasicBlock *BBs[3] = { 1287 createBasicBlock("monotonic", CurFn), 1288 createBasicBlock("release", CurFn), 1289 createBasicBlock("seqcst", CurFn) 1290 }; 1291 llvm::AtomicOrdering Orders[3] = { 1292 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1293 }; 1294 1295 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1296 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1297 1298 for (unsigned i = 0; i < 3; ++i) { 1299 Builder.SetInsertPoint(BBs[i]); 1300 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1301 Store->setAlignment(1); 1302 Store->setOrdering(Orders[i]); 1303 Builder.CreateBr(ContBB); 1304 } 1305 1306 SI->addCase(Builder.getInt32(0), BBs[0]); 1307 SI->addCase(Builder.getInt32(3), BBs[1]); 1308 SI->addCase(Builder.getInt32(5), BBs[2]); 1309 1310 Builder.SetInsertPoint(ContBB); 1311 return RValue::get(nullptr); 1312 } 1313 1314 case Builtin::BI__atomic_thread_fence: 1315 case Builtin::BI__atomic_signal_fence: 1316 case Builtin::BI__c11_atomic_thread_fence: 1317 case Builtin::BI__c11_atomic_signal_fence: { 1318 llvm::SynchronizationScope Scope; 1319 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1320 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1321 Scope = llvm::SingleThread; 1322 else 1323 Scope = llvm::CrossThread; 1324 Value *Order = EmitScalarExpr(E->getArg(0)); 1325 if (isa<llvm::ConstantInt>(Order)) { 1326 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1327 switch (ord) { 1328 case 0: // memory_order_relaxed 1329 default: // invalid order 1330 break; 1331 case 1: // memory_order_consume 1332 case 2: // memory_order_acquire 1333 Builder.CreateFence(llvm::Acquire, Scope); 1334 break; 1335 case 3: // memory_order_release 1336 Builder.CreateFence(llvm::Release, Scope); 1337 break; 1338 case 4: // memory_order_acq_rel 1339 Builder.CreateFence(llvm::AcquireRelease, Scope); 1340 break; 1341 case 5: // memory_order_seq_cst 1342 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1343 break; 1344 } 1345 return RValue::get(nullptr); 1346 } 1347 1348 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1349 AcquireBB = createBasicBlock("acquire", CurFn); 1350 ReleaseBB = createBasicBlock("release", CurFn); 1351 AcqRelBB = createBasicBlock("acqrel", CurFn); 1352 SeqCstBB = createBasicBlock("seqcst", CurFn); 1353 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1354 1355 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1356 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1357 1358 Builder.SetInsertPoint(AcquireBB); 1359 Builder.CreateFence(llvm::Acquire, Scope); 1360 Builder.CreateBr(ContBB); 1361 SI->addCase(Builder.getInt32(1), AcquireBB); 1362 SI->addCase(Builder.getInt32(2), AcquireBB); 1363 1364 Builder.SetInsertPoint(ReleaseBB); 1365 Builder.CreateFence(llvm::Release, Scope); 1366 Builder.CreateBr(ContBB); 1367 SI->addCase(Builder.getInt32(3), ReleaseBB); 1368 1369 Builder.SetInsertPoint(AcqRelBB); 1370 Builder.CreateFence(llvm::AcquireRelease, Scope); 1371 Builder.CreateBr(ContBB); 1372 SI->addCase(Builder.getInt32(4), AcqRelBB); 1373 1374 Builder.SetInsertPoint(SeqCstBB); 1375 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1376 Builder.CreateBr(ContBB); 1377 SI->addCase(Builder.getInt32(5), SeqCstBB); 1378 1379 Builder.SetInsertPoint(ContBB); 1380 return RValue::get(nullptr); 1381 } 1382 1383 // Library functions with special handling. 1384 case Builtin::BIsqrt: 1385 case Builtin::BIsqrtf: 1386 case Builtin::BIsqrtl: { 1387 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1388 // in finite- or unsafe-math mode (the intrinsic has different semantics 1389 // for handling negative numbers compared to the library function, so 1390 // -fmath-errno=0 is not enough). 1391 if (!FD->hasAttr<ConstAttr>()) 1392 break; 1393 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1394 CGM.getCodeGenOpts().NoNaNsFPMath)) 1395 break; 1396 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1397 llvm::Type *ArgType = Arg0->getType(); 1398 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1399 return RValue::get(Builder.CreateCall(F, Arg0)); 1400 } 1401 1402 case Builtin::BI__builtin_pow: 1403 case Builtin::BI__builtin_powf: 1404 case Builtin::BI__builtin_powl: 1405 case Builtin::BIpow: 1406 case Builtin::BIpowf: 1407 case Builtin::BIpowl: { 1408 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1409 if (!FD->hasAttr<ConstAttr>()) 1410 break; 1411 Value *Base = EmitScalarExpr(E->getArg(0)); 1412 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1413 llvm::Type *ArgType = Base->getType(); 1414 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1415 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1416 } 1417 1418 case Builtin::BIfma: 1419 case Builtin::BIfmaf: 1420 case Builtin::BIfmal: 1421 case Builtin::BI__builtin_fma: 1422 case Builtin::BI__builtin_fmaf: 1423 case Builtin::BI__builtin_fmal: { 1424 // Rewrite fma to intrinsic. 1425 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1426 llvm::Type *ArgType = FirstArg->getType(); 1427 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1428 return RValue::get(Builder.CreateCall3(F, FirstArg, 1429 EmitScalarExpr(E->getArg(1)), 1430 EmitScalarExpr(E->getArg(2)))); 1431 } 1432 1433 case Builtin::BI__builtin_signbit: 1434 case Builtin::BI__builtin_signbitf: 1435 case Builtin::BI__builtin_signbitl: { 1436 return RValue::get( 1437 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1438 ConvertType(E->getType()))); 1439 } 1440 case Builtin::BI__builtin_annotation: { 1441 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1442 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1443 AnnVal->getType()); 1444 1445 // Get the annotation string, go through casts. Sema requires this to be a 1446 // non-wide string literal, potentially casted, so the cast<> is safe. 1447 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1448 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1449 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1450 } 1451 case Builtin::BI__builtin_addcb: 1452 case Builtin::BI__builtin_addcs: 1453 case Builtin::BI__builtin_addc: 1454 case Builtin::BI__builtin_addcl: 1455 case Builtin::BI__builtin_addcll: 1456 case Builtin::BI__builtin_subcb: 1457 case Builtin::BI__builtin_subcs: 1458 case Builtin::BI__builtin_subc: 1459 case Builtin::BI__builtin_subcl: 1460 case Builtin::BI__builtin_subcll: { 1461 1462 // We translate all of these builtins from expressions of the form: 1463 // int x = ..., y = ..., carryin = ..., carryout, result; 1464 // result = __builtin_addc(x, y, carryin, &carryout); 1465 // 1466 // to LLVM IR of the form: 1467 // 1468 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1469 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1470 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1471 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1472 // i32 %carryin) 1473 // %result = extractvalue {i32, i1} %tmp2, 0 1474 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1475 // %tmp3 = or i1 %carry1, %carry2 1476 // %tmp4 = zext i1 %tmp3 to i32 1477 // store i32 %tmp4, i32* %carryout 1478 1479 // Scalarize our inputs. 1480 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1481 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1482 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1483 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1484 EmitPointerWithAlignment(E->getArg(3)); 1485 1486 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1487 llvm::Intrinsic::ID IntrinsicId; 1488 switch (BuiltinID) { 1489 default: llvm_unreachable("Unknown multiprecision builtin id."); 1490 case Builtin::BI__builtin_addcb: 1491 case Builtin::BI__builtin_addcs: 1492 case Builtin::BI__builtin_addc: 1493 case Builtin::BI__builtin_addcl: 1494 case Builtin::BI__builtin_addcll: 1495 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1496 break; 1497 case Builtin::BI__builtin_subcb: 1498 case Builtin::BI__builtin_subcs: 1499 case Builtin::BI__builtin_subc: 1500 case Builtin::BI__builtin_subcl: 1501 case Builtin::BI__builtin_subcll: 1502 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1503 break; 1504 } 1505 1506 // Construct our resulting LLVM IR expression. 1507 llvm::Value *Carry1; 1508 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1509 X, Y, Carry1); 1510 llvm::Value *Carry2; 1511 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1512 Sum1, Carryin, Carry2); 1513 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1514 X->getType()); 1515 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1516 CarryOutPtr.first); 1517 CarryOutStore->setAlignment(CarryOutPtr.second); 1518 return RValue::get(Sum2); 1519 } 1520 case Builtin::BI__builtin_uadd_overflow: 1521 case Builtin::BI__builtin_uaddl_overflow: 1522 case Builtin::BI__builtin_uaddll_overflow: 1523 case Builtin::BI__builtin_usub_overflow: 1524 case Builtin::BI__builtin_usubl_overflow: 1525 case Builtin::BI__builtin_usubll_overflow: 1526 case Builtin::BI__builtin_umul_overflow: 1527 case Builtin::BI__builtin_umull_overflow: 1528 case Builtin::BI__builtin_umulll_overflow: 1529 case Builtin::BI__builtin_sadd_overflow: 1530 case Builtin::BI__builtin_saddl_overflow: 1531 case Builtin::BI__builtin_saddll_overflow: 1532 case Builtin::BI__builtin_ssub_overflow: 1533 case Builtin::BI__builtin_ssubl_overflow: 1534 case Builtin::BI__builtin_ssubll_overflow: 1535 case Builtin::BI__builtin_smul_overflow: 1536 case Builtin::BI__builtin_smull_overflow: 1537 case Builtin::BI__builtin_smulll_overflow: { 1538 1539 // We translate all of these builtins directly to the relevant llvm IR node. 1540 1541 // Scalarize our inputs. 1542 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1543 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1544 std::pair<llvm::Value *, unsigned> SumOutPtr = 1545 EmitPointerWithAlignment(E->getArg(2)); 1546 1547 // Decide which of the overflow intrinsics we are lowering to: 1548 llvm::Intrinsic::ID IntrinsicId; 1549 switch (BuiltinID) { 1550 default: llvm_unreachable("Unknown security overflow builtin id."); 1551 case Builtin::BI__builtin_uadd_overflow: 1552 case Builtin::BI__builtin_uaddl_overflow: 1553 case Builtin::BI__builtin_uaddll_overflow: 1554 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1555 break; 1556 case Builtin::BI__builtin_usub_overflow: 1557 case Builtin::BI__builtin_usubl_overflow: 1558 case Builtin::BI__builtin_usubll_overflow: 1559 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1560 break; 1561 case Builtin::BI__builtin_umul_overflow: 1562 case Builtin::BI__builtin_umull_overflow: 1563 case Builtin::BI__builtin_umulll_overflow: 1564 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1565 break; 1566 case Builtin::BI__builtin_sadd_overflow: 1567 case Builtin::BI__builtin_saddl_overflow: 1568 case Builtin::BI__builtin_saddll_overflow: 1569 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1570 break; 1571 case Builtin::BI__builtin_ssub_overflow: 1572 case Builtin::BI__builtin_ssubl_overflow: 1573 case Builtin::BI__builtin_ssubll_overflow: 1574 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1575 break; 1576 case Builtin::BI__builtin_smul_overflow: 1577 case Builtin::BI__builtin_smull_overflow: 1578 case Builtin::BI__builtin_smulll_overflow: 1579 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1580 break; 1581 } 1582 1583 1584 llvm::Value *Carry; 1585 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1586 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1587 SumOutStore->setAlignment(SumOutPtr.second); 1588 1589 return RValue::get(Carry); 1590 } 1591 case Builtin::BI__builtin_addressof: 1592 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1593 case Builtin::BI__builtin_operator_new: 1594 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1595 E->getArg(0), false); 1596 case Builtin::BI__builtin_operator_delete: 1597 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1598 E->getArg(0), true); 1599 case Builtin::BI__noop: 1600 // __noop always evaluates to an integer literal zero. 1601 return RValue::get(ConstantInt::get(IntTy, 0)); 1602 case Builtin::BI__builtin_call_with_static_chain: { 1603 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1604 const Expr *Chain = E->getArg(1); 1605 return EmitCall(Call->getCallee()->getType(), 1606 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1607 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1608 } 1609 case Builtin::BI_InterlockedExchange: 1610 case Builtin::BI_InterlockedExchangePointer: 1611 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1612 case Builtin::BI_InterlockedCompareExchangePointer: { 1613 llvm::Type *RTy; 1614 llvm::IntegerType *IntType = 1615 IntegerType::get(getLLVMContext(), 1616 getContext().getTypeSize(E->getType())); 1617 llvm::Type *IntPtrType = IntType->getPointerTo(); 1618 1619 llvm::Value *Destination = 1620 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1621 1622 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1623 RTy = Exchange->getType(); 1624 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1625 1626 llvm::Value *Comparand = 1627 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1628 1629 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1630 SequentiallyConsistent, 1631 SequentiallyConsistent); 1632 Result->setVolatile(true); 1633 1634 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1635 0), 1636 RTy)); 1637 } 1638 case Builtin::BI_InterlockedCompareExchange: { 1639 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1640 EmitScalarExpr(E->getArg(0)), 1641 EmitScalarExpr(E->getArg(2)), 1642 EmitScalarExpr(E->getArg(1)), 1643 SequentiallyConsistent, 1644 SequentiallyConsistent); 1645 CXI->setVolatile(true); 1646 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1647 } 1648 case Builtin::BI_InterlockedIncrement: { 1649 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1650 AtomicRMWInst::Add, 1651 EmitScalarExpr(E->getArg(0)), 1652 ConstantInt::get(Int32Ty, 1), 1653 llvm::SequentiallyConsistent); 1654 RMWI->setVolatile(true); 1655 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1656 } 1657 case Builtin::BI_InterlockedDecrement: { 1658 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1659 AtomicRMWInst::Sub, 1660 EmitScalarExpr(E->getArg(0)), 1661 ConstantInt::get(Int32Ty, 1), 1662 llvm::SequentiallyConsistent); 1663 RMWI->setVolatile(true); 1664 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1665 } 1666 case Builtin::BI_InterlockedExchangeAdd: { 1667 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1668 AtomicRMWInst::Add, 1669 EmitScalarExpr(E->getArg(0)), 1670 EmitScalarExpr(E->getArg(1)), 1671 llvm::SequentiallyConsistent); 1672 RMWI->setVolatile(true); 1673 return RValue::get(RMWI); 1674 } 1675 case Builtin::BI__readfsdword: { 1676 Value *IntToPtr = 1677 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1678 llvm::PointerType::get(CGM.Int32Ty, 257)); 1679 LoadInst *Load = 1680 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1681 return RValue::get(Load); 1682 } 1683 1684 case Builtin::BI__exception_code: 1685 case Builtin::BI_exception_code: 1686 return RValue::get(EmitSEHExceptionCode()); 1687 case Builtin::BI__exception_info: 1688 case Builtin::BI_exception_info: 1689 return RValue::get(EmitSEHExceptionInfo()); 1690 case Builtin::BI__abnormal_termination: 1691 case Builtin::BI_abnormal_termination: 1692 return RValue::get(EmitSEHAbnormalTermination()); 1693 case Builtin::BI_setjmpex: { 1694 if (getTarget().getTriple().isOSMSVCRT()) { 1695 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1696 llvm::AttributeSet ReturnsTwiceAttr = 1697 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1698 llvm::Attribute::ReturnsTwice); 1699 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1700 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1701 "_setjmpex", ReturnsTwiceAttr); 1702 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1703 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1704 llvm::Value *FrameAddr = 1705 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1706 ConstantInt::get(Int32Ty, 0)); 1707 llvm::Value *Args[] = {Buf, FrameAddr}; 1708 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1709 CS.setAttributes(ReturnsTwiceAttr); 1710 return RValue::get(CS.getInstruction()); 1711 } 1712 break; 1713 } 1714 case Builtin::BI_setjmp: { 1715 if (getTarget().getTriple().isOSMSVCRT()) { 1716 llvm::AttributeSet ReturnsTwiceAttr = 1717 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1718 llvm::Attribute::ReturnsTwice); 1719 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1720 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1721 llvm::CallSite CS; 1722 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1723 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1724 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1725 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1726 "_setjmp3", ReturnsTwiceAttr); 1727 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1728 llvm::Value *Args[] = {Buf, Count}; 1729 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1730 } else { 1731 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1732 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1733 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1734 "_setjmp", ReturnsTwiceAttr); 1735 llvm::Value *FrameAddr = 1736 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1737 ConstantInt::get(Int32Ty, 0)); 1738 llvm::Value *Args[] = {Buf, FrameAddr}; 1739 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1740 } 1741 CS.setAttributes(ReturnsTwiceAttr); 1742 return RValue::get(CS.getInstruction()); 1743 } 1744 break; 1745 } 1746 1747 case Builtin::BI__GetExceptionInfo: { 1748 if (llvm::GlobalVariable *GV = 1749 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1750 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1751 break; 1752 } 1753 } 1754 1755 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1756 // the call using the normal call path, but using the unmangled 1757 // version of the function name. 1758 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1759 return emitLibraryCall(*this, FD, E, 1760 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1761 1762 // If this is a predefined lib function (e.g. malloc), emit the call 1763 // using exactly the normal call path. 1764 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1765 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1766 1767 // See if we have a target specific intrinsic. 1768 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1769 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1770 if (const char *Prefix = 1771 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1772 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1773 // NOTE we dont need to perform a compatibility flag check here since the 1774 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1775 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1776 if (IntrinsicID == Intrinsic::not_intrinsic) 1777 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1778 } 1779 1780 if (IntrinsicID != Intrinsic::not_intrinsic) { 1781 SmallVector<Value*, 16> Args; 1782 1783 // Find out if any arguments are required to be integer constant 1784 // expressions. 1785 unsigned ICEArguments = 0; 1786 ASTContext::GetBuiltinTypeError Error; 1787 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1788 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1789 1790 Function *F = CGM.getIntrinsic(IntrinsicID); 1791 llvm::FunctionType *FTy = F->getFunctionType(); 1792 1793 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1794 Value *ArgValue; 1795 // If this is a normal argument, just emit it as a scalar. 1796 if ((ICEArguments & (1 << i)) == 0) { 1797 ArgValue = EmitScalarExpr(E->getArg(i)); 1798 } else { 1799 // If this is required to be a constant, constant fold it so that we 1800 // know that the generated intrinsic gets a ConstantInt. 1801 llvm::APSInt Result; 1802 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1803 assert(IsConst && "Constant arg isn't actually constant?"); 1804 (void)IsConst; 1805 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1806 } 1807 1808 // If the intrinsic arg type is different from the builtin arg type 1809 // we need to do a bit cast. 1810 llvm::Type *PTy = FTy->getParamType(i); 1811 if (PTy != ArgValue->getType()) { 1812 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1813 "Must be able to losslessly bit cast to param"); 1814 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1815 } 1816 1817 Args.push_back(ArgValue); 1818 } 1819 1820 Value *V = Builder.CreateCall(F, Args); 1821 QualType BuiltinRetType = E->getType(); 1822 1823 llvm::Type *RetTy = VoidTy; 1824 if (!BuiltinRetType->isVoidType()) 1825 RetTy = ConvertType(BuiltinRetType); 1826 1827 if (RetTy != V->getType()) { 1828 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1829 "Must be able to losslessly bit cast result type"); 1830 V = Builder.CreateBitCast(V, RetTy); 1831 } 1832 1833 return RValue::get(V); 1834 } 1835 1836 // See if we have a target specific builtin that needs to be lowered. 1837 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1838 return RValue::get(V); 1839 1840 ErrorUnsupported(E, "builtin function"); 1841 1842 // Unknown builtin, for now just dump it out and return undef. 1843 return GetUndefRValue(E->getType()); 1844 } 1845 1846 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1847 const CallExpr *E) { 1848 switch (getTarget().getTriple().getArch()) { 1849 case llvm::Triple::arm: 1850 case llvm::Triple::armeb: 1851 case llvm::Triple::thumb: 1852 case llvm::Triple::thumbeb: 1853 return EmitARMBuiltinExpr(BuiltinID, E); 1854 case llvm::Triple::aarch64: 1855 case llvm::Triple::aarch64_be: 1856 return EmitAArch64BuiltinExpr(BuiltinID, E); 1857 case llvm::Triple::x86: 1858 case llvm::Triple::x86_64: 1859 return EmitX86BuiltinExpr(BuiltinID, E); 1860 case llvm::Triple::ppc: 1861 case llvm::Triple::ppc64: 1862 case llvm::Triple::ppc64le: 1863 return EmitPPCBuiltinExpr(BuiltinID, E); 1864 case llvm::Triple::r600: 1865 case llvm::Triple::amdgcn: 1866 return EmitR600BuiltinExpr(BuiltinID, E); 1867 default: 1868 return nullptr; 1869 } 1870 } 1871 1872 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1873 NeonTypeFlags TypeFlags, 1874 bool V1Ty=false) { 1875 int IsQuad = TypeFlags.isQuad(); 1876 switch (TypeFlags.getEltType()) { 1877 case NeonTypeFlags::Int8: 1878 case NeonTypeFlags::Poly8: 1879 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1880 case NeonTypeFlags::Int16: 1881 case NeonTypeFlags::Poly16: 1882 case NeonTypeFlags::Float16: 1883 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1884 case NeonTypeFlags::Int32: 1885 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1886 case NeonTypeFlags::Int64: 1887 case NeonTypeFlags::Poly64: 1888 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1889 case NeonTypeFlags::Poly128: 1890 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1891 // There is a lot of i128 and f128 API missing. 1892 // so we use v16i8 to represent poly128 and get pattern matched. 1893 return llvm::VectorType::get(CGF->Int8Ty, 16); 1894 case NeonTypeFlags::Float32: 1895 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1896 case NeonTypeFlags::Float64: 1897 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1898 } 1899 llvm_unreachable("Unknown vector element type!"); 1900 } 1901 1902 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1903 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1904 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1905 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1906 } 1907 1908 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1909 const char *name, 1910 unsigned shift, bool rightshift) { 1911 unsigned j = 0; 1912 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1913 ai != ae; ++ai, ++j) 1914 if (shift > 0 && shift == j) 1915 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1916 else 1917 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1918 1919 return Builder.CreateCall(F, Ops, name); 1920 } 1921 1922 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1923 bool neg) { 1924 int SV = cast<ConstantInt>(V)->getSExtValue(); 1925 1926 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1927 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1928 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1929 } 1930 1931 // \brief Right-shift a vector by a constant. 1932 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1933 llvm::Type *Ty, bool usgn, 1934 const char *name) { 1935 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1936 1937 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1938 int EltSize = VTy->getScalarSizeInBits(); 1939 1940 Vec = Builder.CreateBitCast(Vec, Ty); 1941 1942 // lshr/ashr are undefined when the shift amount is equal to the vector 1943 // element size. 1944 if (ShiftAmt == EltSize) { 1945 if (usgn) { 1946 // Right-shifting an unsigned value by its size yields 0. 1947 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1948 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1949 } else { 1950 // Right-shifting a signed value by its size is equivalent 1951 // to a shift of size-1. 1952 --ShiftAmt; 1953 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1954 } 1955 } 1956 1957 Shift = EmitNeonShiftVector(Shift, Ty, false); 1958 if (usgn) 1959 return Builder.CreateLShr(Vec, Shift, name); 1960 else 1961 return Builder.CreateAShr(Vec, Shift, name); 1962 } 1963 1964 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1965 /// alignment of the type referenced by the pointer. Skip over implicit 1966 /// casts. 1967 std::pair<llvm::Value*, unsigned> 1968 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1969 assert(Addr->getType()->isPointerType()); 1970 Addr = Addr->IgnoreParens(); 1971 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1972 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1973 ICE->getSubExpr()->getType()->isPointerType()) { 1974 std::pair<llvm::Value*, unsigned> Ptr = 1975 EmitPointerWithAlignment(ICE->getSubExpr()); 1976 Ptr.first = Builder.CreateBitCast(Ptr.first, 1977 ConvertType(Addr->getType())); 1978 return Ptr; 1979 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1980 LValue LV = EmitLValue(ICE->getSubExpr()); 1981 unsigned Align = LV.getAlignment().getQuantity(); 1982 if (!Align) { 1983 // FIXME: Once LValues are fixed to always set alignment, 1984 // zap this code. 1985 QualType PtTy = ICE->getSubExpr()->getType(); 1986 if (!PtTy->isIncompleteType()) 1987 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1988 else 1989 Align = 1; 1990 } 1991 return std::make_pair(LV.getAddress(), Align); 1992 } 1993 } 1994 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1995 if (UO->getOpcode() == UO_AddrOf) { 1996 LValue LV = EmitLValue(UO->getSubExpr()); 1997 unsigned Align = LV.getAlignment().getQuantity(); 1998 if (!Align) { 1999 // FIXME: Once LValues are fixed to always set alignment, 2000 // zap this code. 2001 QualType PtTy = UO->getSubExpr()->getType(); 2002 if (!PtTy->isIncompleteType()) 2003 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2004 else 2005 Align = 1; 2006 } 2007 return std::make_pair(LV.getAddress(), Align); 2008 } 2009 } 2010 2011 unsigned Align = 1; 2012 QualType PtTy = Addr->getType()->getPointeeType(); 2013 if (!PtTy->isIncompleteType()) 2014 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2015 2016 return std::make_pair(EmitScalarExpr(Addr), Align); 2017 } 2018 2019 enum { 2020 AddRetType = (1 << 0), 2021 Add1ArgType = (1 << 1), 2022 Add2ArgTypes = (1 << 2), 2023 2024 VectorizeRetType = (1 << 3), 2025 VectorizeArgTypes = (1 << 4), 2026 2027 InventFloatType = (1 << 5), 2028 UnsignedAlts = (1 << 6), 2029 2030 Use64BitVectors = (1 << 7), 2031 Use128BitVectors = (1 << 8), 2032 2033 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2034 VectorRet = AddRetType | VectorizeRetType, 2035 VectorRetGetArgs01 = 2036 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2037 FpCmpzModifiers = 2038 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2039 }; 2040 2041 struct NeonIntrinsicInfo { 2042 unsigned BuiltinID; 2043 unsigned LLVMIntrinsic; 2044 unsigned AltLLVMIntrinsic; 2045 const char *NameHint; 2046 unsigned TypeModifier; 2047 2048 bool operator<(unsigned RHSBuiltinID) const { 2049 return BuiltinID < RHSBuiltinID; 2050 } 2051 }; 2052 2053 #define NEONMAP0(NameBase) \ 2054 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2055 2056 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2057 { NEON:: BI__builtin_neon_ ## NameBase, \ 2058 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2059 2060 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2061 { NEON:: BI__builtin_neon_ ## NameBase, \ 2062 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2063 #NameBase, TypeModifier } 2064 2065 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2066 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2067 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2068 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2069 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2070 NEONMAP0(vaddhn_v), 2071 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2072 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2073 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2074 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2075 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2076 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2077 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2078 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2079 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2080 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2081 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2082 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2083 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2084 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2085 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2086 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2087 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2088 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2089 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2090 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2091 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 2092 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2093 NEONMAP0(vcvt_f32_v), 2094 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2095 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2096 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2097 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2098 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2099 NEONMAP0(vcvt_s32_v), 2100 NEONMAP0(vcvt_s64_v), 2101 NEONMAP0(vcvt_u32_v), 2102 NEONMAP0(vcvt_u64_v), 2103 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2104 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2105 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2106 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2107 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2108 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2109 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2110 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2111 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2112 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2113 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2114 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2115 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2116 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2117 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2118 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2119 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2120 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2121 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2122 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2123 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2124 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2125 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2126 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2127 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2128 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2129 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2130 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2131 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2132 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2133 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2134 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2135 NEONMAP0(vcvtq_f32_v), 2136 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2137 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2138 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2139 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2140 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2141 NEONMAP0(vcvtq_s32_v), 2142 NEONMAP0(vcvtq_s64_v), 2143 NEONMAP0(vcvtq_u32_v), 2144 NEONMAP0(vcvtq_u64_v), 2145 NEONMAP0(vext_v), 2146 NEONMAP0(vextq_v), 2147 NEONMAP0(vfma_v), 2148 NEONMAP0(vfmaq_v), 2149 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2150 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2151 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2152 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2153 NEONMAP0(vld1_dup_v), 2154 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2155 NEONMAP0(vld1q_dup_v), 2156 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2157 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2158 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2159 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2160 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2161 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2162 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2163 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2164 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2165 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2166 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2167 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2168 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2169 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2170 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2171 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2172 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2173 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2174 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2175 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2176 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2177 NEONMAP0(vmovl_v), 2178 NEONMAP0(vmovn_v), 2179 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2180 NEONMAP0(vmull_v), 2181 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2182 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2183 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2184 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2185 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2186 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2187 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2188 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2189 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2190 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2191 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2192 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2193 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2194 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2195 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2196 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2197 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2198 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2199 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2200 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2201 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2202 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2203 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2204 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2205 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2206 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2207 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2208 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2209 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2210 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2211 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2212 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2213 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2214 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2215 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2216 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2217 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2218 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2219 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2220 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2221 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2222 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2223 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2224 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2225 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2226 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2227 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2228 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2229 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2230 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2231 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2232 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2233 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2234 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2235 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2236 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2237 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2238 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2239 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2240 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2241 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2242 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2243 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2244 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2245 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2246 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2247 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2248 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2249 NEONMAP0(vshl_n_v), 2250 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2251 NEONMAP0(vshll_n_v), 2252 NEONMAP0(vshlq_n_v), 2253 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2254 NEONMAP0(vshr_n_v), 2255 NEONMAP0(vshrn_n_v), 2256 NEONMAP0(vshrq_n_v), 2257 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2258 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2259 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2260 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2261 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2262 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2263 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2264 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2265 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2266 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2267 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2268 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2269 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2270 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2271 NEONMAP0(vsubhn_v), 2272 NEONMAP0(vtrn_v), 2273 NEONMAP0(vtrnq_v), 2274 NEONMAP0(vtst_v), 2275 NEONMAP0(vtstq_v), 2276 NEONMAP0(vuzp_v), 2277 NEONMAP0(vuzpq_v), 2278 NEONMAP0(vzip_v), 2279 NEONMAP0(vzipq_v) 2280 }; 2281 2282 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2283 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2284 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2285 NEONMAP0(vaddhn_v), 2286 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2287 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2288 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2289 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2290 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2291 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2292 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2293 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2294 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2295 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2296 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2297 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2298 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2299 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2300 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2301 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2302 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2303 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2304 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2305 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2306 NEONMAP0(vcvt_f32_v), 2307 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2308 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2309 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2310 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2311 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2312 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2313 NEONMAP0(vcvtq_f32_v), 2314 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2315 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2316 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2317 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2318 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2319 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2320 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2321 NEONMAP0(vext_v), 2322 NEONMAP0(vextq_v), 2323 NEONMAP0(vfma_v), 2324 NEONMAP0(vfmaq_v), 2325 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2326 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2327 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2328 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2329 NEONMAP0(vmovl_v), 2330 NEONMAP0(vmovn_v), 2331 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2332 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2333 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2334 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2335 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2336 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2337 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2338 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2339 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2340 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2341 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2342 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2343 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2344 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2345 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2346 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2347 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2348 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2349 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2350 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2351 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2352 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2353 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2354 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2355 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2356 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2357 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2358 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2359 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2360 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2361 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2362 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2363 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2364 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2365 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2366 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2367 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2368 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2369 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2370 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2371 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2372 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2373 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2374 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2375 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2376 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2377 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2378 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2379 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2380 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2381 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2382 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2383 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2384 NEONMAP0(vshl_n_v), 2385 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2386 NEONMAP0(vshll_n_v), 2387 NEONMAP0(vshlq_n_v), 2388 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2389 NEONMAP0(vshr_n_v), 2390 NEONMAP0(vshrn_n_v), 2391 NEONMAP0(vshrq_n_v), 2392 NEONMAP0(vsubhn_v), 2393 NEONMAP0(vtst_v), 2394 NEONMAP0(vtstq_v), 2395 }; 2396 2397 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2398 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2399 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2400 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2401 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2402 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2403 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2404 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2405 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2406 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2407 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2408 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2409 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2410 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2411 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2412 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2413 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2414 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2415 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2416 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2417 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2418 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2419 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2420 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2421 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2422 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2423 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2424 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2425 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2426 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2427 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2428 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2429 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2430 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2431 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2432 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2433 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2434 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2435 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2436 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2437 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2438 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2439 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2440 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2441 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2442 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2443 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2444 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2445 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2446 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2447 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2448 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2449 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2450 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2451 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2452 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2453 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2454 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2455 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2456 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2457 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2458 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2459 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2460 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2461 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2462 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2463 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2464 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2465 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2466 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2467 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2468 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2469 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2470 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2471 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2472 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2473 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2474 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2475 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2476 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2477 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2478 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2479 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2480 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2481 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2482 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2483 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2484 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2485 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2486 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2487 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2488 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2489 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2490 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2491 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2492 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2493 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2494 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2495 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2496 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2497 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2498 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2499 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2500 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2501 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2502 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2503 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2504 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2505 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2506 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2507 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2508 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2509 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2510 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2511 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2512 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2513 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2514 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2515 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2516 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2517 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2518 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2519 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2520 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2521 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2522 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2523 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2524 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2525 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2526 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2527 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2528 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2529 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2530 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2531 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2532 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2533 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2534 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2535 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2536 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2537 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2538 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2539 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2540 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2541 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2542 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2543 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2544 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2545 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2546 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2547 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2548 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2549 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2550 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2551 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2552 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2553 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2554 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2555 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2556 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2557 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2558 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2559 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2560 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2561 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2562 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2563 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2564 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2565 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2566 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2567 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2568 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2569 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2570 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2571 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2572 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2573 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2574 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2575 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2576 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2577 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2578 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2579 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2580 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2581 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2582 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2583 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2584 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2585 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2586 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2587 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2588 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2589 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2590 }; 2591 2592 #undef NEONMAP0 2593 #undef NEONMAP1 2594 #undef NEONMAP2 2595 2596 static bool NEONSIMDIntrinsicsProvenSorted = false; 2597 2598 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2599 static bool AArch64SISDIntrinsicsProvenSorted = false; 2600 2601 2602 static const NeonIntrinsicInfo * 2603 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2604 unsigned BuiltinID, bool &MapProvenSorted) { 2605 2606 #ifndef NDEBUG 2607 if (!MapProvenSorted) { 2608 // FIXME: use std::is_sorted once C++11 is allowed 2609 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2610 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2611 MapProvenSorted = true; 2612 } 2613 #endif 2614 2615 const NeonIntrinsicInfo *Builtin = 2616 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2617 2618 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2619 return Builtin; 2620 2621 return nullptr; 2622 } 2623 2624 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2625 unsigned Modifier, 2626 llvm::Type *ArgType, 2627 const CallExpr *E) { 2628 int VectorSize = 0; 2629 if (Modifier & Use64BitVectors) 2630 VectorSize = 64; 2631 else if (Modifier & Use128BitVectors) 2632 VectorSize = 128; 2633 2634 // Return type. 2635 SmallVector<llvm::Type *, 3> Tys; 2636 if (Modifier & AddRetType) { 2637 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2638 if (Modifier & VectorizeRetType) 2639 Ty = llvm::VectorType::get( 2640 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2641 2642 Tys.push_back(Ty); 2643 } 2644 2645 // Arguments. 2646 if (Modifier & VectorizeArgTypes) { 2647 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2648 ArgType = llvm::VectorType::get(ArgType, Elts); 2649 } 2650 2651 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2652 Tys.push_back(ArgType); 2653 2654 if (Modifier & Add2ArgTypes) 2655 Tys.push_back(ArgType); 2656 2657 if (Modifier & InventFloatType) 2658 Tys.push_back(FloatTy); 2659 2660 return CGM.getIntrinsic(IntrinsicID, Tys); 2661 } 2662 2663 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2664 const NeonIntrinsicInfo &SISDInfo, 2665 SmallVectorImpl<Value *> &Ops, 2666 const CallExpr *E) { 2667 unsigned BuiltinID = SISDInfo.BuiltinID; 2668 unsigned int Int = SISDInfo.LLVMIntrinsic; 2669 unsigned Modifier = SISDInfo.TypeModifier; 2670 const char *s = SISDInfo.NameHint; 2671 2672 switch (BuiltinID) { 2673 case NEON::BI__builtin_neon_vcled_s64: 2674 case NEON::BI__builtin_neon_vcled_u64: 2675 case NEON::BI__builtin_neon_vcles_f32: 2676 case NEON::BI__builtin_neon_vcled_f64: 2677 case NEON::BI__builtin_neon_vcltd_s64: 2678 case NEON::BI__builtin_neon_vcltd_u64: 2679 case NEON::BI__builtin_neon_vclts_f32: 2680 case NEON::BI__builtin_neon_vcltd_f64: 2681 case NEON::BI__builtin_neon_vcales_f32: 2682 case NEON::BI__builtin_neon_vcaled_f64: 2683 case NEON::BI__builtin_neon_vcalts_f32: 2684 case NEON::BI__builtin_neon_vcaltd_f64: 2685 // Only one direction of comparisons actually exist, cmle is actually a cmge 2686 // with swapped operands. The table gives us the right intrinsic but we 2687 // still need to do the swap. 2688 std::swap(Ops[0], Ops[1]); 2689 break; 2690 } 2691 2692 assert(Int && "Generic code assumes a valid intrinsic"); 2693 2694 // Determine the type(s) of this overloaded AArch64 intrinsic. 2695 const Expr *Arg = E->getArg(0); 2696 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2697 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2698 2699 int j = 0; 2700 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2701 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2702 ai != ae; ++ai, ++j) { 2703 llvm::Type *ArgTy = ai->getType(); 2704 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2705 ArgTy->getPrimitiveSizeInBits()) 2706 continue; 2707 2708 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2709 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2710 // it before inserting. 2711 Ops[j] = 2712 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2713 Ops[j] = 2714 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2715 } 2716 2717 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2718 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2719 if (ResultType->getPrimitiveSizeInBits() < 2720 Result->getType()->getPrimitiveSizeInBits()) 2721 return CGF.Builder.CreateExtractElement(Result, C0); 2722 2723 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2724 } 2725 2726 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2727 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2728 const char *NameHint, unsigned Modifier, const CallExpr *E, 2729 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2730 // Get the last argument, which specifies the vector type. 2731 llvm::APSInt NeonTypeConst; 2732 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2733 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2734 return nullptr; 2735 2736 // Determine the type of this overloaded NEON intrinsic. 2737 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2738 bool Usgn = Type.isUnsigned(); 2739 bool Quad = Type.isQuad(); 2740 2741 llvm::VectorType *VTy = GetNeonType(this, Type); 2742 llvm::Type *Ty = VTy; 2743 if (!Ty) 2744 return nullptr; 2745 2746 unsigned Int = LLVMIntrinsic; 2747 if ((Modifier & UnsignedAlts) && !Usgn) 2748 Int = AltLLVMIntrinsic; 2749 2750 switch (BuiltinID) { 2751 default: break; 2752 case NEON::BI__builtin_neon_vabs_v: 2753 case NEON::BI__builtin_neon_vabsq_v: 2754 if (VTy->getElementType()->isFloatingPointTy()) 2755 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2756 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2757 case NEON::BI__builtin_neon_vaddhn_v: { 2758 llvm::VectorType *SrcTy = 2759 llvm::VectorType::getExtendedElementVectorType(VTy); 2760 2761 // %sum = add <4 x i32> %lhs, %rhs 2762 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2763 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2764 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2765 2766 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2767 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2768 SrcTy->getScalarSizeInBits() / 2); 2769 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2770 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2771 2772 // %res = trunc <4 x i32> %high to <4 x i16> 2773 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2774 } 2775 case NEON::BI__builtin_neon_vcale_v: 2776 case NEON::BI__builtin_neon_vcaleq_v: 2777 case NEON::BI__builtin_neon_vcalt_v: 2778 case NEON::BI__builtin_neon_vcaltq_v: 2779 std::swap(Ops[0], Ops[1]); 2780 case NEON::BI__builtin_neon_vcage_v: 2781 case NEON::BI__builtin_neon_vcageq_v: 2782 case NEON::BI__builtin_neon_vcagt_v: 2783 case NEON::BI__builtin_neon_vcagtq_v: { 2784 llvm::Type *VecFlt = llvm::VectorType::get( 2785 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2786 VTy->getNumElements()); 2787 llvm::Type *Tys[] = { VTy, VecFlt }; 2788 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2789 return EmitNeonCall(F, Ops, NameHint); 2790 } 2791 case NEON::BI__builtin_neon_vclz_v: 2792 case NEON::BI__builtin_neon_vclzq_v: 2793 // We generate target-independent intrinsic, which needs a second argument 2794 // for whether or not clz of zero is undefined; on ARM it isn't. 2795 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2796 break; 2797 case NEON::BI__builtin_neon_vcvt_f32_v: 2798 case NEON::BI__builtin_neon_vcvtq_f32_v: 2799 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2800 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2801 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2802 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2803 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2804 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2805 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2806 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2807 bool Double = 2808 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2809 llvm::Type *FloatTy = 2810 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2811 : NeonTypeFlags::Float32, 2812 false, Quad)); 2813 llvm::Type *Tys[2] = { FloatTy, Ty }; 2814 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2815 Function *F = CGM.getIntrinsic(Int, Tys); 2816 return EmitNeonCall(F, Ops, "vcvt_n"); 2817 } 2818 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2819 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2820 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2821 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2822 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2823 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2824 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2825 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2826 bool Double = 2827 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2828 llvm::Type *FloatTy = 2829 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2830 : NeonTypeFlags::Float32, 2831 false, Quad)); 2832 llvm::Type *Tys[2] = { Ty, FloatTy }; 2833 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2834 return EmitNeonCall(F, Ops, "vcvt_n"); 2835 } 2836 case NEON::BI__builtin_neon_vcvt_s32_v: 2837 case NEON::BI__builtin_neon_vcvt_u32_v: 2838 case NEON::BI__builtin_neon_vcvt_s64_v: 2839 case NEON::BI__builtin_neon_vcvt_u64_v: 2840 case NEON::BI__builtin_neon_vcvtq_s32_v: 2841 case NEON::BI__builtin_neon_vcvtq_u32_v: 2842 case NEON::BI__builtin_neon_vcvtq_s64_v: 2843 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2844 bool Double = 2845 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2846 llvm::Type *FloatTy = 2847 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2848 : NeonTypeFlags::Float32, 2849 false, Quad)); 2850 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2851 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2852 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2853 } 2854 case NEON::BI__builtin_neon_vcvta_s32_v: 2855 case NEON::BI__builtin_neon_vcvta_s64_v: 2856 case NEON::BI__builtin_neon_vcvta_u32_v: 2857 case NEON::BI__builtin_neon_vcvta_u64_v: 2858 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2859 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2860 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2861 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2862 case NEON::BI__builtin_neon_vcvtn_s32_v: 2863 case NEON::BI__builtin_neon_vcvtn_s64_v: 2864 case NEON::BI__builtin_neon_vcvtn_u32_v: 2865 case NEON::BI__builtin_neon_vcvtn_u64_v: 2866 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2867 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2868 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2869 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2870 case NEON::BI__builtin_neon_vcvtp_s32_v: 2871 case NEON::BI__builtin_neon_vcvtp_s64_v: 2872 case NEON::BI__builtin_neon_vcvtp_u32_v: 2873 case NEON::BI__builtin_neon_vcvtp_u64_v: 2874 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2875 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2876 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2877 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2878 case NEON::BI__builtin_neon_vcvtm_s32_v: 2879 case NEON::BI__builtin_neon_vcvtm_s64_v: 2880 case NEON::BI__builtin_neon_vcvtm_u32_v: 2881 case NEON::BI__builtin_neon_vcvtm_u64_v: 2882 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2883 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2884 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2885 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2886 bool Double = 2887 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2888 llvm::Type *InTy = 2889 GetNeonType(this, 2890 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2891 : NeonTypeFlags::Float32, false, Quad)); 2892 llvm::Type *Tys[2] = { Ty, InTy }; 2893 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2894 } 2895 case NEON::BI__builtin_neon_vext_v: 2896 case NEON::BI__builtin_neon_vextq_v: { 2897 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2898 SmallVector<Constant*, 16> Indices; 2899 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2900 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2901 2902 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2903 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2904 Value *SV = llvm::ConstantVector::get(Indices); 2905 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2906 } 2907 case NEON::BI__builtin_neon_vfma_v: 2908 case NEON::BI__builtin_neon_vfmaq_v: { 2909 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2910 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2911 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2912 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2913 2914 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2915 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2916 } 2917 case NEON::BI__builtin_neon_vld1_v: 2918 case NEON::BI__builtin_neon_vld1q_v: 2919 Ops.push_back(Align); 2920 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2921 case NEON::BI__builtin_neon_vld2_v: 2922 case NEON::BI__builtin_neon_vld2q_v: 2923 case NEON::BI__builtin_neon_vld3_v: 2924 case NEON::BI__builtin_neon_vld3q_v: 2925 case NEON::BI__builtin_neon_vld4_v: 2926 case NEON::BI__builtin_neon_vld4q_v: { 2927 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2928 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2929 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2930 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2931 return Builder.CreateStore(Ops[1], Ops[0]); 2932 } 2933 case NEON::BI__builtin_neon_vld1_dup_v: 2934 case NEON::BI__builtin_neon_vld1q_dup_v: { 2935 Value *V = UndefValue::get(Ty); 2936 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2937 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2938 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2939 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2940 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2941 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2942 return EmitNeonSplat(Ops[0], CI); 2943 } 2944 case NEON::BI__builtin_neon_vld2_lane_v: 2945 case NEON::BI__builtin_neon_vld2q_lane_v: 2946 case NEON::BI__builtin_neon_vld3_lane_v: 2947 case NEON::BI__builtin_neon_vld3q_lane_v: 2948 case NEON::BI__builtin_neon_vld4_lane_v: 2949 case NEON::BI__builtin_neon_vld4q_lane_v: { 2950 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2951 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2952 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2953 Ops.push_back(Align); 2954 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2955 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2956 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2957 return Builder.CreateStore(Ops[1], Ops[0]); 2958 } 2959 case NEON::BI__builtin_neon_vmovl_v: { 2960 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2961 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2962 if (Usgn) 2963 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2964 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2965 } 2966 case NEON::BI__builtin_neon_vmovn_v: { 2967 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2968 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2969 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2970 } 2971 case NEON::BI__builtin_neon_vmull_v: 2972 // FIXME: the integer vmull operations could be emitted in terms of pure 2973 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2974 // hoisting the exts outside loops. Until global ISel comes along that can 2975 // see through such movement this leads to bad CodeGen. So we need an 2976 // intrinsic for now. 2977 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2978 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2979 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2980 case NEON::BI__builtin_neon_vpadal_v: 2981 case NEON::BI__builtin_neon_vpadalq_v: { 2982 // The source operand type has twice as many elements of half the size. 2983 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2984 llvm::Type *EltTy = 2985 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2986 llvm::Type *NarrowTy = 2987 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2988 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2989 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2990 } 2991 case NEON::BI__builtin_neon_vpaddl_v: 2992 case NEON::BI__builtin_neon_vpaddlq_v: { 2993 // The source operand type has twice as many elements of half the size. 2994 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2995 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2996 llvm::Type *NarrowTy = 2997 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2998 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2999 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3000 } 3001 case NEON::BI__builtin_neon_vqdmlal_v: 3002 case NEON::BI__builtin_neon_vqdmlsl_v: { 3003 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3004 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 3005 MulOps, "vqdmlal"); 3006 3007 SmallVector<Value *, 2> AccumOps; 3008 AccumOps.push_back(Ops[0]); 3009 AccumOps.push_back(Mul); 3010 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 3011 AccumOps, NameHint); 3012 } 3013 case NEON::BI__builtin_neon_vqshl_n_v: 3014 case NEON::BI__builtin_neon_vqshlq_n_v: 3015 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3016 1, false); 3017 case NEON::BI__builtin_neon_vqshlu_n_v: 3018 case NEON::BI__builtin_neon_vqshluq_n_v: 3019 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3020 1, false); 3021 case NEON::BI__builtin_neon_vrecpe_v: 3022 case NEON::BI__builtin_neon_vrecpeq_v: 3023 case NEON::BI__builtin_neon_vrsqrte_v: 3024 case NEON::BI__builtin_neon_vrsqrteq_v: 3025 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3026 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3027 3028 case NEON::BI__builtin_neon_vrshr_n_v: 3029 case NEON::BI__builtin_neon_vrshrq_n_v: 3030 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3031 1, true); 3032 case NEON::BI__builtin_neon_vshl_n_v: 3033 case NEON::BI__builtin_neon_vshlq_n_v: 3034 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3035 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3036 "vshl_n"); 3037 case NEON::BI__builtin_neon_vshll_n_v: { 3038 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3039 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3040 if (Usgn) 3041 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3042 else 3043 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3044 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3045 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3046 } 3047 case NEON::BI__builtin_neon_vshrn_n_v: { 3048 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3049 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3050 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3051 if (Usgn) 3052 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3053 else 3054 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3055 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3056 } 3057 case NEON::BI__builtin_neon_vshr_n_v: 3058 case NEON::BI__builtin_neon_vshrq_n_v: 3059 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3060 case NEON::BI__builtin_neon_vst1_v: 3061 case NEON::BI__builtin_neon_vst1q_v: 3062 case NEON::BI__builtin_neon_vst2_v: 3063 case NEON::BI__builtin_neon_vst2q_v: 3064 case NEON::BI__builtin_neon_vst3_v: 3065 case NEON::BI__builtin_neon_vst3q_v: 3066 case NEON::BI__builtin_neon_vst4_v: 3067 case NEON::BI__builtin_neon_vst4q_v: 3068 case NEON::BI__builtin_neon_vst2_lane_v: 3069 case NEON::BI__builtin_neon_vst2q_lane_v: 3070 case NEON::BI__builtin_neon_vst3_lane_v: 3071 case NEON::BI__builtin_neon_vst3q_lane_v: 3072 case NEON::BI__builtin_neon_vst4_lane_v: 3073 case NEON::BI__builtin_neon_vst4q_lane_v: 3074 Ops.push_back(Align); 3075 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3076 case NEON::BI__builtin_neon_vsubhn_v: { 3077 llvm::VectorType *SrcTy = 3078 llvm::VectorType::getExtendedElementVectorType(VTy); 3079 3080 // %sum = add <4 x i32> %lhs, %rhs 3081 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3082 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3083 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3084 3085 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3086 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 3087 SrcTy->getScalarSizeInBits() / 2); 3088 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 3089 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3090 3091 // %res = trunc <4 x i32> %high to <4 x i16> 3092 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3093 } 3094 case NEON::BI__builtin_neon_vtrn_v: 3095 case NEON::BI__builtin_neon_vtrnq_v: { 3096 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3097 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3098 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3099 Value *SV = nullptr; 3100 3101 for (unsigned vi = 0; vi != 2; ++vi) { 3102 SmallVector<Constant*, 16> Indices; 3103 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3104 Indices.push_back(Builder.getInt32(i+vi)); 3105 Indices.push_back(Builder.getInt32(i+e+vi)); 3106 } 3107 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3108 SV = llvm::ConstantVector::get(Indices); 3109 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3110 SV = Builder.CreateStore(SV, Addr); 3111 } 3112 return SV; 3113 } 3114 case NEON::BI__builtin_neon_vtst_v: 3115 case NEON::BI__builtin_neon_vtstq_v: { 3116 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3117 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3118 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3119 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3120 ConstantAggregateZero::get(Ty)); 3121 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3122 } 3123 case NEON::BI__builtin_neon_vuzp_v: 3124 case NEON::BI__builtin_neon_vuzpq_v: { 3125 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3126 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3127 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3128 Value *SV = nullptr; 3129 3130 for (unsigned vi = 0; vi != 2; ++vi) { 3131 SmallVector<Constant*, 16> Indices; 3132 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3133 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3134 3135 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3136 SV = llvm::ConstantVector::get(Indices); 3137 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3138 SV = Builder.CreateStore(SV, Addr); 3139 } 3140 return SV; 3141 } 3142 case NEON::BI__builtin_neon_vzip_v: 3143 case NEON::BI__builtin_neon_vzipq_v: { 3144 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3145 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3146 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3147 Value *SV = nullptr; 3148 3149 for (unsigned vi = 0; vi != 2; ++vi) { 3150 SmallVector<Constant*, 16> Indices; 3151 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3152 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3153 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3154 } 3155 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3156 SV = llvm::ConstantVector::get(Indices); 3157 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3158 SV = Builder.CreateStore(SV, Addr); 3159 } 3160 return SV; 3161 } 3162 } 3163 3164 assert(Int && "Expected valid intrinsic number"); 3165 3166 // Determine the type(s) of this overloaded AArch64 intrinsic. 3167 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3168 3169 Value *Result = EmitNeonCall(F, Ops, NameHint); 3170 llvm::Type *ResultType = ConvertType(E->getType()); 3171 // AArch64 intrinsic one-element vector type cast to 3172 // scalar type expected by the builtin 3173 return Builder.CreateBitCast(Result, ResultType, NameHint); 3174 } 3175 3176 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3177 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3178 const CmpInst::Predicate Ip, const Twine &Name) { 3179 llvm::Type *OTy = Op->getType(); 3180 3181 // FIXME: this is utterly horrific. We should not be looking at previous 3182 // codegen context to find out what needs doing. Unfortunately TableGen 3183 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3184 // (etc). 3185 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3186 OTy = BI->getOperand(0)->getType(); 3187 3188 Op = Builder.CreateBitCast(Op, OTy); 3189 if (OTy->getScalarType()->isFloatingPointTy()) { 3190 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3191 } else { 3192 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3193 } 3194 return Builder.CreateSExt(Op, Ty, Name); 3195 } 3196 3197 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3198 Value *ExtOp, Value *IndexOp, 3199 llvm::Type *ResTy, unsigned IntID, 3200 const char *Name) { 3201 SmallVector<Value *, 2> TblOps; 3202 if (ExtOp) 3203 TblOps.push_back(ExtOp); 3204 3205 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3206 SmallVector<Constant*, 16> Indices; 3207 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3208 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3209 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3210 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3211 } 3212 Value *SV = llvm::ConstantVector::get(Indices); 3213 3214 int PairPos = 0, End = Ops.size() - 1; 3215 while (PairPos < End) { 3216 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3217 Ops[PairPos+1], SV, Name)); 3218 PairPos += 2; 3219 } 3220 3221 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3222 // of the 128-bit lookup table with zero. 3223 if (PairPos == End) { 3224 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3225 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3226 ZeroTbl, SV, Name)); 3227 } 3228 3229 Function *TblF; 3230 TblOps.push_back(IndexOp); 3231 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3232 3233 return CGF.EmitNeonCall(TblF, TblOps, Name); 3234 } 3235 3236 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3237 switch (BuiltinID) { 3238 default: 3239 return nullptr; 3240 case ARM::BI__builtin_arm_nop: 3241 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3242 llvm::ConstantInt::get(Int32Ty, 0)); 3243 case ARM::BI__builtin_arm_yield: 3244 case ARM::BI__yield: 3245 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3246 llvm::ConstantInt::get(Int32Ty, 1)); 3247 case ARM::BI__builtin_arm_wfe: 3248 case ARM::BI__wfe: 3249 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3250 llvm::ConstantInt::get(Int32Ty, 2)); 3251 case ARM::BI__builtin_arm_wfi: 3252 case ARM::BI__wfi: 3253 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3254 llvm::ConstantInt::get(Int32Ty, 3)); 3255 case ARM::BI__builtin_arm_sev: 3256 case ARM::BI__sev: 3257 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3258 llvm::ConstantInt::get(Int32Ty, 4)); 3259 case ARM::BI__builtin_arm_sevl: 3260 case ARM::BI__sevl: 3261 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3262 llvm::ConstantInt::get(Int32Ty, 5)); 3263 } 3264 } 3265 3266 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3267 const CallExpr *E) { 3268 if (auto Hint = GetValueForARMHint(BuiltinID)) 3269 return Hint; 3270 3271 if (BuiltinID == ARM::BI__emit) { 3272 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3273 llvm::FunctionType *FTy = 3274 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3275 3276 APSInt Value; 3277 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3278 llvm_unreachable("Sema will ensure that the parameter is constant"); 3279 3280 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3281 3282 llvm::InlineAsm *Emit = 3283 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3284 /*SideEffects=*/true) 3285 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3286 /*SideEffects=*/true); 3287 3288 return Builder.CreateCall(Emit); 3289 } 3290 3291 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3292 Value *Option = EmitScalarExpr(E->getArg(0)); 3293 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3294 } 3295 3296 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3297 Value *Address = EmitScalarExpr(E->getArg(0)); 3298 Value *RW = EmitScalarExpr(E->getArg(1)); 3299 Value *IsData = EmitScalarExpr(E->getArg(2)); 3300 3301 // Locality is not supported on ARM target 3302 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3303 3304 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3305 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3306 } 3307 3308 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3309 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3310 EmitScalarExpr(E->getArg(0)), 3311 "rbit"); 3312 } 3313 3314 if (BuiltinID == ARM::BI__clear_cache) { 3315 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3316 const FunctionDecl *FD = E->getDirectCallee(); 3317 SmallVector<Value*, 2> Ops; 3318 for (unsigned i = 0; i < 2; i++) 3319 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3320 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3321 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3322 StringRef Name = FD->getName(); 3323 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3324 } 3325 3326 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3327 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3328 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3329 getContext().getTypeSize(E->getType()) == 64) || 3330 BuiltinID == ARM::BI__ldrexd) { 3331 Function *F; 3332 3333 switch (BuiltinID) { 3334 default: llvm_unreachable("unexpected builtin"); 3335 case ARM::BI__builtin_arm_ldaex: 3336 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3337 break; 3338 case ARM::BI__builtin_arm_ldrexd: 3339 case ARM::BI__builtin_arm_ldrex: 3340 case ARM::BI__ldrexd: 3341 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3342 break; 3343 } 3344 3345 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3346 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3347 "ldrexd"); 3348 3349 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3350 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3351 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3352 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3353 3354 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3355 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3356 Val = Builder.CreateOr(Val, Val1); 3357 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3358 } 3359 3360 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3361 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3362 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3363 3364 QualType Ty = E->getType(); 3365 llvm::Type *RealResTy = ConvertType(Ty); 3366 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3367 getContext().getTypeSize(Ty)); 3368 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3369 3370 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3371 ? Intrinsic::arm_ldaex 3372 : Intrinsic::arm_ldrex, 3373 LoadAddr->getType()); 3374 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3375 3376 if (RealResTy->isPointerTy()) 3377 return Builder.CreateIntToPtr(Val, RealResTy); 3378 else { 3379 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3380 return Builder.CreateBitCast(Val, RealResTy); 3381 } 3382 } 3383 3384 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3385 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3386 BuiltinID == ARM::BI__builtin_arm_strex) && 3387 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3388 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3389 ? Intrinsic::arm_stlexd 3390 : Intrinsic::arm_strexd); 3391 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3392 3393 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3394 Value *Val = EmitScalarExpr(E->getArg(0)); 3395 Builder.CreateStore(Val, Tmp); 3396 3397 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3398 Val = Builder.CreateLoad(LdPtr); 3399 3400 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3401 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3402 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3403 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3404 } 3405 3406 if (BuiltinID == ARM::BI__builtin_arm_strex || 3407 BuiltinID == ARM::BI__builtin_arm_stlex) { 3408 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3409 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3410 3411 QualType Ty = E->getArg(0)->getType(); 3412 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3413 getContext().getTypeSize(Ty)); 3414 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3415 3416 if (StoreVal->getType()->isPointerTy()) 3417 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3418 else { 3419 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3420 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3421 } 3422 3423 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3424 ? Intrinsic::arm_stlex 3425 : Intrinsic::arm_strex, 3426 StoreAddr->getType()); 3427 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3428 } 3429 3430 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3431 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3432 return Builder.CreateCall(F); 3433 } 3434 3435 // CRC32 3436 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3437 switch (BuiltinID) { 3438 case ARM::BI__builtin_arm_crc32b: 3439 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3440 case ARM::BI__builtin_arm_crc32cb: 3441 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3442 case ARM::BI__builtin_arm_crc32h: 3443 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3444 case ARM::BI__builtin_arm_crc32ch: 3445 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3446 case ARM::BI__builtin_arm_crc32w: 3447 case ARM::BI__builtin_arm_crc32d: 3448 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3449 case ARM::BI__builtin_arm_crc32cw: 3450 case ARM::BI__builtin_arm_crc32cd: 3451 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3452 } 3453 3454 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3455 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3456 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3457 3458 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3459 // intrinsics, hence we need different codegen for these cases. 3460 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3461 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3462 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3463 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3464 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3465 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3466 3467 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3468 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3469 return Builder.CreateCall2(F, Res, Arg1b); 3470 } else { 3471 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3472 3473 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3474 return Builder.CreateCall2(F, Arg0, Arg1); 3475 } 3476 } 3477 3478 SmallVector<Value*, 4> Ops; 3479 llvm::Value *Align = nullptr; 3480 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3481 if (i == 0) { 3482 switch (BuiltinID) { 3483 case NEON::BI__builtin_neon_vld1_v: 3484 case NEON::BI__builtin_neon_vld1q_v: 3485 case NEON::BI__builtin_neon_vld1q_lane_v: 3486 case NEON::BI__builtin_neon_vld1_lane_v: 3487 case NEON::BI__builtin_neon_vld1_dup_v: 3488 case NEON::BI__builtin_neon_vld1q_dup_v: 3489 case NEON::BI__builtin_neon_vst1_v: 3490 case NEON::BI__builtin_neon_vst1q_v: 3491 case NEON::BI__builtin_neon_vst1q_lane_v: 3492 case NEON::BI__builtin_neon_vst1_lane_v: 3493 case NEON::BI__builtin_neon_vst2_v: 3494 case NEON::BI__builtin_neon_vst2q_v: 3495 case NEON::BI__builtin_neon_vst2_lane_v: 3496 case NEON::BI__builtin_neon_vst2q_lane_v: 3497 case NEON::BI__builtin_neon_vst3_v: 3498 case NEON::BI__builtin_neon_vst3q_v: 3499 case NEON::BI__builtin_neon_vst3_lane_v: 3500 case NEON::BI__builtin_neon_vst3q_lane_v: 3501 case NEON::BI__builtin_neon_vst4_v: 3502 case NEON::BI__builtin_neon_vst4q_v: 3503 case NEON::BI__builtin_neon_vst4_lane_v: 3504 case NEON::BI__builtin_neon_vst4q_lane_v: 3505 // Get the alignment for the argument in addition to the value; 3506 // we'll use it later. 3507 std::pair<llvm::Value*, unsigned> Src = 3508 EmitPointerWithAlignment(E->getArg(0)); 3509 Ops.push_back(Src.first); 3510 Align = Builder.getInt32(Src.second); 3511 continue; 3512 } 3513 } 3514 if (i == 1) { 3515 switch (BuiltinID) { 3516 case NEON::BI__builtin_neon_vld2_v: 3517 case NEON::BI__builtin_neon_vld2q_v: 3518 case NEON::BI__builtin_neon_vld3_v: 3519 case NEON::BI__builtin_neon_vld3q_v: 3520 case NEON::BI__builtin_neon_vld4_v: 3521 case NEON::BI__builtin_neon_vld4q_v: 3522 case NEON::BI__builtin_neon_vld2_lane_v: 3523 case NEON::BI__builtin_neon_vld2q_lane_v: 3524 case NEON::BI__builtin_neon_vld3_lane_v: 3525 case NEON::BI__builtin_neon_vld3q_lane_v: 3526 case NEON::BI__builtin_neon_vld4_lane_v: 3527 case NEON::BI__builtin_neon_vld4q_lane_v: 3528 case NEON::BI__builtin_neon_vld2_dup_v: 3529 case NEON::BI__builtin_neon_vld3_dup_v: 3530 case NEON::BI__builtin_neon_vld4_dup_v: 3531 // Get the alignment for the argument in addition to the value; 3532 // we'll use it later. 3533 std::pair<llvm::Value*, unsigned> Src = 3534 EmitPointerWithAlignment(E->getArg(1)); 3535 Ops.push_back(Src.first); 3536 Align = Builder.getInt32(Src.second); 3537 continue; 3538 } 3539 } 3540 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3541 } 3542 3543 switch (BuiltinID) { 3544 default: break; 3545 // vget_lane and vset_lane are not overloaded and do not have an extra 3546 // argument that specifies the vector type. 3547 case NEON::BI__builtin_neon_vget_lane_i8: 3548 case NEON::BI__builtin_neon_vget_lane_i16: 3549 case NEON::BI__builtin_neon_vget_lane_i32: 3550 case NEON::BI__builtin_neon_vget_lane_i64: 3551 case NEON::BI__builtin_neon_vget_lane_f32: 3552 case NEON::BI__builtin_neon_vgetq_lane_i8: 3553 case NEON::BI__builtin_neon_vgetq_lane_i16: 3554 case NEON::BI__builtin_neon_vgetq_lane_i32: 3555 case NEON::BI__builtin_neon_vgetq_lane_i64: 3556 case NEON::BI__builtin_neon_vgetq_lane_f32: 3557 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3558 "vget_lane"); 3559 case NEON::BI__builtin_neon_vset_lane_i8: 3560 case NEON::BI__builtin_neon_vset_lane_i16: 3561 case NEON::BI__builtin_neon_vset_lane_i32: 3562 case NEON::BI__builtin_neon_vset_lane_i64: 3563 case NEON::BI__builtin_neon_vset_lane_f32: 3564 case NEON::BI__builtin_neon_vsetq_lane_i8: 3565 case NEON::BI__builtin_neon_vsetq_lane_i16: 3566 case NEON::BI__builtin_neon_vsetq_lane_i32: 3567 case NEON::BI__builtin_neon_vsetq_lane_i64: 3568 case NEON::BI__builtin_neon_vsetq_lane_f32: 3569 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3570 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3571 3572 // Non-polymorphic crypto instructions also not overloaded 3573 case NEON::BI__builtin_neon_vsha1h_u32: 3574 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3575 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3576 "vsha1h"); 3577 case NEON::BI__builtin_neon_vsha1cq_u32: 3578 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3579 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3580 "vsha1h"); 3581 case NEON::BI__builtin_neon_vsha1pq_u32: 3582 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3583 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3584 "vsha1h"); 3585 case NEON::BI__builtin_neon_vsha1mq_u32: 3586 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3587 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3588 "vsha1h"); 3589 } 3590 3591 // Get the last argument, which specifies the vector type. 3592 llvm::APSInt Result; 3593 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3594 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3595 return nullptr; 3596 3597 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3598 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3599 // Determine the overloaded type of this builtin. 3600 llvm::Type *Ty; 3601 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3602 Ty = FloatTy; 3603 else 3604 Ty = DoubleTy; 3605 3606 // Determine whether this is an unsigned conversion or not. 3607 bool usgn = Result.getZExtValue() == 1; 3608 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3609 3610 // Call the appropriate intrinsic. 3611 Function *F = CGM.getIntrinsic(Int, Ty); 3612 return Builder.CreateCall(F, Ops, "vcvtr"); 3613 } 3614 3615 // Determine the type of this overloaded NEON intrinsic. 3616 NeonTypeFlags Type(Result.getZExtValue()); 3617 bool usgn = Type.isUnsigned(); 3618 bool rightShift = false; 3619 3620 llvm::VectorType *VTy = GetNeonType(this, Type); 3621 llvm::Type *Ty = VTy; 3622 if (!Ty) 3623 return nullptr; 3624 3625 // Many NEON builtins have identical semantics and uses in ARM and 3626 // AArch64. Emit these in a single function. 3627 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3628 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3629 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3630 if (Builtin) 3631 return EmitCommonNeonBuiltinExpr( 3632 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3633 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3634 3635 unsigned Int; 3636 switch (BuiltinID) { 3637 default: return nullptr; 3638 case NEON::BI__builtin_neon_vld1q_lane_v: 3639 // Handle 64-bit integer elements as a special case. Use shuffles of 3640 // one-element vectors to avoid poor code for i64 in the backend. 3641 if (VTy->getElementType()->isIntegerTy(64)) { 3642 // Extract the other lane. 3643 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3644 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3645 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3646 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3647 // Load the value as a one-element vector. 3648 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3649 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3650 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3651 // Combine them. 3652 SmallVector<Constant*, 2> Indices; 3653 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3654 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3655 SV = llvm::ConstantVector::get(Indices); 3656 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3657 } 3658 // fall through 3659 case NEON::BI__builtin_neon_vld1_lane_v: { 3660 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3661 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3662 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3663 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3664 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3665 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3666 } 3667 case NEON::BI__builtin_neon_vld2_dup_v: 3668 case NEON::BI__builtin_neon_vld3_dup_v: 3669 case NEON::BI__builtin_neon_vld4_dup_v: { 3670 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3671 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3672 switch (BuiltinID) { 3673 case NEON::BI__builtin_neon_vld2_dup_v: 3674 Int = Intrinsic::arm_neon_vld2; 3675 break; 3676 case NEON::BI__builtin_neon_vld3_dup_v: 3677 Int = Intrinsic::arm_neon_vld3; 3678 break; 3679 case NEON::BI__builtin_neon_vld4_dup_v: 3680 Int = Intrinsic::arm_neon_vld4; 3681 break; 3682 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3683 } 3684 Function *F = CGM.getIntrinsic(Int, Ty); 3685 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3686 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3687 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3688 return Builder.CreateStore(Ops[1], Ops[0]); 3689 } 3690 switch (BuiltinID) { 3691 case NEON::BI__builtin_neon_vld2_dup_v: 3692 Int = Intrinsic::arm_neon_vld2lane; 3693 break; 3694 case NEON::BI__builtin_neon_vld3_dup_v: 3695 Int = Intrinsic::arm_neon_vld3lane; 3696 break; 3697 case NEON::BI__builtin_neon_vld4_dup_v: 3698 Int = Intrinsic::arm_neon_vld4lane; 3699 break; 3700 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3701 } 3702 Function *F = CGM.getIntrinsic(Int, Ty); 3703 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3704 3705 SmallVector<Value*, 6> Args; 3706 Args.push_back(Ops[1]); 3707 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3708 3709 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3710 Args.push_back(CI); 3711 Args.push_back(Align); 3712 3713 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3714 // splat lane 0 to all elts in each vector of the result. 3715 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3716 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3717 Value *Elt = Builder.CreateBitCast(Val, Ty); 3718 Elt = EmitNeonSplat(Elt, CI); 3719 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3720 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3721 } 3722 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3723 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3724 return Builder.CreateStore(Ops[1], Ops[0]); 3725 } 3726 case NEON::BI__builtin_neon_vqrshrn_n_v: 3727 Int = 3728 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3729 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3730 1, true); 3731 case NEON::BI__builtin_neon_vqrshrun_n_v: 3732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3733 Ops, "vqrshrun_n", 1, true); 3734 case NEON::BI__builtin_neon_vqshrn_n_v: 3735 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3736 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3737 1, true); 3738 case NEON::BI__builtin_neon_vqshrun_n_v: 3739 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3740 Ops, "vqshrun_n", 1, true); 3741 case NEON::BI__builtin_neon_vrecpe_v: 3742 case NEON::BI__builtin_neon_vrecpeq_v: 3743 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3744 Ops, "vrecpe"); 3745 case NEON::BI__builtin_neon_vrshrn_n_v: 3746 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3747 Ops, "vrshrn_n", 1, true); 3748 case NEON::BI__builtin_neon_vrsra_n_v: 3749 case NEON::BI__builtin_neon_vrsraq_n_v: 3750 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3751 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3752 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3753 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3754 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3755 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3756 case NEON::BI__builtin_neon_vsri_n_v: 3757 case NEON::BI__builtin_neon_vsriq_n_v: 3758 rightShift = true; 3759 case NEON::BI__builtin_neon_vsli_n_v: 3760 case NEON::BI__builtin_neon_vsliq_n_v: 3761 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3762 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3763 Ops, "vsli_n"); 3764 case NEON::BI__builtin_neon_vsra_n_v: 3765 case NEON::BI__builtin_neon_vsraq_n_v: 3766 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3767 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3768 return Builder.CreateAdd(Ops[0], Ops[1]); 3769 case NEON::BI__builtin_neon_vst1q_lane_v: 3770 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3771 // a one-element vector and avoid poor code for i64 in the backend. 3772 if (VTy->getElementType()->isIntegerTy(64)) { 3773 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3774 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3775 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3776 Ops[2] = Align; 3777 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3778 Ops[1]->getType()), Ops); 3779 } 3780 // fall through 3781 case NEON::BI__builtin_neon_vst1_lane_v: { 3782 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3783 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3784 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3785 StoreInst *St = Builder.CreateStore(Ops[1], 3786 Builder.CreateBitCast(Ops[0], Ty)); 3787 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3788 return St; 3789 } 3790 case NEON::BI__builtin_neon_vtbl1_v: 3791 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3792 Ops, "vtbl1"); 3793 case NEON::BI__builtin_neon_vtbl2_v: 3794 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3795 Ops, "vtbl2"); 3796 case NEON::BI__builtin_neon_vtbl3_v: 3797 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3798 Ops, "vtbl3"); 3799 case NEON::BI__builtin_neon_vtbl4_v: 3800 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3801 Ops, "vtbl4"); 3802 case NEON::BI__builtin_neon_vtbx1_v: 3803 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3804 Ops, "vtbx1"); 3805 case NEON::BI__builtin_neon_vtbx2_v: 3806 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3807 Ops, "vtbx2"); 3808 case NEON::BI__builtin_neon_vtbx3_v: 3809 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3810 Ops, "vtbx3"); 3811 case NEON::BI__builtin_neon_vtbx4_v: 3812 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3813 Ops, "vtbx4"); 3814 } 3815 } 3816 3817 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3818 const CallExpr *E, 3819 SmallVectorImpl<Value *> &Ops) { 3820 unsigned int Int = 0; 3821 const char *s = nullptr; 3822 3823 switch (BuiltinID) { 3824 default: 3825 return nullptr; 3826 case NEON::BI__builtin_neon_vtbl1_v: 3827 case NEON::BI__builtin_neon_vqtbl1_v: 3828 case NEON::BI__builtin_neon_vqtbl1q_v: 3829 case NEON::BI__builtin_neon_vtbl2_v: 3830 case NEON::BI__builtin_neon_vqtbl2_v: 3831 case NEON::BI__builtin_neon_vqtbl2q_v: 3832 case NEON::BI__builtin_neon_vtbl3_v: 3833 case NEON::BI__builtin_neon_vqtbl3_v: 3834 case NEON::BI__builtin_neon_vqtbl3q_v: 3835 case NEON::BI__builtin_neon_vtbl4_v: 3836 case NEON::BI__builtin_neon_vqtbl4_v: 3837 case NEON::BI__builtin_neon_vqtbl4q_v: 3838 break; 3839 case NEON::BI__builtin_neon_vtbx1_v: 3840 case NEON::BI__builtin_neon_vqtbx1_v: 3841 case NEON::BI__builtin_neon_vqtbx1q_v: 3842 case NEON::BI__builtin_neon_vtbx2_v: 3843 case NEON::BI__builtin_neon_vqtbx2_v: 3844 case NEON::BI__builtin_neon_vqtbx2q_v: 3845 case NEON::BI__builtin_neon_vtbx3_v: 3846 case NEON::BI__builtin_neon_vqtbx3_v: 3847 case NEON::BI__builtin_neon_vqtbx3q_v: 3848 case NEON::BI__builtin_neon_vtbx4_v: 3849 case NEON::BI__builtin_neon_vqtbx4_v: 3850 case NEON::BI__builtin_neon_vqtbx4q_v: 3851 break; 3852 } 3853 3854 assert(E->getNumArgs() >= 3); 3855 3856 // Get the last argument, which specifies the vector type. 3857 llvm::APSInt Result; 3858 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3859 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3860 return nullptr; 3861 3862 // Determine the type of this overloaded NEON intrinsic. 3863 NeonTypeFlags Type(Result.getZExtValue()); 3864 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3865 llvm::Type *Ty = VTy; 3866 if (!Ty) 3867 return nullptr; 3868 3869 unsigned nElts = VTy->getNumElements(); 3870 3871 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3872 3873 // AArch64 scalar builtins are not overloaded, they do not have an extra 3874 // argument that specifies the vector type, need to handle each case. 3875 SmallVector<Value *, 2> TblOps; 3876 switch (BuiltinID) { 3877 case NEON::BI__builtin_neon_vtbl1_v: { 3878 TblOps.push_back(Ops[0]); 3879 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3880 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3881 } 3882 case NEON::BI__builtin_neon_vtbl2_v: { 3883 TblOps.push_back(Ops[0]); 3884 TblOps.push_back(Ops[1]); 3885 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3886 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3887 } 3888 case NEON::BI__builtin_neon_vtbl3_v: { 3889 TblOps.push_back(Ops[0]); 3890 TblOps.push_back(Ops[1]); 3891 TblOps.push_back(Ops[2]); 3892 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3893 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3894 } 3895 case NEON::BI__builtin_neon_vtbl4_v: { 3896 TblOps.push_back(Ops[0]); 3897 TblOps.push_back(Ops[1]); 3898 TblOps.push_back(Ops[2]); 3899 TblOps.push_back(Ops[3]); 3900 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3901 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3902 } 3903 case NEON::BI__builtin_neon_vtbx1_v: { 3904 TblOps.push_back(Ops[1]); 3905 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3906 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3907 3908 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3909 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3910 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3911 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3912 3913 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3914 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3915 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3916 } 3917 case NEON::BI__builtin_neon_vtbx2_v: { 3918 TblOps.push_back(Ops[1]); 3919 TblOps.push_back(Ops[2]); 3920 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3921 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3922 } 3923 case NEON::BI__builtin_neon_vtbx3_v: { 3924 TblOps.push_back(Ops[1]); 3925 TblOps.push_back(Ops[2]); 3926 TblOps.push_back(Ops[3]); 3927 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3928 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3929 3930 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3931 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3932 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3933 TwentyFourV); 3934 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3935 3936 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3937 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3938 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3939 } 3940 case NEON::BI__builtin_neon_vtbx4_v: { 3941 TblOps.push_back(Ops[1]); 3942 TblOps.push_back(Ops[2]); 3943 TblOps.push_back(Ops[3]); 3944 TblOps.push_back(Ops[4]); 3945 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3946 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3947 } 3948 case NEON::BI__builtin_neon_vqtbl1_v: 3949 case NEON::BI__builtin_neon_vqtbl1q_v: 3950 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3951 case NEON::BI__builtin_neon_vqtbl2_v: 3952 case NEON::BI__builtin_neon_vqtbl2q_v: { 3953 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3954 case NEON::BI__builtin_neon_vqtbl3_v: 3955 case NEON::BI__builtin_neon_vqtbl3q_v: 3956 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3957 case NEON::BI__builtin_neon_vqtbl4_v: 3958 case NEON::BI__builtin_neon_vqtbl4q_v: 3959 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3960 case NEON::BI__builtin_neon_vqtbx1_v: 3961 case NEON::BI__builtin_neon_vqtbx1q_v: 3962 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3963 case NEON::BI__builtin_neon_vqtbx2_v: 3964 case NEON::BI__builtin_neon_vqtbx2q_v: 3965 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3966 case NEON::BI__builtin_neon_vqtbx3_v: 3967 case NEON::BI__builtin_neon_vqtbx3q_v: 3968 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3969 case NEON::BI__builtin_neon_vqtbx4_v: 3970 case NEON::BI__builtin_neon_vqtbx4q_v: 3971 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3972 } 3973 } 3974 3975 if (!Int) 3976 return nullptr; 3977 3978 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3979 return CGF.EmitNeonCall(F, Ops, s); 3980 } 3981 3982 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3983 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3984 Op = Builder.CreateBitCast(Op, Int16Ty); 3985 Value *V = UndefValue::get(VTy); 3986 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3987 Op = Builder.CreateInsertElement(V, Op, CI); 3988 return Op; 3989 } 3990 3991 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3992 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3993 Op = Builder.CreateBitCast(Op, Int8Ty); 3994 Value *V = UndefValue::get(VTy); 3995 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3996 Op = Builder.CreateInsertElement(V, Op, CI); 3997 return Op; 3998 } 3999 4000 Value *CodeGenFunction:: 4001 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 4002 const char *Name) { 4003 // i8 is not a legal types for AArch64, so we can't just use 4004 // a normal overloaded intrinsic call for these scalar types. Instead 4005 // we'll build 64-bit vectors w/ lane zero being our input values and 4006 // perform the operation on that. The back end can pattern match directly 4007 // to the scalar instruction. 4008 Ops[0] = vectorWrapScalar8(Ops[0]); 4009 Ops[1] = vectorWrapScalar8(Ops[1]); 4010 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4011 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 4012 Constant *CI = ConstantInt::get(SizeTy, 0); 4013 return Builder.CreateExtractElement(V, CI, "lane0"); 4014 } 4015 4016 Value *CodeGenFunction:: 4017 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 4018 const char *Name) { 4019 // i16 is not a legal types for AArch64, so we can't just use 4020 // a normal overloaded intrinsic call for these scalar types. Instead 4021 // we'll build 64-bit vectors w/ lane zero being our input values and 4022 // perform the operation on that. The back end can pattern match directly 4023 // to the scalar instruction. 4024 Ops[0] = vectorWrapScalar16(Ops[0]); 4025 Ops[1] = vectorWrapScalar16(Ops[1]); 4026 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4027 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 4028 Constant *CI = ConstantInt::get(SizeTy, 0); 4029 return Builder.CreateExtractElement(V, CI, "lane0"); 4030 } 4031 4032 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4033 const CallExpr *E) { 4034 unsigned HintID = static_cast<unsigned>(-1); 4035 switch (BuiltinID) { 4036 default: break; 4037 case AArch64::BI__builtin_arm_nop: 4038 HintID = 0; 4039 break; 4040 case AArch64::BI__builtin_arm_yield: 4041 HintID = 1; 4042 break; 4043 case AArch64::BI__builtin_arm_wfe: 4044 HintID = 2; 4045 break; 4046 case AArch64::BI__builtin_arm_wfi: 4047 HintID = 3; 4048 break; 4049 case AArch64::BI__builtin_arm_sev: 4050 HintID = 4; 4051 break; 4052 case AArch64::BI__builtin_arm_sevl: 4053 HintID = 5; 4054 break; 4055 } 4056 4057 if (HintID != static_cast<unsigned>(-1)) { 4058 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4059 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4060 } 4061 4062 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4063 Value *Address = EmitScalarExpr(E->getArg(0)); 4064 Value *RW = EmitScalarExpr(E->getArg(1)); 4065 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4066 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4067 Value *IsData = EmitScalarExpr(E->getArg(4)); 4068 4069 Value *Locality = nullptr; 4070 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4071 // Temporal fetch, needs to convert cache level to locality. 4072 Locality = llvm::ConstantInt::get(Int32Ty, 4073 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4074 } else { 4075 // Streaming fetch. 4076 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4077 } 4078 4079 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4080 // PLDL3STRM or PLDL2STRM. 4081 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4082 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 4083 } 4084 4085 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4086 assert((getContext().getTypeSize(E->getType()) == 32) && 4087 "rbit of unusual size!"); 4088 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4089 return Builder.CreateCall( 4090 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4091 } 4092 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4093 assert((getContext().getTypeSize(E->getType()) == 64) && 4094 "rbit of unusual size!"); 4095 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4096 return Builder.CreateCall( 4097 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4098 } 4099 4100 if (BuiltinID == AArch64::BI__clear_cache) { 4101 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4102 const FunctionDecl *FD = E->getDirectCallee(); 4103 SmallVector<Value*, 2> Ops; 4104 for (unsigned i = 0; i < 2; i++) 4105 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4106 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4107 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4108 StringRef Name = FD->getName(); 4109 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4110 } 4111 4112 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4113 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4114 getContext().getTypeSize(E->getType()) == 128) { 4115 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4116 ? Intrinsic::aarch64_ldaxp 4117 : Intrinsic::aarch64_ldxp); 4118 4119 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4120 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4121 "ldxp"); 4122 4123 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4124 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4125 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4126 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4127 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4128 4129 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4130 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4131 Val = Builder.CreateOr(Val, Val1); 4132 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4133 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4134 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4135 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4136 4137 QualType Ty = E->getType(); 4138 llvm::Type *RealResTy = ConvertType(Ty); 4139 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4140 getContext().getTypeSize(Ty)); 4141 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4142 4143 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4144 ? Intrinsic::aarch64_ldaxr 4145 : Intrinsic::aarch64_ldxr, 4146 LoadAddr->getType()); 4147 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4148 4149 if (RealResTy->isPointerTy()) 4150 return Builder.CreateIntToPtr(Val, RealResTy); 4151 4152 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4153 return Builder.CreateBitCast(Val, RealResTy); 4154 } 4155 4156 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4157 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4158 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4159 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4160 ? Intrinsic::aarch64_stlxp 4161 : Intrinsic::aarch64_stxp); 4162 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4163 4164 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4165 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4166 One); 4167 Value *Val = EmitScalarExpr(E->getArg(0)); 4168 Builder.CreateStore(Val, Tmp); 4169 4170 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4171 Val = Builder.CreateLoad(LdPtr); 4172 4173 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4174 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4175 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4176 Int8PtrTy); 4177 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 4178 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 4179 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4180 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4181 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4182 4183 QualType Ty = E->getArg(0)->getType(); 4184 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4185 getContext().getTypeSize(Ty)); 4186 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4187 4188 if (StoreVal->getType()->isPointerTy()) 4189 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4190 else { 4191 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4192 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4193 } 4194 4195 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4196 ? Intrinsic::aarch64_stlxr 4197 : Intrinsic::aarch64_stxr, 4198 StoreAddr->getType()); 4199 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4200 } 4201 4202 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4203 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4204 return Builder.CreateCall(F); 4205 } 4206 4207 // CRC32 4208 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4209 switch (BuiltinID) { 4210 case AArch64::BI__builtin_arm_crc32b: 4211 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4212 case AArch64::BI__builtin_arm_crc32cb: 4213 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4214 case AArch64::BI__builtin_arm_crc32h: 4215 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4216 case AArch64::BI__builtin_arm_crc32ch: 4217 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4218 case AArch64::BI__builtin_arm_crc32w: 4219 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4220 case AArch64::BI__builtin_arm_crc32cw: 4221 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4222 case AArch64::BI__builtin_arm_crc32d: 4223 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4224 case AArch64::BI__builtin_arm_crc32cd: 4225 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4226 } 4227 4228 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4229 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4230 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4231 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4232 4233 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4234 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4235 4236 return Builder.CreateCall2(F, Arg0, Arg1); 4237 } 4238 4239 llvm::SmallVector<Value*, 4> Ops; 4240 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4241 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4242 4243 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4244 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4245 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4246 4247 if (Builtin) { 4248 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4249 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4250 assert(Result && "SISD intrinsic should have been handled"); 4251 return Result; 4252 } 4253 4254 llvm::APSInt Result; 4255 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4256 NeonTypeFlags Type(0); 4257 if (Arg->isIntegerConstantExpr(Result, getContext())) 4258 // Determine the type of this overloaded NEON intrinsic. 4259 Type = NeonTypeFlags(Result.getZExtValue()); 4260 4261 bool usgn = Type.isUnsigned(); 4262 bool quad = Type.isQuad(); 4263 4264 // Handle non-overloaded intrinsics first. 4265 switch (BuiltinID) { 4266 default: break; 4267 case NEON::BI__builtin_neon_vldrq_p128: { 4268 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4269 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4270 return Builder.CreateLoad(Ptr); 4271 } 4272 case NEON::BI__builtin_neon_vstrq_p128: { 4273 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4274 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4275 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4276 } 4277 case NEON::BI__builtin_neon_vcvts_u32_f32: 4278 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4279 usgn = true; 4280 // FALL THROUGH 4281 case NEON::BI__builtin_neon_vcvts_s32_f32: 4282 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4283 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4284 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4285 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4286 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4287 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4288 if (usgn) 4289 return Builder.CreateFPToUI(Ops[0], InTy); 4290 return Builder.CreateFPToSI(Ops[0], InTy); 4291 } 4292 case NEON::BI__builtin_neon_vcvts_f32_u32: 4293 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4294 usgn = true; 4295 // FALL THROUGH 4296 case NEON::BI__builtin_neon_vcvts_f32_s32: 4297 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4298 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4299 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4300 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4301 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4302 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4303 if (usgn) 4304 return Builder.CreateUIToFP(Ops[0], FTy); 4305 return Builder.CreateSIToFP(Ops[0], FTy); 4306 } 4307 case NEON::BI__builtin_neon_vpaddd_s64: { 4308 llvm::Type *Ty = 4309 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4310 Value *Vec = EmitScalarExpr(E->getArg(0)); 4311 // The vector is v2f64, so make sure it's bitcast to that. 4312 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4313 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4314 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4315 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4316 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4317 // Pairwise addition of a v2f64 into a scalar f64. 4318 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4319 } 4320 case NEON::BI__builtin_neon_vpaddd_f64: { 4321 llvm::Type *Ty = 4322 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4323 Value *Vec = EmitScalarExpr(E->getArg(0)); 4324 // The vector is v2f64, so make sure it's bitcast to that. 4325 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4326 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4327 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4328 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4329 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4330 // Pairwise addition of a v2f64 into a scalar f64. 4331 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4332 } 4333 case NEON::BI__builtin_neon_vpadds_f32: { 4334 llvm::Type *Ty = 4335 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4336 Value *Vec = EmitScalarExpr(E->getArg(0)); 4337 // The vector is v2f32, so make sure it's bitcast to that. 4338 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4339 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4340 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4341 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4342 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4343 // Pairwise addition of a v2f32 into a scalar f32. 4344 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4345 } 4346 case NEON::BI__builtin_neon_vceqzd_s64: 4347 case NEON::BI__builtin_neon_vceqzd_f64: 4348 case NEON::BI__builtin_neon_vceqzs_f32: 4349 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4350 return EmitAArch64CompareBuiltinExpr( 4351 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4352 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4353 case NEON::BI__builtin_neon_vcgezd_s64: 4354 case NEON::BI__builtin_neon_vcgezd_f64: 4355 case NEON::BI__builtin_neon_vcgezs_f32: 4356 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4357 return EmitAArch64CompareBuiltinExpr( 4358 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4359 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4360 case NEON::BI__builtin_neon_vclezd_s64: 4361 case NEON::BI__builtin_neon_vclezd_f64: 4362 case NEON::BI__builtin_neon_vclezs_f32: 4363 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4364 return EmitAArch64CompareBuiltinExpr( 4365 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4366 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4367 case NEON::BI__builtin_neon_vcgtzd_s64: 4368 case NEON::BI__builtin_neon_vcgtzd_f64: 4369 case NEON::BI__builtin_neon_vcgtzs_f32: 4370 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4371 return EmitAArch64CompareBuiltinExpr( 4372 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4373 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4374 case NEON::BI__builtin_neon_vcltzd_s64: 4375 case NEON::BI__builtin_neon_vcltzd_f64: 4376 case NEON::BI__builtin_neon_vcltzs_f32: 4377 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4378 return EmitAArch64CompareBuiltinExpr( 4379 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4380 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4381 4382 case NEON::BI__builtin_neon_vceqzd_u64: { 4383 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4384 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4385 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4386 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4387 llvm::Constant::getNullValue(Ty)); 4388 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4389 } 4390 case NEON::BI__builtin_neon_vceqd_f64: 4391 case NEON::BI__builtin_neon_vcled_f64: 4392 case NEON::BI__builtin_neon_vcltd_f64: 4393 case NEON::BI__builtin_neon_vcged_f64: 4394 case NEON::BI__builtin_neon_vcgtd_f64: { 4395 llvm::CmpInst::Predicate P; 4396 switch (BuiltinID) { 4397 default: llvm_unreachable("missing builtin ID in switch!"); 4398 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4399 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4400 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4401 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4402 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4403 } 4404 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4405 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4406 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4407 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4408 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4409 } 4410 case NEON::BI__builtin_neon_vceqs_f32: 4411 case NEON::BI__builtin_neon_vcles_f32: 4412 case NEON::BI__builtin_neon_vclts_f32: 4413 case NEON::BI__builtin_neon_vcges_f32: 4414 case NEON::BI__builtin_neon_vcgts_f32: { 4415 llvm::CmpInst::Predicate P; 4416 switch (BuiltinID) { 4417 default: llvm_unreachable("missing builtin ID in switch!"); 4418 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4419 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4420 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4421 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4422 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4423 } 4424 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4425 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4426 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4427 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4428 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4429 } 4430 case NEON::BI__builtin_neon_vceqd_s64: 4431 case NEON::BI__builtin_neon_vceqd_u64: 4432 case NEON::BI__builtin_neon_vcgtd_s64: 4433 case NEON::BI__builtin_neon_vcgtd_u64: 4434 case NEON::BI__builtin_neon_vcltd_s64: 4435 case NEON::BI__builtin_neon_vcltd_u64: 4436 case NEON::BI__builtin_neon_vcged_u64: 4437 case NEON::BI__builtin_neon_vcged_s64: 4438 case NEON::BI__builtin_neon_vcled_u64: 4439 case NEON::BI__builtin_neon_vcled_s64: { 4440 llvm::CmpInst::Predicate P; 4441 switch (BuiltinID) { 4442 default: llvm_unreachable("missing builtin ID in switch!"); 4443 case NEON::BI__builtin_neon_vceqd_s64: 4444 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4445 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4446 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4447 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4448 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4449 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4450 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4451 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4452 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4453 } 4454 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4455 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4456 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4457 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4458 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4459 } 4460 case NEON::BI__builtin_neon_vtstd_s64: 4461 case NEON::BI__builtin_neon_vtstd_u64: { 4462 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4463 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4464 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4465 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4466 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4467 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4468 llvm::Constant::getNullValue(Ty)); 4469 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4470 } 4471 case NEON::BI__builtin_neon_vset_lane_i8: 4472 case NEON::BI__builtin_neon_vset_lane_i16: 4473 case NEON::BI__builtin_neon_vset_lane_i32: 4474 case NEON::BI__builtin_neon_vset_lane_i64: 4475 case NEON::BI__builtin_neon_vset_lane_f32: 4476 case NEON::BI__builtin_neon_vsetq_lane_i8: 4477 case NEON::BI__builtin_neon_vsetq_lane_i16: 4478 case NEON::BI__builtin_neon_vsetq_lane_i32: 4479 case NEON::BI__builtin_neon_vsetq_lane_i64: 4480 case NEON::BI__builtin_neon_vsetq_lane_f32: 4481 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4482 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4483 case NEON::BI__builtin_neon_vset_lane_f64: 4484 // The vector type needs a cast for the v1f64 variant. 4485 Ops[1] = Builder.CreateBitCast(Ops[1], 4486 llvm::VectorType::get(DoubleTy, 1)); 4487 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4488 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4489 case NEON::BI__builtin_neon_vsetq_lane_f64: 4490 // The vector type needs a cast for the v2f64 variant. 4491 Ops[1] = Builder.CreateBitCast(Ops[1], 4492 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4493 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4494 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4495 4496 case NEON::BI__builtin_neon_vget_lane_i8: 4497 case NEON::BI__builtin_neon_vdupb_lane_i8: 4498 Ops[0] = Builder.CreateBitCast(Ops[0], 4499 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4500 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4501 "vget_lane"); 4502 case NEON::BI__builtin_neon_vgetq_lane_i8: 4503 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4504 Ops[0] = Builder.CreateBitCast(Ops[0], 4505 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4506 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4507 "vgetq_lane"); 4508 case NEON::BI__builtin_neon_vget_lane_i16: 4509 case NEON::BI__builtin_neon_vduph_lane_i16: 4510 Ops[0] = Builder.CreateBitCast(Ops[0], 4511 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4512 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4513 "vget_lane"); 4514 case NEON::BI__builtin_neon_vgetq_lane_i16: 4515 case NEON::BI__builtin_neon_vduph_laneq_i16: 4516 Ops[0] = Builder.CreateBitCast(Ops[0], 4517 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4518 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4519 "vgetq_lane"); 4520 case NEON::BI__builtin_neon_vget_lane_i32: 4521 case NEON::BI__builtin_neon_vdups_lane_i32: 4522 Ops[0] = Builder.CreateBitCast( 4523 Ops[0], 4524 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4525 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4526 "vget_lane"); 4527 case NEON::BI__builtin_neon_vdups_lane_f32: 4528 Ops[0] = Builder.CreateBitCast(Ops[0], 4529 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4530 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4531 "vdups_lane"); 4532 case NEON::BI__builtin_neon_vgetq_lane_i32: 4533 case NEON::BI__builtin_neon_vdups_laneq_i32: 4534 Ops[0] = Builder.CreateBitCast(Ops[0], 4535 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4536 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4537 "vgetq_lane"); 4538 case NEON::BI__builtin_neon_vget_lane_i64: 4539 case NEON::BI__builtin_neon_vdupd_lane_i64: 4540 Ops[0] = Builder.CreateBitCast(Ops[0], 4541 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4542 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4543 "vget_lane"); 4544 case NEON::BI__builtin_neon_vdupd_lane_f64: 4545 Ops[0] = Builder.CreateBitCast(Ops[0], 4546 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4547 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4548 "vdupd_lane"); 4549 case NEON::BI__builtin_neon_vgetq_lane_i64: 4550 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4551 Ops[0] = Builder.CreateBitCast(Ops[0], 4552 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4553 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4554 "vgetq_lane"); 4555 case NEON::BI__builtin_neon_vget_lane_f32: 4556 Ops[0] = Builder.CreateBitCast(Ops[0], 4557 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4558 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4559 "vget_lane"); 4560 case NEON::BI__builtin_neon_vget_lane_f64: 4561 Ops[0] = Builder.CreateBitCast(Ops[0], 4562 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4563 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4564 "vget_lane"); 4565 case NEON::BI__builtin_neon_vgetq_lane_f32: 4566 case NEON::BI__builtin_neon_vdups_laneq_f32: 4567 Ops[0] = Builder.CreateBitCast(Ops[0], 4568 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4569 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4570 "vgetq_lane"); 4571 case NEON::BI__builtin_neon_vgetq_lane_f64: 4572 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4573 Ops[0] = Builder.CreateBitCast(Ops[0], 4574 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4575 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4576 "vgetq_lane"); 4577 case NEON::BI__builtin_neon_vaddd_s64: 4578 case NEON::BI__builtin_neon_vaddd_u64: 4579 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4580 case NEON::BI__builtin_neon_vsubd_s64: 4581 case NEON::BI__builtin_neon_vsubd_u64: 4582 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4583 case NEON::BI__builtin_neon_vqdmlalh_s16: 4584 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4585 SmallVector<Value *, 2> ProductOps; 4586 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4587 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4588 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4589 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4590 ProductOps, "vqdmlXl"); 4591 Constant *CI = ConstantInt::get(SizeTy, 0); 4592 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4593 4594 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4595 ? Intrinsic::aarch64_neon_sqadd 4596 : Intrinsic::aarch64_neon_sqsub; 4597 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4598 } 4599 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4600 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4601 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4602 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4603 Ops, "vqshlu_n"); 4604 } 4605 case NEON::BI__builtin_neon_vqshld_n_u64: 4606 case NEON::BI__builtin_neon_vqshld_n_s64: { 4607 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4608 ? Intrinsic::aarch64_neon_uqshl 4609 : Intrinsic::aarch64_neon_sqshl; 4610 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4611 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4612 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4613 } 4614 case NEON::BI__builtin_neon_vrshrd_n_u64: 4615 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4616 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4617 ? Intrinsic::aarch64_neon_urshl 4618 : Intrinsic::aarch64_neon_srshl; 4619 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4620 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4621 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4622 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4623 } 4624 case NEON::BI__builtin_neon_vrsrad_n_u64: 4625 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4626 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4627 ? Intrinsic::aarch64_neon_urshl 4628 : Intrinsic::aarch64_neon_srshl; 4629 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4630 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4631 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4632 Builder.CreateSExt(Ops[2], Int64Ty)); 4633 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4634 } 4635 case NEON::BI__builtin_neon_vshld_n_s64: 4636 case NEON::BI__builtin_neon_vshld_n_u64: { 4637 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4638 return Builder.CreateShl( 4639 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4640 } 4641 case NEON::BI__builtin_neon_vshrd_n_s64: { 4642 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4643 return Builder.CreateAShr( 4644 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4645 Amt->getZExtValue())), 4646 "shrd_n"); 4647 } 4648 case NEON::BI__builtin_neon_vshrd_n_u64: { 4649 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4650 uint64_t ShiftAmt = Amt->getZExtValue(); 4651 // Right-shifting an unsigned value by its size yields 0. 4652 if (ShiftAmt == 64) 4653 return ConstantInt::get(Int64Ty, 0); 4654 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4655 "shrd_n"); 4656 } 4657 case NEON::BI__builtin_neon_vsrad_n_s64: { 4658 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4659 Ops[1] = Builder.CreateAShr( 4660 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4661 Amt->getZExtValue())), 4662 "shrd_n"); 4663 return Builder.CreateAdd(Ops[0], Ops[1]); 4664 } 4665 case NEON::BI__builtin_neon_vsrad_n_u64: { 4666 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4667 uint64_t ShiftAmt = Amt->getZExtValue(); 4668 // Right-shifting an unsigned value by its size yields 0. 4669 // As Op + 0 = Op, return Ops[0] directly. 4670 if (ShiftAmt == 64) 4671 return Ops[0]; 4672 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4673 "shrd_n"); 4674 return Builder.CreateAdd(Ops[0], Ops[1]); 4675 } 4676 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4677 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4678 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4679 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4680 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4681 "lane"); 4682 SmallVector<Value *, 2> ProductOps; 4683 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4684 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4685 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4686 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4687 ProductOps, "vqdmlXl"); 4688 Constant *CI = ConstantInt::get(SizeTy, 0); 4689 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4690 Ops.pop_back(); 4691 4692 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4693 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4694 ? Intrinsic::aarch64_neon_sqadd 4695 : Intrinsic::aarch64_neon_sqsub; 4696 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4697 } 4698 case NEON::BI__builtin_neon_vqdmlals_s32: 4699 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4700 SmallVector<Value *, 2> ProductOps; 4701 ProductOps.push_back(Ops[1]); 4702 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4703 Ops[1] = 4704 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4705 ProductOps, "vqdmlXl"); 4706 4707 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4708 ? Intrinsic::aarch64_neon_sqadd 4709 : Intrinsic::aarch64_neon_sqsub; 4710 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4711 } 4712 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4713 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4714 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4715 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4716 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4717 "lane"); 4718 SmallVector<Value *, 2> ProductOps; 4719 ProductOps.push_back(Ops[1]); 4720 ProductOps.push_back(Ops[2]); 4721 Ops[1] = 4722 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4723 ProductOps, "vqdmlXl"); 4724 Ops.pop_back(); 4725 4726 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4727 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4728 ? Intrinsic::aarch64_neon_sqadd 4729 : Intrinsic::aarch64_neon_sqsub; 4730 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4731 } 4732 } 4733 4734 llvm::VectorType *VTy = GetNeonType(this, Type); 4735 llvm::Type *Ty = VTy; 4736 if (!Ty) 4737 return nullptr; 4738 4739 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4740 // defer to common code if it's been added to our special map. 4741 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4742 AArch64SIMDIntrinsicsProvenSorted); 4743 4744 if (Builtin) 4745 return EmitCommonNeonBuiltinExpr( 4746 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4747 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4748 4749 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4750 return V; 4751 4752 unsigned Int; 4753 switch (BuiltinID) { 4754 default: return nullptr; 4755 case NEON::BI__builtin_neon_vbsl_v: 4756 case NEON::BI__builtin_neon_vbslq_v: { 4757 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4758 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4759 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4760 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4761 4762 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4763 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4764 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4765 return Builder.CreateBitCast(Ops[0], Ty); 4766 } 4767 case NEON::BI__builtin_neon_vfma_lane_v: 4768 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4769 // The ARM builtins (and instructions) have the addend as the first 4770 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4771 Value *Addend = Ops[0]; 4772 Value *Multiplicand = Ops[1]; 4773 Value *LaneSource = Ops[2]; 4774 Ops[0] = Multiplicand; 4775 Ops[1] = LaneSource; 4776 Ops[2] = Addend; 4777 4778 // Now adjust things to handle the lane access. 4779 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4780 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4781 VTy; 4782 llvm::Constant *cst = cast<Constant>(Ops[3]); 4783 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4784 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4785 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4786 4787 Ops.pop_back(); 4788 Int = Intrinsic::fma; 4789 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4790 } 4791 case NEON::BI__builtin_neon_vfma_laneq_v: { 4792 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4793 // v1f64 fma should be mapped to Neon scalar f64 fma 4794 if (VTy && VTy->getElementType() == DoubleTy) { 4795 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4796 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4797 llvm::Type *VTy = GetNeonType(this, 4798 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4799 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4800 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4801 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4802 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4803 return Builder.CreateBitCast(Result, Ty); 4804 } 4805 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4806 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4807 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4808 4809 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4810 VTy->getNumElements() * 2); 4811 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4812 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4813 cast<ConstantInt>(Ops[3])); 4814 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4815 4816 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4817 } 4818 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4819 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4820 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4821 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4822 4823 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4824 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4825 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4826 } 4827 case NEON::BI__builtin_neon_vfmas_lane_f32: 4828 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4829 case NEON::BI__builtin_neon_vfmad_lane_f64: 4830 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4831 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4832 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4833 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4834 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4835 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4836 } 4837 case NEON::BI__builtin_neon_vfms_v: 4838 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4839 // FIXME: probably remove when we no longer support aarch64_simd.h 4840 // (arm_neon.h delegates to vfma). 4841 4842 // The ARM builtins (and instructions) have the addend as the first 4843 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4844 Value *Subtrahend = Ops[0]; 4845 Value *Multiplicand = Ops[2]; 4846 Ops[0] = Multiplicand; 4847 Ops[2] = Subtrahend; 4848 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4849 Ops[1] = Builder.CreateFNeg(Ops[1]); 4850 Int = Intrinsic::fma; 4851 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4852 } 4853 case NEON::BI__builtin_neon_vmull_v: 4854 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4855 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4856 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4857 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4858 case NEON::BI__builtin_neon_vmax_v: 4859 case NEON::BI__builtin_neon_vmaxq_v: 4860 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4861 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4862 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4863 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4864 case NEON::BI__builtin_neon_vmin_v: 4865 case NEON::BI__builtin_neon_vminq_v: 4866 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4867 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4868 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4869 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4870 case NEON::BI__builtin_neon_vabd_v: 4871 case NEON::BI__builtin_neon_vabdq_v: 4872 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4873 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4874 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4875 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4876 case NEON::BI__builtin_neon_vpadal_v: 4877 case NEON::BI__builtin_neon_vpadalq_v: { 4878 unsigned ArgElts = VTy->getNumElements(); 4879 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4880 unsigned BitWidth = EltTy->getBitWidth(); 4881 llvm::Type *ArgTy = llvm::VectorType::get( 4882 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4883 llvm::Type* Tys[2] = { VTy, ArgTy }; 4884 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4885 SmallVector<llvm::Value*, 1> TmpOps; 4886 TmpOps.push_back(Ops[1]); 4887 Function *F = CGM.getIntrinsic(Int, Tys); 4888 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4889 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4890 return Builder.CreateAdd(tmp, addend); 4891 } 4892 case NEON::BI__builtin_neon_vpmin_v: 4893 case NEON::BI__builtin_neon_vpminq_v: 4894 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4895 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4896 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4897 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4898 case NEON::BI__builtin_neon_vpmax_v: 4899 case NEON::BI__builtin_neon_vpmaxq_v: 4900 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4901 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4902 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4903 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4904 case NEON::BI__builtin_neon_vminnm_v: 4905 case NEON::BI__builtin_neon_vminnmq_v: 4906 Int = Intrinsic::aarch64_neon_fminnm; 4907 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4908 case NEON::BI__builtin_neon_vmaxnm_v: 4909 case NEON::BI__builtin_neon_vmaxnmq_v: 4910 Int = Intrinsic::aarch64_neon_fmaxnm; 4911 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4912 case NEON::BI__builtin_neon_vrecpss_f32: { 4913 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4914 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4915 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4916 Ops, "vrecps"); 4917 } 4918 case NEON::BI__builtin_neon_vrecpsd_f64: { 4919 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4920 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4921 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4922 Ops, "vrecps"); 4923 } 4924 case NEON::BI__builtin_neon_vqshrun_n_v: 4925 Int = Intrinsic::aarch64_neon_sqshrun; 4926 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4927 case NEON::BI__builtin_neon_vqrshrun_n_v: 4928 Int = Intrinsic::aarch64_neon_sqrshrun; 4929 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4930 case NEON::BI__builtin_neon_vqshrn_n_v: 4931 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4933 case NEON::BI__builtin_neon_vrshrn_n_v: 4934 Int = Intrinsic::aarch64_neon_rshrn; 4935 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4936 case NEON::BI__builtin_neon_vqrshrn_n_v: 4937 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4938 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4939 case NEON::BI__builtin_neon_vrnda_v: 4940 case NEON::BI__builtin_neon_vrndaq_v: { 4941 Int = Intrinsic::round; 4942 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4943 } 4944 case NEON::BI__builtin_neon_vrndi_v: 4945 case NEON::BI__builtin_neon_vrndiq_v: { 4946 Int = Intrinsic::nearbyint; 4947 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4948 } 4949 case NEON::BI__builtin_neon_vrndm_v: 4950 case NEON::BI__builtin_neon_vrndmq_v: { 4951 Int = Intrinsic::floor; 4952 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4953 } 4954 case NEON::BI__builtin_neon_vrndn_v: 4955 case NEON::BI__builtin_neon_vrndnq_v: { 4956 Int = Intrinsic::aarch64_neon_frintn; 4957 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4958 } 4959 case NEON::BI__builtin_neon_vrndp_v: 4960 case NEON::BI__builtin_neon_vrndpq_v: { 4961 Int = Intrinsic::ceil; 4962 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4963 } 4964 case NEON::BI__builtin_neon_vrndx_v: 4965 case NEON::BI__builtin_neon_vrndxq_v: { 4966 Int = Intrinsic::rint; 4967 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4968 } 4969 case NEON::BI__builtin_neon_vrnd_v: 4970 case NEON::BI__builtin_neon_vrndq_v: { 4971 Int = Intrinsic::trunc; 4972 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4973 } 4974 case NEON::BI__builtin_neon_vceqz_v: 4975 case NEON::BI__builtin_neon_vceqzq_v: 4976 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4977 ICmpInst::ICMP_EQ, "vceqz"); 4978 case NEON::BI__builtin_neon_vcgez_v: 4979 case NEON::BI__builtin_neon_vcgezq_v: 4980 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4981 ICmpInst::ICMP_SGE, "vcgez"); 4982 case NEON::BI__builtin_neon_vclez_v: 4983 case NEON::BI__builtin_neon_vclezq_v: 4984 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4985 ICmpInst::ICMP_SLE, "vclez"); 4986 case NEON::BI__builtin_neon_vcgtz_v: 4987 case NEON::BI__builtin_neon_vcgtzq_v: 4988 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4989 ICmpInst::ICMP_SGT, "vcgtz"); 4990 case NEON::BI__builtin_neon_vcltz_v: 4991 case NEON::BI__builtin_neon_vcltzq_v: 4992 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4993 ICmpInst::ICMP_SLT, "vcltz"); 4994 case NEON::BI__builtin_neon_vcvt_f64_v: 4995 case NEON::BI__builtin_neon_vcvtq_f64_v: 4996 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4997 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4998 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4999 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5000 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5001 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5002 "unexpected vcvt_f64_f32 builtin"); 5003 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5004 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5005 5006 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5007 } 5008 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5009 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5010 "unexpected vcvt_f32_f64 builtin"); 5011 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5012 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5013 5014 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5015 } 5016 case NEON::BI__builtin_neon_vcvt_s32_v: 5017 case NEON::BI__builtin_neon_vcvt_u32_v: 5018 case NEON::BI__builtin_neon_vcvt_s64_v: 5019 case NEON::BI__builtin_neon_vcvt_u64_v: 5020 case NEON::BI__builtin_neon_vcvtq_s32_v: 5021 case NEON::BI__builtin_neon_vcvtq_u32_v: 5022 case NEON::BI__builtin_neon_vcvtq_s64_v: 5023 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5024 bool Double = 5025 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5026 llvm::Type *InTy = 5027 GetNeonType(this, 5028 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5029 : NeonTypeFlags::Float32, false, quad)); 5030 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5031 if (usgn) 5032 return Builder.CreateFPToUI(Ops[0], Ty); 5033 return Builder.CreateFPToSI(Ops[0], Ty); 5034 } 5035 case NEON::BI__builtin_neon_vcvta_s32_v: 5036 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5037 case NEON::BI__builtin_neon_vcvta_u32_v: 5038 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5039 case NEON::BI__builtin_neon_vcvta_s64_v: 5040 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5041 case NEON::BI__builtin_neon_vcvta_u64_v: 5042 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5043 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5044 bool Double = 5045 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5046 llvm::Type *InTy = 5047 GetNeonType(this, 5048 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5049 : NeonTypeFlags::Float32, false, quad)); 5050 llvm::Type *Tys[2] = { Ty, InTy }; 5051 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5052 } 5053 case NEON::BI__builtin_neon_vcvtm_s32_v: 5054 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5055 case NEON::BI__builtin_neon_vcvtm_u32_v: 5056 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5057 case NEON::BI__builtin_neon_vcvtm_s64_v: 5058 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5059 case NEON::BI__builtin_neon_vcvtm_u64_v: 5060 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5061 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5062 bool Double = 5063 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5064 llvm::Type *InTy = 5065 GetNeonType(this, 5066 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5067 : NeonTypeFlags::Float32, false, quad)); 5068 llvm::Type *Tys[2] = { Ty, InTy }; 5069 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5070 } 5071 case NEON::BI__builtin_neon_vcvtn_s32_v: 5072 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5073 case NEON::BI__builtin_neon_vcvtn_u32_v: 5074 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5075 case NEON::BI__builtin_neon_vcvtn_s64_v: 5076 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5077 case NEON::BI__builtin_neon_vcvtn_u64_v: 5078 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5079 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5080 bool Double = 5081 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5082 llvm::Type *InTy = 5083 GetNeonType(this, 5084 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5085 : NeonTypeFlags::Float32, false, quad)); 5086 llvm::Type *Tys[2] = { Ty, InTy }; 5087 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5088 } 5089 case NEON::BI__builtin_neon_vcvtp_s32_v: 5090 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5091 case NEON::BI__builtin_neon_vcvtp_u32_v: 5092 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5093 case NEON::BI__builtin_neon_vcvtp_s64_v: 5094 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5095 case NEON::BI__builtin_neon_vcvtp_u64_v: 5096 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5097 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5098 bool Double = 5099 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5100 llvm::Type *InTy = 5101 GetNeonType(this, 5102 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5103 : NeonTypeFlags::Float32, false, quad)); 5104 llvm::Type *Tys[2] = { Ty, InTy }; 5105 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5106 } 5107 case NEON::BI__builtin_neon_vmulx_v: 5108 case NEON::BI__builtin_neon_vmulxq_v: { 5109 Int = Intrinsic::aarch64_neon_fmulx; 5110 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5111 } 5112 case NEON::BI__builtin_neon_vmul_lane_v: 5113 case NEON::BI__builtin_neon_vmul_laneq_v: { 5114 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5115 bool Quad = false; 5116 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5117 Quad = true; 5118 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5119 llvm::Type *VTy = GetNeonType(this, 5120 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5121 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5122 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5123 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5124 return Builder.CreateBitCast(Result, Ty); 5125 } 5126 case NEON::BI__builtin_neon_vnegd_s64: 5127 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5128 case NEON::BI__builtin_neon_vpmaxnm_v: 5129 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5130 Int = Intrinsic::aarch64_neon_fmaxnmp; 5131 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5132 } 5133 case NEON::BI__builtin_neon_vpminnm_v: 5134 case NEON::BI__builtin_neon_vpminnmq_v: { 5135 Int = Intrinsic::aarch64_neon_fminnmp; 5136 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5137 } 5138 case NEON::BI__builtin_neon_vsqrt_v: 5139 case NEON::BI__builtin_neon_vsqrtq_v: { 5140 Int = Intrinsic::sqrt; 5141 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5142 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5143 } 5144 case NEON::BI__builtin_neon_vrbit_v: 5145 case NEON::BI__builtin_neon_vrbitq_v: { 5146 Int = Intrinsic::aarch64_neon_rbit; 5147 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5148 } 5149 case NEON::BI__builtin_neon_vaddv_u8: 5150 // FIXME: These are handled by the AArch64 scalar code. 5151 usgn = true; 5152 // FALLTHROUGH 5153 case NEON::BI__builtin_neon_vaddv_s8: { 5154 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5155 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5156 VTy = 5157 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5158 llvm::Type *Tys[2] = { Ty, VTy }; 5159 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5160 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5161 return Builder.CreateTrunc(Ops[0], 5162 llvm::IntegerType::get(getLLVMContext(), 8)); 5163 } 5164 case NEON::BI__builtin_neon_vaddv_u16: 5165 usgn = true; 5166 // FALLTHROUGH 5167 case NEON::BI__builtin_neon_vaddv_s16: { 5168 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5169 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5170 VTy = 5171 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5172 llvm::Type *Tys[2] = { Ty, VTy }; 5173 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5174 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5175 return Builder.CreateTrunc(Ops[0], 5176 llvm::IntegerType::get(getLLVMContext(), 16)); 5177 } 5178 case NEON::BI__builtin_neon_vaddvq_u8: 5179 usgn = true; 5180 // FALLTHROUGH 5181 case NEON::BI__builtin_neon_vaddvq_s8: { 5182 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5183 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5184 VTy = 5185 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5186 llvm::Type *Tys[2] = { Ty, VTy }; 5187 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5188 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5189 return Builder.CreateTrunc(Ops[0], 5190 llvm::IntegerType::get(getLLVMContext(), 8)); 5191 } 5192 case NEON::BI__builtin_neon_vaddvq_u16: 5193 usgn = true; 5194 // FALLTHROUGH 5195 case NEON::BI__builtin_neon_vaddvq_s16: { 5196 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5197 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5198 VTy = 5199 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5200 llvm::Type *Tys[2] = { Ty, VTy }; 5201 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5202 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5203 return Builder.CreateTrunc(Ops[0], 5204 llvm::IntegerType::get(getLLVMContext(), 16)); 5205 } 5206 case NEON::BI__builtin_neon_vmaxv_u8: { 5207 Int = Intrinsic::aarch64_neon_umaxv; 5208 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5209 VTy = 5210 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5211 llvm::Type *Tys[2] = { Ty, VTy }; 5212 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5213 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5214 return Builder.CreateTrunc(Ops[0], 5215 llvm::IntegerType::get(getLLVMContext(), 8)); 5216 } 5217 case NEON::BI__builtin_neon_vmaxv_u16: { 5218 Int = Intrinsic::aarch64_neon_umaxv; 5219 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5220 VTy = 5221 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5222 llvm::Type *Tys[2] = { Ty, VTy }; 5223 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5224 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5225 return Builder.CreateTrunc(Ops[0], 5226 llvm::IntegerType::get(getLLVMContext(), 16)); 5227 } 5228 case NEON::BI__builtin_neon_vmaxvq_u8: { 5229 Int = Intrinsic::aarch64_neon_umaxv; 5230 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5231 VTy = 5232 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5233 llvm::Type *Tys[2] = { Ty, VTy }; 5234 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5235 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5236 return Builder.CreateTrunc(Ops[0], 5237 llvm::IntegerType::get(getLLVMContext(), 8)); 5238 } 5239 case NEON::BI__builtin_neon_vmaxvq_u16: { 5240 Int = Intrinsic::aarch64_neon_umaxv; 5241 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5242 VTy = 5243 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5244 llvm::Type *Tys[2] = { Ty, VTy }; 5245 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5246 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5247 return Builder.CreateTrunc(Ops[0], 5248 llvm::IntegerType::get(getLLVMContext(), 16)); 5249 } 5250 case NEON::BI__builtin_neon_vmaxv_s8: { 5251 Int = Intrinsic::aarch64_neon_smaxv; 5252 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5253 VTy = 5254 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5255 llvm::Type *Tys[2] = { Ty, VTy }; 5256 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5257 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5258 return Builder.CreateTrunc(Ops[0], 5259 llvm::IntegerType::get(getLLVMContext(), 8)); 5260 } 5261 case NEON::BI__builtin_neon_vmaxv_s16: { 5262 Int = Intrinsic::aarch64_neon_smaxv; 5263 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5264 VTy = 5265 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5266 llvm::Type *Tys[2] = { Ty, VTy }; 5267 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5268 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5269 return Builder.CreateTrunc(Ops[0], 5270 llvm::IntegerType::get(getLLVMContext(), 16)); 5271 } 5272 case NEON::BI__builtin_neon_vmaxvq_s8: { 5273 Int = Intrinsic::aarch64_neon_smaxv; 5274 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5275 VTy = 5276 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5277 llvm::Type *Tys[2] = { Ty, VTy }; 5278 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5279 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5280 return Builder.CreateTrunc(Ops[0], 5281 llvm::IntegerType::get(getLLVMContext(), 8)); 5282 } 5283 case NEON::BI__builtin_neon_vmaxvq_s16: { 5284 Int = Intrinsic::aarch64_neon_smaxv; 5285 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5286 VTy = 5287 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5288 llvm::Type *Tys[2] = { Ty, VTy }; 5289 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5290 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5291 return Builder.CreateTrunc(Ops[0], 5292 llvm::IntegerType::get(getLLVMContext(), 16)); 5293 } 5294 case NEON::BI__builtin_neon_vminv_u8: { 5295 Int = Intrinsic::aarch64_neon_uminv; 5296 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5297 VTy = 5298 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5299 llvm::Type *Tys[2] = { Ty, VTy }; 5300 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5301 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5302 return Builder.CreateTrunc(Ops[0], 5303 llvm::IntegerType::get(getLLVMContext(), 8)); 5304 } 5305 case NEON::BI__builtin_neon_vminv_u16: { 5306 Int = Intrinsic::aarch64_neon_uminv; 5307 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5308 VTy = 5309 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5310 llvm::Type *Tys[2] = { Ty, VTy }; 5311 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5312 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5313 return Builder.CreateTrunc(Ops[0], 5314 llvm::IntegerType::get(getLLVMContext(), 16)); 5315 } 5316 case NEON::BI__builtin_neon_vminvq_u8: { 5317 Int = Intrinsic::aarch64_neon_uminv; 5318 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5319 VTy = 5320 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5321 llvm::Type *Tys[2] = { Ty, VTy }; 5322 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5323 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5324 return Builder.CreateTrunc(Ops[0], 5325 llvm::IntegerType::get(getLLVMContext(), 8)); 5326 } 5327 case NEON::BI__builtin_neon_vminvq_u16: { 5328 Int = Intrinsic::aarch64_neon_uminv; 5329 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5330 VTy = 5331 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5332 llvm::Type *Tys[2] = { Ty, VTy }; 5333 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5334 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5335 return Builder.CreateTrunc(Ops[0], 5336 llvm::IntegerType::get(getLLVMContext(), 16)); 5337 } 5338 case NEON::BI__builtin_neon_vminv_s8: { 5339 Int = Intrinsic::aarch64_neon_sminv; 5340 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5341 VTy = 5342 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5343 llvm::Type *Tys[2] = { Ty, VTy }; 5344 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5345 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5346 return Builder.CreateTrunc(Ops[0], 5347 llvm::IntegerType::get(getLLVMContext(), 8)); 5348 } 5349 case NEON::BI__builtin_neon_vminv_s16: { 5350 Int = Intrinsic::aarch64_neon_sminv; 5351 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5352 VTy = 5353 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5354 llvm::Type *Tys[2] = { Ty, VTy }; 5355 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5356 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5357 return Builder.CreateTrunc(Ops[0], 5358 llvm::IntegerType::get(getLLVMContext(), 16)); 5359 } 5360 case NEON::BI__builtin_neon_vminvq_s8: { 5361 Int = Intrinsic::aarch64_neon_sminv; 5362 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5363 VTy = 5364 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5365 llvm::Type *Tys[2] = { Ty, VTy }; 5366 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5367 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5368 return Builder.CreateTrunc(Ops[0], 5369 llvm::IntegerType::get(getLLVMContext(), 8)); 5370 } 5371 case NEON::BI__builtin_neon_vminvq_s16: { 5372 Int = Intrinsic::aarch64_neon_sminv; 5373 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5374 VTy = 5375 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5376 llvm::Type *Tys[2] = { Ty, VTy }; 5377 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5378 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5379 return Builder.CreateTrunc(Ops[0], 5380 llvm::IntegerType::get(getLLVMContext(), 16)); 5381 } 5382 case NEON::BI__builtin_neon_vmul_n_f64: { 5383 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5384 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5385 return Builder.CreateFMul(Ops[0], RHS); 5386 } 5387 case NEON::BI__builtin_neon_vaddlv_u8: { 5388 Int = Intrinsic::aarch64_neon_uaddlv; 5389 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5390 VTy = 5391 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5392 llvm::Type *Tys[2] = { Ty, VTy }; 5393 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5394 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5395 return Builder.CreateTrunc(Ops[0], 5396 llvm::IntegerType::get(getLLVMContext(), 16)); 5397 } 5398 case NEON::BI__builtin_neon_vaddlv_u16: { 5399 Int = Intrinsic::aarch64_neon_uaddlv; 5400 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5401 VTy = 5402 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5403 llvm::Type *Tys[2] = { Ty, VTy }; 5404 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5405 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5406 } 5407 case NEON::BI__builtin_neon_vaddlvq_u8: { 5408 Int = Intrinsic::aarch64_neon_uaddlv; 5409 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5410 VTy = 5411 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5412 llvm::Type *Tys[2] = { Ty, VTy }; 5413 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5414 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5415 return Builder.CreateTrunc(Ops[0], 5416 llvm::IntegerType::get(getLLVMContext(), 16)); 5417 } 5418 case NEON::BI__builtin_neon_vaddlvq_u16: { 5419 Int = Intrinsic::aarch64_neon_uaddlv; 5420 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5421 VTy = 5422 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5423 llvm::Type *Tys[2] = { Ty, VTy }; 5424 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5425 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5426 } 5427 case NEON::BI__builtin_neon_vaddlv_s8: { 5428 Int = Intrinsic::aarch64_neon_saddlv; 5429 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5430 VTy = 5431 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5432 llvm::Type *Tys[2] = { Ty, VTy }; 5433 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5434 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5435 return Builder.CreateTrunc(Ops[0], 5436 llvm::IntegerType::get(getLLVMContext(), 16)); 5437 } 5438 case NEON::BI__builtin_neon_vaddlv_s16: { 5439 Int = Intrinsic::aarch64_neon_saddlv; 5440 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5441 VTy = 5442 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5443 llvm::Type *Tys[2] = { Ty, VTy }; 5444 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5445 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5446 } 5447 case NEON::BI__builtin_neon_vaddlvq_s8: { 5448 Int = Intrinsic::aarch64_neon_saddlv; 5449 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5450 VTy = 5451 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5452 llvm::Type *Tys[2] = { Ty, VTy }; 5453 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5454 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5455 return Builder.CreateTrunc(Ops[0], 5456 llvm::IntegerType::get(getLLVMContext(), 16)); 5457 } 5458 case NEON::BI__builtin_neon_vaddlvq_s16: { 5459 Int = Intrinsic::aarch64_neon_saddlv; 5460 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5461 VTy = 5462 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5463 llvm::Type *Tys[2] = { Ty, VTy }; 5464 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5465 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5466 } 5467 case NEON::BI__builtin_neon_vsri_n_v: 5468 case NEON::BI__builtin_neon_vsriq_n_v: { 5469 Int = Intrinsic::aarch64_neon_vsri; 5470 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5471 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5472 } 5473 case NEON::BI__builtin_neon_vsli_n_v: 5474 case NEON::BI__builtin_neon_vsliq_n_v: { 5475 Int = Intrinsic::aarch64_neon_vsli; 5476 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5477 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5478 } 5479 case NEON::BI__builtin_neon_vsra_n_v: 5480 case NEON::BI__builtin_neon_vsraq_n_v: 5481 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5482 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5483 return Builder.CreateAdd(Ops[0], Ops[1]); 5484 case NEON::BI__builtin_neon_vrsra_n_v: 5485 case NEON::BI__builtin_neon_vrsraq_n_v: { 5486 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5487 SmallVector<llvm::Value*,2> TmpOps; 5488 TmpOps.push_back(Ops[1]); 5489 TmpOps.push_back(Ops[2]); 5490 Function* F = CGM.getIntrinsic(Int, Ty); 5491 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5492 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5493 return Builder.CreateAdd(Ops[0], tmp); 5494 } 5495 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5496 // of an Align parameter here. 5497 case NEON::BI__builtin_neon_vld1_x2_v: 5498 case NEON::BI__builtin_neon_vld1q_x2_v: 5499 case NEON::BI__builtin_neon_vld1_x3_v: 5500 case NEON::BI__builtin_neon_vld1q_x3_v: 5501 case NEON::BI__builtin_neon_vld1_x4_v: 5502 case NEON::BI__builtin_neon_vld1q_x4_v: { 5503 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5504 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5505 llvm::Type *Tys[2] = { VTy, PTy }; 5506 unsigned Int; 5507 switch (BuiltinID) { 5508 case NEON::BI__builtin_neon_vld1_x2_v: 5509 case NEON::BI__builtin_neon_vld1q_x2_v: 5510 Int = Intrinsic::aarch64_neon_ld1x2; 5511 break; 5512 case NEON::BI__builtin_neon_vld1_x3_v: 5513 case NEON::BI__builtin_neon_vld1q_x3_v: 5514 Int = Intrinsic::aarch64_neon_ld1x3; 5515 break; 5516 case NEON::BI__builtin_neon_vld1_x4_v: 5517 case NEON::BI__builtin_neon_vld1q_x4_v: 5518 Int = Intrinsic::aarch64_neon_ld1x4; 5519 break; 5520 } 5521 Function *F = CGM.getIntrinsic(Int, Tys); 5522 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5523 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5524 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5525 return Builder.CreateStore(Ops[1], Ops[0]); 5526 } 5527 case NEON::BI__builtin_neon_vst1_x2_v: 5528 case NEON::BI__builtin_neon_vst1q_x2_v: 5529 case NEON::BI__builtin_neon_vst1_x3_v: 5530 case NEON::BI__builtin_neon_vst1q_x3_v: 5531 case NEON::BI__builtin_neon_vst1_x4_v: 5532 case NEON::BI__builtin_neon_vst1q_x4_v: { 5533 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5534 llvm::Type *Tys[2] = { VTy, PTy }; 5535 unsigned Int; 5536 switch (BuiltinID) { 5537 case NEON::BI__builtin_neon_vst1_x2_v: 5538 case NEON::BI__builtin_neon_vst1q_x2_v: 5539 Int = Intrinsic::aarch64_neon_st1x2; 5540 break; 5541 case NEON::BI__builtin_neon_vst1_x3_v: 5542 case NEON::BI__builtin_neon_vst1q_x3_v: 5543 Int = Intrinsic::aarch64_neon_st1x3; 5544 break; 5545 case NEON::BI__builtin_neon_vst1_x4_v: 5546 case NEON::BI__builtin_neon_vst1q_x4_v: 5547 Int = Intrinsic::aarch64_neon_st1x4; 5548 break; 5549 } 5550 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5551 IntOps.push_back(Ops[0]); 5552 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5553 } 5554 case NEON::BI__builtin_neon_vld1_v: 5555 case NEON::BI__builtin_neon_vld1q_v: 5556 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5557 return Builder.CreateLoad(Ops[0]); 5558 case NEON::BI__builtin_neon_vst1_v: 5559 case NEON::BI__builtin_neon_vst1q_v: 5560 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5561 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5562 return Builder.CreateStore(Ops[1], Ops[0]); 5563 case NEON::BI__builtin_neon_vld1_lane_v: 5564 case NEON::BI__builtin_neon_vld1q_lane_v: 5565 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5566 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5567 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5568 Ops[0] = Builder.CreateLoad(Ops[0]); 5569 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5570 case NEON::BI__builtin_neon_vld1_dup_v: 5571 case NEON::BI__builtin_neon_vld1q_dup_v: { 5572 Value *V = UndefValue::get(Ty); 5573 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5574 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5575 Ops[0] = Builder.CreateLoad(Ops[0]); 5576 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5577 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5578 return EmitNeonSplat(Ops[0], CI); 5579 } 5580 case NEON::BI__builtin_neon_vst1_lane_v: 5581 case NEON::BI__builtin_neon_vst1q_lane_v: 5582 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5583 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5584 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5585 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5586 case NEON::BI__builtin_neon_vld2_v: 5587 case NEON::BI__builtin_neon_vld2q_v: { 5588 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5589 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5590 llvm::Type *Tys[2] = { VTy, PTy }; 5591 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5592 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5593 Ops[0] = Builder.CreateBitCast(Ops[0], 5594 llvm::PointerType::getUnqual(Ops[1]->getType())); 5595 return Builder.CreateStore(Ops[1], Ops[0]); 5596 } 5597 case NEON::BI__builtin_neon_vld3_v: 5598 case NEON::BI__builtin_neon_vld3q_v: { 5599 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5600 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5601 llvm::Type *Tys[2] = { VTy, PTy }; 5602 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5603 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5604 Ops[0] = Builder.CreateBitCast(Ops[0], 5605 llvm::PointerType::getUnqual(Ops[1]->getType())); 5606 return Builder.CreateStore(Ops[1], Ops[0]); 5607 } 5608 case NEON::BI__builtin_neon_vld4_v: 5609 case NEON::BI__builtin_neon_vld4q_v: { 5610 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5611 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5612 llvm::Type *Tys[2] = { VTy, PTy }; 5613 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5614 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5615 Ops[0] = Builder.CreateBitCast(Ops[0], 5616 llvm::PointerType::getUnqual(Ops[1]->getType())); 5617 return Builder.CreateStore(Ops[1], Ops[0]); 5618 } 5619 case NEON::BI__builtin_neon_vld2_dup_v: 5620 case NEON::BI__builtin_neon_vld2q_dup_v: { 5621 llvm::Type *PTy = 5622 llvm::PointerType::getUnqual(VTy->getElementType()); 5623 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5624 llvm::Type *Tys[2] = { VTy, PTy }; 5625 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5626 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5627 Ops[0] = Builder.CreateBitCast(Ops[0], 5628 llvm::PointerType::getUnqual(Ops[1]->getType())); 5629 return Builder.CreateStore(Ops[1], Ops[0]); 5630 } 5631 case NEON::BI__builtin_neon_vld3_dup_v: 5632 case NEON::BI__builtin_neon_vld3q_dup_v: { 5633 llvm::Type *PTy = 5634 llvm::PointerType::getUnqual(VTy->getElementType()); 5635 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5636 llvm::Type *Tys[2] = { VTy, PTy }; 5637 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5638 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5639 Ops[0] = Builder.CreateBitCast(Ops[0], 5640 llvm::PointerType::getUnqual(Ops[1]->getType())); 5641 return Builder.CreateStore(Ops[1], Ops[0]); 5642 } 5643 case NEON::BI__builtin_neon_vld4_dup_v: 5644 case NEON::BI__builtin_neon_vld4q_dup_v: { 5645 llvm::Type *PTy = 5646 llvm::PointerType::getUnqual(VTy->getElementType()); 5647 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5648 llvm::Type *Tys[2] = { VTy, PTy }; 5649 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5650 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5651 Ops[0] = Builder.CreateBitCast(Ops[0], 5652 llvm::PointerType::getUnqual(Ops[1]->getType())); 5653 return Builder.CreateStore(Ops[1], Ops[0]); 5654 } 5655 case NEON::BI__builtin_neon_vld2_lane_v: 5656 case NEON::BI__builtin_neon_vld2q_lane_v: { 5657 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5658 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5659 Ops.push_back(Ops[1]); 5660 Ops.erase(Ops.begin()+1); 5661 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5662 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5663 Ops[3] = Builder.CreateZExt(Ops[3], 5664 llvm::IntegerType::get(getLLVMContext(), 64)); 5665 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5666 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5667 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5668 return Builder.CreateStore(Ops[1], Ops[0]); 5669 } 5670 case NEON::BI__builtin_neon_vld3_lane_v: 5671 case NEON::BI__builtin_neon_vld3q_lane_v: { 5672 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5673 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5674 Ops.push_back(Ops[1]); 5675 Ops.erase(Ops.begin()+1); 5676 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5677 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5678 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5679 Ops[4] = Builder.CreateZExt(Ops[4], 5680 llvm::IntegerType::get(getLLVMContext(), 64)); 5681 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5682 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5683 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5684 return Builder.CreateStore(Ops[1], Ops[0]); 5685 } 5686 case NEON::BI__builtin_neon_vld4_lane_v: 5687 case NEON::BI__builtin_neon_vld4q_lane_v: { 5688 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5689 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5690 Ops.push_back(Ops[1]); 5691 Ops.erase(Ops.begin()+1); 5692 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5693 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5694 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5695 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5696 Ops[5] = Builder.CreateZExt(Ops[5], 5697 llvm::IntegerType::get(getLLVMContext(), 64)); 5698 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5699 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5700 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5701 return Builder.CreateStore(Ops[1], Ops[0]); 5702 } 5703 case NEON::BI__builtin_neon_vst2_v: 5704 case NEON::BI__builtin_neon_vst2q_v: { 5705 Ops.push_back(Ops[0]); 5706 Ops.erase(Ops.begin()); 5707 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5708 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5709 Ops, ""); 5710 } 5711 case NEON::BI__builtin_neon_vst2_lane_v: 5712 case NEON::BI__builtin_neon_vst2q_lane_v: { 5713 Ops.push_back(Ops[0]); 5714 Ops.erase(Ops.begin()); 5715 Ops[2] = Builder.CreateZExt(Ops[2], 5716 llvm::IntegerType::get(getLLVMContext(), 64)); 5717 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5718 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5719 Ops, ""); 5720 } 5721 case NEON::BI__builtin_neon_vst3_v: 5722 case NEON::BI__builtin_neon_vst3q_v: { 5723 Ops.push_back(Ops[0]); 5724 Ops.erase(Ops.begin()); 5725 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5726 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5727 Ops, ""); 5728 } 5729 case NEON::BI__builtin_neon_vst3_lane_v: 5730 case NEON::BI__builtin_neon_vst3q_lane_v: { 5731 Ops.push_back(Ops[0]); 5732 Ops.erase(Ops.begin()); 5733 Ops[3] = Builder.CreateZExt(Ops[3], 5734 llvm::IntegerType::get(getLLVMContext(), 64)); 5735 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5736 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5737 Ops, ""); 5738 } 5739 case NEON::BI__builtin_neon_vst4_v: 5740 case NEON::BI__builtin_neon_vst4q_v: { 5741 Ops.push_back(Ops[0]); 5742 Ops.erase(Ops.begin()); 5743 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5744 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5745 Ops, ""); 5746 } 5747 case NEON::BI__builtin_neon_vst4_lane_v: 5748 case NEON::BI__builtin_neon_vst4q_lane_v: { 5749 Ops.push_back(Ops[0]); 5750 Ops.erase(Ops.begin()); 5751 Ops[4] = Builder.CreateZExt(Ops[4], 5752 llvm::IntegerType::get(getLLVMContext(), 64)); 5753 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5754 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5755 Ops, ""); 5756 } 5757 case NEON::BI__builtin_neon_vtrn_v: 5758 case NEON::BI__builtin_neon_vtrnq_v: { 5759 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5760 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5761 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5762 Value *SV = nullptr; 5763 5764 for (unsigned vi = 0; vi != 2; ++vi) { 5765 SmallVector<Constant*, 16> Indices; 5766 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5767 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5768 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5769 } 5770 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5771 SV = llvm::ConstantVector::get(Indices); 5772 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5773 SV = Builder.CreateStore(SV, Addr); 5774 } 5775 return SV; 5776 } 5777 case NEON::BI__builtin_neon_vuzp_v: 5778 case NEON::BI__builtin_neon_vuzpq_v: { 5779 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5780 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5781 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5782 Value *SV = nullptr; 5783 5784 for (unsigned vi = 0; vi != 2; ++vi) { 5785 SmallVector<Constant*, 16> Indices; 5786 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5787 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5788 5789 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5790 SV = llvm::ConstantVector::get(Indices); 5791 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5792 SV = Builder.CreateStore(SV, Addr); 5793 } 5794 return SV; 5795 } 5796 case NEON::BI__builtin_neon_vzip_v: 5797 case NEON::BI__builtin_neon_vzipq_v: { 5798 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5799 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5800 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5801 Value *SV = nullptr; 5802 5803 for (unsigned vi = 0; vi != 2; ++vi) { 5804 SmallVector<Constant*, 16> Indices; 5805 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5806 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5807 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5808 } 5809 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5810 SV = llvm::ConstantVector::get(Indices); 5811 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5812 SV = Builder.CreateStore(SV, Addr); 5813 } 5814 return SV; 5815 } 5816 case NEON::BI__builtin_neon_vqtbl1q_v: { 5817 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5818 Ops, "vtbl1"); 5819 } 5820 case NEON::BI__builtin_neon_vqtbl2q_v: { 5821 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5822 Ops, "vtbl2"); 5823 } 5824 case NEON::BI__builtin_neon_vqtbl3q_v: { 5825 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5826 Ops, "vtbl3"); 5827 } 5828 case NEON::BI__builtin_neon_vqtbl4q_v: { 5829 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5830 Ops, "vtbl4"); 5831 } 5832 case NEON::BI__builtin_neon_vqtbx1q_v: { 5833 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5834 Ops, "vtbx1"); 5835 } 5836 case NEON::BI__builtin_neon_vqtbx2q_v: { 5837 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5838 Ops, "vtbx2"); 5839 } 5840 case NEON::BI__builtin_neon_vqtbx3q_v: { 5841 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5842 Ops, "vtbx3"); 5843 } 5844 case NEON::BI__builtin_neon_vqtbx4q_v: { 5845 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5846 Ops, "vtbx4"); 5847 } 5848 case NEON::BI__builtin_neon_vsqadd_v: 5849 case NEON::BI__builtin_neon_vsqaddq_v: { 5850 Int = Intrinsic::aarch64_neon_usqadd; 5851 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5852 } 5853 case NEON::BI__builtin_neon_vuqadd_v: 5854 case NEON::BI__builtin_neon_vuqaddq_v: { 5855 Int = Intrinsic::aarch64_neon_suqadd; 5856 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5857 } 5858 } 5859 } 5860 5861 llvm::Value *CodeGenFunction:: 5862 BuildVector(ArrayRef<llvm::Value*> Ops) { 5863 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5864 "Not a power-of-two sized vector!"); 5865 bool AllConstants = true; 5866 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5867 AllConstants &= isa<Constant>(Ops[i]); 5868 5869 // If this is a constant vector, create a ConstantVector. 5870 if (AllConstants) { 5871 SmallVector<llvm::Constant*, 16> CstOps; 5872 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5873 CstOps.push_back(cast<Constant>(Ops[i])); 5874 return llvm::ConstantVector::get(CstOps); 5875 } 5876 5877 // Otherwise, insertelement the values to build the vector. 5878 Value *Result = 5879 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5880 5881 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5882 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5883 5884 return Result; 5885 } 5886 5887 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5888 const CallExpr *E) { 5889 SmallVector<Value*, 4> Ops; 5890 5891 // Find out if any arguments are required to be integer constant expressions. 5892 unsigned ICEArguments = 0; 5893 ASTContext::GetBuiltinTypeError Error; 5894 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5895 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5896 5897 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5898 // If this is a normal argument, just emit it as a scalar. 5899 if ((ICEArguments & (1 << i)) == 0) { 5900 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5901 continue; 5902 } 5903 5904 // If this is required to be a constant, constant fold it so that we know 5905 // that the generated intrinsic gets a ConstantInt. 5906 llvm::APSInt Result; 5907 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5908 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5909 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5910 } 5911 5912 switch (BuiltinID) { 5913 default: return nullptr; 5914 case X86::BI_mm_prefetch: { 5915 Value *Address = EmitScalarExpr(E->getArg(0)); 5916 Value *RW = ConstantInt::get(Int32Ty, 0); 5917 Value *Locality = EmitScalarExpr(E->getArg(1)); 5918 Value *Data = ConstantInt::get(Int32Ty, 1); 5919 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5920 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5921 } 5922 case X86::BI__builtin_ia32_vec_init_v8qi: 5923 case X86::BI__builtin_ia32_vec_init_v4hi: 5924 case X86::BI__builtin_ia32_vec_init_v2si: 5925 return Builder.CreateBitCast(BuildVector(Ops), 5926 llvm::Type::getX86_MMXTy(getLLVMContext())); 5927 case X86::BI__builtin_ia32_vec_ext_v2si: 5928 return Builder.CreateExtractElement(Ops[0], 5929 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5930 case X86::BI__builtin_ia32_ldmxcsr: { 5931 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5932 Builder.CreateStore(Ops[0], Tmp); 5933 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5934 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5935 } 5936 case X86::BI__builtin_ia32_stmxcsr: { 5937 Value *Tmp = CreateMemTemp(E->getType()); 5938 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5939 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5940 return Builder.CreateLoad(Tmp, "stmxcsr"); 5941 } 5942 case X86::BI__builtin_ia32_storehps: 5943 case X86::BI__builtin_ia32_storelps: { 5944 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5945 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5946 5947 // cast val v2i64 5948 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5949 5950 // extract (0, 1) 5951 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5952 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5953 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5954 5955 // cast pointer to i64 & store 5956 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5957 return Builder.CreateStore(Ops[1], Ops[0]); 5958 } 5959 case X86::BI__builtin_ia32_palignr128: 5960 case X86::BI__builtin_ia32_palignr256: { 5961 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5962 5963 unsigned NumElts = 5964 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 5965 assert(NumElts % 16 == 0); 5966 unsigned NumLanes = NumElts / 16; 5967 unsigned NumLaneElts = NumElts / NumLanes; 5968 5969 // If palignr is shifting the pair of vectors more than the size of two 5970 // lanes, emit zero. 5971 if (ShiftVal >= (2 * NumLaneElts)) 5972 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5973 5974 // If palignr is shifting the pair of input vectors more than one lane, 5975 // but less than two lanes, convert to shifting in zeroes. 5976 if (ShiftVal > NumLaneElts) { 5977 ShiftVal -= NumLaneElts; 5978 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 5979 } 5980 5981 SmallVector<llvm::Constant*, 32> Indices; 5982 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5983 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 5984 for (unsigned i = 0; i != NumLaneElts; ++i) { 5985 unsigned Idx = ShiftVal + i; 5986 if (Idx >= NumLaneElts) 5987 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 5988 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 5989 } 5990 } 5991 5992 Value* SV = llvm::ConstantVector::get(Indices); 5993 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5994 } 5995 case X86::BI__builtin_ia32_pslldqi256: { 5996 // Shift value is in bits so divide by 8. 5997 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 5998 5999 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6000 if (shiftVal >= 16) 6001 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6002 6003 SmallVector<llvm::Constant*, 32> Indices; 6004 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6005 for (unsigned l = 0; l != 32; l += 16) { 6006 for (unsigned i = 0; i != 16; ++i) { 6007 unsigned Idx = 32 + i - shiftVal; 6008 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6009 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6010 } 6011 } 6012 6013 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6014 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6015 Value *Zero = llvm::Constant::getNullValue(VecTy); 6016 6017 Value *SV = llvm::ConstantVector::get(Indices); 6018 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6019 llvm::Type *ResultType = ConvertType(E->getType()); 6020 return Builder.CreateBitCast(SV, ResultType, "cast"); 6021 } 6022 case X86::BI__builtin_ia32_psrldqi256: { 6023 // Shift value is in bits so divide by 8. 6024 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6025 6026 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6027 if (shiftVal >= 16) 6028 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6029 6030 SmallVector<llvm::Constant*, 32> Indices; 6031 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6032 for (unsigned l = 0; l != 32; l += 16) { 6033 for (unsigned i = 0; i != 16; ++i) { 6034 unsigned Idx = i + shiftVal; 6035 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6036 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6037 } 6038 } 6039 6040 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6041 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6042 Value *Zero = llvm::Constant::getNullValue(VecTy); 6043 6044 Value *SV = llvm::ConstantVector::get(Indices); 6045 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6046 llvm::Type *ResultType = ConvertType(E->getType()); 6047 return Builder.CreateBitCast(SV, ResultType, "cast"); 6048 } 6049 case X86::BI__builtin_ia32_movntps: 6050 case X86::BI__builtin_ia32_movntps256: 6051 case X86::BI__builtin_ia32_movntpd: 6052 case X86::BI__builtin_ia32_movntpd256: 6053 case X86::BI__builtin_ia32_movntdq: 6054 case X86::BI__builtin_ia32_movntdq256: 6055 case X86::BI__builtin_ia32_movnti: 6056 case X86::BI__builtin_ia32_movnti64: { 6057 llvm::MDNode *Node = llvm::MDNode::get( 6058 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6059 6060 // Convert the type of the pointer to a pointer to the stored type. 6061 Value *BC = Builder.CreateBitCast(Ops[0], 6062 llvm::PointerType::getUnqual(Ops[1]->getType()), 6063 "cast"); 6064 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6065 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6066 6067 // If the operand is an integer, we can't assume alignment. Otherwise, 6068 // assume natural alignment. 6069 QualType ArgTy = E->getArg(1)->getType(); 6070 unsigned Align; 6071 if (ArgTy->isIntegerType()) 6072 Align = 1; 6073 else 6074 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6075 SI->setAlignment(Align); 6076 return SI; 6077 } 6078 // 3DNow! 6079 case X86::BI__builtin_ia32_pswapdsf: 6080 case X86::BI__builtin_ia32_pswapdsi: { 6081 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6082 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6083 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6084 return Builder.CreateCall(F, Ops, "pswapd"); 6085 } 6086 case X86::BI__builtin_ia32_rdrand16_step: 6087 case X86::BI__builtin_ia32_rdrand32_step: 6088 case X86::BI__builtin_ia32_rdrand64_step: 6089 case X86::BI__builtin_ia32_rdseed16_step: 6090 case X86::BI__builtin_ia32_rdseed32_step: 6091 case X86::BI__builtin_ia32_rdseed64_step: { 6092 Intrinsic::ID ID; 6093 switch (BuiltinID) { 6094 default: llvm_unreachable("Unsupported intrinsic!"); 6095 case X86::BI__builtin_ia32_rdrand16_step: 6096 ID = Intrinsic::x86_rdrand_16; 6097 break; 6098 case X86::BI__builtin_ia32_rdrand32_step: 6099 ID = Intrinsic::x86_rdrand_32; 6100 break; 6101 case X86::BI__builtin_ia32_rdrand64_step: 6102 ID = Intrinsic::x86_rdrand_64; 6103 break; 6104 case X86::BI__builtin_ia32_rdseed16_step: 6105 ID = Intrinsic::x86_rdseed_16; 6106 break; 6107 case X86::BI__builtin_ia32_rdseed32_step: 6108 ID = Intrinsic::x86_rdseed_32; 6109 break; 6110 case X86::BI__builtin_ia32_rdseed64_step: 6111 ID = Intrinsic::x86_rdseed_64; 6112 break; 6113 } 6114 6115 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6116 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6117 return Builder.CreateExtractValue(Call, 1); 6118 } 6119 // SSE comparison intrisics 6120 case X86::BI__builtin_ia32_cmpeqps: 6121 case X86::BI__builtin_ia32_cmpltps: 6122 case X86::BI__builtin_ia32_cmpleps: 6123 case X86::BI__builtin_ia32_cmpunordps: 6124 case X86::BI__builtin_ia32_cmpneqps: 6125 case X86::BI__builtin_ia32_cmpnltps: 6126 case X86::BI__builtin_ia32_cmpnleps: 6127 case X86::BI__builtin_ia32_cmpordps: 6128 case X86::BI__builtin_ia32_cmpeqss: 6129 case X86::BI__builtin_ia32_cmpltss: 6130 case X86::BI__builtin_ia32_cmpless: 6131 case X86::BI__builtin_ia32_cmpunordss: 6132 case X86::BI__builtin_ia32_cmpneqss: 6133 case X86::BI__builtin_ia32_cmpnltss: 6134 case X86::BI__builtin_ia32_cmpnless: 6135 case X86::BI__builtin_ia32_cmpordss: 6136 case X86::BI__builtin_ia32_cmpeqpd: 6137 case X86::BI__builtin_ia32_cmpltpd: 6138 case X86::BI__builtin_ia32_cmplepd: 6139 case X86::BI__builtin_ia32_cmpunordpd: 6140 case X86::BI__builtin_ia32_cmpneqpd: 6141 case X86::BI__builtin_ia32_cmpnltpd: 6142 case X86::BI__builtin_ia32_cmpnlepd: 6143 case X86::BI__builtin_ia32_cmpordpd: 6144 case X86::BI__builtin_ia32_cmpeqsd: 6145 case X86::BI__builtin_ia32_cmpltsd: 6146 case X86::BI__builtin_ia32_cmplesd: 6147 case X86::BI__builtin_ia32_cmpunordsd: 6148 case X86::BI__builtin_ia32_cmpneqsd: 6149 case X86::BI__builtin_ia32_cmpnltsd: 6150 case X86::BI__builtin_ia32_cmpnlesd: 6151 case X86::BI__builtin_ia32_cmpordsd: 6152 // These exist so that the builtin that takes an immediate can be bounds 6153 // checked by clang to avoid passing bad immediates to the backend. Since 6154 // AVX has a larger immediate than SSE we would need separate builtins to 6155 // do the different bounds checking. Rather than create a clang specific 6156 // SSE only builtin, this implements eight separate builtins to match gcc 6157 // implementation. 6158 6159 // Choose the immediate. 6160 unsigned Imm; 6161 switch (BuiltinID) { 6162 default: llvm_unreachable("Unsupported intrinsic!"); 6163 case X86::BI__builtin_ia32_cmpeqps: 6164 case X86::BI__builtin_ia32_cmpeqss: 6165 case X86::BI__builtin_ia32_cmpeqpd: 6166 case X86::BI__builtin_ia32_cmpeqsd: 6167 Imm = 0; 6168 break; 6169 case X86::BI__builtin_ia32_cmpltps: 6170 case X86::BI__builtin_ia32_cmpltss: 6171 case X86::BI__builtin_ia32_cmpltpd: 6172 case X86::BI__builtin_ia32_cmpltsd: 6173 Imm = 1; 6174 break; 6175 case X86::BI__builtin_ia32_cmpleps: 6176 case X86::BI__builtin_ia32_cmpless: 6177 case X86::BI__builtin_ia32_cmplepd: 6178 case X86::BI__builtin_ia32_cmplesd: 6179 Imm = 2; 6180 break; 6181 case X86::BI__builtin_ia32_cmpunordps: 6182 case X86::BI__builtin_ia32_cmpunordss: 6183 case X86::BI__builtin_ia32_cmpunordpd: 6184 case X86::BI__builtin_ia32_cmpunordsd: 6185 Imm = 3; 6186 break; 6187 case X86::BI__builtin_ia32_cmpneqps: 6188 case X86::BI__builtin_ia32_cmpneqss: 6189 case X86::BI__builtin_ia32_cmpneqpd: 6190 case X86::BI__builtin_ia32_cmpneqsd: 6191 Imm = 4; 6192 break; 6193 case X86::BI__builtin_ia32_cmpnltps: 6194 case X86::BI__builtin_ia32_cmpnltss: 6195 case X86::BI__builtin_ia32_cmpnltpd: 6196 case X86::BI__builtin_ia32_cmpnltsd: 6197 Imm = 5; 6198 break; 6199 case X86::BI__builtin_ia32_cmpnleps: 6200 case X86::BI__builtin_ia32_cmpnless: 6201 case X86::BI__builtin_ia32_cmpnlepd: 6202 case X86::BI__builtin_ia32_cmpnlesd: 6203 Imm = 6; 6204 break; 6205 case X86::BI__builtin_ia32_cmpordps: 6206 case X86::BI__builtin_ia32_cmpordss: 6207 case X86::BI__builtin_ia32_cmpordpd: 6208 case X86::BI__builtin_ia32_cmpordsd: 6209 Imm = 7; 6210 break; 6211 } 6212 6213 // Choose the intrinsic ID. 6214 const char *name; 6215 Intrinsic::ID ID; 6216 switch (BuiltinID) { 6217 default: llvm_unreachable("Unsupported intrinsic!"); 6218 case X86::BI__builtin_ia32_cmpeqps: 6219 case X86::BI__builtin_ia32_cmpltps: 6220 case X86::BI__builtin_ia32_cmpleps: 6221 case X86::BI__builtin_ia32_cmpunordps: 6222 case X86::BI__builtin_ia32_cmpneqps: 6223 case X86::BI__builtin_ia32_cmpnltps: 6224 case X86::BI__builtin_ia32_cmpnleps: 6225 case X86::BI__builtin_ia32_cmpordps: 6226 name = "cmpps"; 6227 ID = Intrinsic::x86_sse_cmp_ps; 6228 break; 6229 case X86::BI__builtin_ia32_cmpeqss: 6230 case X86::BI__builtin_ia32_cmpltss: 6231 case X86::BI__builtin_ia32_cmpless: 6232 case X86::BI__builtin_ia32_cmpunordss: 6233 case X86::BI__builtin_ia32_cmpneqss: 6234 case X86::BI__builtin_ia32_cmpnltss: 6235 case X86::BI__builtin_ia32_cmpnless: 6236 case X86::BI__builtin_ia32_cmpordss: 6237 name = "cmpss"; 6238 ID = Intrinsic::x86_sse_cmp_ss; 6239 break; 6240 case X86::BI__builtin_ia32_cmpeqpd: 6241 case X86::BI__builtin_ia32_cmpltpd: 6242 case X86::BI__builtin_ia32_cmplepd: 6243 case X86::BI__builtin_ia32_cmpunordpd: 6244 case X86::BI__builtin_ia32_cmpneqpd: 6245 case X86::BI__builtin_ia32_cmpnltpd: 6246 case X86::BI__builtin_ia32_cmpnlepd: 6247 case X86::BI__builtin_ia32_cmpordpd: 6248 name = "cmppd"; 6249 ID = Intrinsic::x86_sse2_cmp_pd; 6250 break; 6251 case X86::BI__builtin_ia32_cmpeqsd: 6252 case X86::BI__builtin_ia32_cmpltsd: 6253 case X86::BI__builtin_ia32_cmplesd: 6254 case X86::BI__builtin_ia32_cmpunordsd: 6255 case X86::BI__builtin_ia32_cmpneqsd: 6256 case X86::BI__builtin_ia32_cmpnltsd: 6257 case X86::BI__builtin_ia32_cmpnlesd: 6258 case X86::BI__builtin_ia32_cmpordsd: 6259 name = "cmpsd"; 6260 ID = Intrinsic::x86_sse2_cmp_sd; 6261 break; 6262 } 6263 6264 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6265 llvm::Function *F = CGM.getIntrinsic(ID); 6266 return Builder.CreateCall(F, Ops, name); 6267 } 6268 } 6269 6270 6271 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6272 const CallExpr *E) { 6273 SmallVector<Value*, 4> Ops; 6274 6275 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6276 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6277 6278 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6279 6280 switch (BuiltinID) { 6281 default: return nullptr; 6282 6283 // vec_ld, vec_lvsl, vec_lvsr 6284 case PPC::BI__builtin_altivec_lvx: 6285 case PPC::BI__builtin_altivec_lvxl: 6286 case PPC::BI__builtin_altivec_lvebx: 6287 case PPC::BI__builtin_altivec_lvehx: 6288 case PPC::BI__builtin_altivec_lvewx: 6289 case PPC::BI__builtin_altivec_lvsl: 6290 case PPC::BI__builtin_altivec_lvsr: 6291 case PPC::BI__builtin_vsx_lxvd2x: 6292 case PPC::BI__builtin_vsx_lxvw4x: 6293 { 6294 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6295 6296 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6297 Ops.pop_back(); 6298 6299 switch (BuiltinID) { 6300 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6301 case PPC::BI__builtin_altivec_lvx: 6302 ID = Intrinsic::ppc_altivec_lvx; 6303 break; 6304 case PPC::BI__builtin_altivec_lvxl: 6305 ID = Intrinsic::ppc_altivec_lvxl; 6306 break; 6307 case PPC::BI__builtin_altivec_lvebx: 6308 ID = Intrinsic::ppc_altivec_lvebx; 6309 break; 6310 case PPC::BI__builtin_altivec_lvehx: 6311 ID = Intrinsic::ppc_altivec_lvehx; 6312 break; 6313 case PPC::BI__builtin_altivec_lvewx: 6314 ID = Intrinsic::ppc_altivec_lvewx; 6315 break; 6316 case PPC::BI__builtin_altivec_lvsl: 6317 ID = Intrinsic::ppc_altivec_lvsl; 6318 break; 6319 case PPC::BI__builtin_altivec_lvsr: 6320 ID = Intrinsic::ppc_altivec_lvsr; 6321 break; 6322 case PPC::BI__builtin_vsx_lxvd2x: 6323 ID = Intrinsic::ppc_vsx_lxvd2x; 6324 break; 6325 case PPC::BI__builtin_vsx_lxvw4x: 6326 ID = Intrinsic::ppc_vsx_lxvw4x; 6327 break; 6328 } 6329 llvm::Function *F = CGM.getIntrinsic(ID); 6330 return Builder.CreateCall(F, Ops, ""); 6331 } 6332 6333 // vec_st 6334 case PPC::BI__builtin_altivec_stvx: 6335 case PPC::BI__builtin_altivec_stvxl: 6336 case PPC::BI__builtin_altivec_stvebx: 6337 case PPC::BI__builtin_altivec_stvehx: 6338 case PPC::BI__builtin_altivec_stvewx: 6339 case PPC::BI__builtin_vsx_stxvd2x: 6340 case PPC::BI__builtin_vsx_stxvw4x: 6341 { 6342 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6343 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6344 Ops.pop_back(); 6345 6346 switch (BuiltinID) { 6347 default: llvm_unreachable("Unsupported st intrinsic!"); 6348 case PPC::BI__builtin_altivec_stvx: 6349 ID = Intrinsic::ppc_altivec_stvx; 6350 break; 6351 case PPC::BI__builtin_altivec_stvxl: 6352 ID = Intrinsic::ppc_altivec_stvxl; 6353 break; 6354 case PPC::BI__builtin_altivec_stvebx: 6355 ID = Intrinsic::ppc_altivec_stvebx; 6356 break; 6357 case PPC::BI__builtin_altivec_stvehx: 6358 ID = Intrinsic::ppc_altivec_stvehx; 6359 break; 6360 case PPC::BI__builtin_altivec_stvewx: 6361 ID = Intrinsic::ppc_altivec_stvewx; 6362 break; 6363 case PPC::BI__builtin_vsx_stxvd2x: 6364 ID = Intrinsic::ppc_vsx_stxvd2x; 6365 break; 6366 case PPC::BI__builtin_vsx_stxvw4x: 6367 ID = Intrinsic::ppc_vsx_stxvw4x; 6368 break; 6369 } 6370 llvm::Function *F = CGM.getIntrinsic(ID); 6371 return Builder.CreateCall(F, Ops, ""); 6372 } 6373 // P8 Crypto builtins 6374 case PPC::BI__builtin_altivec_crypto_vshasigmaw: 6375 case PPC::BI__builtin_altivec_crypto_vshasigmad: 6376 { 6377 ConstantInt *CI1 = dyn_cast<ConstantInt>(Ops[1]); 6378 ConstantInt *CI2 = dyn_cast<ConstantInt>(Ops[2]); 6379 assert(CI1 && CI2); 6380 if (CI1->getZExtValue() > 1) { 6381 CGM.Error(E->getArg(1)->getExprLoc(), 6382 "argument out of range (should be 0-1)."); 6383 return llvm::UndefValue::get(Ops[0]->getType()); 6384 } 6385 if (CI2->getZExtValue() > 15) { 6386 CGM.Error(E->getArg(2)->getExprLoc(), 6387 "argument out of range (should be 0-15)."); 6388 return llvm::UndefValue::get(Ops[0]->getType()); 6389 } 6390 switch (BuiltinID) { 6391 default: llvm_unreachable("Unsupported sigma intrinsic!"); 6392 case PPC::BI__builtin_altivec_crypto_vshasigmaw: 6393 ID = Intrinsic::ppc_altivec_crypto_vshasigmaw; 6394 break; 6395 case PPC::BI__builtin_altivec_crypto_vshasigmad: 6396 ID = Intrinsic::ppc_altivec_crypto_vshasigmad; 6397 break; 6398 } 6399 llvm::Function *F = CGM.getIntrinsic(ID); 6400 return Builder.CreateCall(F, Ops, ""); 6401 } 6402 } 6403 } 6404 6405 // Emit an intrinsic that has 1 float or double. 6406 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6407 const CallExpr *E, 6408 unsigned IntrinsicID) { 6409 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6410 6411 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6412 return CGF.Builder.CreateCall(F, Src0); 6413 } 6414 6415 // Emit an intrinsic that has 3 float or double operands. 6416 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6417 const CallExpr *E, 6418 unsigned IntrinsicID) { 6419 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6420 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6421 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6422 6423 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6424 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6425 } 6426 6427 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6428 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6429 const CallExpr *E, 6430 unsigned IntrinsicID) { 6431 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6432 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6433 6434 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6435 return CGF.Builder.CreateCall2(F, Src0, Src1); 6436 } 6437 6438 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6439 const CallExpr *E) { 6440 switch (BuiltinID) { 6441 case R600::BI__builtin_amdgpu_div_scale: 6442 case R600::BI__builtin_amdgpu_div_scalef: { 6443 // Translate from the intrinsics's struct return to the builtin's out 6444 // argument. 6445 6446 std::pair<llvm::Value *, unsigned> FlagOutPtr 6447 = EmitPointerWithAlignment(E->getArg(3)); 6448 6449 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6450 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6451 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6452 6453 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6454 X->getType()); 6455 6456 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6457 6458 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6459 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6460 6461 llvm::Type *RealFlagType 6462 = FlagOutPtr.first->getType()->getPointerElementType(); 6463 6464 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6465 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6466 FlagStore->setAlignment(FlagOutPtr.second); 6467 return Result; 6468 } 6469 case R600::BI__builtin_amdgpu_div_fmas: 6470 case R600::BI__builtin_amdgpu_div_fmasf: { 6471 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6472 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6473 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6474 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6475 6476 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6477 Src0->getType()); 6478 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6479 return Builder.CreateCall4(F, Src0, Src1, Src2, Src3ToBool); 6480 } 6481 case R600::BI__builtin_amdgpu_div_fixup: 6482 case R600::BI__builtin_amdgpu_div_fixupf: 6483 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6484 case R600::BI__builtin_amdgpu_trig_preop: 6485 case R600::BI__builtin_amdgpu_trig_preopf: 6486 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6487 case R600::BI__builtin_amdgpu_rcp: 6488 case R600::BI__builtin_amdgpu_rcpf: 6489 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6490 case R600::BI__builtin_amdgpu_rsq: 6491 case R600::BI__builtin_amdgpu_rsqf: 6492 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6493 case R600::BI__builtin_amdgpu_rsq_clamped: 6494 case R600::BI__builtin_amdgpu_rsq_clampedf: 6495 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6496 case R600::BI__builtin_amdgpu_ldexp: 6497 case R600::BI__builtin_amdgpu_ldexpf: 6498 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6499 case R600::BI__builtin_amdgpu_class: 6500 case R600::BI__builtin_amdgpu_classf: 6501 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6502 default: 6503 return nullptr; 6504 } 6505 } 6506