1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGCXXABI.h" 16 #include "CGObjCRuntime.h" 17 #include "CodeGenModule.h" 18 #include "TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "clang/Basic/TargetInfo.h" 23 #include "clang/CodeGen/CGFunctionInfo.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include <sstream> 30 31 using namespace clang; 32 using namespace CodeGen; 33 using namespace llvm; 34 35 /// getBuiltinLibFunction - Given a builtin id for a function like 36 /// "__builtin_fabsf", return a Function* for "fabsf". 37 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 38 unsigned BuiltinID) { 39 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 40 41 // Get the name, skip over the __builtin_ prefix (if necessary). 42 StringRef Name; 43 GlobalDecl D(FD); 44 45 // If the builtin has been declared explicitly with an assembler label, 46 // use the mangled name. This differs from the plain label on platforms 47 // that prefix labels. 48 if (FD->hasAttr<AsmLabelAttr>()) 49 Name = getMangledName(D); 50 else 51 Name = Context.BuiltinInfo.getName(BuiltinID) + 10; 52 53 llvm::FunctionType *Ty = 54 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 55 56 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 57 } 58 59 /// Emit the conversions required to turn the given value into an 60 /// integer of the given size. 61 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 62 QualType T, llvm::IntegerType *IntType) { 63 V = CGF.EmitToMemory(V, T); 64 65 if (V->getType()->isPointerTy()) 66 return CGF.Builder.CreatePtrToInt(V, IntType); 67 68 assert(V->getType() == IntType); 69 return V; 70 } 71 72 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 73 QualType T, llvm::Type *ResultType) { 74 V = CGF.EmitFromMemory(V, T); 75 76 if (ResultType->isPointerTy()) 77 return CGF.Builder.CreateIntToPtr(V, ResultType); 78 79 assert(V->getType() == ResultType); 80 return V; 81 } 82 83 /// Utility to insert an atomic instruction based on Instrinsic::ID 84 /// and the expression node. 85 static Value *MakeBinaryAtomicValue(CodeGenFunction &CGF, 86 llvm::AtomicRMWInst::BinOp Kind, 87 const CallExpr *E) { 88 QualType T = E->getType(); 89 assert(E->getArg(0)->getType()->isPointerType()); 90 assert(CGF.getContext().hasSameUnqualifiedType(T, 91 E->getArg(0)->getType()->getPointeeType())); 92 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 93 94 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 95 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 96 97 llvm::IntegerType *IntType = 98 llvm::IntegerType::get(CGF.getLLVMContext(), 99 CGF.getContext().getTypeSize(T)); 100 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 101 102 llvm::Value *Args[2]; 103 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 104 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 105 llvm::Type *ValueType = Args[1]->getType(); 106 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 107 108 llvm::Value *Result = 109 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 110 llvm::SequentiallyConsistent); 111 return EmitFromInt(CGF, Result, T, ValueType); 112 } 113 114 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 115 llvm::AtomicRMWInst::BinOp Kind, 116 const CallExpr *E) { 117 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); 118 } 119 120 /// Utility to insert an atomic instruction based Instrinsic::ID and 121 /// the expression node, where the return value is the result of the 122 /// operation. 123 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 124 llvm::AtomicRMWInst::BinOp Kind, 125 const CallExpr *E, 126 Instruction::BinaryOps Op, 127 bool Invert = false) { 128 QualType T = E->getType(); 129 assert(E->getArg(0)->getType()->isPointerType()); 130 assert(CGF.getContext().hasSameUnqualifiedType(T, 131 E->getArg(0)->getType()->getPointeeType())); 132 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 133 134 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 135 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 136 137 llvm::IntegerType *IntType = 138 llvm::IntegerType::get(CGF.getLLVMContext(), 139 CGF.getContext().getTypeSize(T)); 140 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 141 142 llvm::Value *Args[2]; 143 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 144 llvm::Type *ValueType = Args[1]->getType(); 145 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 146 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 147 148 llvm::Value *Result = 149 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 150 llvm::SequentiallyConsistent); 151 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 152 if (Invert) 153 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 154 llvm::ConstantInt::get(IntType, -1)); 155 Result = EmitFromInt(CGF, Result, T, ValueType); 156 return RValue::get(Result); 157 } 158 159 /// @brief Utility to insert an atomic cmpxchg instruction. 160 /// 161 /// @param CGF The current codegen function. 162 /// @param E Builtin call expression to convert to cmpxchg. 163 /// arg0 - address to operate on 164 /// arg1 - value to compare with 165 /// arg2 - new value 166 /// @param ReturnBool Specifies whether to return success flag of 167 /// cmpxchg result or the old value. 168 /// 169 /// @returns result of cmpxchg, according to ReturnBool 170 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, 171 bool ReturnBool) { 172 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); 173 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 174 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 175 176 llvm::IntegerType *IntType = llvm::IntegerType::get( 177 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); 178 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 179 180 Value *Args[3]; 181 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 182 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 183 llvm::Type *ValueType = Args[1]->getType(); 184 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 185 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); 186 187 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 188 llvm::SequentiallyConsistent, 189 llvm::SequentiallyConsistent); 190 if (ReturnBool) 191 // Extract boolean success flag and zext it to int. 192 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), 193 CGF.ConvertType(E->getType())); 194 else 195 // Extract old value and emit it using the same type as compare value. 196 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, 197 ValueType); 198 } 199 200 /// EmitFAbs - Emit a call to @llvm.fabs(). 201 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 202 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 203 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 204 Call->setDoesNotAccessMemory(); 205 return Call; 206 } 207 208 /// Emit the computation of the sign bit for a floating point value. Returns 209 /// the i1 sign bit value. 210 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 211 LLVMContext &C = CGF.CGM.getLLVMContext(); 212 213 llvm::Type *Ty = V->getType(); 214 int Width = Ty->getPrimitiveSizeInBits(); 215 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 216 V = CGF.Builder.CreateBitCast(V, IntTy); 217 if (Ty->isPPC_FP128Ty()) { 218 // The higher-order double comes first, and so we need to truncate the 219 // pair to extract the overall sign. The order of the pair is the same 220 // in both little- and big-Endian modes. 221 Width >>= 1; 222 IntTy = llvm::IntegerType::get(C, Width); 223 V = CGF.Builder.CreateTrunc(V, IntTy); 224 } 225 Value *Zero = llvm::Constant::getNullValue(IntTy); 226 return CGF.Builder.CreateICmpSLT(V, Zero); 227 } 228 229 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 230 const CallExpr *E, llvm::Value *calleeValue) { 231 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 232 ReturnValueSlot(), Fn); 233 } 234 235 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 236 /// depending on IntrinsicID. 237 /// 238 /// \arg CGF The current codegen function. 239 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 240 /// \arg X The first argument to the llvm.*.with.overflow.*. 241 /// \arg Y The second argument to the llvm.*.with.overflow.*. 242 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 243 /// \returns The result (i.e. sum/product) returned by the intrinsic. 244 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 245 const llvm::Intrinsic::ID IntrinsicID, 246 llvm::Value *X, llvm::Value *Y, 247 llvm::Value *&Carry) { 248 // Make sure we have integers of the same width. 249 assert(X->getType() == Y->getType() && 250 "Arguments must be the same type. (Did you forget to make sure both " 251 "arguments have the same integer width?)"); 252 253 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 254 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 255 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 256 return CGF.Builder.CreateExtractValue(Tmp, 0); 257 } 258 259 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 260 unsigned BuiltinID, const CallExpr *E, 261 ReturnValueSlot ReturnValue) { 262 // See if we can constant fold this builtin. If so, don't emit it at all. 263 Expr::EvalResult Result; 264 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 265 !Result.hasSideEffects()) { 266 if (Result.Val.isInt()) 267 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 268 Result.Val.getInt())); 269 if (Result.Val.isFloat()) 270 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 271 Result.Val.getFloat())); 272 } 273 274 switch (BuiltinID) { 275 default: break; // Handle intrinsics and libm functions below. 276 case Builtin::BI__builtin___CFStringMakeConstantString: 277 case Builtin::BI__builtin___NSStringMakeConstantString: 278 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 279 case Builtin::BI__builtin_stdarg_start: 280 case Builtin::BI__builtin_va_start: 281 case Builtin::BI__va_start: 282 case Builtin::BI__builtin_va_end: { 283 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 284 ? EmitScalarExpr(E->getArg(0)) 285 : EmitVAListRef(E->getArg(0)); 286 llvm::Type *DestType = Int8PtrTy; 287 if (ArgValue->getType() != DestType) 288 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 289 ArgValue->getName().data()); 290 291 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 292 Intrinsic::vaend : Intrinsic::vastart; 293 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 294 } 295 case Builtin::BI__builtin_va_copy: { 296 Value *DstPtr = EmitVAListRef(E->getArg(0)); 297 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 298 299 llvm::Type *Type = Int8PtrTy; 300 301 DstPtr = Builder.CreateBitCast(DstPtr, Type); 302 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 303 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 304 {DstPtr, SrcPtr})); 305 } 306 case Builtin::BI__builtin_abs: 307 case Builtin::BI__builtin_labs: 308 case Builtin::BI__builtin_llabs: { 309 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 310 311 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 312 Value *CmpResult = 313 Builder.CreateICmpSGE(ArgValue, 314 llvm::Constant::getNullValue(ArgValue->getType()), 315 "abscond"); 316 Value *Result = 317 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 318 319 return RValue::get(Result); 320 } 321 case Builtin::BI__builtin_fabs: 322 case Builtin::BI__builtin_fabsf: 323 case Builtin::BI__builtin_fabsl: { 324 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 325 Value *Result = EmitFAbs(*this, Arg1); 326 return RValue::get(Result); 327 } 328 case Builtin::BI__builtin_fmod: 329 case Builtin::BI__builtin_fmodf: 330 case Builtin::BI__builtin_fmodl: { 331 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 332 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 333 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 334 return RValue::get(Result); 335 } 336 337 case Builtin::BI__builtin_conj: 338 case Builtin::BI__builtin_conjf: 339 case Builtin::BI__builtin_conjl: { 340 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 341 Value *Real = ComplexVal.first; 342 Value *Imag = ComplexVal.second; 343 Value *Zero = 344 Imag->getType()->isFPOrFPVectorTy() 345 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 346 : llvm::Constant::getNullValue(Imag->getType()); 347 348 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 349 return RValue::getComplex(std::make_pair(Real, Imag)); 350 } 351 case Builtin::BI__builtin_creal: 352 case Builtin::BI__builtin_crealf: 353 case Builtin::BI__builtin_creall: 354 case Builtin::BIcreal: 355 case Builtin::BIcrealf: 356 case Builtin::BIcreall: { 357 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 358 return RValue::get(ComplexVal.first); 359 } 360 361 case Builtin::BI__builtin_cimag: 362 case Builtin::BI__builtin_cimagf: 363 case Builtin::BI__builtin_cimagl: 364 case Builtin::BIcimag: 365 case Builtin::BIcimagf: 366 case Builtin::BIcimagl: { 367 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 368 return RValue::get(ComplexVal.second); 369 } 370 371 case Builtin::BI__builtin_ctzs: 372 case Builtin::BI__builtin_ctz: 373 case Builtin::BI__builtin_ctzl: 374 case Builtin::BI__builtin_ctzll: { 375 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 376 377 llvm::Type *ArgType = ArgValue->getType(); 378 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 379 380 llvm::Type *ResultType = ConvertType(E->getType()); 381 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 382 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 383 if (Result->getType() != ResultType) 384 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 385 "cast"); 386 return RValue::get(Result); 387 } 388 case Builtin::BI__builtin_clzs: 389 case Builtin::BI__builtin_clz: 390 case Builtin::BI__builtin_clzl: 391 case Builtin::BI__builtin_clzll: { 392 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 393 394 llvm::Type *ArgType = ArgValue->getType(); 395 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 396 397 llvm::Type *ResultType = ConvertType(E->getType()); 398 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 399 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 400 if (Result->getType() != ResultType) 401 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 402 "cast"); 403 return RValue::get(Result); 404 } 405 case Builtin::BI__builtin_ffs: 406 case Builtin::BI__builtin_ffsl: 407 case Builtin::BI__builtin_ffsll: { 408 // ffs(x) -> x ? cttz(x) + 1 : 0 409 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 410 411 llvm::Type *ArgType = ArgValue->getType(); 412 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 413 414 llvm::Type *ResultType = ConvertType(E->getType()); 415 Value *Tmp = 416 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 417 llvm::ConstantInt::get(ArgType, 1)); 418 Value *Zero = llvm::Constant::getNullValue(ArgType); 419 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 420 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 421 if (Result->getType() != ResultType) 422 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 423 "cast"); 424 return RValue::get(Result); 425 } 426 case Builtin::BI__builtin_parity: 427 case Builtin::BI__builtin_parityl: 428 case Builtin::BI__builtin_parityll: { 429 // parity(x) -> ctpop(x) & 1 430 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 431 432 llvm::Type *ArgType = ArgValue->getType(); 433 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 434 435 llvm::Type *ResultType = ConvertType(E->getType()); 436 Value *Tmp = Builder.CreateCall(F, ArgValue); 437 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 438 if (Result->getType() != ResultType) 439 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 440 "cast"); 441 return RValue::get(Result); 442 } 443 case Builtin::BI__builtin_popcount: 444 case Builtin::BI__builtin_popcountl: 445 case Builtin::BI__builtin_popcountll: { 446 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 447 448 llvm::Type *ArgType = ArgValue->getType(); 449 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 450 451 llvm::Type *ResultType = ConvertType(E->getType()); 452 Value *Result = Builder.CreateCall(F, ArgValue); 453 if (Result->getType() != ResultType) 454 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 455 "cast"); 456 return RValue::get(Result); 457 } 458 case Builtin::BI__builtin_expect: { 459 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 460 llvm::Type *ArgType = ArgValue->getType(); 461 462 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 463 // Don't generate llvm.expect on -O0 as the backend won't use it for 464 // anything. 465 // Note, we still IRGen ExpectedValue because it could have side-effects. 466 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 467 return RValue::get(ArgValue); 468 469 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 470 Value *Result = 471 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 472 return RValue::get(Result); 473 } 474 case Builtin::BI__builtin_assume_aligned: { 475 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 476 Value *OffsetValue = 477 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 478 479 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 480 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 481 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 482 483 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 484 return RValue::get(PtrValue); 485 } 486 case Builtin::BI__assume: 487 case Builtin::BI__builtin_assume: { 488 if (E->getArg(0)->HasSideEffects(getContext())) 489 return RValue::get(nullptr); 490 491 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 492 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 493 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 494 } 495 case Builtin::BI__builtin_bswap16: 496 case Builtin::BI__builtin_bswap32: 497 case Builtin::BI__builtin_bswap64: { 498 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 499 llvm::Type *ArgType = ArgValue->getType(); 500 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 501 return RValue::get(Builder.CreateCall(F, ArgValue)); 502 } 503 case Builtin::BI__builtin_object_size: { 504 // We rely on constant folding to deal with expressions with side effects. 505 assert(!E->getArg(0)->HasSideEffects(getContext()) && 506 "should have been constant folded"); 507 508 // We pass this builtin onto the optimizer so that it can 509 // figure out the object size in more complex cases. 510 llvm::Type *ResType = ConvertType(E->getType()); 511 512 // LLVM only supports 0 and 2, make sure that we pass along that 513 // as a boolean. 514 Value *Ty = EmitScalarExpr(E->getArg(1)); 515 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 516 assert(CI); 517 uint64_t val = CI->getZExtValue(); 518 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 519 // FIXME: Get right address space. 520 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 521 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 522 return RValue::get( 523 Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0)), CI})); 524 } 525 case Builtin::BI__builtin_prefetch: { 526 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 527 // FIXME: Technically these constants should of type 'int', yes? 528 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 529 llvm::ConstantInt::get(Int32Ty, 0); 530 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 531 llvm::ConstantInt::get(Int32Ty, 3); 532 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 533 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 534 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 535 } 536 case Builtin::BI__builtin_readcyclecounter: { 537 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 538 return RValue::get(Builder.CreateCall(F)); 539 } 540 case Builtin::BI__builtin___clear_cache: { 541 Value *Begin = EmitScalarExpr(E->getArg(0)); 542 Value *End = EmitScalarExpr(E->getArg(1)); 543 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 544 return RValue::get(Builder.CreateCall(F, {Begin, End})); 545 } 546 case Builtin::BI__builtin_trap: 547 return RValue::get(EmitTrapCall(Intrinsic::trap)); 548 case Builtin::BI__debugbreak: 549 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 550 case Builtin::BI__builtin_unreachable: { 551 if (SanOpts.has(SanitizerKind::Unreachable)) { 552 SanitizerScope SanScope(this); 553 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 554 SanitizerKind::Unreachable), 555 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 556 None); 557 } else 558 Builder.CreateUnreachable(); 559 560 // We do need to preserve an insertion point. 561 EmitBlock(createBasicBlock("unreachable.cont")); 562 563 return RValue::get(nullptr); 564 } 565 566 case Builtin::BI__builtin_powi: 567 case Builtin::BI__builtin_powif: 568 case Builtin::BI__builtin_powil: { 569 Value *Base = EmitScalarExpr(E->getArg(0)); 570 Value *Exponent = EmitScalarExpr(E->getArg(1)); 571 llvm::Type *ArgType = Base->getType(); 572 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 573 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 574 } 575 576 case Builtin::BI__builtin_isgreater: 577 case Builtin::BI__builtin_isgreaterequal: 578 case Builtin::BI__builtin_isless: 579 case Builtin::BI__builtin_islessequal: 580 case Builtin::BI__builtin_islessgreater: 581 case Builtin::BI__builtin_isunordered: { 582 // Ordered comparisons: we know the arguments to these are matching scalar 583 // floating point values. 584 Value *LHS = EmitScalarExpr(E->getArg(0)); 585 Value *RHS = EmitScalarExpr(E->getArg(1)); 586 587 switch (BuiltinID) { 588 default: llvm_unreachable("Unknown ordered comparison"); 589 case Builtin::BI__builtin_isgreater: 590 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 591 break; 592 case Builtin::BI__builtin_isgreaterequal: 593 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 594 break; 595 case Builtin::BI__builtin_isless: 596 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 597 break; 598 case Builtin::BI__builtin_islessequal: 599 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 600 break; 601 case Builtin::BI__builtin_islessgreater: 602 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 603 break; 604 case Builtin::BI__builtin_isunordered: 605 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 606 break; 607 } 608 // ZExt bool to int type. 609 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 610 } 611 case Builtin::BI__builtin_isnan: { 612 Value *V = EmitScalarExpr(E->getArg(0)); 613 V = Builder.CreateFCmpUNO(V, V, "cmp"); 614 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 615 } 616 617 case Builtin::BI__builtin_isinf: { 618 // isinf(x) --> fabs(x) == infinity 619 Value *V = EmitScalarExpr(E->getArg(0)); 620 V = EmitFAbs(*this, V); 621 622 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 623 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 624 } 625 626 case Builtin::BI__builtin_isinf_sign: { 627 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 628 Value *Arg = EmitScalarExpr(E->getArg(0)); 629 Value *AbsArg = EmitFAbs(*this, Arg); 630 Value *IsInf = Builder.CreateFCmpOEQ( 631 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 632 Value *IsNeg = EmitSignBit(*this, Arg); 633 634 llvm::Type *IntTy = ConvertType(E->getType()); 635 Value *Zero = Constant::getNullValue(IntTy); 636 Value *One = ConstantInt::get(IntTy, 1); 637 Value *NegativeOne = ConstantInt::get(IntTy, -1); 638 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 639 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 640 return RValue::get(Result); 641 } 642 643 case Builtin::BI__builtin_isnormal: { 644 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 645 Value *V = EmitScalarExpr(E->getArg(0)); 646 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 647 648 Value *Abs = EmitFAbs(*this, V); 649 Value *IsLessThanInf = 650 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 651 APFloat Smallest = APFloat::getSmallestNormalized( 652 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 653 Value *IsNormal = 654 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 655 "isnormal"); 656 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 657 V = Builder.CreateAnd(V, IsNormal, "and"); 658 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 659 } 660 661 case Builtin::BI__builtin_isfinite: { 662 // isfinite(x) --> x == x && fabs(x) != infinity; 663 Value *V = EmitScalarExpr(E->getArg(0)); 664 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 665 666 Value *Abs = EmitFAbs(*this, V); 667 Value *IsNotInf = 668 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 669 670 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 671 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 672 } 673 674 case Builtin::BI__builtin_fpclassify: { 675 Value *V = EmitScalarExpr(E->getArg(5)); 676 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 677 678 // Create Result 679 BasicBlock *Begin = Builder.GetInsertBlock(); 680 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 681 Builder.SetInsertPoint(End); 682 PHINode *Result = 683 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 684 "fpclassify_result"); 685 686 // if (V==0) return FP_ZERO 687 Builder.SetInsertPoint(Begin); 688 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 689 "iszero"); 690 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 691 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 692 Builder.CreateCondBr(IsZero, End, NotZero); 693 Result->addIncoming(ZeroLiteral, Begin); 694 695 // if (V != V) return FP_NAN 696 Builder.SetInsertPoint(NotZero); 697 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 698 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 699 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 700 Builder.CreateCondBr(IsNan, End, NotNan); 701 Result->addIncoming(NanLiteral, NotZero); 702 703 // if (fabs(V) == infinity) return FP_INFINITY 704 Builder.SetInsertPoint(NotNan); 705 Value *VAbs = EmitFAbs(*this, V); 706 Value *IsInf = 707 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 708 "isinf"); 709 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 710 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 711 Builder.CreateCondBr(IsInf, End, NotInf); 712 Result->addIncoming(InfLiteral, NotNan); 713 714 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 715 Builder.SetInsertPoint(NotInf); 716 APFloat Smallest = APFloat::getSmallestNormalized( 717 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 718 Value *IsNormal = 719 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 720 "isnormal"); 721 Value *NormalResult = 722 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 723 EmitScalarExpr(E->getArg(3))); 724 Builder.CreateBr(End); 725 Result->addIncoming(NormalResult, NotInf); 726 727 // return Result 728 Builder.SetInsertPoint(End); 729 return RValue::get(Result); 730 } 731 732 case Builtin::BIalloca: 733 case Builtin::BI_alloca: 734 case Builtin::BI__builtin_alloca: { 735 Value *Size = EmitScalarExpr(E->getArg(0)); 736 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 737 } 738 case Builtin::BIbzero: 739 case Builtin::BI__builtin_bzero: { 740 std::pair<llvm::Value*, unsigned> Dest = 741 EmitPointerWithAlignment(E->getArg(0)); 742 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 743 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 744 E->getArg(0)->getExprLoc(), FD, 0); 745 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 746 Dest.second, false); 747 return RValue::get(Dest.first); 748 } 749 case Builtin::BImemcpy: 750 case Builtin::BI__builtin_memcpy: { 751 std::pair<llvm::Value*, unsigned> Dest = 752 EmitPointerWithAlignment(E->getArg(0)); 753 std::pair<llvm::Value*, unsigned> Src = 754 EmitPointerWithAlignment(E->getArg(1)); 755 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 756 unsigned Align = std::min(Dest.second, Src.second); 757 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 758 E->getArg(0)->getExprLoc(), FD, 0); 759 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 760 E->getArg(1)->getExprLoc(), FD, 1); 761 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 762 return RValue::get(Dest.first); 763 } 764 765 case Builtin::BI__builtin___memcpy_chk: { 766 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 767 llvm::APSInt Size, DstSize; 768 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 769 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 770 break; 771 if (Size.ugt(DstSize)) 772 break; 773 std::pair<llvm::Value*, unsigned> Dest = 774 EmitPointerWithAlignment(E->getArg(0)); 775 std::pair<llvm::Value*, unsigned> Src = 776 EmitPointerWithAlignment(E->getArg(1)); 777 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 778 unsigned Align = std::min(Dest.second, Src.second); 779 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 780 return RValue::get(Dest.first); 781 } 782 783 case Builtin::BI__builtin_objc_memmove_collectable: { 784 Value *Address = EmitScalarExpr(E->getArg(0)); 785 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 786 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 787 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 788 Address, SrcAddr, SizeVal); 789 return RValue::get(Address); 790 } 791 792 case Builtin::BI__builtin___memmove_chk: { 793 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 794 llvm::APSInt Size, DstSize; 795 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 796 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 797 break; 798 if (Size.ugt(DstSize)) 799 break; 800 std::pair<llvm::Value*, unsigned> Dest = 801 EmitPointerWithAlignment(E->getArg(0)); 802 std::pair<llvm::Value*, unsigned> Src = 803 EmitPointerWithAlignment(E->getArg(1)); 804 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 805 unsigned Align = std::min(Dest.second, Src.second); 806 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 807 return RValue::get(Dest.first); 808 } 809 810 case Builtin::BImemmove: 811 case Builtin::BI__builtin_memmove: { 812 std::pair<llvm::Value*, unsigned> Dest = 813 EmitPointerWithAlignment(E->getArg(0)); 814 std::pair<llvm::Value*, unsigned> Src = 815 EmitPointerWithAlignment(E->getArg(1)); 816 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 817 unsigned Align = std::min(Dest.second, Src.second); 818 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 819 E->getArg(0)->getExprLoc(), FD, 0); 820 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 821 E->getArg(1)->getExprLoc(), FD, 1); 822 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 823 return RValue::get(Dest.first); 824 } 825 case Builtin::BImemset: 826 case Builtin::BI__builtin_memset: { 827 std::pair<llvm::Value*, unsigned> Dest = 828 EmitPointerWithAlignment(E->getArg(0)); 829 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 830 Builder.getInt8Ty()); 831 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 832 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 833 E->getArg(0)->getExprLoc(), FD, 0); 834 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 835 return RValue::get(Dest.first); 836 } 837 case Builtin::BI__builtin___memset_chk: { 838 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 839 llvm::APSInt Size, DstSize; 840 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 841 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 842 break; 843 if (Size.ugt(DstSize)) 844 break; 845 std::pair<llvm::Value*, unsigned> Dest = 846 EmitPointerWithAlignment(E->getArg(0)); 847 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 848 Builder.getInt8Ty()); 849 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 850 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 851 return RValue::get(Dest.first); 852 } 853 case Builtin::BI__builtin_dwarf_cfa: { 854 // The offset in bytes from the first argument to the CFA. 855 // 856 // Why on earth is this in the frontend? Is there any reason at 857 // all that the backend can't reasonably determine this while 858 // lowering llvm.eh.dwarf.cfa()? 859 // 860 // TODO: If there's a satisfactory reason, add a target hook for 861 // this instead of hard-coding 0, which is correct for most targets. 862 int32_t Offset = 0; 863 864 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 865 return RValue::get(Builder.CreateCall(F, 866 llvm::ConstantInt::get(Int32Ty, Offset))); 867 } 868 case Builtin::BI__builtin_return_address: { 869 Value *Depth = 870 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 871 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 872 return RValue::get(Builder.CreateCall(F, Depth)); 873 } 874 case Builtin::BI__builtin_frame_address: { 875 Value *Depth = 876 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 877 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 878 return RValue::get(Builder.CreateCall(F, Depth)); 879 } 880 case Builtin::BI__builtin_extract_return_addr: { 881 Value *Address = EmitScalarExpr(E->getArg(0)); 882 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 883 return RValue::get(Result); 884 } 885 case Builtin::BI__builtin_frob_return_addr: { 886 Value *Address = EmitScalarExpr(E->getArg(0)); 887 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 888 return RValue::get(Result); 889 } 890 case Builtin::BI__builtin_dwarf_sp_column: { 891 llvm::IntegerType *Ty 892 = cast<llvm::IntegerType>(ConvertType(E->getType())); 893 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 894 if (Column == -1) { 895 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 896 return RValue::get(llvm::UndefValue::get(Ty)); 897 } 898 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 899 } 900 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 901 Value *Address = EmitScalarExpr(E->getArg(0)); 902 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 903 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 904 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 905 } 906 case Builtin::BI__builtin_eh_return: { 907 Value *Int = EmitScalarExpr(E->getArg(0)); 908 Value *Ptr = EmitScalarExpr(E->getArg(1)); 909 910 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 911 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 912 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 913 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 914 ? Intrinsic::eh_return_i32 915 : Intrinsic::eh_return_i64); 916 Builder.CreateCall(F, {Int, Ptr}); 917 Builder.CreateUnreachable(); 918 919 // We do need to preserve an insertion point. 920 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 921 922 return RValue::get(nullptr); 923 } 924 case Builtin::BI__builtin_unwind_init: { 925 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 926 return RValue::get(Builder.CreateCall(F)); 927 } 928 case Builtin::BI__builtin_extend_pointer: { 929 // Extends a pointer to the size of an _Unwind_Word, which is 930 // uint64_t on all platforms. Generally this gets poked into a 931 // register and eventually used as an address, so if the 932 // addressing registers are wider than pointers and the platform 933 // doesn't implicitly ignore high-order bits when doing 934 // addressing, we need to make sure we zext / sext based on 935 // the platform's expectations. 936 // 937 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 938 939 // Cast the pointer to intptr_t. 940 Value *Ptr = EmitScalarExpr(E->getArg(0)); 941 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 942 943 // If that's 64 bits, we're done. 944 if (IntPtrTy->getBitWidth() == 64) 945 return RValue::get(Result); 946 947 // Otherwise, ask the codegen data what to do. 948 if (getTargetHooks().extendPointerWithSExt()) 949 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 950 else 951 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 952 } 953 case Builtin::BI__builtin_setjmp: { 954 // Buffer is a void**. 955 Value *Buf = EmitScalarExpr(E->getArg(0)); 956 957 // Store the frame pointer to the setjmp buffer. 958 Value *FrameAddr = 959 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 960 ConstantInt::get(Int32Ty, 0)); 961 Builder.CreateStore(FrameAddr, Buf); 962 963 // Store the stack pointer to the setjmp buffer. 964 Value *StackAddr = 965 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 966 Value *StackSaveSlot = 967 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 968 Builder.CreateStore(StackAddr, StackSaveSlot); 969 970 // Call LLVM's EH setjmp, which is lightweight. 971 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 972 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 973 return RValue::get(Builder.CreateCall(F, Buf)); 974 } 975 case Builtin::BI__builtin_longjmp: { 976 Value *Buf = EmitScalarExpr(E->getArg(0)); 977 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 978 979 // Call LLVM's EH longjmp, which is lightweight. 980 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 981 982 // longjmp doesn't return; mark this as unreachable. 983 Builder.CreateUnreachable(); 984 985 // We do need to preserve an insertion point. 986 EmitBlock(createBasicBlock("longjmp.cont")); 987 988 return RValue::get(nullptr); 989 } 990 case Builtin::BI__sync_fetch_and_add: 991 case Builtin::BI__sync_fetch_and_sub: 992 case Builtin::BI__sync_fetch_and_or: 993 case Builtin::BI__sync_fetch_and_and: 994 case Builtin::BI__sync_fetch_and_xor: 995 case Builtin::BI__sync_fetch_and_nand: 996 case Builtin::BI__sync_add_and_fetch: 997 case Builtin::BI__sync_sub_and_fetch: 998 case Builtin::BI__sync_and_and_fetch: 999 case Builtin::BI__sync_or_and_fetch: 1000 case Builtin::BI__sync_xor_and_fetch: 1001 case Builtin::BI__sync_nand_and_fetch: 1002 case Builtin::BI__sync_val_compare_and_swap: 1003 case Builtin::BI__sync_bool_compare_and_swap: 1004 case Builtin::BI__sync_lock_test_and_set: 1005 case Builtin::BI__sync_lock_release: 1006 case Builtin::BI__sync_swap: 1007 llvm_unreachable("Shouldn't make it through sema"); 1008 case Builtin::BI__sync_fetch_and_add_1: 1009 case Builtin::BI__sync_fetch_and_add_2: 1010 case Builtin::BI__sync_fetch_and_add_4: 1011 case Builtin::BI__sync_fetch_and_add_8: 1012 case Builtin::BI__sync_fetch_and_add_16: 1013 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1014 case Builtin::BI__sync_fetch_and_sub_1: 1015 case Builtin::BI__sync_fetch_and_sub_2: 1016 case Builtin::BI__sync_fetch_and_sub_4: 1017 case Builtin::BI__sync_fetch_and_sub_8: 1018 case Builtin::BI__sync_fetch_and_sub_16: 1019 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1020 case Builtin::BI__sync_fetch_and_or_1: 1021 case Builtin::BI__sync_fetch_and_or_2: 1022 case Builtin::BI__sync_fetch_and_or_4: 1023 case Builtin::BI__sync_fetch_and_or_8: 1024 case Builtin::BI__sync_fetch_and_or_16: 1025 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1026 case Builtin::BI__sync_fetch_and_and_1: 1027 case Builtin::BI__sync_fetch_and_and_2: 1028 case Builtin::BI__sync_fetch_and_and_4: 1029 case Builtin::BI__sync_fetch_and_and_8: 1030 case Builtin::BI__sync_fetch_and_and_16: 1031 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1032 case Builtin::BI__sync_fetch_and_xor_1: 1033 case Builtin::BI__sync_fetch_and_xor_2: 1034 case Builtin::BI__sync_fetch_and_xor_4: 1035 case Builtin::BI__sync_fetch_and_xor_8: 1036 case Builtin::BI__sync_fetch_and_xor_16: 1037 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1038 case Builtin::BI__sync_fetch_and_nand_1: 1039 case Builtin::BI__sync_fetch_and_nand_2: 1040 case Builtin::BI__sync_fetch_and_nand_4: 1041 case Builtin::BI__sync_fetch_and_nand_8: 1042 case Builtin::BI__sync_fetch_and_nand_16: 1043 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1044 1045 // Clang extensions: not overloaded yet. 1046 case Builtin::BI__sync_fetch_and_min: 1047 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1048 case Builtin::BI__sync_fetch_and_max: 1049 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1050 case Builtin::BI__sync_fetch_and_umin: 1051 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1052 case Builtin::BI__sync_fetch_and_umax: 1053 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1054 1055 case Builtin::BI__sync_add_and_fetch_1: 1056 case Builtin::BI__sync_add_and_fetch_2: 1057 case Builtin::BI__sync_add_and_fetch_4: 1058 case Builtin::BI__sync_add_and_fetch_8: 1059 case Builtin::BI__sync_add_and_fetch_16: 1060 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1061 llvm::Instruction::Add); 1062 case Builtin::BI__sync_sub_and_fetch_1: 1063 case Builtin::BI__sync_sub_and_fetch_2: 1064 case Builtin::BI__sync_sub_and_fetch_4: 1065 case Builtin::BI__sync_sub_and_fetch_8: 1066 case Builtin::BI__sync_sub_and_fetch_16: 1067 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1068 llvm::Instruction::Sub); 1069 case Builtin::BI__sync_and_and_fetch_1: 1070 case Builtin::BI__sync_and_and_fetch_2: 1071 case Builtin::BI__sync_and_and_fetch_4: 1072 case Builtin::BI__sync_and_and_fetch_8: 1073 case Builtin::BI__sync_and_and_fetch_16: 1074 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1075 llvm::Instruction::And); 1076 case Builtin::BI__sync_or_and_fetch_1: 1077 case Builtin::BI__sync_or_and_fetch_2: 1078 case Builtin::BI__sync_or_and_fetch_4: 1079 case Builtin::BI__sync_or_and_fetch_8: 1080 case Builtin::BI__sync_or_and_fetch_16: 1081 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1082 llvm::Instruction::Or); 1083 case Builtin::BI__sync_xor_and_fetch_1: 1084 case Builtin::BI__sync_xor_and_fetch_2: 1085 case Builtin::BI__sync_xor_and_fetch_4: 1086 case Builtin::BI__sync_xor_and_fetch_8: 1087 case Builtin::BI__sync_xor_and_fetch_16: 1088 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1089 llvm::Instruction::Xor); 1090 case Builtin::BI__sync_nand_and_fetch_1: 1091 case Builtin::BI__sync_nand_and_fetch_2: 1092 case Builtin::BI__sync_nand_and_fetch_4: 1093 case Builtin::BI__sync_nand_and_fetch_8: 1094 case Builtin::BI__sync_nand_and_fetch_16: 1095 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1096 llvm::Instruction::And, true); 1097 1098 case Builtin::BI__sync_val_compare_and_swap_1: 1099 case Builtin::BI__sync_val_compare_and_swap_2: 1100 case Builtin::BI__sync_val_compare_and_swap_4: 1101 case Builtin::BI__sync_val_compare_and_swap_8: 1102 case Builtin::BI__sync_val_compare_and_swap_16: 1103 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1104 1105 case Builtin::BI__sync_bool_compare_and_swap_1: 1106 case Builtin::BI__sync_bool_compare_and_swap_2: 1107 case Builtin::BI__sync_bool_compare_and_swap_4: 1108 case Builtin::BI__sync_bool_compare_and_swap_8: 1109 case Builtin::BI__sync_bool_compare_and_swap_16: 1110 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1111 1112 case Builtin::BI__sync_swap_1: 1113 case Builtin::BI__sync_swap_2: 1114 case Builtin::BI__sync_swap_4: 1115 case Builtin::BI__sync_swap_8: 1116 case Builtin::BI__sync_swap_16: 1117 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1118 1119 case Builtin::BI__sync_lock_test_and_set_1: 1120 case Builtin::BI__sync_lock_test_and_set_2: 1121 case Builtin::BI__sync_lock_test_and_set_4: 1122 case Builtin::BI__sync_lock_test_and_set_8: 1123 case Builtin::BI__sync_lock_test_and_set_16: 1124 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1125 1126 case Builtin::BI__sync_lock_release_1: 1127 case Builtin::BI__sync_lock_release_2: 1128 case Builtin::BI__sync_lock_release_4: 1129 case Builtin::BI__sync_lock_release_8: 1130 case Builtin::BI__sync_lock_release_16: { 1131 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1132 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1133 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1134 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1135 StoreSize.getQuantity() * 8); 1136 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1137 llvm::StoreInst *Store = 1138 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1139 Store->setAlignment(StoreSize.getQuantity()); 1140 Store->setAtomic(llvm::Release); 1141 return RValue::get(nullptr); 1142 } 1143 1144 case Builtin::BI__sync_synchronize: { 1145 // We assume this is supposed to correspond to a C++0x-style 1146 // sequentially-consistent fence (i.e. this is only usable for 1147 // synchonization, not device I/O or anything like that). This intrinsic 1148 // is really badly designed in the sense that in theory, there isn't 1149 // any way to safely use it... but in practice, it mostly works 1150 // to use it with non-atomic loads and stores to get acquire/release 1151 // semantics. 1152 Builder.CreateFence(llvm::SequentiallyConsistent); 1153 return RValue::get(nullptr); 1154 } 1155 1156 case Builtin::BI__c11_atomic_is_lock_free: 1157 case Builtin::BI__atomic_is_lock_free: { 1158 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1159 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1160 // _Atomic(T) is always properly-aligned. 1161 const char *LibCallName = "__atomic_is_lock_free"; 1162 CallArgList Args; 1163 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1164 getContext().getSizeType()); 1165 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1166 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1167 getContext().VoidPtrTy); 1168 else 1169 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1170 getContext().VoidPtrTy); 1171 const CGFunctionInfo &FuncInfo = 1172 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1173 FunctionType::ExtInfo(), 1174 RequiredArgs::All); 1175 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1176 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1177 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1178 } 1179 1180 case Builtin::BI__atomic_test_and_set: { 1181 // Look at the argument type to determine whether this is a volatile 1182 // operation. The parameter type is always volatile. 1183 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1184 bool Volatile = 1185 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1186 1187 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1188 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1189 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1190 Value *NewVal = Builder.getInt8(1); 1191 Value *Order = EmitScalarExpr(E->getArg(1)); 1192 if (isa<llvm::ConstantInt>(Order)) { 1193 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1194 AtomicRMWInst *Result = nullptr; 1195 switch (ord) { 1196 case 0: // memory_order_relaxed 1197 default: // invalid order 1198 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1199 Ptr, NewVal, 1200 llvm::Monotonic); 1201 break; 1202 case 1: // memory_order_consume 1203 case 2: // memory_order_acquire 1204 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1205 Ptr, NewVal, 1206 llvm::Acquire); 1207 break; 1208 case 3: // memory_order_release 1209 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1210 Ptr, NewVal, 1211 llvm::Release); 1212 break; 1213 case 4: // memory_order_acq_rel 1214 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1215 Ptr, NewVal, 1216 llvm::AcquireRelease); 1217 break; 1218 case 5: // memory_order_seq_cst 1219 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1220 Ptr, NewVal, 1221 llvm::SequentiallyConsistent); 1222 break; 1223 } 1224 Result->setVolatile(Volatile); 1225 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1226 } 1227 1228 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1229 1230 llvm::BasicBlock *BBs[5] = { 1231 createBasicBlock("monotonic", CurFn), 1232 createBasicBlock("acquire", CurFn), 1233 createBasicBlock("release", CurFn), 1234 createBasicBlock("acqrel", CurFn), 1235 createBasicBlock("seqcst", CurFn) 1236 }; 1237 llvm::AtomicOrdering Orders[5] = { 1238 llvm::Monotonic, llvm::Acquire, llvm::Release, 1239 llvm::AcquireRelease, llvm::SequentiallyConsistent 1240 }; 1241 1242 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1243 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1244 1245 Builder.SetInsertPoint(ContBB); 1246 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1247 1248 for (unsigned i = 0; i < 5; ++i) { 1249 Builder.SetInsertPoint(BBs[i]); 1250 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1251 Ptr, NewVal, Orders[i]); 1252 RMW->setVolatile(Volatile); 1253 Result->addIncoming(RMW, BBs[i]); 1254 Builder.CreateBr(ContBB); 1255 } 1256 1257 SI->addCase(Builder.getInt32(0), BBs[0]); 1258 SI->addCase(Builder.getInt32(1), BBs[1]); 1259 SI->addCase(Builder.getInt32(2), BBs[1]); 1260 SI->addCase(Builder.getInt32(3), BBs[2]); 1261 SI->addCase(Builder.getInt32(4), BBs[3]); 1262 SI->addCase(Builder.getInt32(5), BBs[4]); 1263 1264 Builder.SetInsertPoint(ContBB); 1265 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1266 } 1267 1268 case Builtin::BI__atomic_clear: { 1269 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1270 bool Volatile = 1271 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1272 1273 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1274 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1275 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1276 Value *NewVal = Builder.getInt8(0); 1277 Value *Order = EmitScalarExpr(E->getArg(1)); 1278 if (isa<llvm::ConstantInt>(Order)) { 1279 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1280 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1281 Store->setAlignment(1); 1282 switch (ord) { 1283 case 0: // memory_order_relaxed 1284 default: // invalid order 1285 Store->setOrdering(llvm::Monotonic); 1286 break; 1287 case 3: // memory_order_release 1288 Store->setOrdering(llvm::Release); 1289 break; 1290 case 5: // memory_order_seq_cst 1291 Store->setOrdering(llvm::SequentiallyConsistent); 1292 break; 1293 } 1294 return RValue::get(nullptr); 1295 } 1296 1297 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1298 1299 llvm::BasicBlock *BBs[3] = { 1300 createBasicBlock("monotonic", CurFn), 1301 createBasicBlock("release", CurFn), 1302 createBasicBlock("seqcst", CurFn) 1303 }; 1304 llvm::AtomicOrdering Orders[3] = { 1305 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1306 }; 1307 1308 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1309 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1310 1311 for (unsigned i = 0; i < 3; ++i) { 1312 Builder.SetInsertPoint(BBs[i]); 1313 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1314 Store->setAlignment(1); 1315 Store->setOrdering(Orders[i]); 1316 Builder.CreateBr(ContBB); 1317 } 1318 1319 SI->addCase(Builder.getInt32(0), BBs[0]); 1320 SI->addCase(Builder.getInt32(3), BBs[1]); 1321 SI->addCase(Builder.getInt32(5), BBs[2]); 1322 1323 Builder.SetInsertPoint(ContBB); 1324 return RValue::get(nullptr); 1325 } 1326 1327 case Builtin::BI__atomic_thread_fence: 1328 case Builtin::BI__atomic_signal_fence: 1329 case Builtin::BI__c11_atomic_thread_fence: 1330 case Builtin::BI__c11_atomic_signal_fence: { 1331 llvm::SynchronizationScope Scope; 1332 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1333 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1334 Scope = llvm::SingleThread; 1335 else 1336 Scope = llvm::CrossThread; 1337 Value *Order = EmitScalarExpr(E->getArg(0)); 1338 if (isa<llvm::ConstantInt>(Order)) { 1339 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1340 switch (ord) { 1341 case 0: // memory_order_relaxed 1342 default: // invalid order 1343 break; 1344 case 1: // memory_order_consume 1345 case 2: // memory_order_acquire 1346 Builder.CreateFence(llvm::Acquire, Scope); 1347 break; 1348 case 3: // memory_order_release 1349 Builder.CreateFence(llvm::Release, Scope); 1350 break; 1351 case 4: // memory_order_acq_rel 1352 Builder.CreateFence(llvm::AcquireRelease, Scope); 1353 break; 1354 case 5: // memory_order_seq_cst 1355 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1356 break; 1357 } 1358 return RValue::get(nullptr); 1359 } 1360 1361 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1362 AcquireBB = createBasicBlock("acquire", CurFn); 1363 ReleaseBB = createBasicBlock("release", CurFn); 1364 AcqRelBB = createBasicBlock("acqrel", CurFn); 1365 SeqCstBB = createBasicBlock("seqcst", CurFn); 1366 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1367 1368 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1369 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1370 1371 Builder.SetInsertPoint(AcquireBB); 1372 Builder.CreateFence(llvm::Acquire, Scope); 1373 Builder.CreateBr(ContBB); 1374 SI->addCase(Builder.getInt32(1), AcquireBB); 1375 SI->addCase(Builder.getInt32(2), AcquireBB); 1376 1377 Builder.SetInsertPoint(ReleaseBB); 1378 Builder.CreateFence(llvm::Release, Scope); 1379 Builder.CreateBr(ContBB); 1380 SI->addCase(Builder.getInt32(3), ReleaseBB); 1381 1382 Builder.SetInsertPoint(AcqRelBB); 1383 Builder.CreateFence(llvm::AcquireRelease, Scope); 1384 Builder.CreateBr(ContBB); 1385 SI->addCase(Builder.getInt32(4), AcqRelBB); 1386 1387 Builder.SetInsertPoint(SeqCstBB); 1388 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1389 Builder.CreateBr(ContBB); 1390 SI->addCase(Builder.getInt32(5), SeqCstBB); 1391 1392 Builder.SetInsertPoint(ContBB); 1393 return RValue::get(nullptr); 1394 } 1395 1396 // Library functions with special handling. 1397 case Builtin::BIsqrt: 1398 case Builtin::BIsqrtf: 1399 case Builtin::BIsqrtl: { 1400 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1401 // in finite- or unsafe-math mode (the intrinsic has different semantics 1402 // for handling negative numbers compared to the library function, so 1403 // -fmath-errno=0 is not enough). 1404 if (!FD->hasAttr<ConstAttr>()) 1405 break; 1406 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1407 CGM.getCodeGenOpts().NoNaNsFPMath)) 1408 break; 1409 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1410 llvm::Type *ArgType = Arg0->getType(); 1411 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1412 return RValue::get(Builder.CreateCall(F, Arg0)); 1413 } 1414 1415 case Builtin::BI__builtin_pow: 1416 case Builtin::BI__builtin_powf: 1417 case Builtin::BI__builtin_powl: 1418 case Builtin::BIpow: 1419 case Builtin::BIpowf: 1420 case Builtin::BIpowl: { 1421 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1422 if (!FD->hasAttr<ConstAttr>()) 1423 break; 1424 Value *Base = EmitScalarExpr(E->getArg(0)); 1425 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1426 llvm::Type *ArgType = Base->getType(); 1427 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1428 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1429 } 1430 1431 case Builtin::BIfma: 1432 case Builtin::BIfmaf: 1433 case Builtin::BIfmal: 1434 case Builtin::BI__builtin_fma: 1435 case Builtin::BI__builtin_fmaf: 1436 case Builtin::BI__builtin_fmal: { 1437 // Rewrite fma to intrinsic. 1438 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1439 llvm::Type *ArgType = FirstArg->getType(); 1440 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1441 return RValue::get( 1442 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1443 EmitScalarExpr(E->getArg(2))})); 1444 } 1445 1446 case Builtin::BI__builtin_signbit: 1447 case Builtin::BI__builtin_signbitf: 1448 case Builtin::BI__builtin_signbitl: { 1449 return RValue::get( 1450 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1451 ConvertType(E->getType()))); 1452 } 1453 case Builtin::BI__builtin_annotation: { 1454 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1455 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1456 AnnVal->getType()); 1457 1458 // Get the annotation string, go through casts. Sema requires this to be a 1459 // non-wide string literal, potentially casted, so the cast<> is safe. 1460 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1461 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1462 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1463 } 1464 case Builtin::BI__builtin_addcb: 1465 case Builtin::BI__builtin_addcs: 1466 case Builtin::BI__builtin_addc: 1467 case Builtin::BI__builtin_addcl: 1468 case Builtin::BI__builtin_addcll: 1469 case Builtin::BI__builtin_subcb: 1470 case Builtin::BI__builtin_subcs: 1471 case Builtin::BI__builtin_subc: 1472 case Builtin::BI__builtin_subcl: 1473 case Builtin::BI__builtin_subcll: { 1474 1475 // We translate all of these builtins from expressions of the form: 1476 // int x = ..., y = ..., carryin = ..., carryout, result; 1477 // result = __builtin_addc(x, y, carryin, &carryout); 1478 // 1479 // to LLVM IR of the form: 1480 // 1481 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1482 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1483 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1484 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1485 // i32 %carryin) 1486 // %result = extractvalue {i32, i1} %tmp2, 0 1487 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1488 // %tmp3 = or i1 %carry1, %carry2 1489 // %tmp4 = zext i1 %tmp3 to i32 1490 // store i32 %tmp4, i32* %carryout 1491 1492 // Scalarize our inputs. 1493 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1494 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1495 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1496 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1497 EmitPointerWithAlignment(E->getArg(3)); 1498 1499 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1500 llvm::Intrinsic::ID IntrinsicId; 1501 switch (BuiltinID) { 1502 default: llvm_unreachable("Unknown multiprecision builtin id."); 1503 case Builtin::BI__builtin_addcb: 1504 case Builtin::BI__builtin_addcs: 1505 case Builtin::BI__builtin_addc: 1506 case Builtin::BI__builtin_addcl: 1507 case Builtin::BI__builtin_addcll: 1508 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1509 break; 1510 case Builtin::BI__builtin_subcb: 1511 case Builtin::BI__builtin_subcs: 1512 case Builtin::BI__builtin_subc: 1513 case Builtin::BI__builtin_subcl: 1514 case Builtin::BI__builtin_subcll: 1515 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1516 break; 1517 } 1518 1519 // Construct our resulting LLVM IR expression. 1520 llvm::Value *Carry1; 1521 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1522 X, Y, Carry1); 1523 llvm::Value *Carry2; 1524 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1525 Sum1, Carryin, Carry2); 1526 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1527 X->getType()); 1528 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1529 CarryOutPtr.first); 1530 CarryOutStore->setAlignment(CarryOutPtr.second); 1531 return RValue::get(Sum2); 1532 } 1533 case Builtin::BI__builtin_uadd_overflow: 1534 case Builtin::BI__builtin_uaddl_overflow: 1535 case Builtin::BI__builtin_uaddll_overflow: 1536 case Builtin::BI__builtin_usub_overflow: 1537 case Builtin::BI__builtin_usubl_overflow: 1538 case Builtin::BI__builtin_usubll_overflow: 1539 case Builtin::BI__builtin_umul_overflow: 1540 case Builtin::BI__builtin_umull_overflow: 1541 case Builtin::BI__builtin_umulll_overflow: 1542 case Builtin::BI__builtin_sadd_overflow: 1543 case Builtin::BI__builtin_saddl_overflow: 1544 case Builtin::BI__builtin_saddll_overflow: 1545 case Builtin::BI__builtin_ssub_overflow: 1546 case Builtin::BI__builtin_ssubl_overflow: 1547 case Builtin::BI__builtin_ssubll_overflow: 1548 case Builtin::BI__builtin_smul_overflow: 1549 case Builtin::BI__builtin_smull_overflow: 1550 case Builtin::BI__builtin_smulll_overflow: { 1551 1552 // We translate all of these builtins directly to the relevant llvm IR node. 1553 1554 // Scalarize our inputs. 1555 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1556 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1557 std::pair<llvm::Value *, unsigned> SumOutPtr = 1558 EmitPointerWithAlignment(E->getArg(2)); 1559 1560 // Decide which of the overflow intrinsics we are lowering to: 1561 llvm::Intrinsic::ID IntrinsicId; 1562 switch (BuiltinID) { 1563 default: llvm_unreachable("Unknown security overflow builtin id."); 1564 case Builtin::BI__builtin_uadd_overflow: 1565 case Builtin::BI__builtin_uaddl_overflow: 1566 case Builtin::BI__builtin_uaddll_overflow: 1567 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1568 break; 1569 case Builtin::BI__builtin_usub_overflow: 1570 case Builtin::BI__builtin_usubl_overflow: 1571 case Builtin::BI__builtin_usubll_overflow: 1572 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1573 break; 1574 case Builtin::BI__builtin_umul_overflow: 1575 case Builtin::BI__builtin_umull_overflow: 1576 case Builtin::BI__builtin_umulll_overflow: 1577 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1578 break; 1579 case Builtin::BI__builtin_sadd_overflow: 1580 case Builtin::BI__builtin_saddl_overflow: 1581 case Builtin::BI__builtin_saddll_overflow: 1582 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1583 break; 1584 case Builtin::BI__builtin_ssub_overflow: 1585 case Builtin::BI__builtin_ssubl_overflow: 1586 case Builtin::BI__builtin_ssubll_overflow: 1587 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1588 break; 1589 case Builtin::BI__builtin_smul_overflow: 1590 case Builtin::BI__builtin_smull_overflow: 1591 case Builtin::BI__builtin_smulll_overflow: 1592 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1593 break; 1594 } 1595 1596 1597 llvm::Value *Carry; 1598 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1599 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1600 SumOutStore->setAlignment(SumOutPtr.second); 1601 1602 return RValue::get(Carry); 1603 } 1604 case Builtin::BI__builtin_addressof: 1605 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1606 case Builtin::BI__builtin_operator_new: 1607 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1608 E->getArg(0), false); 1609 case Builtin::BI__builtin_operator_delete: 1610 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1611 E->getArg(0), true); 1612 case Builtin::BI__noop: 1613 // __noop always evaluates to an integer literal zero. 1614 return RValue::get(ConstantInt::get(IntTy, 0)); 1615 case Builtin::BI__builtin_call_with_static_chain: { 1616 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1617 const Expr *Chain = E->getArg(1); 1618 return EmitCall(Call->getCallee()->getType(), 1619 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1620 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1621 } 1622 case Builtin::BI_InterlockedExchange: 1623 case Builtin::BI_InterlockedExchangePointer: 1624 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1625 case Builtin::BI_InterlockedCompareExchangePointer: { 1626 llvm::Type *RTy; 1627 llvm::IntegerType *IntType = 1628 IntegerType::get(getLLVMContext(), 1629 getContext().getTypeSize(E->getType())); 1630 llvm::Type *IntPtrType = IntType->getPointerTo(); 1631 1632 llvm::Value *Destination = 1633 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1634 1635 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1636 RTy = Exchange->getType(); 1637 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1638 1639 llvm::Value *Comparand = 1640 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1641 1642 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1643 SequentiallyConsistent, 1644 SequentiallyConsistent); 1645 Result->setVolatile(true); 1646 1647 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1648 0), 1649 RTy)); 1650 } 1651 case Builtin::BI_InterlockedCompareExchange: { 1652 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1653 EmitScalarExpr(E->getArg(0)), 1654 EmitScalarExpr(E->getArg(2)), 1655 EmitScalarExpr(E->getArg(1)), 1656 SequentiallyConsistent, 1657 SequentiallyConsistent); 1658 CXI->setVolatile(true); 1659 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1660 } 1661 case Builtin::BI_InterlockedIncrement: { 1662 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1663 AtomicRMWInst::Add, 1664 EmitScalarExpr(E->getArg(0)), 1665 ConstantInt::get(Int32Ty, 1), 1666 llvm::SequentiallyConsistent); 1667 RMWI->setVolatile(true); 1668 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1669 } 1670 case Builtin::BI_InterlockedDecrement: { 1671 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1672 AtomicRMWInst::Sub, 1673 EmitScalarExpr(E->getArg(0)), 1674 ConstantInt::get(Int32Ty, 1), 1675 llvm::SequentiallyConsistent); 1676 RMWI->setVolatile(true); 1677 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1678 } 1679 case Builtin::BI_InterlockedExchangeAdd: { 1680 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1681 AtomicRMWInst::Add, 1682 EmitScalarExpr(E->getArg(0)), 1683 EmitScalarExpr(E->getArg(1)), 1684 llvm::SequentiallyConsistent); 1685 RMWI->setVolatile(true); 1686 return RValue::get(RMWI); 1687 } 1688 case Builtin::BI__readfsdword: { 1689 Value *IntToPtr = 1690 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1691 llvm::PointerType::get(CGM.Int32Ty, 257)); 1692 LoadInst *Load = 1693 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1694 return RValue::get(Load); 1695 } 1696 1697 case Builtin::BI__exception_code: 1698 case Builtin::BI_exception_code: 1699 return RValue::get(EmitSEHExceptionCode()); 1700 case Builtin::BI__exception_info: 1701 case Builtin::BI_exception_info: 1702 return RValue::get(EmitSEHExceptionInfo()); 1703 case Builtin::BI__abnormal_termination: 1704 case Builtin::BI_abnormal_termination: 1705 return RValue::get(EmitSEHAbnormalTermination()); 1706 case Builtin::BI_setjmpex: { 1707 if (getTarget().getTriple().isOSMSVCRT()) { 1708 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1709 llvm::AttributeSet ReturnsTwiceAttr = 1710 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1711 llvm::Attribute::ReturnsTwice); 1712 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1713 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1714 "_setjmpex", ReturnsTwiceAttr); 1715 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1716 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1717 llvm::Value *FrameAddr = 1718 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1719 ConstantInt::get(Int32Ty, 0)); 1720 llvm::Value *Args[] = {Buf, FrameAddr}; 1721 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1722 CS.setAttributes(ReturnsTwiceAttr); 1723 return RValue::get(CS.getInstruction()); 1724 } 1725 break; 1726 } 1727 case Builtin::BI_setjmp: { 1728 if (getTarget().getTriple().isOSMSVCRT()) { 1729 llvm::AttributeSet ReturnsTwiceAttr = 1730 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1731 llvm::Attribute::ReturnsTwice); 1732 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1733 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1734 llvm::CallSite CS; 1735 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1736 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1737 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1738 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1739 "_setjmp3", ReturnsTwiceAttr); 1740 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1741 llvm::Value *Args[] = {Buf, Count}; 1742 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1743 } else { 1744 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1745 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1746 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1747 "_setjmp", ReturnsTwiceAttr); 1748 llvm::Value *FrameAddr = 1749 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1750 ConstantInt::get(Int32Ty, 0)); 1751 llvm::Value *Args[] = {Buf, FrameAddr}; 1752 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1753 } 1754 CS.setAttributes(ReturnsTwiceAttr); 1755 return RValue::get(CS.getInstruction()); 1756 } 1757 break; 1758 } 1759 1760 case Builtin::BI__GetExceptionInfo: { 1761 if (llvm::GlobalVariable *GV = 1762 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1763 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1764 break; 1765 } 1766 } 1767 1768 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1769 // the call using the normal call path, but using the unmangled 1770 // version of the function name. 1771 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1772 return emitLibraryCall(*this, FD, E, 1773 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1774 1775 // If this is a predefined lib function (e.g. malloc), emit the call 1776 // using exactly the normal call path. 1777 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1778 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1779 1780 // See if we have a target specific intrinsic. 1781 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 1782 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1783 if (const char *Prefix = 1784 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1785 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1786 // NOTE we dont need to perform a compatibility flag check here since the 1787 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1788 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1789 if (IntrinsicID == Intrinsic::not_intrinsic) 1790 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1791 } 1792 1793 if (IntrinsicID != Intrinsic::not_intrinsic) { 1794 SmallVector<Value*, 16> Args; 1795 1796 // Find out if any arguments are required to be integer constant 1797 // expressions. 1798 unsigned ICEArguments = 0; 1799 ASTContext::GetBuiltinTypeError Error; 1800 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1801 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1802 1803 Function *F = CGM.getIntrinsic(IntrinsicID); 1804 llvm::FunctionType *FTy = F->getFunctionType(); 1805 1806 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1807 Value *ArgValue; 1808 // If this is a normal argument, just emit it as a scalar. 1809 if ((ICEArguments & (1 << i)) == 0) { 1810 ArgValue = EmitScalarExpr(E->getArg(i)); 1811 } else { 1812 // If this is required to be a constant, constant fold it so that we 1813 // know that the generated intrinsic gets a ConstantInt. 1814 llvm::APSInt Result; 1815 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1816 assert(IsConst && "Constant arg isn't actually constant?"); 1817 (void)IsConst; 1818 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1819 } 1820 1821 // If the intrinsic arg type is different from the builtin arg type 1822 // we need to do a bit cast. 1823 llvm::Type *PTy = FTy->getParamType(i); 1824 if (PTy != ArgValue->getType()) { 1825 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1826 "Must be able to losslessly bit cast to param"); 1827 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1828 } 1829 1830 Args.push_back(ArgValue); 1831 } 1832 1833 Value *V = Builder.CreateCall(F, Args); 1834 QualType BuiltinRetType = E->getType(); 1835 1836 llvm::Type *RetTy = VoidTy; 1837 if (!BuiltinRetType->isVoidType()) 1838 RetTy = ConvertType(BuiltinRetType); 1839 1840 if (RetTy != V->getType()) { 1841 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1842 "Must be able to losslessly bit cast result type"); 1843 V = Builder.CreateBitCast(V, RetTy); 1844 } 1845 1846 return RValue::get(V); 1847 } 1848 1849 // See if we have a target specific builtin that needs to be lowered. 1850 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1851 return RValue::get(V); 1852 1853 ErrorUnsupported(E, "builtin function"); 1854 1855 // Unknown builtin, for now just dump it out and return undef. 1856 return GetUndefRValue(E->getType()); 1857 } 1858 1859 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1860 const CallExpr *E) { 1861 switch (getTarget().getTriple().getArch()) { 1862 case llvm::Triple::arm: 1863 case llvm::Triple::armeb: 1864 case llvm::Triple::thumb: 1865 case llvm::Triple::thumbeb: 1866 return EmitARMBuiltinExpr(BuiltinID, E); 1867 case llvm::Triple::aarch64: 1868 case llvm::Triple::aarch64_be: 1869 return EmitAArch64BuiltinExpr(BuiltinID, E); 1870 case llvm::Triple::x86: 1871 case llvm::Triple::x86_64: 1872 return EmitX86BuiltinExpr(BuiltinID, E); 1873 case llvm::Triple::ppc: 1874 case llvm::Triple::ppc64: 1875 case llvm::Triple::ppc64le: 1876 return EmitPPCBuiltinExpr(BuiltinID, E); 1877 case llvm::Triple::r600: 1878 case llvm::Triple::amdgcn: 1879 return EmitAMDGPUBuiltinExpr(BuiltinID, E); 1880 case llvm::Triple::systemz: 1881 return EmitSystemZBuiltinExpr(BuiltinID, E); 1882 case llvm::Triple::nvptx: 1883 case llvm::Triple::nvptx64: 1884 return EmitNVPTXBuiltinExpr(BuiltinID, E); 1885 default: 1886 return nullptr; 1887 } 1888 } 1889 1890 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1891 NeonTypeFlags TypeFlags, 1892 bool V1Ty=false) { 1893 int IsQuad = TypeFlags.isQuad(); 1894 switch (TypeFlags.getEltType()) { 1895 case NeonTypeFlags::Int8: 1896 case NeonTypeFlags::Poly8: 1897 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1898 case NeonTypeFlags::Int16: 1899 case NeonTypeFlags::Poly16: 1900 case NeonTypeFlags::Float16: 1901 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1902 case NeonTypeFlags::Int32: 1903 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1904 case NeonTypeFlags::Int64: 1905 case NeonTypeFlags::Poly64: 1906 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1907 case NeonTypeFlags::Poly128: 1908 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1909 // There is a lot of i128 and f128 API missing. 1910 // so we use v16i8 to represent poly128 and get pattern matched. 1911 return llvm::VectorType::get(CGF->Int8Ty, 16); 1912 case NeonTypeFlags::Float32: 1913 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1914 case NeonTypeFlags::Float64: 1915 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1916 } 1917 llvm_unreachable("Unknown vector element type!"); 1918 } 1919 1920 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 1921 NeonTypeFlags IntTypeFlags) { 1922 int IsQuad = IntTypeFlags.isQuad(); 1923 switch (IntTypeFlags.getEltType()) { 1924 case NeonTypeFlags::Int32: 1925 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 1926 case NeonTypeFlags::Int64: 1927 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 1928 default: 1929 llvm_unreachable("Type can't be converted to floating-point!"); 1930 } 1931 } 1932 1933 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1934 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1935 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1936 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1937 } 1938 1939 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1940 const char *name, 1941 unsigned shift, bool rightshift) { 1942 unsigned j = 0; 1943 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1944 ai != ae; ++ai, ++j) 1945 if (shift > 0 && shift == j) 1946 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1947 else 1948 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1949 1950 return Builder.CreateCall(F, Ops, name); 1951 } 1952 1953 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1954 bool neg) { 1955 int SV = cast<ConstantInt>(V)->getSExtValue(); 1956 return ConstantInt::get(Ty, neg ? -SV : SV); 1957 } 1958 1959 // \brief Right-shift a vector by a constant. 1960 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1961 llvm::Type *Ty, bool usgn, 1962 const char *name) { 1963 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1964 1965 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1966 int EltSize = VTy->getScalarSizeInBits(); 1967 1968 Vec = Builder.CreateBitCast(Vec, Ty); 1969 1970 // lshr/ashr are undefined when the shift amount is equal to the vector 1971 // element size. 1972 if (ShiftAmt == EltSize) { 1973 if (usgn) { 1974 // Right-shifting an unsigned value by its size yields 0. 1975 return llvm::ConstantAggregateZero::get(VTy); 1976 } else { 1977 // Right-shifting a signed value by its size is equivalent 1978 // to a shift of size-1. 1979 --ShiftAmt; 1980 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1981 } 1982 } 1983 1984 Shift = EmitNeonShiftVector(Shift, Ty, false); 1985 if (usgn) 1986 return Builder.CreateLShr(Vec, Shift, name); 1987 else 1988 return Builder.CreateAShr(Vec, Shift, name); 1989 } 1990 1991 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1992 /// alignment of the type referenced by the pointer. Skip over implicit 1993 /// casts. 1994 std::pair<llvm::Value*, unsigned> 1995 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1996 assert(Addr->getType()->isPointerType()); 1997 Addr = Addr->IgnoreParens(); 1998 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1999 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 2000 ICE->getSubExpr()->getType()->isPointerType()) { 2001 std::pair<llvm::Value*, unsigned> Ptr = 2002 EmitPointerWithAlignment(ICE->getSubExpr()); 2003 Ptr.first = Builder.CreateBitCast(Ptr.first, 2004 ConvertType(Addr->getType())); 2005 return Ptr; 2006 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 2007 LValue LV = EmitLValue(ICE->getSubExpr()); 2008 unsigned Align = LV.getAlignment().getQuantity(); 2009 if (!Align) { 2010 // FIXME: Once LValues are fixed to always set alignment, 2011 // zap this code. 2012 QualType PtTy = ICE->getSubExpr()->getType(); 2013 if (!PtTy->isIncompleteType()) 2014 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2015 else 2016 Align = 1; 2017 } 2018 return std::make_pair(LV.getAddress(), Align); 2019 } 2020 } 2021 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 2022 if (UO->getOpcode() == UO_AddrOf) { 2023 LValue LV = EmitLValue(UO->getSubExpr()); 2024 unsigned Align = LV.getAlignment().getQuantity(); 2025 if (!Align) { 2026 // FIXME: Once LValues are fixed to always set alignment, 2027 // zap this code. 2028 QualType PtTy = UO->getSubExpr()->getType(); 2029 if (!PtTy->isIncompleteType()) 2030 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2031 else 2032 Align = 1; 2033 } 2034 return std::make_pair(LV.getAddress(), Align); 2035 } 2036 } 2037 2038 unsigned Align = 1; 2039 QualType PtTy = Addr->getType()->getPointeeType(); 2040 if (!PtTy->isIncompleteType()) 2041 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2042 2043 return std::make_pair(EmitScalarExpr(Addr), Align); 2044 } 2045 2046 enum { 2047 AddRetType = (1 << 0), 2048 Add1ArgType = (1 << 1), 2049 Add2ArgTypes = (1 << 2), 2050 2051 VectorizeRetType = (1 << 3), 2052 VectorizeArgTypes = (1 << 4), 2053 2054 InventFloatType = (1 << 5), 2055 UnsignedAlts = (1 << 6), 2056 2057 Use64BitVectors = (1 << 7), 2058 Use128BitVectors = (1 << 8), 2059 2060 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2061 VectorRet = AddRetType | VectorizeRetType, 2062 VectorRetGetArgs01 = 2063 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2064 FpCmpzModifiers = 2065 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2066 }; 2067 2068 struct NeonIntrinsicInfo { 2069 unsigned BuiltinID; 2070 unsigned LLVMIntrinsic; 2071 unsigned AltLLVMIntrinsic; 2072 const char *NameHint; 2073 unsigned TypeModifier; 2074 2075 bool operator<(unsigned RHSBuiltinID) const { 2076 return BuiltinID < RHSBuiltinID; 2077 } 2078 }; 2079 2080 #define NEONMAP0(NameBase) \ 2081 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2082 2083 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2084 { NEON:: BI__builtin_neon_ ## NameBase, \ 2085 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2086 2087 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2088 { NEON:: BI__builtin_neon_ ## NameBase, \ 2089 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2090 #NameBase, TypeModifier } 2091 2092 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2093 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2094 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2095 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2096 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2097 NEONMAP0(vaddhn_v), 2098 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2099 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2100 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2101 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2102 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2103 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2104 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2105 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2106 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2107 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2108 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2109 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2110 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2111 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2112 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2113 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2114 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2115 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2116 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2117 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2118 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2119 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2120 NEONMAP0(vcvt_f32_v), 2121 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2122 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2123 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2124 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2125 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2126 NEONMAP0(vcvt_s32_v), 2127 NEONMAP0(vcvt_s64_v), 2128 NEONMAP0(vcvt_u32_v), 2129 NEONMAP0(vcvt_u64_v), 2130 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2131 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2132 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2133 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2134 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2135 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2136 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2137 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2138 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2139 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2140 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2141 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2142 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2143 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2144 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2145 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2146 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2147 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2148 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2149 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2150 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2151 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2152 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2153 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2154 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2155 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2156 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2157 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2158 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2159 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2160 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2161 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2162 NEONMAP0(vcvtq_f32_v), 2163 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2164 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2165 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2166 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2167 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2168 NEONMAP0(vcvtq_s32_v), 2169 NEONMAP0(vcvtq_s64_v), 2170 NEONMAP0(vcvtq_u32_v), 2171 NEONMAP0(vcvtq_u64_v), 2172 NEONMAP0(vext_v), 2173 NEONMAP0(vextq_v), 2174 NEONMAP0(vfma_v), 2175 NEONMAP0(vfmaq_v), 2176 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2177 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2178 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2179 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2180 NEONMAP0(vld1_dup_v), 2181 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2182 NEONMAP0(vld1q_dup_v), 2183 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2184 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2185 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2186 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2187 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2188 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2189 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2190 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2191 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2192 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2193 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2194 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2195 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2196 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2197 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2198 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2199 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2200 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2201 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2202 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2203 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2204 NEONMAP0(vmovl_v), 2205 NEONMAP0(vmovn_v), 2206 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2207 NEONMAP0(vmull_v), 2208 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2209 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2210 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2211 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2212 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2213 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2214 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2215 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2216 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2217 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2218 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2219 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2220 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2221 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2222 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2223 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2224 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2225 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2226 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2227 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2228 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2229 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2230 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2231 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2232 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2233 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2234 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2235 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2236 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2237 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2238 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2239 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2240 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2241 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2242 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2243 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2244 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2245 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2246 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2247 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2248 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2249 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2250 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2251 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2252 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2253 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2254 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2255 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2256 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2257 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2258 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2259 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2260 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2261 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2262 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2263 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2264 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2265 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2266 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2267 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2268 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2269 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2270 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2271 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2272 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2273 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2274 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2275 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2276 NEONMAP0(vshl_n_v), 2277 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2278 NEONMAP0(vshll_n_v), 2279 NEONMAP0(vshlq_n_v), 2280 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2281 NEONMAP0(vshr_n_v), 2282 NEONMAP0(vshrn_n_v), 2283 NEONMAP0(vshrq_n_v), 2284 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2285 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2286 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2287 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2288 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2289 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2290 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2291 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2292 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2293 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2294 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2295 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2296 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2297 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2298 NEONMAP0(vsubhn_v), 2299 NEONMAP0(vtrn_v), 2300 NEONMAP0(vtrnq_v), 2301 NEONMAP0(vtst_v), 2302 NEONMAP0(vtstq_v), 2303 NEONMAP0(vuzp_v), 2304 NEONMAP0(vuzpq_v), 2305 NEONMAP0(vzip_v), 2306 NEONMAP0(vzipq_v) 2307 }; 2308 2309 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2310 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2311 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2312 NEONMAP0(vaddhn_v), 2313 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2314 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2315 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2316 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2317 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2318 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2319 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2320 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2321 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2322 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2323 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2324 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2325 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2326 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2327 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2328 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2329 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2330 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2331 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2332 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2333 NEONMAP0(vcvt_f32_v), 2334 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2335 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2336 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2337 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2338 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2339 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2340 NEONMAP0(vcvtq_f32_v), 2341 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2342 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2343 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2344 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2345 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2346 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2347 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2348 NEONMAP0(vext_v), 2349 NEONMAP0(vextq_v), 2350 NEONMAP0(vfma_v), 2351 NEONMAP0(vfmaq_v), 2352 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2353 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2354 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2355 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2356 NEONMAP0(vmovl_v), 2357 NEONMAP0(vmovn_v), 2358 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2359 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2360 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2361 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2362 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2363 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2364 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2365 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2366 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2367 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2368 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2369 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2370 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2371 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2372 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2373 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2374 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2375 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2376 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2377 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2378 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2379 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2380 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2381 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2382 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2383 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2384 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2385 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2386 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2387 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2388 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2389 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2390 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2391 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2392 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2393 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2394 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2395 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2396 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2397 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2398 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2399 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2400 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2401 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2402 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2403 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2404 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2405 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2406 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2407 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2408 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2409 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2410 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2411 NEONMAP0(vshl_n_v), 2412 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2413 NEONMAP0(vshll_n_v), 2414 NEONMAP0(vshlq_n_v), 2415 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2416 NEONMAP0(vshr_n_v), 2417 NEONMAP0(vshrn_n_v), 2418 NEONMAP0(vshrq_n_v), 2419 NEONMAP0(vsubhn_v), 2420 NEONMAP0(vtst_v), 2421 NEONMAP0(vtstq_v), 2422 }; 2423 2424 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2425 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2426 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2427 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2428 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2429 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2430 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2431 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2432 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2433 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2434 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2435 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2436 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2437 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2438 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2439 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2440 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2441 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2442 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2443 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2444 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2445 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2446 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2447 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2448 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2449 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2450 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2451 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2452 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2453 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2454 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2455 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2456 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2457 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2458 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2459 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2460 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2461 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2462 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2463 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2464 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2465 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2466 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2467 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2468 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2469 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2470 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2471 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2472 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2473 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2474 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2475 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2476 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2477 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2478 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2479 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2480 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2481 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2482 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2483 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2484 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2485 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2486 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2487 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2488 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2489 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2490 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2491 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2492 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2493 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2494 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2495 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2496 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2497 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2498 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2499 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2500 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2501 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2502 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2503 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2504 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2505 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2506 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2507 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2508 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2509 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2510 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2511 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2512 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2513 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2514 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2515 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2516 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2517 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2518 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2519 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2520 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2521 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2522 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2523 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2524 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2525 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2526 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2527 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2528 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2529 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2530 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2531 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2532 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2533 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2534 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2535 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2536 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2537 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2538 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2539 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2540 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2541 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2542 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2543 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2544 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2545 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2546 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2547 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2548 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2549 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2550 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2551 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2552 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2553 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2554 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2555 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2556 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2557 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2558 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2559 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2560 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2561 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2562 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2563 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2564 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2565 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2566 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2567 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2568 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2569 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2570 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2571 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2572 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2573 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2574 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2575 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2576 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2577 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2578 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2579 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2580 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2581 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2582 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2583 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2584 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2585 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2586 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2587 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2588 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2589 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2590 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2591 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2592 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2593 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2594 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2595 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2596 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2597 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2598 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2599 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2600 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2601 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2602 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2603 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2604 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2605 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2606 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2607 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2608 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2609 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2610 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2611 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2612 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2613 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2614 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2615 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2616 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2617 }; 2618 2619 #undef NEONMAP0 2620 #undef NEONMAP1 2621 #undef NEONMAP2 2622 2623 static bool NEONSIMDIntrinsicsProvenSorted = false; 2624 2625 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2626 static bool AArch64SISDIntrinsicsProvenSorted = false; 2627 2628 2629 static const NeonIntrinsicInfo * 2630 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2631 unsigned BuiltinID, bool &MapProvenSorted) { 2632 2633 #ifndef NDEBUG 2634 if (!MapProvenSorted) { 2635 // FIXME: use std::is_sorted once C++11 is allowed 2636 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2637 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2638 MapProvenSorted = true; 2639 } 2640 #endif 2641 2642 const NeonIntrinsicInfo *Builtin = 2643 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2644 2645 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2646 return Builtin; 2647 2648 return nullptr; 2649 } 2650 2651 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2652 unsigned Modifier, 2653 llvm::Type *ArgType, 2654 const CallExpr *E) { 2655 int VectorSize = 0; 2656 if (Modifier & Use64BitVectors) 2657 VectorSize = 64; 2658 else if (Modifier & Use128BitVectors) 2659 VectorSize = 128; 2660 2661 // Return type. 2662 SmallVector<llvm::Type *, 3> Tys; 2663 if (Modifier & AddRetType) { 2664 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2665 if (Modifier & VectorizeRetType) 2666 Ty = llvm::VectorType::get( 2667 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2668 2669 Tys.push_back(Ty); 2670 } 2671 2672 // Arguments. 2673 if (Modifier & VectorizeArgTypes) { 2674 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2675 ArgType = llvm::VectorType::get(ArgType, Elts); 2676 } 2677 2678 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2679 Tys.push_back(ArgType); 2680 2681 if (Modifier & Add2ArgTypes) 2682 Tys.push_back(ArgType); 2683 2684 if (Modifier & InventFloatType) 2685 Tys.push_back(FloatTy); 2686 2687 return CGM.getIntrinsic(IntrinsicID, Tys); 2688 } 2689 2690 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2691 const NeonIntrinsicInfo &SISDInfo, 2692 SmallVectorImpl<Value *> &Ops, 2693 const CallExpr *E) { 2694 unsigned BuiltinID = SISDInfo.BuiltinID; 2695 unsigned int Int = SISDInfo.LLVMIntrinsic; 2696 unsigned Modifier = SISDInfo.TypeModifier; 2697 const char *s = SISDInfo.NameHint; 2698 2699 switch (BuiltinID) { 2700 case NEON::BI__builtin_neon_vcled_s64: 2701 case NEON::BI__builtin_neon_vcled_u64: 2702 case NEON::BI__builtin_neon_vcles_f32: 2703 case NEON::BI__builtin_neon_vcled_f64: 2704 case NEON::BI__builtin_neon_vcltd_s64: 2705 case NEON::BI__builtin_neon_vcltd_u64: 2706 case NEON::BI__builtin_neon_vclts_f32: 2707 case NEON::BI__builtin_neon_vcltd_f64: 2708 case NEON::BI__builtin_neon_vcales_f32: 2709 case NEON::BI__builtin_neon_vcaled_f64: 2710 case NEON::BI__builtin_neon_vcalts_f32: 2711 case NEON::BI__builtin_neon_vcaltd_f64: 2712 // Only one direction of comparisons actually exist, cmle is actually a cmge 2713 // with swapped operands. The table gives us the right intrinsic but we 2714 // still need to do the swap. 2715 std::swap(Ops[0], Ops[1]); 2716 break; 2717 } 2718 2719 assert(Int && "Generic code assumes a valid intrinsic"); 2720 2721 // Determine the type(s) of this overloaded AArch64 intrinsic. 2722 const Expr *Arg = E->getArg(0); 2723 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2724 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2725 2726 int j = 0; 2727 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2728 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2729 ai != ae; ++ai, ++j) { 2730 llvm::Type *ArgTy = ai->getType(); 2731 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2732 ArgTy->getPrimitiveSizeInBits()) 2733 continue; 2734 2735 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2736 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2737 // it before inserting. 2738 Ops[j] = 2739 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2740 Ops[j] = 2741 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2742 } 2743 2744 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2745 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2746 if (ResultType->getPrimitiveSizeInBits() < 2747 Result->getType()->getPrimitiveSizeInBits()) 2748 return CGF.Builder.CreateExtractElement(Result, C0); 2749 2750 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2751 } 2752 2753 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2754 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2755 const char *NameHint, unsigned Modifier, const CallExpr *E, 2756 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2757 // Get the last argument, which specifies the vector type. 2758 llvm::APSInt NeonTypeConst; 2759 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2760 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2761 return nullptr; 2762 2763 // Determine the type of this overloaded NEON intrinsic. 2764 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2765 bool Usgn = Type.isUnsigned(); 2766 bool Quad = Type.isQuad(); 2767 2768 llvm::VectorType *VTy = GetNeonType(this, Type); 2769 llvm::Type *Ty = VTy; 2770 if (!Ty) 2771 return nullptr; 2772 2773 unsigned Int = LLVMIntrinsic; 2774 if ((Modifier & UnsignedAlts) && !Usgn) 2775 Int = AltLLVMIntrinsic; 2776 2777 switch (BuiltinID) { 2778 default: break; 2779 case NEON::BI__builtin_neon_vabs_v: 2780 case NEON::BI__builtin_neon_vabsq_v: 2781 if (VTy->getElementType()->isFloatingPointTy()) 2782 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2783 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2784 case NEON::BI__builtin_neon_vaddhn_v: { 2785 llvm::VectorType *SrcTy = 2786 llvm::VectorType::getExtendedElementVectorType(VTy); 2787 2788 // %sum = add <4 x i32> %lhs, %rhs 2789 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2790 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2791 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2792 2793 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2794 Constant *ShiftAmt = 2795 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 2796 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2797 2798 // %res = trunc <4 x i32> %high to <4 x i16> 2799 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2800 } 2801 case NEON::BI__builtin_neon_vcale_v: 2802 case NEON::BI__builtin_neon_vcaleq_v: 2803 case NEON::BI__builtin_neon_vcalt_v: 2804 case NEON::BI__builtin_neon_vcaltq_v: 2805 std::swap(Ops[0], Ops[1]); 2806 case NEON::BI__builtin_neon_vcage_v: 2807 case NEON::BI__builtin_neon_vcageq_v: 2808 case NEON::BI__builtin_neon_vcagt_v: 2809 case NEON::BI__builtin_neon_vcagtq_v: { 2810 llvm::Type *VecFlt = llvm::VectorType::get( 2811 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2812 VTy->getNumElements()); 2813 llvm::Type *Tys[] = { VTy, VecFlt }; 2814 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2815 return EmitNeonCall(F, Ops, NameHint); 2816 } 2817 case NEON::BI__builtin_neon_vclz_v: 2818 case NEON::BI__builtin_neon_vclzq_v: 2819 // We generate target-independent intrinsic, which needs a second argument 2820 // for whether or not clz of zero is undefined; on ARM it isn't. 2821 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2822 break; 2823 case NEON::BI__builtin_neon_vcvt_f32_v: 2824 case NEON::BI__builtin_neon_vcvtq_f32_v: 2825 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2826 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2827 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2828 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2829 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2830 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2831 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2832 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2833 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 2834 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2835 Function *F = CGM.getIntrinsic(Int, Tys); 2836 return EmitNeonCall(F, Ops, "vcvt_n"); 2837 } 2838 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2839 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2840 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2841 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2842 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2843 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2844 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2845 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2846 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 2847 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2848 return EmitNeonCall(F, Ops, "vcvt_n"); 2849 } 2850 case NEON::BI__builtin_neon_vcvt_s32_v: 2851 case NEON::BI__builtin_neon_vcvt_u32_v: 2852 case NEON::BI__builtin_neon_vcvt_s64_v: 2853 case NEON::BI__builtin_neon_vcvt_u64_v: 2854 case NEON::BI__builtin_neon_vcvtq_s32_v: 2855 case NEON::BI__builtin_neon_vcvtq_u32_v: 2856 case NEON::BI__builtin_neon_vcvtq_s64_v: 2857 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2858 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 2859 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2860 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2861 } 2862 case NEON::BI__builtin_neon_vcvta_s32_v: 2863 case NEON::BI__builtin_neon_vcvta_s64_v: 2864 case NEON::BI__builtin_neon_vcvta_u32_v: 2865 case NEON::BI__builtin_neon_vcvta_u64_v: 2866 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2867 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2868 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2869 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2870 case NEON::BI__builtin_neon_vcvtn_s32_v: 2871 case NEON::BI__builtin_neon_vcvtn_s64_v: 2872 case NEON::BI__builtin_neon_vcvtn_u32_v: 2873 case NEON::BI__builtin_neon_vcvtn_u64_v: 2874 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2875 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2876 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2877 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2878 case NEON::BI__builtin_neon_vcvtp_s32_v: 2879 case NEON::BI__builtin_neon_vcvtp_s64_v: 2880 case NEON::BI__builtin_neon_vcvtp_u32_v: 2881 case NEON::BI__builtin_neon_vcvtp_u64_v: 2882 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2883 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2884 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2885 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2886 case NEON::BI__builtin_neon_vcvtm_s32_v: 2887 case NEON::BI__builtin_neon_vcvtm_s64_v: 2888 case NEON::BI__builtin_neon_vcvtm_u32_v: 2889 case NEON::BI__builtin_neon_vcvtm_u64_v: 2890 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2891 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2892 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2893 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2894 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 2895 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2896 } 2897 case NEON::BI__builtin_neon_vext_v: 2898 case NEON::BI__builtin_neon_vextq_v: { 2899 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2900 SmallVector<Constant*, 16> Indices; 2901 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2902 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2903 2904 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2905 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2906 Value *SV = llvm::ConstantVector::get(Indices); 2907 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2908 } 2909 case NEON::BI__builtin_neon_vfma_v: 2910 case NEON::BI__builtin_neon_vfmaq_v: { 2911 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2912 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2913 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2914 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2915 2916 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2917 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 2918 } 2919 case NEON::BI__builtin_neon_vld1_v: 2920 case NEON::BI__builtin_neon_vld1q_v: 2921 Ops.push_back(Align); 2922 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2923 case NEON::BI__builtin_neon_vld2_v: 2924 case NEON::BI__builtin_neon_vld2q_v: 2925 case NEON::BI__builtin_neon_vld3_v: 2926 case NEON::BI__builtin_neon_vld3q_v: 2927 case NEON::BI__builtin_neon_vld4_v: 2928 case NEON::BI__builtin_neon_vld4q_v: { 2929 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2930 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 2931 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2932 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2933 return Builder.CreateStore(Ops[1], Ops[0]); 2934 } 2935 case NEON::BI__builtin_neon_vld1_dup_v: 2936 case NEON::BI__builtin_neon_vld1q_dup_v: { 2937 Value *V = UndefValue::get(Ty); 2938 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2939 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2940 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2941 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2942 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2943 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2944 return EmitNeonSplat(Ops[0], CI); 2945 } 2946 case NEON::BI__builtin_neon_vld2_lane_v: 2947 case NEON::BI__builtin_neon_vld2q_lane_v: 2948 case NEON::BI__builtin_neon_vld3_lane_v: 2949 case NEON::BI__builtin_neon_vld3q_lane_v: 2950 case NEON::BI__builtin_neon_vld4_lane_v: 2951 case NEON::BI__builtin_neon_vld4q_lane_v: { 2952 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2953 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2954 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2955 Ops.push_back(Align); 2956 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2957 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2958 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2959 return Builder.CreateStore(Ops[1], Ops[0]); 2960 } 2961 case NEON::BI__builtin_neon_vmovl_v: { 2962 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2963 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2964 if (Usgn) 2965 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2966 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2967 } 2968 case NEON::BI__builtin_neon_vmovn_v: { 2969 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2970 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2971 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2972 } 2973 case NEON::BI__builtin_neon_vmull_v: 2974 // FIXME: the integer vmull operations could be emitted in terms of pure 2975 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2976 // hoisting the exts outside loops. Until global ISel comes along that can 2977 // see through such movement this leads to bad CodeGen. So we need an 2978 // intrinsic for now. 2979 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2980 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2981 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2982 case NEON::BI__builtin_neon_vpadal_v: 2983 case NEON::BI__builtin_neon_vpadalq_v: { 2984 // The source operand type has twice as many elements of half the size. 2985 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2986 llvm::Type *EltTy = 2987 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2988 llvm::Type *NarrowTy = 2989 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2990 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2991 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2992 } 2993 case NEON::BI__builtin_neon_vpaddl_v: 2994 case NEON::BI__builtin_neon_vpaddlq_v: { 2995 // The source operand type has twice as many elements of half the size. 2996 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2997 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2998 llvm::Type *NarrowTy = 2999 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3000 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3001 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3002 } 3003 case NEON::BI__builtin_neon_vqdmlal_v: 3004 case NEON::BI__builtin_neon_vqdmlsl_v: { 3005 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3006 Ops[1] = 3007 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3008 Ops.resize(2); 3009 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3010 } 3011 case NEON::BI__builtin_neon_vqshl_n_v: 3012 case NEON::BI__builtin_neon_vqshlq_n_v: 3013 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3014 1, false); 3015 case NEON::BI__builtin_neon_vqshlu_n_v: 3016 case NEON::BI__builtin_neon_vqshluq_n_v: 3017 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3018 1, false); 3019 case NEON::BI__builtin_neon_vrecpe_v: 3020 case NEON::BI__builtin_neon_vrecpeq_v: 3021 case NEON::BI__builtin_neon_vrsqrte_v: 3022 case NEON::BI__builtin_neon_vrsqrteq_v: 3023 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3024 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3025 3026 case NEON::BI__builtin_neon_vrshr_n_v: 3027 case NEON::BI__builtin_neon_vrshrq_n_v: 3028 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3029 1, true); 3030 case NEON::BI__builtin_neon_vshl_n_v: 3031 case NEON::BI__builtin_neon_vshlq_n_v: 3032 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3033 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3034 "vshl_n"); 3035 case NEON::BI__builtin_neon_vshll_n_v: { 3036 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3037 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3038 if (Usgn) 3039 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3040 else 3041 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3042 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3043 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3044 } 3045 case NEON::BI__builtin_neon_vshrn_n_v: { 3046 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3047 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3048 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3049 if (Usgn) 3050 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3051 else 3052 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3053 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3054 } 3055 case NEON::BI__builtin_neon_vshr_n_v: 3056 case NEON::BI__builtin_neon_vshrq_n_v: 3057 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3058 case NEON::BI__builtin_neon_vst1_v: 3059 case NEON::BI__builtin_neon_vst1q_v: 3060 case NEON::BI__builtin_neon_vst2_v: 3061 case NEON::BI__builtin_neon_vst2q_v: 3062 case NEON::BI__builtin_neon_vst3_v: 3063 case NEON::BI__builtin_neon_vst3q_v: 3064 case NEON::BI__builtin_neon_vst4_v: 3065 case NEON::BI__builtin_neon_vst4q_v: 3066 case NEON::BI__builtin_neon_vst2_lane_v: 3067 case NEON::BI__builtin_neon_vst2q_lane_v: 3068 case NEON::BI__builtin_neon_vst3_lane_v: 3069 case NEON::BI__builtin_neon_vst3q_lane_v: 3070 case NEON::BI__builtin_neon_vst4_lane_v: 3071 case NEON::BI__builtin_neon_vst4q_lane_v: 3072 Ops.push_back(Align); 3073 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3074 case NEON::BI__builtin_neon_vsubhn_v: { 3075 llvm::VectorType *SrcTy = 3076 llvm::VectorType::getExtendedElementVectorType(VTy); 3077 3078 // %sum = add <4 x i32> %lhs, %rhs 3079 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3080 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3081 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3082 3083 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3084 Constant *ShiftAmt = 3085 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3086 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3087 3088 // %res = trunc <4 x i32> %high to <4 x i16> 3089 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3090 } 3091 case NEON::BI__builtin_neon_vtrn_v: 3092 case NEON::BI__builtin_neon_vtrnq_v: { 3093 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3094 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3095 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3096 Value *SV = nullptr; 3097 3098 for (unsigned vi = 0; vi != 2; ++vi) { 3099 SmallVector<Constant*, 16> Indices; 3100 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3101 Indices.push_back(Builder.getInt32(i+vi)); 3102 Indices.push_back(Builder.getInt32(i+e+vi)); 3103 } 3104 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3105 SV = llvm::ConstantVector::get(Indices); 3106 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3107 SV = Builder.CreateStore(SV, Addr); 3108 } 3109 return SV; 3110 } 3111 case NEON::BI__builtin_neon_vtst_v: 3112 case NEON::BI__builtin_neon_vtstq_v: { 3113 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3114 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3115 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3116 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3117 ConstantAggregateZero::get(Ty)); 3118 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3119 } 3120 case NEON::BI__builtin_neon_vuzp_v: 3121 case NEON::BI__builtin_neon_vuzpq_v: { 3122 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3123 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3124 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3125 Value *SV = nullptr; 3126 3127 for (unsigned vi = 0; vi != 2; ++vi) { 3128 SmallVector<Constant*, 16> Indices; 3129 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3130 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3131 3132 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3133 SV = llvm::ConstantVector::get(Indices); 3134 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3135 SV = Builder.CreateStore(SV, Addr); 3136 } 3137 return SV; 3138 } 3139 case NEON::BI__builtin_neon_vzip_v: 3140 case NEON::BI__builtin_neon_vzipq_v: { 3141 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3142 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3143 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3144 Value *SV = nullptr; 3145 3146 for (unsigned vi = 0; vi != 2; ++vi) { 3147 SmallVector<Constant*, 16> Indices; 3148 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3149 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3150 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3151 } 3152 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3153 SV = llvm::ConstantVector::get(Indices); 3154 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3155 SV = Builder.CreateStore(SV, Addr); 3156 } 3157 return SV; 3158 } 3159 } 3160 3161 assert(Int && "Expected valid intrinsic number"); 3162 3163 // Determine the type(s) of this overloaded AArch64 intrinsic. 3164 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3165 3166 Value *Result = EmitNeonCall(F, Ops, NameHint); 3167 llvm::Type *ResultType = ConvertType(E->getType()); 3168 // AArch64 intrinsic one-element vector type cast to 3169 // scalar type expected by the builtin 3170 return Builder.CreateBitCast(Result, ResultType, NameHint); 3171 } 3172 3173 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3174 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3175 const CmpInst::Predicate Ip, const Twine &Name) { 3176 llvm::Type *OTy = Op->getType(); 3177 3178 // FIXME: this is utterly horrific. We should not be looking at previous 3179 // codegen context to find out what needs doing. Unfortunately TableGen 3180 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3181 // (etc). 3182 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3183 OTy = BI->getOperand(0)->getType(); 3184 3185 Op = Builder.CreateBitCast(Op, OTy); 3186 if (OTy->getScalarType()->isFloatingPointTy()) { 3187 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3188 } else { 3189 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3190 } 3191 return Builder.CreateSExt(Op, Ty, Name); 3192 } 3193 3194 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3195 Value *ExtOp, Value *IndexOp, 3196 llvm::Type *ResTy, unsigned IntID, 3197 const char *Name) { 3198 SmallVector<Value *, 2> TblOps; 3199 if (ExtOp) 3200 TblOps.push_back(ExtOp); 3201 3202 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3203 SmallVector<Constant*, 16> Indices; 3204 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3205 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3206 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3207 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3208 } 3209 Value *SV = llvm::ConstantVector::get(Indices); 3210 3211 int PairPos = 0, End = Ops.size() - 1; 3212 while (PairPos < End) { 3213 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3214 Ops[PairPos+1], SV, Name)); 3215 PairPos += 2; 3216 } 3217 3218 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3219 // of the 128-bit lookup table with zero. 3220 if (PairPos == End) { 3221 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3222 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3223 ZeroTbl, SV, Name)); 3224 } 3225 3226 Function *TblF; 3227 TblOps.push_back(IndexOp); 3228 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3229 3230 return CGF.EmitNeonCall(TblF, TblOps, Name); 3231 } 3232 3233 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3234 unsigned Value; 3235 switch (BuiltinID) { 3236 default: 3237 return nullptr; 3238 case ARM::BI__builtin_arm_nop: 3239 Value = 0; 3240 break; 3241 case ARM::BI__builtin_arm_yield: 3242 case ARM::BI__yield: 3243 Value = 1; 3244 break; 3245 case ARM::BI__builtin_arm_wfe: 3246 case ARM::BI__wfe: 3247 Value = 2; 3248 break; 3249 case ARM::BI__builtin_arm_wfi: 3250 case ARM::BI__wfi: 3251 Value = 3; 3252 break; 3253 case ARM::BI__builtin_arm_sev: 3254 case ARM::BI__sev: 3255 Value = 4; 3256 break; 3257 case ARM::BI__builtin_arm_sevl: 3258 case ARM::BI__sevl: 3259 Value = 5; 3260 break; 3261 } 3262 3263 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3264 llvm::ConstantInt::get(Int32Ty, Value)); 3265 } 3266 3267 // Generates the IR for the read/write special register builtin, 3268 // ValueType is the type of the value that is to be written or read, 3269 // RegisterType is the type of the register being written to or read from. 3270 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3271 const CallExpr *E, 3272 llvm::Type *RegisterType, 3273 llvm::Type *ValueType, bool IsRead) { 3274 // write and register intrinsics only support 32 and 64 bit operations. 3275 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3276 && "Unsupported size for register."); 3277 3278 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3279 CodeGen::CodeGenModule &CGM = CGF.CGM; 3280 LLVMContext &Context = CGM.getLLVMContext(); 3281 3282 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3283 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3284 3285 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3286 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3287 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3288 3289 llvm::Type *Types[] = { RegisterType }; 3290 3291 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3292 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3293 && "Can't fit 64-bit value in 32-bit register"); 3294 3295 if (IsRead) { 3296 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3297 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3298 3299 if (MixedTypes) 3300 // Read into 64 bit register and then truncate result to 32 bit. 3301 return Builder.CreateTrunc(Call, ValueType); 3302 3303 if (ValueType->isPointerTy()) 3304 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3305 return Builder.CreateIntToPtr(Call, ValueType); 3306 3307 return Call; 3308 } 3309 3310 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3311 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3312 if (MixedTypes) { 3313 // Extend 32 bit write value to 64 bit to pass to write. 3314 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3315 return Builder.CreateCall(F, { Metadata, ArgValue }); 3316 } 3317 3318 if (ValueType->isPointerTy()) { 3319 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3320 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3321 return Builder.CreateCall(F, { Metadata, ArgValue }); 3322 } 3323 3324 return Builder.CreateCall(F, { Metadata, ArgValue }); 3325 } 3326 3327 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3328 /// argument that specifies the vector type. 3329 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3330 switch (BuiltinID) { 3331 default: break; 3332 case NEON::BI__builtin_neon_vget_lane_i8: 3333 case NEON::BI__builtin_neon_vget_lane_i16: 3334 case NEON::BI__builtin_neon_vget_lane_i32: 3335 case NEON::BI__builtin_neon_vget_lane_i64: 3336 case NEON::BI__builtin_neon_vget_lane_f32: 3337 case NEON::BI__builtin_neon_vgetq_lane_i8: 3338 case NEON::BI__builtin_neon_vgetq_lane_i16: 3339 case NEON::BI__builtin_neon_vgetq_lane_i32: 3340 case NEON::BI__builtin_neon_vgetq_lane_i64: 3341 case NEON::BI__builtin_neon_vgetq_lane_f32: 3342 case NEON::BI__builtin_neon_vset_lane_i8: 3343 case NEON::BI__builtin_neon_vset_lane_i16: 3344 case NEON::BI__builtin_neon_vset_lane_i32: 3345 case NEON::BI__builtin_neon_vset_lane_i64: 3346 case NEON::BI__builtin_neon_vset_lane_f32: 3347 case NEON::BI__builtin_neon_vsetq_lane_i8: 3348 case NEON::BI__builtin_neon_vsetq_lane_i16: 3349 case NEON::BI__builtin_neon_vsetq_lane_i32: 3350 case NEON::BI__builtin_neon_vsetq_lane_i64: 3351 case NEON::BI__builtin_neon_vsetq_lane_f32: 3352 case NEON::BI__builtin_neon_vsha1h_u32: 3353 case NEON::BI__builtin_neon_vsha1cq_u32: 3354 case NEON::BI__builtin_neon_vsha1pq_u32: 3355 case NEON::BI__builtin_neon_vsha1mq_u32: 3356 case ARM::BI_MoveToCoprocessor: 3357 case ARM::BI_MoveToCoprocessor2: 3358 return false; 3359 } 3360 return true; 3361 } 3362 3363 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3364 const CallExpr *E) { 3365 if (auto Hint = GetValueForARMHint(BuiltinID)) 3366 return Hint; 3367 3368 if (BuiltinID == ARM::BI__emit) { 3369 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3370 llvm::FunctionType *FTy = 3371 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3372 3373 APSInt Value; 3374 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3375 llvm_unreachable("Sema will ensure that the parameter is constant"); 3376 3377 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3378 3379 llvm::InlineAsm *Emit = 3380 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3381 /*SideEffects=*/true) 3382 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3383 /*SideEffects=*/true); 3384 3385 return Builder.CreateCall(Emit); 3386 } 3387 3388 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3389 Value *Option = EmitScalarExpr(E->getArg(0)); 3390 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3391 } 3392 3393 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3394 Value *Address = EmitScalarExpr(E->getArg(0)); 3395 Value *RW = EmitScalarExpr(E->getArg(1)); 3396 Value *IsData = EmitScalarExpr(E->getArg(2)); 3397 3398 // Locality is not supported on ARM target 3399 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3400 3401 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3402 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3403 } 3404 3405 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3406 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3407 EmitScalarExpr(E->getArg(0)), 3408 "rbit"); 3409 } 3410 3411 if (BuiltinID == ARM::BI__clear_cache) { 3412 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3413 const FunctionDecl *FD = E->getDirectCallee(); 3414 Value *Ops[2]; 3415 for (unsigned i = 0; i < 2; i++) 3416 Ops[i] = EmitScalarExpr(E->getArg(i)); 3417 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3418 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3419 StringRef Name = FD->getName(); 3420 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3421 } 3422 3423 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3424 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3425 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3426 getContext().getTypeSize(E->getType()) == 64) || 3427 BuiltinID == ARM::BI__ldrexd) { 3428 Function *F; 3429 3430 switch (BuiltinID) { 3431 default: llvm_unreachable("unexpected builtin"); 3432 case ARM::BI__builtin_arm_ldaex: 3433 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3434 break; 3435 case ARM::BI__builtin_arm_ldrexd: 3436 case ARM::BI__builtin_arm_ldrex: 3437 case ARM::BI__ldrexd: 3438 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3439 break; 3440 } 3441 3442 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3443 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3444 "ldrexd"); 3445 3446 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3447 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3448 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3449 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3450 3451 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3452 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3453 Val = Builder.CreateOr(Val, Val1); 3454 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3455 } 3456 3457 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3458 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3459 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3460 3461 QualType Ty = E->getType(); 3462 llvm::Type *RealResTy = ConvertType(Ty); 3463 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3464 getContext().getTypeSize(Ty)); 3465 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3466 3467 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3468 ? Intrinsic::arm_ldaex 3469 : Intrinsic::arm_ldrex, 3470 LoadAddr->getType()); 3471 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3472 3473 if (RealResTy->isPointerTy()) 3474 return Builder.CreateIntToPtr(Val, RealResTy); 3475 else { 3476 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3477 return Builder.CreateBitCast(Val, RealResTy); 3478 } 3479 } 3480 3481 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3482 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3483 BuiltinID == ARM::BI__builtin_arm_strex) && 3484 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3485 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3486 ? Intrinsic::arm_stlexd 3487 : Intrinsic::arm_strexd); 3488 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3489 3490 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3491 Value *Val = EmitScalarExpr(E->getArg(0)); 3492 Builder.CreateStore(Val, Tmp); 3493 3494 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3495 Val = Builder.CreateLoad(LdPtr); 3496 3497 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3498 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3499 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3500 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3501 } 3502 3503 if (BuiltinID == ARM::BI__builtin_arm_strex || 3504 BuiltinID == ARM::BI__builtin_arm_stlex) { 3505 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3506 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3507 3508 QualType Ty = E->getArg(0)->getType(); 3509 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3510 getContext().getTypeSize(Ty)); 3511 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3512 3513 if (StoreVal->getType()->isPointerTy()) 3514 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3515 else { 3516 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3517 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3518 } 3519 3520 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3521 ? Intrinsic::arm_stlex 3522 : Intrinsic::arm_strex, 3523 StoreAddr->getType()); 3524 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3525 } 3526 3527 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3528 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3529 return Builder.CreateCall(F); 3530 } 3531 3532 // CRC32 3533 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3534 switch (BuiltinID) { 3535 case ARM::BI__builtin_arm_crc32b: 3536 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3537 case ARM::BI__builtin_arm_crc32cb: 3538 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3539 case ARM::BI__builtin_arm_crc32h: 3540 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3541 case ARM::BI__builtin_arm_crc32ch: 3542 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3543 case ARM::BI__builtin_arm_crc32w: 3544 case ARM::BI__builtin_arm_crc32d: 3545 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3546 case ARM::BI__builtin_arm_crc32cw: 3547 case ARM::BI__builtin_arm_crc32cd: 3548 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3549 } 3550 3551 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3552 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3553 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3554 3555 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3556 // intrinsics, hence we need different codegen for these cases. 3557 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3558 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3559 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3560 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3561 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3562 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3563 3564 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3565 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3566 return Builder.CreateCall(F, {Res, Arg1b}); 3567 } else { 3568 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3569 3570 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3571 return Builder.CreateCall(F, {Arg0, Arg1}); 3572 } 3573 } 3574 3575 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3576 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3577 BuiltinID == ARM::BI__builtin_arm_rsrp || 3578 BuiltinID == ARM::BI__builtin_arm_wsr || 3579 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3580 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3581 3582 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3583 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3584 BuiltinID == ARM::BI__builtin_arm_rsrp; 3585 3586 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3587 BuiltinID == ARM::BI__builtin_arm_wsrp; 3588 3589 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3590 BuiltinID == ARM::BI__builtin_arm_wsr64; 3591 3592 llvm::Type *ValueType; 3593 llvm::Type *RegisterType; 3594 if (IsPointerBuiltin) { 3595 ValueType = VoidPtrTy; 3596 RegisterType = Int32Ty; 3597 } else if (Is64Bit) { 3598 ValueType = RegisterType = Int64Ty; 3599 } else { 3600 ValueType = RegisterType = Int32Ty; 3601 } 3602 3603 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3604 } 3605 3606 // Find out if any arguments are required to be integer constant 3607 // expressions. 3608 unsigned ICEArguments = 0; 3609 ASTContext::GetBuiltinTypeError Error; 3610 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3611 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3612 3613 SmallVector<Value*, 4> Ops; 3614 llvm::Value *Align = nullptr; 3615 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3616 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3617 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3618 if (i == 0) { 3619 switch (BuiltinID) { 3620 case NEON::BI__builtin_neon_vld1_v: 3621 case NEON::BI__builtin_neon_vld1q_v: 3622 case NEON::BI__builtin_neon_vld1q_lane_v: 3623 case NEON::BI__builtin_neon_vld1_lane_v: 3624 case NEON::BI__builtin_neon_vld1_dup_v: 3625 case NEON::BI__builtin_neon_vld1q_dup_v: 3626 case NEON::BI__builtin_neon_vst1_v: 3627 case NEON::BI__builtin_neon_vst1q_v: 3628 case NEON::BI__builtin_neon_vst1q_lane_v: 3629 case NEON::BI__builtin_neon_vst1_lane_v: 3630 case NEON::BI__builtin_neon_vst2_v: 3631 case NEON::BI__builtin_neon_vst2q_v: 3632 case NEON::BI__builtin_neon_vst2_lane_v: 3633 case NEON::BI__builtin_neon_vst2q_lane_v: 3634 case NEON::BI__builtin_neon_vst3_v: 3635 case NEON::BI__builtin_neon_vst3q_v: 3636 case NEON::BI__builtin_neon_vst3_lane_v: 3637 case NEON::BI__builtin_neon_vst3q_lane_v: 3638 case NEON::BI__builtin_neon_vst4_v: 3639 case NEON::BI__builtin_neon_vst4q_v: 3640 case NEON::BI__builtin_neon_vst4_lane_v: 3641 case NEON::BI__builtin_neon_vst4q_lane_v: 3642 // Get the alignment for the argument in addition to the value; 3643 // we'll use it later. 3644 std::pair<llvm::Value*, unsigned> Src = 3645 EmitPointerWithAlignment(E->getArg(0)); 3646 Ops.push_back(Src.first); 3647 Align = Builder.getInt32(Src.second); 3648 continue; 3649 } 3650 } 3651 if (i == 1) { 3652 switch (BuiltinID) { 3653 case NEON::BI__builtin_neon_vld2_v: 3654 case NEON::BI__builtin_neon_vld2q_v: 3655 case NEON::BI__builtin_neon_vld3_v: 3656 case NEON::BI__builtin_neon_vld3q_v: 3657 case NEON::BI__builtin_neon_vld4_v: 3658 case NEON::BI__builtin_neon_vld4q_v: 3659 case NEON::BI__builtin_neon_vld2_lane_v: 3660 case NEON::BI__builtin_neon_vld2q_lane_v: 3661 case NEON::BI__builtin_neon_vld3_lane_v: 3662 case NEON::BI__builtin_neon_vld3q_lane_v: 3663 case NEON::BI__builtin_neon_vld4_lane_v: 3664 case NEON::BI__builtin_neon_vld4q_lane_v: 3665 case NEON::BI__builtin_neon_vld2_dup_v: 3666 case NEON::BI__builtin_neon_vld3_dup_v: 3667 case NEON::BI__builtin_neon_vld4_dup_v: 3668 // Get the alignment for the argument in addition to the value; 3669 // we'll use it later. 3670 std::pair<llvm::Value*, unsigned> Src = 3671 EmitPointerWithAlignment(E->getArg(1)); 3672 Ops.push_back(Src.first); 3673 Align = Builder.getInt32(Src.second); 3674 continue; 3675 } 3676 } 3677 3678 if ((ICEArguments & (1 << i)) == 0) { 3679 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3680 } else { 3681 // If this is required to be a constant, constant fold it so that we know 3682 // that the generated intrinsic gets a ConstantInt. 3683 llvm::APSInt Result; 3684 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 3685 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 3686 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 3687 } 3688 } 3689 3690 switch (BuiltinID) { 3691 default: break; 3692 3693 case NEON::BI__builtin_neon_vget_lane_i8: 3694 case NEON::BI__builtin_neon_vget_lane_i16: 3695 case NEON::BI__builtin_neon_vget_lane_i32: 3696 case NEON::BI__builtin_neon_vget_lane_i64: 3697 case NEON::BI__builtin_neon_vget_lane_f32: 3698 case NEON::BI__builtin_neon_vgetq_lane_i8: 3699 case NEON::BI__builtin_neon_vgetq_lane_i16: 3700 case NEON::BI__builtin_neon_vgetq_lane_i32: 3701 case NEON::BI__builtin_neon_vgetq_lane_i64: 3702 case NEON::BI__builtin_neon_vgetq_lane_f32: 3703 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 3704 3705 case NEON::BI__builtin_neon_vset_lane_i8: 3706 case NEON::BI__builtin_neon_vset_lane_i16: 3707 case NEON::BI__builtin_neon_vset_lane_i32: 3708 case NEON::BI__builtin_neon_vset_lane_i64: 3709 case NEON::BI__builtin_neon_vset_lane_f32: 3710 case NEON::BI__builtin_neon_vsetq_lane_i8: 3711 case NEON::BI__builtin_neon_vsetq_lane_i16: 3712 case NEON::BI__builtin_neon_vsetq_lane_i32: 3713 case NEON::BI__builtin_neon_vsetq_lane_i64: 3714 case NEON::BI__builtin_neon_vsetq_lane_f32: 3715 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3716 3717 case NEON::BI__builtin_neon_vsha1h_u32: 3718 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3719 "vsha1h"); 3720 case NEON::BI__builtin_neon_vsha1cq_u32: 3721 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3722 "vsha1h"); 3723 case NEON::BI__builtin_neon_vsha1pq_u32: 3724 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3725 "vsha1h"); 3726 case NEON::BI__builtin_neon_vsha1mq_u32: 3727 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3728 "vsha1h"); 3729 3730 // The ARM _MoveToCoprocessor builtins put the input register value as 3731 // the first argument, but the LLVM intrinsic expects it as the third one. 3732 case ARM::BI_MoveToCoprocessor: 3733 case ARM::BI_MoveToCoprocessor2: { 3734 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 3735 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 3736 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 3737 Ops[3], Ops[4], Ops[5]}); 3738 } 3739 } 3740 3741 // Get the last argument, which specifies the vector type. 3742 assert(HasExtraArg); 3743 llvm::APSInt Result; 3744 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3745 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3746 return nullptr; 3747 3748 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3749 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3750 // Determine the overloaded type of this builtin. 3751 llvm::Type *Ty; 3752 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3753 Ty = FloatTy; 3754 else 3755 Ty = DoubleTy; 3756 3757 // Determine whether this is an unsigned conversion or not. 3758 bool usgn = Result.getZExtValue() == 1; 3759 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3760 3761 // Call the appropriate intrinsic. 3762 Function *F = CGM.getIntrinsic(Int, Ty); 3763 return Builder.CreateCall(F, Ops, "vcvtr"); 3764 } 3765 3766 // Determine the type of this overloaded NEON intrinsic. 3767 NeonTypeFlags Type(Result.getZExtValue()); 3768 bool usgn = Type.isUnsigned(); 3769 bool rightShift = false; 3770 3771 llvm::VectorType *VTy = GetNeonType(this, Type); 3772 llvm::Type *Ty = VTy; 3773 if (!Ty) 3774 return nullptr; 3775 3776 // Many NEON builtins have identical semantics and uses in ARM and 3777 // AArch64. Emit these in a single function. 3778 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3779 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3780 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3781 if (Builtin) 3782 return EmitCommonNeonBuiltinExpr( 3783 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3784 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3785 3786 unsigned Int; 3787 switch (BuiltinID) { 3788 default: return nullptr; 3789 case NEON::BI__builtin_neon_vld1q_lane_v: 3790 // Handle 64-bit integer elements as a special case. Use shuffles of 3791 // one-element vectors to avoid poor code for i64 in the backend. 3792 if (VTy->getElementType()->isIntegerTy(64)) { 3793 // Extract the other lane. 3794 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3795 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3796 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3797 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3798 // Load the value as a one-element vector. 3799 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3800 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3801 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3802 // Combine them. 3803 uint32_t Indices[] = {1 - Lane, Lane}; 3804 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 3805 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3806 } 3807 // fall through 3808 case NEON::BI__builtin_neon_vld1_lane_v: { 3809 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3810 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3811 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3812 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3813 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3814 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3815 } 3816 case NEON::BI__builtin_neon_vld2_dup_v: 3817 case NEON::BI__builtin_neon_vld3_dup_v: 3818 case NEON::BI__builtin_neon_vld4_dup_v: { 3819 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3820 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3821 switch (BuiltinID) { 3822 case NEON::BI__builtin_neon_vld2_dup_v: 3823 Int = Intrinsic::arm_neon_vld2; 3824 break; 3825 case NEON::BI__builtin_neon_vld3_dup_v: 3826 Int = Intrinsic::arm_neon_vld3; 3827 break; 3828 case NEON::BI__builtin_neon_vld4_dup_v: 3829 Int = Intrinsic::arm_neon_vld4; 3830 break; 3831 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3832 } 3833 Function *F = CGM.getIntrinsic(Int, Ty); 3834 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3835 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3836 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3837 return Builder.CreateStore(Ops[1], Ops[0]); 3838 } 3839 switch (BuiltinID) { 3840 case NEON::BI__builtin_neon_vld2_dup_v: 3841 Int = Intrinsic::arm_neon_vld2lane; 3842 break; 3843 case NEON::BI__builtin_neon_vld3_dup_v: 3844 Int = Intrinsic::arm_neon_vld3lane; 3845 break; 3846 case NEON::BI__builtin_neon_vld4_dup_v: 3847 Int = Intrinsic::arm_neon_vld4lane; 3848 break; 3849 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3850 } 3851 Function *F = CGM.getIntrinsic(Int, Ty); 3852 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3853 3854 SmallVector<Value*, 6> Args; 3855 Args.push_back(Ops[1]); 3856 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3857 3858 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3859 Args.push_back(CI); 3860 Args.push_back(Align); 3861 3862 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3863 // splat lane 0 to all elts in each vector of the result. 3864 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3865 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3866 Value *Elt = Builder.CreateBitCast(Val, Ty); 3867 Elt = EmitNeonSplat(Elt, CI); 3868 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3869 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3870 } 3871 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3872 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3873 return Builder.CreateStore(Ops[1], Ops[0]); 3874 } 3875 case NEON::BI__builtin_neon_vqrshrn_n_v: 3876 Int = 3877 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3878 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3879 1, true); 3880 case NEON::BI__builtin_neon_vqrshrun_n_v: 3881 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3882 Ops, "vqrshrun_n", 1, true); 3883 case NEON::BI__builtin_neon_vqshrn_n_v: 3884 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3885 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3886 1, true); 3887 case NEON::BI__builtin_neon_vqshrun_n_v: 3888 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3889 Ops, "vqshrun_n", 1, true); 3890 case NEON::BI__builtin_neon_vrecpe_v: 3891 case NEON::BI__builtin_neon_vrecpeq_v: 3892 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3893 Ops, "vrecpe"); 3894 case NEON::BI__builtin_neon_vrshrn_n_v: 3895 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3896 Ops, "vrshrn_n", 1, true); 3897 case NEON::BI__builtin_neon_vrsra_n_v: 3898 case NEON::BI__builtin_neon_vrsraq_n_v: 3899 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3900 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3901 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3902 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3903 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 3904 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3905 case NEON::BI__builtin_neon_vsri_n_v: 3906 case NEON::BI__builtin_neon_vsriq_n_v: 3907 rightShift = true; 3908 case NEON::BI__builtin_neon_vsli_n_v: 3909 case NEON::BI__builtin_neon_vsliq_n_v: 3910 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3911 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3912 Ops, "vsli_n"); 3913 case NEON::BI__builtin_neon_vsra_n_v: 3914 case NEON::BI__builtin_neon_vsraq_n_v: 3915 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3916 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3917 return Builder.CreateAdd(Ops[0], Ops[1]); 3918 case NEON::BI__builtin_neon_vst1q_lane_v: 3919 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3920 // a one-element vector and avoid poor code for i64 in the backend. 3921 if (VTy->getElementType()->isIntegerTy(64)) { 3922 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3923 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3924 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3925 Ops[2] = Align; 3926 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3927 Ops[1]->getType()), Ops); 3928 } 3929 // fall through 3930 case NEON::BI__builtin_neon_vst1_lane_v: { 3931 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3932 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3933 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3934 StoreInst *St = Builder.CreateStore(Ops[1], 3935 Builder.CreateBitCast(Ops[0], Ty)); 3936 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3937 return St; 3938 } 3939 case NEON::BI__builtin_neon_vtbl1_v: 3940 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3941 Ops, "vtbl1"); 3942 case NEON::BI__builtin_neon_vtbl2_v: 3943 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3944 Ops, "vtbl2"); 3945 case NEON::BI__builtin_neon_vtbl3_v: 3946 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3947 Ops, "vtbl3"); 3948 case NEON::BI__builtin_neon_vtbl4_v: 3949 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3950 Ops, "vtbl4"); 3951 case NEON::BI__builtin_neon_vtbx1_v: 3952 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3953 Ops, "vtbx1"); 3954 case NEON::BI__builtin_neon_vtbx2_v: 3955 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3956 Ops, "vtbx2"); 3957 case NEON::BI__builtin_neon_vtbx3_v: 3958 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3959 Ops, "vtbx3"); 3960 case NEON::BI__builtin_neon_vtbx4_v: 3961 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3962 Ops, "vtbx4"); 3963 } 3964 } 3965 3966 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3967 const CallExpr *E, 3968 SmallVectorImpl<Value *> &Ops) { 3969 unsigned int Int = 0; 3970 const char *s = nullptr; 3971 3972 switch (BuiltinID) { 3973 default: 3974 return nullptr; 3975 case NEON::BI__builtin_neon_vtbl1_v: 3976 case NEON::BI__builtin_neon_vqtbl1_v: 3977 case NEON::BI__builtin_neon_vqtbl1q_v: 3978 case NEON::BI__builtin_neon_vtbl2_v: 3979 case NEON::BI__builtin_neon_vqtbl2_v: 3980 case NEON::BI__builtin_neon_vqtbl2q_v: 3981 case NEON::BI__builtin_neon_vtbl3_v: 3982 case NEON::BI__builtin_neon_vqtbl3_v: 3983 case NEON::BI__builtin_neon_vqtbl3q_v: 3984 case NEON::BI__builtin_neon_vtbl4_v: 3985 case NEON::BI__builtin_neon_vqtbl4_v: 3986 case NEON::BI__builtin_neon_vqtbl4q_v: 3987 break; 3988 case NEON::BI__builtin_neon_vtbx1_v: 3989 case NEON::BI__builtin_neon_vqtbx1_v: 3990 case NEON::BI__builtin_neon_vqtbx1q_v: 3991 case NEON::BI__builtin_neon_vtbx2_v: 3992 case NEON::BI__builtin_neon_vqtbx2_v: 3993 case NEON::BI__builtin_neon_vqtbx2q_v: 3994 case NEON::BI__builtin_neon_vtbx3_v: 3995 case NEON::BI__builtin_neon_vqtbx3_v: 3996 case NEON::BI__builtin_neon_vqtbx3q_v: 3997 case NEON::BI__builtin_neon_vtbx4_v: 3998 case NEON::BI__builtin_neon_vqtbx4_v: 3999 case NEON::BI__builtin_neon_vqtbx4q_v: 4000 break; 4001 } 4002 4003 assert(E->getNumArgs() >= 3); 4004 4005 // Get the last argument, which specifies the vector type. 4006 llvm::APSInt Result; 4007 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4008 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4009 return nullptr; 4010 4011 // Determine the type of this overloaded NEON intrinsic. 4012 NeonTypeFlags Type(Result.getZExtValue()); 4013 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4014 if (!Ty) 4015 return nullptr; 4016 4017 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4018 4019 // AArch64 scalar builtins are not overloaded, they do not have an extra 4020 // argument that specifies the vector type, need to handle each case. 4021 switch (BuiltinID) { 4022 case NEON::BI__builtin_neon_vtbl1_v: { 4023 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4024 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4025 "vtbl1"); 4026 } 4027 case NEON::BI__builtin_neon_vtbl2_v: { 4028 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4029 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4030 "vtbl1"); 4031 } 4032 case NEON::BI__builtin_neon_vtbl3_v: { 4033 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4034 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4035 "vtbl2"); 4036 } 4037 case NEON::BI__builtin_neon_vtbl4_v: { 4038 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4039 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4040 "vtbl2"); 4041 } 4042 case NEON::BI__builtin_neon_vtbx1_v: { 4043 Value *TblRes = 4044 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4045 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4046 4047 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4048 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4049 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4050 4051 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4052 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4053 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4054 } 4055 case NEON::BI__builtin_neon_vtbx2_v: { 4056 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4057 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4058 "vtbx1"); 4059 } 4060 case NEON::BI__builtin_neon_vtbx3_v: { 4061 Value *TblRes = 4062 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4063 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4064 4065 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4066 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4067 TwentyFourV); 4068 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4069 4070 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4071 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4072 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4073 } 4074 case NEON::BI__builtin_neon_vtbx4_v: { 4075 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4076 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4077 "vtbx2"); 4078 } 4079 case NEON::BI__builtin_neon_vqtbl1_v: 4080 case NEON::BI__builtin_neon_vqtbl1q_v: 4081 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4082 case NEON::BI__builtin_neon_vqtbl2_v: 4083 case NEON::BI__builtin_neon_vqtbl2q_v: { 4084 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4085 case NEON::BI__builtin_neon_vqtbl3_v: 4086 case NEON::BI__builtin_neon_vqtbl3q_v: 4087 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4088 case NEON::BI__builtin_neon_vqtbl4_v: 4089 case NEON::BI__builtin_neon_vqtbl4q_v: 4090 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4091 case NEON::BI__builtin_neon_vqtbx1_v: 4092 case NEON::BI__builtin_neon_vqtbx1q_v: 4093 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4094 case NEON::BI__builtin_neon_vqtbx2_v: 4095 case NEON::BI__builtin_neon_vqtbx2q_v: 4096 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4097 case NEON::BI__builtin_neon_vqtbx3_v: 4098 case NEON::BI__builtin_neon_vqtbx3q_v: 4099 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4100 case NEON::BI__builtin_neon_vqtbx4_v: 4101 case NEON::BI__builtin_neon_vqtbx4q_v: 4102 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4103 } 4104 } 4105 4106 if (!Int) 4107 return nullptr; 4108 4109 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4110 return CGF.EmitNeonCall(F, Ops, s); 4111 } 4112 4113 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4114 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4115 Op = Builder.CreateBitCast(Op, Int16Ty); 4116 Value *V = UndefValue::get(VTy); 4117 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4118 Op = Builder.CreateInsertElement(V, Op, CI); 4119 return Op; 4120 } 4121 4122 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 4123 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4124 Op = Builder.CreateBitCast(Op, Int8Ty); 4125 Value *V = UndefValue::get(VTy); 4126 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4127 Op = Builder.CreateInsertElement(V, Op, CI); 4128 return Op; 4129 } 4130 4131 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4132 const CallExpr *E) { 4133 unsigned HintID = static_cast<unsigned>(-1); 4134 switch (BuiltinID) { 4135 default: break; 4136 case AArch64::BI__builtin_arm_nop: 4137 HintID = 0; 4138 break; 4139 case AArch64::BI__builtin_arm_yield: 4140 HintID = 1; 4141 break; 4142 case AArch64::BI__builtin_arm_wfe: 4143 HintID = 2; 4144 break; 4145 case AArch64::BI__builtin_arm_wfi: 4146 HintID = 3; 4147 break; 4148 case AArch64::BI__builtin_arm_sev: 4149 HintID = 4; 4150 break; 4151 case AArch64::BI__builtin_arm_sevl: 4152 HintID = 5; 4153 break; 4154 } 4155 4156 if (HintID != static_cast<unsigned>(-1)) { 4157 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4158 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4159 } 4160 4161 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4162 Value *Address = EmitScalarExpr(E->getArg(0)); 4163 Value *RW = EmitScalarExpr(E->getArg(1)); 4164 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4165 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4166 Value *IsData = EmitScalarExpr(E->getArg(4)); 4167 4168 Value *Locality = nullptr; 4169 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4170 // Temporal fetch, needs to convert cache level to locality. 4171 Locality = llvm::ConstantInt::get(Int32Ty, 4172 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4173 } else { 4174 // Streaming fetch. 4175 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4176 } 4177 4178 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4179 // PLDL3STRM or PLDL2STRM. 4180 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4181 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4182 } 4183 4184 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4185 assert((getContext().getTypeSize(E->getType()) == 32) && 4186 "rbit of unusual size!"); 4187 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4188 return Builder.CreateCall( 4189 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4190 } 4191 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4192 assert((getContext().getTypeSize(E->getType()) == 64) && 4193 "rbit of unusual size!"); 4194 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4195 return Builder.CreateCall( 4196 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4197 } 4198 4199 if (BuiltinID == AArch64::BI__clear_cache) { 4200 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4201 const FunctionDecl *FD = E->getDirectCallee(); 4202 Value *Ops[2]; 4203 for (unsigned i = 0; i < 2; i++) 4204 Ops[i] = EmitScalarExpr(E->getArg(i)); 4205 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4206 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4207 StringRef Name = FD->getName(); 4208 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4209 } 4210 4211 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4212 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4213 getContext().getTypeSize(E->getType()) == 128) { 4214 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4215 ? Intrinsic::aarch64_ldaxp 4216 : Intrinsic::aarch64_ldxp); 4217 4218 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4219 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4220 "ldxp"); 4221 4222 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4223 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4224 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4225 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4226 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4227 4228 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4229 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4230 Val = Builder.CreateOr(Val, Val1); 4231 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4232 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4233 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4234 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4235 4236 QualType Ty = E->getType(); 4237 llvm::Type *RealResTy = ConvertType(Ty); 4238 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4239 getContext().getTypeSize(Ty)); 4240 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4241 4242 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4243 ? Intrinsic::aarch64_ldaxr 4244 : Intrinsic::aarch64_ldxr, 4245 LoadAddr->getType()); 4246 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4247 4248 if (RealResTy->isPointerTy()) 4249 return Builder.CreateIntToPtr(Val, RealResTy); 4250 4251 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4252 return Builder.CreateBitCast(Val, RealResTy); 4253 } 4254 4255 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4256 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4257 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4258 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4259 ? Intrinsic::aarch64_stlxp 4260 : Intrinsic::aarch64_stxp); 4261 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4262 4263 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4264 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4265 One); 4266 Value *Val = EmitScalarExpr(E->getArg(0)); 4267 Builder.CreateStore(Val, Tmp); 4268 4269 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4270 Val = Builder.CreateLoad(LdPtr); 4271 4272 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4273 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4274 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4275 Int8PtrTy); 4276 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4277 } 4278 4279 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4280 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4281 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4282 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4283 4284 QualType Ty = E->getArg(0)->getType(); 4285 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4286 getContext().getTypeSize(Ty)); 4287 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4288 4289 if (StoreVal->getType()->isPointerTy()) 4290 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4291 else { 4292 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4293 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4294 } 4295 4296 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4297 ? Intrinsic::aarch64_stlxr 4298 : Intrinsic::aarch64_stxr, 4299 StoreAddr->getType()); 4300 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4301 } 4302 4303 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4304 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4305 return Builder.CreateCall(F); 4306 } 4307 4308 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4309 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4310 return Builder.CreateCall(F); 4311 } 4312 4313 // CRC32 4314 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4315 switch (BuiltinID) { 4316 case AArch64::BI__builtin_arm_crc32b: 4317 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4318 case AArch64::BI__builtin_arm_crc32cb: 4319 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4320 case AArch64::BI__builtin_arm_crc32h: 4321 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4322 case AArch64::BI__builtin_arm_crc32ch: 4323 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4324 case AArch64::BI__builtin_arm_crc32w: 4325 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4326 case AArch64::BI__builtin_arm_crc32cw: 4327 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4328 case AArch64::BI__builtin_arm_crc32d: 4329 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4330 case AArch64::BI__builtin_arm_crc32cd: 4331 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4332 } 4333 4334 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4335 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4336 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4337 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4338 4339 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4340 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4341 4342 return Builder.CreateCall(F, {Arg0, Arg1}); 4343 } 4344 4345 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4346 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4347 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4348 BuiltinID == AArch64::BI__builtin_arm_wsr || 4349 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4350 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4351 4352 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4353 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4354 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4355 4356 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4357 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4358 4359 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4360 BuiltinID != AArch64::BI__builtin_arm_wsr; 4361 4362 llvm::Type *ValueType; 4363 llvm::Type *RegisterType = Int64Ty; 4364 if (IsPointerBuiltin) { 4365 ValueType = VoidPtrTy; 4366 } else if (Is64Bit) { 4367 ValueType = Int64Ty; 4368 } else { 4369 ValueType = Int32Ty; 4370 } 4371 4372 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4373 } 4374 4375 // Find out if any arguments are required to be integer constant 4376 // expressions. 4377 unsigned ICEArguments = 0; 4378 ASTContext::GetBuiltinTypeError Error; 4379 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4380 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4381 4382 llvm::SmallVector<Value*, 4> Ops; 4383 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4384 if ((ICEArguments & (1 << i)) == 0) { 4385 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4386 } else { 4387 // If this is required to be a constant, constant fold it so that we know 4388 // that the generated intrinsic gets a ConstantInt. 4389 llvm::APSInt Result; 4390 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4391 assert(IsConst && "Constant arg isn't actually constant?"); 4392 (void)IsConst; 4393 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4394 } 4395 } 4396 4397 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4398 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4399 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4400 4401 if (Builtin) { 4402 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4403 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4404 assert(Result && "SISD intrinsic should have been handled"); 4405 return Result; 4406 } 4407 4408 llvm::APSInt Result; 4409 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4410 NeonTypeFlags Type(0); 4411 if (Arg->isIntegerConstantExpr(Result, getContext())) 4412 // Determine the type of this overloaded NEON intrinsic. 4413 Type = NeonTypeFlags(Result.getZExtValue()); 4414 4415 bool usgn = Type.isUnsigned(); 4416 bool quad = Type.isQuad(); 4417 4418 // Handle non-overloaded intrinsics first. 4419 switch (BuiltinID) { 4420 default: break; 4421 case NEON::BI__builtin_neon_vldrq_p128: { 4422 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4423 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4424 return Builder.CreateLoad(Ptr); 4425 } 4426 case NEON::BI__builtin_neon_vstrq_p128: { 4427 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4428 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4429 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4430 } 4431 case NEON::BI__builtin_neon_vcvts_u32_f32: 4432 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4433 usgn = true; 4434 // FALL THROUGH 4435 case NEON::BI__builtin_neon_vcvts_s32_f32: 4436 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4437 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4438 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4439 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4440 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4441 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4442 if (usgn) 4443 return Builder.CreateFPToUI(Ops[0], InTy); 4444 return Builder.CreateFPToSI(Ops[0], InTy); 4445 } 4446 case NEON::BI__builtin_neon_vcvts_f32_u32: 4447 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4448 usgn = true; 4449 // FALL THROUGH 4450 case NEON::BI__builtin_neon_vcvts_f32_s32: 4451 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4452 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4453 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4454 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4455 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4456 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4457 if (usgn) 4458 return Builder.CreateUIToFP(Ops[0], FTy); 4459 return Builder.CreateSIToFP(Ops[0], FTy); 4460 } 4461 case NEON::BI__builtin_neon_vpaddd_s64: { 4462 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4463 Value *Vec = EmitScalarExpr(E->getArg(0)); 4464 // The vector is v2f64, so make sure it's bitcast to that. 4465 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4466 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4467 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4468 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4469 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4470 // Pairwise addition of a v2f64 into a scalar f64. 4471 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4472 } 4473 case NEON::BI__builtin_neon_vpaddd_f64: { 4474 llvm::Type *Ty = 4475 llvm::VectorType::get(DoubleTy, 2); 4476 Value *Vec = EmitScalarExpr(E->getArg(0)); 4477 // The vector is v2f64, so make sure it's bitcast to that. 4478 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4479 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4480 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4481 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4482 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4483 // Pairwise addition of a v2f64 into a scalar f64. 4484 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4485 } 4486 case NEON::BI__builtin_neon_vpadds_f32: { 4487 llvm::Type *Ty = 4488 llvm::VectorType::get(FloatTy, 2); 4489 Value *Vec = EmitScalarExpr(E->getArg(0)); 4490 // The vector is v2f32, so make sure it's bitcast to that. 4491 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4492 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4493 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4494 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4495 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4496 // Pairwise addition of a v2f32 into a scalar f32. 4497 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4498 } 4499 case NEON::BI__builtin_neon_vceqzd_s64: 4500 case NEON::BI__builtin_neon_vceqzd_f64: 4501 case NEON::BI__builtin_neon_vceqzs_f32: 4502 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4503 return EmitAArch64CompareBuiltinExpr( 4504 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4505 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4506 case NEON::BI__builtin_neon_vcgezd_s64: 4507 case NEON::BI__builtin_neon_vcgezd_f64: 4508 case NEON::BI__builtin_neon_vcgezs_f32: 4509 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4510 return EmitAArch64CompareBuiltinExpr( 4511 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4512 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4513 case NEON::BI__builtin_neon_vclezd_s64: 4514 case NEON::BI__builtin_neon_vclezd_f64: 4515 case NEON::BI__builtin_neon_vclezs_f32: 4516 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4517 return EmitAArch64CompareBuiltinExpr( 4518 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4519 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4520 case NEON::BI__builtin_neon_vcgtzd_s64: 4521 case NEON::BI__builtin_neon_vcgtzd_f64: 4522 case NEON::BI__builtin_neon_vcgtzs_f32: 4523 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4524 return EmitAArch64CompareBuiltinExpr( 4525 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4526 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4527 case NEON::BI__builtin_neon_vcltzd_s64: 4528 case NEON::BI__builtin_neon_vcltzd_f64: 4529 case NEON::BI__builtin_neon_vcltzs_f32: 4530 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4531 return EmitAArch64CompareBuiltinExpr( 4532 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4533 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4534 4535 case NEON::BI__builtin_neon_vceqzd_u64: { 4536 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4537 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4538 Ops[0] = 4539 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4540 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4541 } 4542 case NEON::BI__builtin_neon_vceqd_f64: 4543 case NEON::BI__builtin_neon_vcled_f64: 4544 case NEON::BI__builtin_neon_vcltd_f64: 4545 case NEON::BI__builtin_neon_vcged_f64: 4546 case NEON::BI__builtin_neon_vcgtd_f64: { 4547 llvm::CmpInst::Predicate P; 4548 switch (BuiltinID) { 4549 default: llvm_unreachable("missing builtin ID in switch!"); 4550 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4551 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4552 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4553 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4554 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4555 } 4556 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4557 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4558 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4559 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4560 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4561 } 4562 case NEON::BI__builtin_neon_vceqs_f32: 4563 case NEON::BI__builtin_neon_vcles_f32: 4564 case NEON::BI__builtin_neon_vclts_f32: 4565 case NEON::BI__builtin_neon_vcges_f32: 4566 case NEON::BI__builtin_neon_vcgts_f32: { 4567 llvm::CmpInst::Predicate P; 4568 switch (BuiltinID) { 4569 default: llvm_unreachable("missing builtin ID in switch!"); 4570 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4571 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4572 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4573 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4574 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4575 } 4576 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4577 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4578 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4579 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4580 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4581 } 4582 case NEON::BI__builtin_neon_vceqd_s64: 4583 case NEON::BI__builtin_neon_vceqd_u64: 4584 case NEON::BI__builtin_neon_vcgtd_s64: 4585 case NEON::BI__builtin_neon_vcgtd_u64: 4586 case NEON::BI__builtin_neon_vcltd_s64: 4587 case NEON::BI__builtin_neon_vcltd_u64: 4588 case NEON::BI__builtin_neon_vcged_u64: 4589 case NEON::BI__builtin_neon_vcged_s64: 4590 case NEON::BI__builtin_neon_vcled_u64: 4591 case NEON::BI__builtin_neon_vcled_s64: { 4592 llvm::CmpInst::Predicate P; 4593 switch (BuiltinID) { 4594 default: llvm_unreachable("missing builtin ID in switch!"); 4595 case NEON::BI__builtin_neon_vceqd_s64: 4596 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4597 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4598 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4599 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4600 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4601 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4602 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4603 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4604 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4605 } 4606 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4607 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4608 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4609 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4610 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4611 } 4612 case NEON::BI__builtin_neon_vtstd_s64: 4613 case NEON::BI__builtin_neon_vtstd_u64: { 4614 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4615 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4616 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4617 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4618 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4619 llvm::Constant::getNullValue(Int64Ty)); 4620 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4621 } 4622 case NEON::BI__builtin_neon_vset_lane_i8: 4623 case NEON::BI__builtin_neon_vset_lane_i16: 4624 case NEON::BI__builtin_neon_vset_lane_i32: 4625 case NEON::BI__builtin_neon_vset_lane_i64: 4626 case NEON::BI__builtin_neon_vset_lane_f32: 4627 case NEON::BI__builtin_neon_vsetq_lane_i8: 4628 case NEON::BI__builtin_neon_vsetq_lane_i16: 4629 case NEON::BI__builtin_neon_vsetq_lane_i32: 4630 case NEON::BI__builtin_neon_vsetq_lane_i64: 4631 case NEON::BI__builtin_neon_vsetq_lane_f32: 4632 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4633 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4634 case NEON::BI__builtin_neon_vset_lane_f64: 4635 // The vector type needs a cast for the v1f64 variant. 4636 Ops[1] = Builder.CreateBitCast(Ops[1], 4637 llvm::VectorType::get(DoubleTy, 1)); 4638 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4639 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4640 case NEON::BI__builtin_neon_vsetq_lane_f64: 4641 // The vector type needs a cast for the v2f64 variant. 4642 Ops[1] = Builder.CreateBitCast(Ops[1], 4643 llvm::VectorType::get(DoubleTy, 2)); 4644 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4645 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4646 4647 case NEON::BI__builtin_neon_vget_lane_i8: 4648 case NEON::BI__builtin_neon_vdupb_lane_i8: 4649 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4650 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4651 "vget_lane"); 4652 case NEON::BI__builtin_neon_vgetq_lane_i8: 4653 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4654 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 4655 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4656 "vgetq_lane"); 4657 case NEON::BI__builtin_neon_vget_lane_i16: 4658 case NEON::BI__builtin_neon_vduph_lane_i16: 4659 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 4660 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4661 "vget_lane"); 4662 case NEON::BI__builtin_neon_vgetq_lane_i16: 4663 case NEON::BI__builtin_neon_vduph_laneq_i16: 4664 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 4665 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4666 "vgetq_lane"); 4667 case NEON::BI__builtin_neon_vget_lane_i32: 4668 case NEON::BI__builtin_neon_vdups_lane_i32: 4669 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 4670 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4671 "vget_lane"); 4672 case NEON::BI__builtin_neon_vdups_lane_f32: 4673 Ops[0] = Builder.CreateBitCast(Ops[0], 4674 llvm::VectorType::get(FloatTy, 2)); 4675 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4676 "vdups_lane"); 4677 case NEON::BI__builtin_neon_vgetq_lane_i32: 4678 case NEON::BI__builtin_neon_vdups_laneq_i32: 4679 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 4680 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4681 "vgetq_lane"); 4682 case NEON::BI__builtin_neon_vget_lane_i64: 4683 case NEON::BI__builtin_neon_vdupd_lane_i64: 4684 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 4685 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4686 "vget_lane"); 4687 case NEON::BI__builtin_neon_vdupd_lane_f64: 4688 Ops[0] = Builder.CreateBitCast(Ops[0], 4689 llvm::VectorType::get(DoubleTy, 1)); 4690 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4691 "vdupd_lane"); 4692 case NEON::BI__builtin_neon_vgetq_lane_i64: 4693 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4694 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 4695 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4696 "vgetq_lane"); 4697 case NEON::BI__builtin_neon_vget_lane_f32: 4698 Ops[0] = Builder.CreateBitCast(Ops[0], 4699 llvm::VectorType::get(FloatTy, 2)); 4700 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4701 "vget_lane"); 4702 case NEON::BI__builtin_neon_vget_lane_f64: 4703 Ops[0] = Builder.CreateBitCast(Ops[0], 4704 llvm::VectorType::get(DoubleTy, 1)); 4705 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4706 "vget_lane"); 4707 case NEON::BI__builtin_neon_vgetq_lane_f32: 4708 case NEON::BI__builtin_neon_vdups_laneq_f32: 4709 Ops[0] = Builder.CreateBitCast(Ops[0], 4710 llvm::VectorType::get(FloatTy, 4)); 4711 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4712 "vgetq_lane"); 4713 case NEON::BI__builtin_neon_vgetq_lane_f64: 4714 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4715 Ops[0] = Builder.CreateBitCast(Ops[0], 4716 llvm::VectorType::get(DoubleTy, 2)); 4717 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4718 "vgetq_lane"); 4719 case NEON::BI__builtin_neon_vaddd_s64: 4720 case NEON::BI__builtin_neon_vaddd_u64: 4721 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4722 case NEON::BI__builtin_neon_vsubd_s64: 4723 case NEON::BI__builtin_neon_vsubd_u64: 4724 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4725 case NEON::BI__builtin_neon_vqdmlalh_s16: 4726 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4727 SmallVector<Value *, 2> ProductOps; 4728 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4729 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4730 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4731 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4732 ProductOps, "vqdmlXl"); 4733 Constant *CI = ConstantInt::get(SizeTy, 0); 4734 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4735 4736 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4737 ? Intrinsic::aarch64_neon_sqadd 4738 : Intrinsic::aarch64_neon_sqsub; 4739 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4740 } 4741 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4742 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4743 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4744 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4745 Ops, "vqshlu_n"); 4746 } 4747 case NEON::BI__builtin_neon_vqshld_n_u64: 4748 case NEON::BI__builtin_neon_vqshld_n_s64: { 4749 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4750 ? Intrinsic::aarch64_neon_uqshl 4751 : Intrinsic::aarch64_neon_sqshl; 4752 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4753 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4754 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4755 } 4756 case NEON::BI__builtin_neon_vrshrd_n_u64: 4757 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4758 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4759 ? Intrinsic::aarch64_neon_urshl 4760 : Intrinsic::aarch64_neon_srshl; 4761 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4762 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4763 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4764 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4765 } 4766 case NEON::BI__builtin_neon_vrsrad_n_u64: 4767 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4768 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4769 ? Intrinsic::aarch64_neon_urshl 4770 : Intrinsic::aarch64_neon_srshl; 4771 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4772 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4773 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4774 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4775 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4776 } 4777 case NEON::BI__builtin_neon_vshld_n_s64: 4778 case NEON::BI__builtin_neon_vshld_n_u64: { 4779 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4780 return Builder.CreateShl( 4781 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4782 } 4783 case NEON::BI__builtin_neon_vshrd_n_s64: { 4784 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4785 return Builder.CreateAShr( 4786 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4787 Amt->getZExtValue())), 4788 "shrd_n"); 4789 } 4790 case NEON::BI__builtin_neon_vshrd_n_u64: { 4791 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4792 uint64_t ShiftAmt = Amt->getZExtValue(); 4793 // Right-shifting an unsigned value by its size yields 0. 4794 if (ShiftAmt == 64) 4795 return ConstantInt::get(Int64Ty, 0); 4796 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4797 "shrd_n"); 4798 } 4799 case NEON::BI__builtin_neon_vsrad_n_s64: { 4800 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4801 Ops[1] = Builder.CreateAShr( 4802 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4803 Amt->getZExtValue())), 4804 "shrd_n"); 4805 return Builder.CreateAdd(Ops[0], Ops[1]); 4806 } 4807 case NEON::BI__builtin_neon_vsrad_n_u64: { 4808 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4809 uint64_t ShiftAmt = Amt->getZExtValue(); 4810 // Right-shifting an unsigned value by its size yields 0. 4811 // As Op + 0 = Op, return Ops[0] directly. 4812 if (ShiftAmt == 64) 4813 return Ops[0]; 4814 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4815 "shrd_n"); 4816 return Builder.CreateAdd(Ops[0], Ops[1]); 4817 } 4818 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4819 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4820 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4821 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4822 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4823 "lane"); 4824 SmallVector<Value *, 2> ProductOps; 4825 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4826 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4827 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4828 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4829 ProductOps, "vqdmlXl"); 4830 Constant *CI = ConstantInt::get(SizeTy, 0); 4831 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4832 Ops.pop_back(); 4833 4834 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4835 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4836 ? Intrinsic::aarch64_neon_sqadd 4837 : Intrinsic::aarch64_neon_sqsub; 4838 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4839 } 4840 case NEON::BI__builtin_neon_vqdmlals_s32: 4841 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4842 SmallVector<Value *, 2> ProductOps; 4843 ProductOps.push_back(Ops[1]); 4844 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4845 Ops[1] = 4846 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4847 ProductOps, "vqdmlXl"); 4848 4849 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4850 ? Intrinsic::aarch64_neon_sqadd 4851 : Intrinsic::aarch64_neon_sqsub; 4852 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4853 } 4854 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4855 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4856 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4857 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4858 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4859 "lane"); 4860 SmallVector<Value *, 2> ProductOps; 4861 ProductOps.push_back(Ops[1]); 4862 ProductOps.push_back(Ops[2]); 4863 Ops[1] = 4864 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4865 ProductOps, "vqdmlXl"); 4866 Ops.pop_back(); 4867 4868 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4869 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4870 ? Intrinsic::aarch64_neon_sqadd 4871 : Intrinsic::aarch64_neon_sqsub; 4872 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4873 } 4874 } 4875 4876 llvm::VectorType *VTy = GetNeonType(this, Type); 4877 llvm::Type *Ty = VTy; 4878 if (!Ty) 4879 return nullptr; 4880 4881 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4882 // defer to common code if it's been added to our special map. 4883 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4884 AArch64SIMDIntrinsicsProvenSorted); 4885 4886 if (Builtin) 4887 return EmitCommonNeonBuiltinExpr( 4888 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4889 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4890 4891 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4892 return V; 4893 4894 unsigned Int; 4895 switch (BuiltinID) { 4896 default: return nullptr; 4897 case NEON::BI__builtin_neon_vbsl_v: 4898 case NEON::BI__builtin_neon_vbslq_v: { 4899 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4900 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4901 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4902 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4903 4904 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4905 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4906 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4907 return Builder.CreateBitCast(Ops[0], Ty); 4908 } 4909 case NEON::BI__builtin_neon_vfma_lane_v: 4910 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4911 // The ARM builtins (and instructions) have the addend as the first 4912 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4913 Value *Addend = Ops[0]; 4914 Value *Multiplicand = Ops[1]; 4915 Value *LaneSource = Ops[2]; 4916 Ops[0] = Multiplicand; 4917 Ops[1] = LaneSource; 4918 Ops[2] = Addend; 4919 4920 // Now adjust things to handle the lane access. 4921 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4922 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4923 VTy; 4924 llvm::Constant *cst = cast<Constant>(Ops[3]); 4925 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4926 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4927 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4928 4929 Ops.pop_back(); 4930 Int = Intrinsic::fma; 4931 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4932 } 4933 case NEON::BI__builtin_neon_vfma_laneq_v: { 4934 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4935 // v1f64 fma should be mapped to Neon scalar f64 fma 4936 if (VTy && VTy->getElementType() == DoubleTy) { 4937 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4938 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4939 llvm::Type *VTy = GetNeonType(this, 4940 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4941 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4942 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4943 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4944 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4945 return Builder.CreateBitCast(Result, Ty); 4946 } 4947 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4948 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4949 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4950 4951 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4952 VTy->getNumElements() * 2); 4953 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4954 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4955 cast<ConstantInt>(Ops[3])); 4956 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4957 4958 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4959 } 4960 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4961 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4962 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4963 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4964 4965 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4966 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4967 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4968 } 4969 case NEON::BI__builtin_neon_vfmas_lane_f32: 4970 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4971 case NEON::BI__builtin_neon_vfmad_lane_f64: 4972 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4973 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4974 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4975 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4976 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4977 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4978 } 4979 case NEON::BI__builtin_neon_vfms_v: 4980 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4981 // FIXME: probably remove when we no longer support aarch64_simd.h 4982 // (arm_neon.h delegates to vfma). 4983 4984 // The ARM builtins (and instructions) have the addend as the first 4985 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4986 Value *Subtrahend = Ops[0]; 4987 Value *Multiplicand = Ops[2]; 4988 Ops[0] = Multiplicand; 4989 Ops[2] = Subtrahend; 4990 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4991 Ops[1] = Builder.CreateFNeg(Ops[1]); 4992 Int = Intrinsic::fma; 4993 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4994 } 4995 case NEON::BI__builtin_neon_vmull_v: 4996 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4997 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4998 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4999 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5000 case NEON::BI__builtin_neon_vmax_v: 5001 case NEON::BI__builtin_neon_vmaxq_v: 5002 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5003 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5004 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5005 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5006 case NEON::BI__builtin_neon_vmin_v: 5007 case NEON::BI__builtin_neon_vminq_v: 5008 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5009 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5010 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5011 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5012 case NEON::BI__builtin_neon_vabd_v: 5013 case NEON::BI__builtin_neon_vabdq_v: 5014 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5015 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5016 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5017 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5018 case NEON::BI__builtin_neon_vpadal_v: 5019 case NEON::BI__builtin_neon_vpadalq_v: { 5020 unsigned ArgElts = VTy->getNumElements(); 5021 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5022 unsigned BitWidth = EltTy->getBitWidth(); 5023 llvm::Type *ArgTy = llvm::VectorType::get( 5024 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5025 llvm::Type* Tys[2] = { VTy, ArgTy }; 5026 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5027 SmallVector<llvm::Value*, 1> TmpOps; 5028 TmpOps.push_back(Ops[1]); 5029 Function *F = CGM.getIntrinsic(Int, Tys); 5030 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5031 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5032 return Builder.CreateAdd(tmp, addend); 5033 } 5034 case NEON::BI__builtin_neon_vpmin_v: 5035 case NEON::BI__builtin_neon_vpminq_v: 5036 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5037 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5038 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5039 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5040 case NEON::BI__builtin_neon_vpmax_v: 5041 case NEON::BI__builtin_neon_vpmaxq_v: 5042 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5043 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5044 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5045 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5046 case NEON::BI__builtin_neon_vminnm_v: 5047 case NEON::BI__builtin_neon_vminnmq_v: 5048 Int = Intrinsic::aarch64_neon_fminnm; 5049 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5050 case NEON::BI__builtin_neon_vmaxnm_v: 5051 case NEON::BI__builtin_neon_vmaxnmq_v: 5052 Int = Intrinsic::aarch64_neon_fmaxnm; 5053 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5054 case NEON::BI__builtin_neon_vrecpss_f32: { 5055 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5056 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5057 Ops, "vrecps"); 5058 } 5059 case NEON::BI__builtin_neon_vrecpsd_f64: { 5060 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5061 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5062 Ops, "vrecps"); 5063 } 5064 case NEON::BI__builtin_neon_vqshrun_n_v: 5065 Int = Intrinsic::aarch64_neon_sqshrun; 5066 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5067 case NEON::BI__builtin_neon_vqrshrun_n_v: 5068 Int = Intrinsic::aarch64_neon_sqrshrun; 5069 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5070 case NEON::BI__builtin_neon_vqshrn_n_v: 5071 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5072 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5073 case NEON::BI__builtin_neon_vrshrn_n_v: 5074 Int = Intrinsic::aarch64_neon_rshrn; 5075 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5076 case NEON::BI__builtin_neon_vqrshrn_n_v: 5077 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5078 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5079 case NEON::BI__builtin_neon_vrnda_v: 5080 case NEON::BI__builtin_neon_vrndaq_v: { 5081 Int = Intrinsic::round; 5082 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5083 } 5084 case NEON::BI__builtin_neon_vrndi_v: 5085 case NEON::BI__builtin_neon_vrndiq_v: { 5086 Int = Intrinsic::nearbyint; 5087 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5088 } 5089 case NEON::BI__builtin_neon_vrndm_v: 5090 case NEON::BI__builtin_neon_vrndmq_v: { 5091 Int = Intrinsic::floor; 5092 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5093 } 5094 case NEON::BI__builtin_neon_vrndn_v: 5095 case NEON::BI__builtin_neon_vrndnq_v: { 5096 Int = Intrinsic::aarch64_neon_frintn; 5097 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5098 } 5099 case NEON::BI__builtin_neon_vrndp_v: 5100 case NEON::BI__builtin_neon_vrndpq_v: { 5101 Int = Intrinsic::ceil; 5102 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5103 } 5104 case NEON::BI__builtin_neon_vrndx_v: 5105 case NEON::BI__builtin_neon_vrndxq_v: { 5106 Int = Intrinsic::rint; 5107 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5108 } 5109 case NEON::BI__builtin_neon_vrnd_v: 5110 case NEON::BI__builtin_neon_vrndq_v: { 5111 Int = Intrinsic::trunc; 5112 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5113 } 5114 case NEON::BI__builtin_neon_vceqz_v: 5115 case NEON::BI__builtin_neon_vceqzq_v: 5116 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5117 ICmpInst::ICMP_EQ, "vceqz"); 5118 case NEON::BI__builtin_neon_vcgez_v: 5119 case NEON::BI__builtin_neon_vcgezq_v: 5120 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5121 ICmpInst::ICMP_SGE, "vcgez"); 5122 case NEON::BI__builtin_neon_vclez_v: 5123 case NEON::BI__builtin_neon_vclezq_v: 5124 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5125 ICmpInst::ICMP_SLE, "vclez"); 5126 case NEON::BI__builtin_neon_vcgtz_v: 5127 case NEON::BI__builtin_neon_vcgtzq_v: 5128 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5129 ICmpInst::ICMP_SGT, "vcgtz"); 5130 case NEON::BI__builtin_neon_vcltz_v: 5131 case NEON::BI__builtin_neon_vcltzq_v: 5132 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5133 ICmpInst::ICMP_SLT, "vcltz"); 5134 case NEON::BI__builtin_neon_vcvt_f64_v: 5135 case NEON::BI__builtin_neon_vcvtq_f64_v: 5136 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5137 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5138 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5139 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5140 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5141 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5142 "unexpected vcvt_f64_f32 builtin"); 5143 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5144 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5145 5146 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5147 } 5148 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5149 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5150 "unexpected vcvt_f32_f64 builtin"); 5151 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5152 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5153 5154 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5155 } 5156 case NEON::BI__builtin_neon_vcvt_s32_v: 5157 case NEON::BI__builtin_neon_vcvt_u32_v: 5158 case NEON::BI__builtin_neon_vcvt_s64_v: 5159 case NEON::BI__builtin_neon_vcvt_u64_v: 5160 case NEON::BI__builtin_neon_vcvtq_s32_v: 5161 case NEON::BI__builtin_neon_vcvtq_u32_v: 5162 case NEON::BI__builtin_neon_vcvtq_s64_v: 5163 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5164 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5165 if (usgn) 5166 return Builder.CreateFPToUI(Ops[0], Ty); 5167 return Builder.CreateFPToSI(Ops[0], Ty); 5168 } 5169 case NEON::BI__builtin_neon_vcvta_s32_v: 5170 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5171 case NEON::BI__builtin_neon_vcvta_u32_v: 5172 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5173 case NEON::BI__builtin_neon_vcvta_s64_v: 5174 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5175 case NEON::BI__builtin_neon_vcvta_u64_v: 5176 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5177 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5178 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5179 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5180 } 5181 case NEON::BI__builtin_neon_vcvtm_s32_v: 5182 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5183 case NEON::BI__builtin_neon_vcvtm_u32_v: 5184 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5185 case NEON::BI__builtin_neon_vcvtm_s64_v: 5186 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5187 case NEON::BI__builtin_neon_vcvtm_u64_v: 5188 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5189 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5190 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5191 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5192 } 5193 case NEON::BI__builtin_neon_vcvtn_s32_v: 5194 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5195 case NEON::BI__builtin_neon_vcvtn_u32_v: 5196 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5197 case NEON::BI__builtin_neon_vcvtn_s64_v: 5198 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5199 case NEON::BI__builtin_neon_vcvtn_u64_v: 5200 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5201 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5202 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5203 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5204 } 5205 case NEON::BI__builtin_neon_vcvtp_s32_v: 5206 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5207 case NEON::BI__builtin_neon_vcvtp_u32_v: 5208 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5209 case NEON::BI__builtin_neon_vcvtp_s64_v: 5210 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5211 case NEON::BI__builtin_neon_vcvtp_u64_v: 5212 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5213 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5214 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5215 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5216 } 5217 case NEON::BI__builtin_neon_vmulx_v: 5218 case NEON::BI__builtin_neon_vmulxq_v: { 5219 Int = Intrinsic::aarch64_neon_fmulx; 5220 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5221 } 5222 case NEON::BI__builtin_neon_vmul_lane_v: 5223 case NEON::BI__builtin_neon_vmul_laneq_v: { 5224 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5225 bool Quad = false; 5226 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5227 Quad = true; 5228 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5229 llvm::Type *VTy = GetNeonType(this, 5230 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5231 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5232 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5233 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5234 return Builder.CreateBitCast(Result, Ty); 5235 } 5236 case NEON::BI__builtin_neon_vnegd_s64: 5237 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5238 case NEON::BI__builtin_neon_vpmaxnm_v: 5239 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5240 Int = Intrinsic::aarch64_neon_fmaxnmp; 5241 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5242 } 5243 case NEON::BI__builtin_neon_vpminnm_v: 5244 case NEON::BI__builtin_neon_vpminnmq_v: { 5245 Int = Intrinsic::aarch64_neon_fminnmp; 5246 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5247 } 5248 case NEON::BI__builtin_neon_vsqrt_v: 5249 case NEON::BI__builtin_neon_vsqrtq_v: { 5250 Int = Intrinsic::sqrt; 5251 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5252 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5253 } 5254 case NEON::BI__builtin_neon_vrbit_v: 5255 case NEON::BI__builtin_neon_vrbitq_v: { 5256 Int = Intrinsic::aarch64_neon_rbit; 5257 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5258 } 5259 case NEON::BI__builtin_neon_vaddv_u8: 5260 // FIXME: These are handled by the AArch64 scalar code. 5261 usgn = true; 5262 // FALLTHROUGH 5263 case NEON::BI__builtin_neon_vaddv_s8: { 5264 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5265 Ty = Int32Ty; 5266 VTy = llvm::VectorType::get(Int8Ty, 8); 5267 llvm::Type *Tys[2] = { Ty, VTy }; 5268 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5269 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5270 return Builder.CreateTrunc(Ops[0], Int8Ty); 5271 } 5272 case NEON::BI__builtin_neon_vaddv_u16: 5273 usgn = true; 5274 // FALLTHROUGH 5275 case NEON::BI__builtin_neon_vaddv_s16: { 5276 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5277 Ty = Int32Ty; 5278 VTy = llvm::VectorType::get(Int16Ty, 4); 5279 llvm::Type *Tys[2] = { Ty, VTy }; 5280 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5281 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5282 return Builder.CreateTrunc(Ops[0], Int16Ty); 5283 } 5284 case NEON::BI__builtin_neon_vaddvq_u8: 5285 usgn = true; 5286 // FALLTHROUGH 5287 case NEON::BI__builtin_neon_vaddvq_s8: { 5288 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5289 Ty = Int32Ty; 5290 VTy = llvm::VectorType::get(Int8Ty, 16); 5291 llvm::Type *Tys[2] = { Ty, VTy }; 5292 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5293 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5294 return Builder.CreateTrunc(Ops[0], Int8Ty); 5295 } 5296 case NEON::BI__builtin_neon_vaddvq_u16: 5297 usgn = true; 5298 // FALLTHROUGH 5299 case NEON::BI__builtin_neon_vaddvq_s16: { 5300 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5301 Ty = Int32Ty; 5302 VTy = llvm::VectorType::get(Int16Ty, 8); 5303 llvm::Type *Tys[2] = { Ty, VTy }; 5304 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5305 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5306 return Builder.CreateTrunc(Ops[0], Int16Ty); 5307 } 5308 case NEON::BI__builtin_neon_vmaxv_u8: { 5309 Int = Intrinsic::aarch64_neon_umaxv; 5310 Ty = Int32Ty; 5311 VTy = llvm::VectorType::get(Int8Ty, 8); 5312 llvm::Type *Tys[2] = { Ty, VTy }; 5313 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5314 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5315 return Builder.CreateTrunc(Ops[0], Int8Ty); 5316 } 5317 case NEON::BI__builtin_neon_vmaxv_u16: { 5318 Int = Intrinsic::aarch64_neon_umaxv; 5319 Ty = Int32Ty; 5320 VTy = llvm::VectorType::get(Int16Ty, 4); 5321 llvm::Type *Tys[2] = { Ty, VTy }; 5322 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5323 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5324 return Builder.CreateTrunc(Ops[0], Int16Ty); 5325 } 5326 case NEON::BI__builtin_neon_vmaxvq_u8: { 5327 Int = Intrinsic::aarch64_neon_umaxv; 5328 Ty = Int32Ty; 5329 VTy = llvm::VectorType::get(Int8Ty, 16); 5330 llvm::Type *Tys[2] = { Ty, VTy }; 5331 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5332 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5333 return Builder.CreateTrunc(Ops[0], Int8Ty); 5334 } 5335 case NEON::BI__builtin_neon_vmaxvq_u16: { 5336 Int = Intrinsic::aarch64_neon_umaxv; 5337 Ty = Int32Ty; 5338 VTy = llvm::VectorType::get(Int16Ty, 8); 5339 llvm::Type *Tys[2] = { Ty, VTy }; 5340 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5341 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5342 return Builder.CreateTrunc(Ops[0], Int16Ty); 5343 } 5344 case NEON::BI__builtin_neon_vmaxv_s8: { 5345 Int = Intrinsic::aarch64_neon_smaxv; 5346 Ty = Int32Ty; 5347 VTy = llvm::VectorType::get(Int8Ty, 8); 5348 llvm::Type *Tys[2] = { Ty, VTy }; 5349 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5350 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5351 return Builder.CreateTrunc(Ops[0], Int8Ty); 5352 } 5353 case NEON::BI__builtin_neon_vmaxv_s16: { 5354 Int = Intrinsic::aarch64_neon_smaxv; 5355 Ty = Int32Ty; 5356 VTy = llvm::VectorType::get(Int16Ty, 4); 5357 llvm::Type *Tys[2] = { Ty, VTy }; 5358 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5359 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5360 return Builder.CreateTrunc(Ops[0], Int16Ty); 5361 } 5362 case NEON::BI__builtin_neon_vmaxvq_s8: { 5363 Int = Intrinsic::aarch64_neon_smaxv; 5364 Ty = Int32Ty; 5365 VTy = llvm::VectorType::get(Int8Ty, 16); 5366 llvm::Type *Tys[2] = { Ty, VTy }; 5367 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5368 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5369 return Builder.CreateTrunc(Ops[0], Int8Ty); 5370 } 5371 case NEON::BI__builtin_neon_vmaxvq_s16: { 5372 Int = Intrinsic::aarch64_neon_smaxv; 5373 Ty = Int32Ty; 5374 VTy = llvm::VectorType::get(Int16Ty, 8); 5375 llvm::Type *Tys[2] = { Ty, VTy }; 5376 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5377 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5378 return Builder.CreateTrunc(Ops[0], Int16Ty); 5379 } 5380 case NEON::BI__builtin_neon_vminv_u8: { 5381 Int = Intrinsic::aarch64_neon_uminv; 5382 Ty = Int32Ty; 5383 VTy = llvm::VectorType::get(Int8Ty, 8); 5384 llvm::Type *Tys[2] = { Ty, VTy }; 5385 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5386 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5387 return Builder.CreateTrunc(Ops[0], Int8Ty); 5388 } 5389 case NEON::BI__builtin_neon_vminv_u16: { 5390 Int = Intrinsic::aarch64_neon_uminv; 5391 Ty = Int32Ty; 5392 VTy = llvm::VectorType::get(Int16Ty, 4); 5393 llvm::Type *Tys[2] = { Ty, VTy }; 5394 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5395 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5396 return Builder.CreateTrunc(Ops[0], Int16Ty); 5397 } 5398 case NEON::BI__builtin_neon_vminvq_u8: { 5399 Int = Intrinsic::aarch64_neon_uminv; 5400 Ty = Int32Ty; 5401 VTy = llvm::VectorType::get(Int8Ty, 16); 5402 llvm::Type *Tys[2] = { Ty, VTy }; 5403 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5404 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5405 return Builder.CreateTrunc(Ops[0], Int8Ty); 5406 } 5407 case NEON::BI__builtin_neon_vminvq_u16: { 5408 Int = Intrinsic::aarch64_neon_uminv; 5409 Ty = Int32Ty; 5410 VTy = llvm::VectorType::get(Int16Ty, 8); 5411 llvm::Type *Tys[2] = { Ty, VTy }; 5412 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5413 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5414 return Builder.CreateTrunc(Ops[0], Int16Ty); 5415 } 5416 case NEON::BI__builtin_neon_vminv_s8: { 5417 Int = Intrinsic::aarch64_neon_sminv; 5418 Ty = Int32Ty; 5419 VTy = llvm::VectorType::get(Int8Ty, 8); 5420 llvm::Type *Tys[2] = { Ty, VTy }; 5421 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5422 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5423 return Builder.CreateTrunc(Ops[0], Int8Ty); 5424 } 5425 case NEON::BI__builtin_neon_vminv_s16: { 5426 Int = Intrinsic::aarch64_neon_sminv; 5427 Ty = Int32Ty; 5428 VTy = llvm::VectorType::get(Int16Ty, 4); 5429 llvm::Type *Tys[2] = { Ty, VTy }; 5430 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5431 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5432 return Builder.CreateTrunc(Ops[0], Int16Ty); 5433 } 5434 case NEON::BI__builtin_neon_vminvq_s8: { 5435 Int = Intrinsic::aarch64_neon_sminv; 5436 Ty = Int32Ty; 5437 VTy = llvm::VectorType::get(Int8Ty, 16); 5438 llvm::Type *Tys[2] = { Ty, VTy }; 5439 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5440 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5441 return Builder.CreateTrunc(Ops[0], Int8Ty); 5442 } 5443 case NEON::BI__builtin_neon_vminvq_s16: { 5444 Int = Intrinsic::aarch64_neon_sminv; 5445 Ty = Int32Ty; 5446 VTy = llvm::VectorType::get(Int16Ty, 8); 5447 llvm::Type *Tys[2] = { Ty, VTy }; 5448 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5449 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5450 return Builder.CreateTrunc(Ops[0], Int16Ty); 5451 } 5452 case NEON::BI__builtin_neon_vmul_n_f64: { 5453 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5454 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5455 return Builder.CreateFMul(Ops[0], RHS); 5456 } 5457 case NEON::BI__builtin_neon_vaddlv_u8: { 5458 Int = Intrinsic::aarch64_neon_uaddlv; 5459 Ty = Int32Ty; 5460 VTy = llvm::VectorType::get(Int8Ty, 8); 5461 llvm::Type *Tys[2] = { Ty, VTy }; 5462 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5463 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5464 return Builder.CreateTrunc(Ops[0], Int16Ty); 5465 } 5466 case NEON::BI__builtin_neon_vaddlv_u16: { 5467 Int = Intrinsic::aarch64_neon_uaddlv; 5468 Ty = Int32Ty; 5469 VTy = llvm::VectorType::get(Int16Ty, 4); 5470 llvm::Type *Tys[2] = { Ty, VTy }; 5471 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5472 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5473 } 5474 case NEON::BI__builtin_neon_vaddlvq_u8: { 5475 Int = Intrinsic::aarch64_neon_uaddlv; 5476 Ty = Int32Ty; 5477 VTy = llvm::VectorType::get(Int8Ty, 16); 5478 llvm::Type *Tys[2] = { Ty, VTy }; 5479 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5480 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5481 return Builder.CreateTrunc(Ops[0], Int16Ty); 5482 } 5483 case NEON::BI__builtin_neon_vaddlvq_u16: { 5484 Int = Intrinsic::aarch64_neon_uaddlv; 5485 Ty = Int32Ty; 5486 VTy = llvm::VectorType::get(Int16Ty, 8); 5487 llvm::Type *Tys[2] = { Ty, VTy }; 5488 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5489 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5490 } 5491 case NEON::BI__builtin_neon_vaddlv_s8: { 5492 Int = Intrinsic::aarch64_neon_saddlv; 5493 Ty = Int32Ty; 5494 VTy = llvm::VectorType::get(Int8Ty, 8); 5495 llvm::Type *Tys[2] = { Ty, VTy }; 5496 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5497 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5498 return Builder.CreateTrunc(Ops[0], Int16Ty); 5499 } 5500 case NEON::BI__builtin_neon_vaddlv_s16: { 5501 Int = Intrinsic::aarch64_neon_saddlv; 5502 Ty = Int32Ty; 5503 VTy = llvm::VectorType::get(Int16Ty, 4); 5504 llvm::Type *Tys[2] = { Ty, VTy }; 5505 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5506 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5507 } 5508 case NEON::BI__builtin_neon_vaddlvq_s8: { 5509 Int = Intrinsic::aarch64_neon_saddlv; 5510 Ty = Int32Ty; 5511 VTy = llvm::VectorType::get(Int8Ty, 16); 5512 llvm::Type *Tys[2] = { Ty, VTy }; 5513 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5514 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5515 return Builder.CreateTrunc(Ops[0], Int16Ty); 5516 } 5517 case NEON::BI__builtin_neon_vaddlvq_s16: { 5518 Int = Intrinsic::aarch64_neon_saddlv; 5519 Ty = Int32Ty; 5520 VTy = llvm::VectorType::get(Int16Ty, 8); 5521 llvm::Type *Tys[2] = { Ty, VTy }; 5522 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5523 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5524 } 5525 case NEON::BI__builtin_neon_vsri_n_v: 5526 case NEON::BI__builtin_neon_vsriq_n_v: { 5527 Int = Intrinsic::aarch64_neon_vsri; 5528 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5529 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5530 } 5531 case NEON::BI__builtin_neon_vsli_n_v: 5532 case NEON::BI__builtin_neon_vsliq_n_v: { 5533 Int = Intrinsic::aarch64_neon_vsli; 5534 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5535 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5536 } 5537 case NEON::BI__builtin_neon_vsra_n_v: 5538 case NEON::BI__builtin_neon_vsraq_n_v: 5539 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5540 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5541 return Builder.CreateAdd(Ops[0], Ops[1]); 5542 case NEON::BI__builtin_neon_vrsra_n_v: 5543 case NEON::BI__builtin_neon_vrsraq_n_v: { 5544 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5545 SmallVector<llvm::Value*,2> TmpOps; 5546 TmpOps.push_back(Ops[1]); 5547 TmpOps.push_back(Ops[2]); 5548 Function* F = CGM.getIntrinsic(Int, Ty); 5549 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5550 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5551 return Builder.CreateAdd(Ops[0], tmp); 5552 } 5553 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5554 // of an Align parameter here. 5555 case NEON::BI__builtin_neon_vld1_x2_v: 5556 case NEON::BI__builtin_neon_vld1q_x2_v: 5557 case NEON::BI__builtin_neon_vld1_x3_v: 5558 case NEON::BI__builtin_neon_vld1q_x3_v: 5559 case NEON::BI__builtin_neon_vld1_x4_v: 5560 case NEON::BI__builtin_neon_vld1q_x4_v: { 5561 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5562 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5563 llvm::Type *Tys[2] = { VTy, PTy }; 5564 unsigned Int; 5565 switch (BuiltinID) { 5566 case NEON::BI__builtin_neon_vld1_x2_v: 5567 case NEON::BI__builtin_neon_vld1q_x2_v: 5568 Int = Intrinsic::aarch64_neon_ld1x2; 5569 break; 5570 case NEON::BI__builtin_neon_vld1_x3_v: 5571 case NEON::BI__builtin_neon_vld1q_x3_v: 5572 Int = Intrinsic::aarch64_neon_ld1x3; 5573 break; 5574 case NEON::BI__builtin_neon_vld1_x4_v: 5575 case NEON::BI__builtin_neon_vld1q_x4_v: 5576 Int = Intrinsic::aarch64_neon_ld1x4; 5577 break; 5578 } 5579 Function *F = CGM.getIntrinsic(Int, Tys); 5580 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5581 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5582 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5583 return Builder.CreateStore(Ops[1], Ops[0]); 5584 } 5585 case NEON::BI__builtin_neon_vst1_x2_v: 5586 case NEON::BI__builtin_neon_vst1q_x2_v: 5587 case NEON::BI__builtin_neon_vst1_x3_v: 5588 case NEON::BI__builtin_neon_vst1q_x3_v: 5589 case NEON::BI__builtin_neon_vst1_x4_v: 5590 case NEON::BI__builtin_neon_vst1q_x4_v: { 5591 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5592 llvm::Type *Tys[2] = { VTy, PTy }; 5593 unsigned Int; 5594 switch (BuiltinID) { 5595 case NEON::BI__builtin_neon_vst1_x2_v: 5596 case NEON::BI__builtin_neon_vst1q_x2_v: 5597 Int = Intrinsic::aarch64_neon_st1x2; 5598 break; 5599 case NEON::BI__builtin_neon_vst1_x3_v: 5600 case NEON::BI__builtin_neon_vst1q_x3_v: 5601 Int = Intrinsic::aarch64_neon_st1x3; 5602 break; 5603 case NEON::BI__builtin_neon_vst1_x4_v: 5604 case NEON::BI__builtin_neon_vst1q_x4_v: 5605 Int = Intrinsic::aarch64_neon_st1x4; 5606 break; 5607 } 5608 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5609 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5610 } 5611 case NEON::BI__builtin_neon_vld1_v: 5612 case NEON::BI__builtin_neon_vld1q_v: 5613 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5614 return Builder.CreateLoad(Ops[0]); 5615 case NEON::BI__builtin_neon_vst1_v: 5616 case NEON::BI__builtin_neon_vst1q_v: 5617 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5618 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5619 return Builder.CreateStore(Ops[1], Ops[0]); 5620 case NEON::BI__builtin_neon_vld1_lane_v: 5621 case NEON::BI__builtin_neon_vld1q_lane_v: 5622 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5623 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5624 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5625 Ops[0] = Builder.CreateLoad(Ops[0]); 5626 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5627 case NEON::BI__builtin_neon_vld1_dup_v: 5628 case NEON::BI__builtin_neon_vld1q_dup_v: { 5629 Value *V = UndefValue::get(Ty); 5630 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5631 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5632 Ops[0] = Builder.CreateLoad(Ops[0]); 5633 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5634 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5635 return EmitNeonSplat(Ops[0], CI); 5636 } 5637 case NEON::BI__builtin_neon_vst1_lane_v: 5638 case NEON::BI__builtin_neon_vst1q_lane_v: 5639 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5640 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5641 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5642 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5643 case NEON::BI__builtin_neon_vld2_v: 5644 case NEON::BI__builtin_neon_vld2q_v: { 5645 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5646 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5647 llvm::Type *Tys[2] = { VTy, PTy }; 5648 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5649 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5650 Ops[0] = Builder.CreateBitCast(Ops[0], 5651 llvm::PointerType::getUnqual(Ops[1]->getType())); 5652 return Builder.CreateStore(Ops[1], Ops[0]); 5653 } 5654 case NEON::BI__builtin_neon_vld3_v: 5655 case NEON::BI__builtin_neon_vld3q_v: { 5656 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5657 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5658 llvm::Type *Tys[2] = { VTy, PTy }; 5659 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5660 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5661 Ops[0] = Builder.CreateBitCast(Ops[0], 5662 llvm::PointerType::getUnqual(Ops[1]->getType())); 5663 return Builder.CreateStore(Ops[1], Ops[0]); 5664 } 5665 case NEON::BI__builtin_neon_vld4_v: 5666 case NEON::BI__builtin_neon_vld4q_v: { 5667 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5668 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5669 llvm::Type *Tys[2] = { VTy, PTy }; 5670 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5671 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5672 Ops[0] = Builder.CreateBitCast(Ops[0], 5673 llvm::PointerType::getUnqual(Ops[1]->getType())); 5674 return Builder.CreateStore(Ops[1], Ops[0]); 5675 } 5676 case NEON::BI__builtin_neon_vld2_dup_v: 5677 case NEON::BI__builtin_neon_vld2q_dup_v: { 5678 llvm::Type *PTy = 5679 llvm::PointerType::getUnqual(VTy->getElementType()); 5680 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5681 llvm::Type *Tys[2] = { VTy, PTy }; 5682 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5683 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5684 Ops[0] = Builder.CreateBitCast(Ops[0], 5685 llvm::PointerType::getUnqual(Ops[1]->getType())); 5686 return Builder.CreateStore(Ops[1], Ops[0]); 5687 } 5688 case NEON::BI__builtin_neon_vld3_dup_v: 5689 case NEON::BI__builtin_neon_vld3q_dup_v: { 5690 llvm::Type *PTy = 5691 llvm::PointerType::getUnqual(VTy->getElementType()); 5692 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5693 llvm::Type *Tys[2] = { VTy, PTy }; 5694 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5695 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5696 Ops[0] = Builder.CreateBitCast(Ops[0], 5697 llvm::PointerType::getUnqual(Ops[1]->getType())); 5698 return Builder.CreateStore(Ops[1], Ops[0]); 5699 } 5700 case NEON::BI__builtin_neon_vld4_dup_v: 5701 case NEON::BI__builtin_neon_vld4q_dup_v: { 5702 llvm::Type *PTy = 5703 llvm::PointerType::getUnqual(VTy->getElementType()); 5704 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5705 llvm::Type *Tys[2] = { VTy, PTy }; 5706 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5707 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5708 Ops[0] = Builder.CreateBitCast(Ops[0], 5709 llvm::PointerType::getUnqual(Ops[1]->getType())); 5710 return Builder.CreateStore(Ops[1], Ops[0]); 5711 } 5712 case NEON::BI__builtin_neon_vld2_lane_v: 5713 case NEON::BI__builtin_neon_vld2q_lane_v: { 5714 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5715 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5716 Ops.push_back(Ops[1]); 5717 Ops.erase(Ops.begin()+1); 5718 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5719 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5720 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5721 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5722 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5723 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5724 return Builder.CreateStore(Ops[1], Ops[0]); 5725 } 5726 case NEON::BI__builtin_neon_vld3_lane_v: 5727 case NEON::BI__builtin_neon_vld3q_lane_v: { 5728 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5729 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5730 Ops.push_back(Ops[1]); 5731 Ops.erase(Ops.begin()+1); 5732 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5733 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5734 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5735 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5736 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5737 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5738 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5739 return Builder.CreateStore(Ops[1], Ops[0]); 5740 } 5741 case NEON::BI__builtin_neon_vld4_lane_v: 5742 case NEON::BI__builtin_neon_vld4q_lane_v: { 5743 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5744 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5745 Ops.push_back(Ops[1]); 5746 Ops.erase(Ops.begin()+1); 5747 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5748 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5749 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5750 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5751 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 5752 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5753 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5754 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5755 return Builder.CreateStore(Ops[1], Ops[0]); 5756 } 5757 case NEON::BI__builtin_neon_vst2_v: 5758 case NEON::BI__builtin_neon_vst2q_v: { 5759 Ops.push_back(Ops[0]); 5760 Ops.erase(Ops.begin()); 5761 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5762 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5763 Ops, ""); 5764 } 5765 case NEON::BI__builtin_neon_vst2_lane_v: 5766 case NEON::BI__builtin_neon_vst2q_lane_v: { 5767 Ops.push_back(Ops[0]); 5768 Ops.erase(Ops.begin()); 5769 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 5770 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5771 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5772 Ops, ""); 5773 } 5774 case NEON::BI__builtin_neon_vst3_v: 5775 case NEON::BI__builtin_neon_vst3q_v: { 5776 Ops.push_back(Ops[0]); 5777 Ops.erase(Ops.begin()); 5778 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5779 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5780 Ops, ""); 5781 } 5782 case NEON::BI__builtin_neon_vst3_lane_v: 5783 case NEON::BI__builtin_neon_vst3q_lane_v: { 5784 Ops.push_back(Ops[0]); 5785 Ops.erase(Ops.begin()); 5786 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5787 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5788 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5789 Ops, ""); 5790 } 5791 case NEON::BI__builtin_neon_vst4_v: 5792 case NEON::BI__builtin_neon_vst4q_v: { 5793 Ops.push_back(Ops[0]); 5794 Ops.erase(Ops.begin()); 5795 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5796 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5797 Ops, ""); 5798 } 5799 case NEON::BI__builtin_neon_vst4_lane_v: 5800 case NEON::BI__builtin_neon_vst4q_lane_v: { 5801 Ops.push_back(Ops[0]); 5802 Ops.erase(Ops.begin()); 5803 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5804 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5805 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5806 Ops, ""); 5807 } 5808 case NEON::BI__builtin_neon_vtrn_v: 5809 case NEON::BI__builtin_neon_vtrnq_v: { 5810 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5811 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5812 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5813 Value *SV = nullptr; 5814 5815 for (unsigned vi = 0; vi != 2; ++vi) { 5816 SmallVector<Constant*, 16> Indices; 5817 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5818 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5819 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5820 } 5821 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5822 SV = llvm::ConstantVector::get(Indices); 5823 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5824 SV = Builder.CreateStore(SV, Addr); 5825 } 5826 return SV; 5827 } 5828 case NEON::BI__builtin_neon_vuzp_v: 5829 case NEON::BI__builtin_neon_vuzpq_v: { 5830 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5831 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5832 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5833 Value *SV = nullptr; 5834 5835 for (unsigned vi = 0; vi != 2; ++vi) { 5836 SmallVector<Constant*, 16> Indices; 5837 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5838 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5839 5840 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5841 SV = llvm::ConstantVector::get(Indices); 5842 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5843 SV = Builder.CreateStore(SV, Addr); 5844 } 5845 return SV; 5846 } 5847 case NEON::BI__builtin_neon_vzip_v: 5848 case NEON::BI__builtin_neon_vzipq_v: { 5849 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5850 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5851 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5852 Value *SV = nullptr; 5853 5854 for (unsigned vi = 0; vi != 2; ++vi) { 5855 SmallVector<Constant*, 16> Indices; 5856 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5857 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5858 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5859 } 5860 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5861 SV = llvm::ConstantVector::get(Indices); 5862 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5863 SV = Builder.CreateStore(SV, Addr); 5864 } 5865 return SV; 5866 } 5867 case NEON::BI__builtin_neon_vqtbl1q_v: { 5868 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5869 Ops, "vtbl1"); 5870 } 5871 case NEON::BI__builtin_neon_vqtbl2q_v: { 5872 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5873 Ops, "vtbl2"); 5874 } 5875 case NEON::BI__builtin_neon_vqtbl3q_v: { 5876 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5877 Ops, "vtbl3"); 5878 } 5879 case NEON::BI__builtin_neon_vqtbl4q_v: { 5880 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5881 Ops, "vtbl4"); 5882 } 5883 case NEON::BI__builtin_neon_vqtbx1q_v: { 5884 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5885 Ops, "vtbx1"); 5886 } 5887 case NEON::BI__builtin_neon_vqtbx2q_v: { 5888 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5889 Ops, "vtbx2"); 5890 } 5891 case NEON::BI__builtin_neon_vqtbx3q_v: { 5892 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5893 Ops, "vtbx3"); 5894 } 5895 case NEON::BI__builtin_neon_vqtbx4q_v: { 5896 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5897 Ops, "vtbx4"); 5898 } 5899 case NEON::BI__builtin_neon_vsqadd_v: 5900 case NEON::BI__builtin_neon_vsqaddq_v: { 5901 Int = Intrinsic::aarch64_neon_usqadd; 5902 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5903 } 5904 case NEON::BI__builtin_neon_vuqadd_v: 5905 case NEON::BI__builtin_neon_vuqaddq_v: { 5906 Int = Intrinsic::aarch64_neon_suqadd; 5907 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5908 } 5909 } 5910 } 5911 5912 llvm::Value *CodeGenFunction:: 5913 BuildVector(ArrayRef<llvm::Value*> Ops) { 5914 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5915 "Not a power-of-two sized vector!"); 5916 bool AllConstants = true; 5917 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5918 AllConstants &= isa<Constant>(Ops[i]); 5919 5920 // If this is a constant vector, create a ConstantVector. 5921 if (AllConstants) { 5922 SmallVector<llvm::Constant*, 16> CstOps; 5923 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5924 CstOps.push_back(cast<Constant>(Ops[i])); 5925 return llvm::ConstantVector::get(CstOps); 5926 } 5927 5928 // Otherwise, insertelement the values to build the vector. 5929 Value *Result = 5930 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5931 5932 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5933 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5934 5935 return Result; 5936 } 5937 5938 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5939 const CallExpr *E) { 5940 SmallVector<Value*, 4> Ops; 5941 5942 // Find out if any arguments are required to be integer constant expressions. 5943 unsigned ICEArguments = 0; 5944 ASTContext::GetBuiltinTypeError Error; 5945 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5946 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5947 5948 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5949 // If this is a normal argument, just emit it as a scalar. 5950 if ((ICEArguments & (1 << i)) == 0) { 5951 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5952 continue; 5953 } 5954 5955 // If this is required to be a constant, constant fold it so that we know 5956 // that the generated intrinsic gets a ConstantInt. 5957 llvm::APSInt Result; 5958 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5959 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5960 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5961 } 5962 5963 switch (BuiltinID) { 5964 default: return nullptr; 5965 case X86::BI__builtin_cpu_supports: { 5966 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 5967 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 5968 5969 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 5970 // based mapping. 5971 // Processor features and mapping to processor feature value. 5972 enum X86Features { 5973 CMOV = 0, 5974 MMX, 5975 POPCNT, 5976 SSE, 5977 SSE2, 5978 SSE3, 5979 SSSE3, 5980 SSE4_1, 5981 SSE4_2, 5982 AVX, 5983 AVX2, 5984 SSE4_A, 5985 FMA4, 5986 XOP, 5987 FMA, 5988 AVX512F, 5989 BMI, 5990 BMI2, 5991 MAX 5992 }; 5993 5994 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 5995 .Case("cmov", X86Features::CMOV) 5996 .Case("mmx", X86Features::MMX) 5997 .Case("popcnt", X86Features::POPCNT) 5998 .Case("sse", X86Features::SSE) 5999 .Case("sse2", X86Features::SSE2) 6000 .Case("sse3", X86Features::SSE3) 6001 .Case("sse4.1", X86Features::SSE4_1) 6002 .Case("sse4.2", X86Features::SSE4_2) 6003 .Case("avx", X86Features::AVX) 6004 .Case("avx2", X86Features::AVX2) 6005 .Case("sse4a", X86Features::SSE4_A) 6006 .Case("fma4", X86Features::FMA4) 6007 .Case("xop", X86Features::XOP) 6008 .Case("fma", X86Features::FMA) 6009 .Case("avx512f", X86Features::AVX512F) 6010 .Case("bmi", X86Features::BMI) 6011 .Case("bmi2", X86Features::BMI2) 6012 .Default(X86Features::MAX); 6013 assert(Feature != X86Features::MAX && "Invalid feature!"); 6014 6015 // Matching the struct layout from the compiler-rt/libgcc structure that is 6016 // filled in: 6017 // unsigned int __cpu_vendor; 6018 // unsigned int __cpu_type; 6019 // unsigned int __cpu_subtype; 6020 // unsigned int __cpu_features[1]; 6021 llvm::Type *STy = llvm::StructType::get( 6022 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6023 6024 // Grab the global __cpu_model. 6025 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6026 6027 // Grab the first (0th) element from the field __cpu_features off of the 6028 // global in the struct STy. 6029 Value *Idxs[] = { 6030 ConstantInt::get(Int32Ty, 0), 6031 ConstantInt::get(Int32Ty, 3), 6032 ConstantInt::get(Int32Ty, 0) 6033 }; 6034 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6035 Value *Features = Builder.CreateLoad(CpuFeatures); 6036 6037 // Check the value of the bit corresponding to the feature requested. 6038 Value *Bitset = Builder.CreateAnd( 6039 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6040 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6041 } 6042 case X86::BI_mm_prefetch: { 6043 Value *Address = EmitScalarExpr(E->getArg(0)); 6044 Value *RW = ConstantInt::get(Int32Ty, 0); 6045 Value *Locality = EmitScalarExpr(E->getArg(1)); 6046 Value *Data = ConstantInt::get(Int32Ty, 1); 6047 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6048 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6049 } 6050 case X86::BI__builtin_ia32_undef128: 6051 case X86::BI__builtin_ia32_undef256: 6052 case X86::BI__builtin_ia32_undef512: 6053 return UndefValue::get(ConvertType(E->getType())); 6054 case X86::BI__builtin_ia32_vec_init_v8qi: 6055 case X86::BI__builtin_ia32_vec_init_v4hi: 6056 case X86::BI__builtin_ia32_vec_init_v2si: 6057 return Builder.CreateBitCast(BuildVector(Ops), 6058 llvm::Type::getX86_MMXTy(getLLVMContext())); 6059 case X86::BI__builtin_ia32_vec_ext_v2si: 6060 return Builder.CreateExtractElement(Ops[0], 6061 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6062 case X86::BI__builtin_ia32_ldmxcsr: { 6063 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 6064 Builder.CreateStore(Ops[0], Tmp); 6065 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6066 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6067 } 6068 case X86::BI__builtin_ia32_stmxcsr: { 6069 Value *Tmp = CreateMemTemp(E->getType()); 6070 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6071 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6072 return Builder.CreateLoad(Tmp, "stmxcsr"); 6073 } 6074 case X86::BI__builtin_ia32_storehps: 6075 case X86::BI__builtin_ia32_storelps: { 6076 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6077 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6078 6079 // cast val v2i64 6080 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6081 6082 // extract (0, 1) 6083 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6084 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6085 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6086 6087 // cast pointer to i64 & store 6088 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6089 return Builder.CreateStore(Ops[1], Ops[0]); 6090 } 6091 case X86::BI__builtin_ia32_palignr128: 6092 case X86::BI__builtin_ia32_palignr256: { 6093 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6094 6095 unsigned NumElts = 6096 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6097 assert(NumElts % 16 == 0); 6098 unsigned NumLanes = NumElts / 16; 6099 unsigned NumLaneElts = NumElts / NumLanes; 6100 6101 // If palignr is shifting the pair of vectors more than the size of two 6102 // lanes, emit zero. 6103 if (ShiftVal >= (2 * NumLaneElts)) 6104 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6105 6106 // If palignr is shifting the pair of input vectors more than one lane, 6107 // but less than two lanes, convert to shifting in zeroes. 6108 if (ShiftVal > NumLaneElts) { 6109 ShiftVal -= NumLaneElts; 6110 Ops[1] = Ops[0]; 6111 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6112 } 6113 6114 uint32_t Indices[32]; 6115 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6116 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6117 for (unsigned i = 0; i != NumLaneElts; ++i) { 6118 unsigned Idx = ShiftVal + i; 6119 if (Idx >= NumLaneElts) 6120 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6121 Indices[l + i] = Idx + l; 6122 } 6123 } 6124 6125 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6126 makeArrayRef(Indices, NumElts)); 6127 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6128 } 6129 case X86::BI__builtin_ia32_pslldqi256: { 6130 // Shift value is in bits so divide by 8. 6131 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6132 6133 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6134 if (shiftVal >= 16) 6135 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6136 6137 uint32_t Indices[32]; 6138 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6139 for (unsigned l = 0; l != 32; l += 16) { 6140 for (unsigned i = 0; i != 16; ++i) { 6141 unsigned Idx = 32 + i - shiftVal; 6142 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6143 Indices[l + i] = Idx + l; 6144 } 6145 } 6146 6147 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6148 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6149 Value *Zero = llvm::Constant::getNullValue(VecTy); 6150 6151 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6152 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6153 llvm::Type *ResultType = ConvertType(E->getType()); 6154 return Builder.CreateBitCast(SV, ResultType, "cast"); 6155 } 6156 case X86::BI__builtin_ia32_psrldqi256: { 6157 // Shift value is in bits so divide by 8. 6158 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6159 6160 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6161 if (shiftVal >= 16) 6162 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6163 6164 uint32_t Indices[32]; 6165 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6166 for (unsigned l = 0; l != 32; l += 16) { 6167 for (unsigned i = 0; i != 16; ++i) { 6168 unsigned Idx = i + shiftVal; 6169 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6170 Indices[l + i] = Idx + l; 6171 } 6172 } 6173 6174 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6175 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6176 Value *Zero = llvm::Constant::getNullValue(VecTy); 6177 6178 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6179 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6180 llvm::Type *ResultType = ConvertType(E->getType()); 6181 return Builder.CreateBitCast(SV, ResultType, "cast"); 6182 } 6183 case X86::BI__builtin_ia32_movntps: 6184 case X86::BI__builtin_ia32_movntps256: 6185 case X86::BI__builtin_ia32_movntpd: 6186 case X86::BI__builtin_ia32_movntpd256: 6187 case X86::BI__builtin_ia32_movntdq: 6188 case X86::BI__builtin_ia32_movntdq256: 6189 case X86::BI__builtin_ia32_movnti: 6190 case X86::BI__builtin_ia32_movnti64: { 6191 llvm::MDNode *Node = llvm::MDNode::get( 6192 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6193 6194 // Convert the type of the pointer to a pointer to the stored type. 6195 Value *BC = Builder.CreateBitCast(Ops[0], 6196 llvm::PointerType::getUnqual(Ops[1]->getType()), 6197 "cast"); 6198 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6199 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6200 6201 // If the operand is an integer, we can't assume alignment. Otherwise, 6202 // assume natural alignment. 6203 QualType ArgTy = E->getArg(1)->getType(); 6204 unsigned Align; 6205 if (ArgTy->isIntegerType()) 6206 Align = 1; 6207 else 6208 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6209 SI->setAlignment(Align); 6210 return SI; 6211 } 6212 // 3DNow! 6213 case X86::BI__builtin_ia32_pswapdsf: 6214 case X86::BI__builtin_ia32_pswapdsi: { 6215 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6216 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6217 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6218 return Builder.CreateCall(F, Ops, "pswapd"); 6219 } 6220 case X86::BI__builtin_ia32_rdrand16_step: 6221 case X86::BI__builtin_ia32_rdrand32_step: 6222 case X86::BI__builtin_ia32_rdrand64_step: 6223 case X86::BI__builtin_ia32_rdseed16_step: 6224 case X86::BI__builtin_ia32_rdseed32_step: 6225 case X86::BI__builtin_ia32_rdseed64_step: { 6226 Intrinsic::ID ID; 6227 switch (BuiltinID) { 6228 default: llvm_unreachable("Unsupported intrinsic!"); 6229 case X86::BI__builtin_ia32_rdrand16_step: 6230 ID = Intrinsic::x86_rdrand_16; 6231 break; 6232 case X86::BI__builtin_ia32_rdrand32_step: 6233 ID = Intrinsic::x86_rdrand_32; 6234 break; 6235 case X86::BI__builtin_ia32_rdrand64_step: 6236 ID = Intrinsic::x86_rdrand_64; 6237 break; 6238 case X86::BI__builtin_ia32_rdseed16_step: 6239 ID = Intrinsic::x86_rdseed_16; 6240 break; 6241 case X86::BI__builtin_ia32_rdseed32_step: 6242 ID = Intrinsic::x86_rdseed_32; 6243 break; 6244 case X86::BI__builtin_ia32_rdseed64_step: 6245 ID = Intrinsic::x86_rdseed_64; 6246 break; 6247 } 6248 6249 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6250 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6251 return Builder.CreateExtractValue(Call, 1); 6252 } 6253 // SSE comparison intrisics 6254 case X86::BI__builtin_ia32_cmpeqps: 6255 case X86::BI__builtin_ia32_cmpltps: 6256 case X86::BI__builtin_ia32_cmpleps: 6257 case X86::BI__builtin_ia32_cmpunordps: 6258 case X86::BI__builtin_ia32_cmpneqps: 6259 case X86::BI__builtin_ia32_cmpnltps: 6260 case X86::BI__builtin_ia32_cmpnleps: 6261 case X86::BI__builtin_ia32_cmpordps: 6262 case X86::BI__builtin_ia32_cmpeqss: 6263 case X86::BI__builtin_ia32_cmpltss: 6264 case X86::BI__builtin_ia32_cmpless: 6265 case X86::BI__builtin_ia32_cmpunordss: 6266 case X86::BI__builtin_ia32_cmpneqss: 6267 case X86::BI__builtin_ia32_cmpnltss: 6268 case X86::BI__builtin_ia32_cmpnless: 6269 case X86::BI__builtin_ia32_cmpordss: 6270 case X86::BI__builtin_ia32_cmpeqpd: 6271 case X86::BI__builtin_ia32_cmpltpd: 6272 case X86::BI__builtin_ia32_cmplepd: 6273 case X86::BI__builtin_ia32_cmpunordpd: 6274 case X86::BI__builtin_ia32_cmpneqpd: 6275 case X86::BI__builtin_ia32_cmpnltpd: 6276 case X86::BI__builtin_ia32_cmpnlepd: 6277 case X86::BI__builtin_ia32_cmpordpd: 6278 case X86::BI__builtin_ia32_cmpeqsd: 6279 case X86::BI__builtin_ia32_cmpltsd: 6280 case X86::BI__builtin_ia32_cmplesd: 6281 case X86::BI__builtin_ia32_cmpunordsd: 6282 case X86::BI__builtin_ia32_cmpneqsd: 6283 case X86::BI__builtin_ia32_cmpnltsd: 6284 case X86::BI__builtin_ia32_cmpnlesd: 6285 case X86::BI__builtin_ia32_cmpordsd: 6286 // These exist so that the builtin that takes an immediate can be bounds 6287 // checked by clang to avoid passing bad immediates to the backend. Since 6288 // AVX has a larger immediate than SSE we would need separate builtins to 6289 // do the different bounds checking. Rather than create a clang specific 6290 // SSE only builtin, this implements eight separate builtins to match gcc 6291 // implementation. 6292 6293 // Choose the immediate. 6294 unsigned Imm; 6295 switch (BuiltinID) { 6296 default: llvm_unreachable("Unsupported intrinsic!"); 6297 case X86::BI__builtin_ia32_cmpeqps: 6298 case X86::BI__builtin_ia32_cmpeqss: 6299 case X86::BI__builtin_ia32_cmpeqpd: 6300 case X86::BI__builtin_ia32_cmpeqsd: 6301 Imm = 0; 6302 break; 6303 case X86::BI__builtin_ia32_cmpltps: 6304 case X86::BI__builtin_ia32_cmpltss: 6305 case X86::BI__builtin_ia32_cmpltpd: 6306 case X86::BI__builtin_ia32_cmpltsd: 6307 Imm = 1; 6308 break; 6309 case X86::BI__builtin_ia32_cmpleps: 6310 case X86::BI__builtin_ia32_cmpless: 6311 case X86::BI__builtin_ia32_cmplepd: 6312 case X86::BI__builtin_ia32_cmplesd: 6313 Imm = 2; 6314 break; 6315 case X86::BI__builtin_ia32_cmpunordps: 6316 case X86::BI__builtin_ia32_cmpunordss: 6317 case X86::BI__builtin_ia32_cmpunordpd: 6318 case X86::BI__builtin_ia32_cmpunordsd: 6319 Imm = 3; 6320 break; 6321 case X86::BI__builtin_ia32_cmpneqps: 6322 case X86::BI__builtin_ia32_cmpneqss: 6323 case X86::BI__builtin_ia32_cmpneqpd: 6324 case X86::BI__builtin_ia32_cmpneqsd: 6325 Imm = 4; 6326 break; 6327 case X86::BI__builtin_ia32_cmpnltps: 6328 case X86::BI__builtin_ia32_cmpnltss: 6329 case X86::BI__builtin_ia32_cmpnltpd: 6330 case X86::BI__builtin_ia32_cmpnltsd: 6331 Imm = 5; 6332 break; 6333 case X86::BI__builtin_ia32_cmpnleps: 6334 case X86::BI__builtin_ia32_cmpnless: 6335 case X86::BI__builtin_ia32_cmpnlepd: 6336 case X86::BI__builtin_ia32_cmpnlesd: 6337 Imm = 6; 6338 break; 6339 case X86::BI__builtin_ia32_cmpordps: 6340 case X86::BI__builtin_ia32_cmpordss: 6341 case X86::BI__builtin_ia32_cmpordpd: 6342 case X86::BI__builtin_ia32_cmpordsd: 6343 Imm = 7; 6344 break; 6345 } 6346 6347 // Choose the intrinsic ID. 6348 const char *name; 6349 Intrinsic::ID ID; 6350 switch (BuiltinID) { 6351 default: llvm_unreachable("Unsupported intrinsic!"); 6352 case X86::BI__builtin_ia32_cmpeqps: 6353 case X86::BI__builtin_ia32_cmpltps: 6354 case X86::BI__builtin_ia32_cmpleps: 6355 case X86::BI__builtin_ia32_cmpunordps: 6356 case X86::BI__builtin_ia32_cmpneqps: 6357 case X86::BI__builtin_ia32_cmpnltps: 6358 case X86::BI__builtin_ia32_cmpnleps: 6359 case X86::BI__builtin_ia32_cmpordps: 6360 name = "cmpps"; 6361 ID = Intrinsic::x86_sse_cmp_ps; 6362 break; 6363 case X86::BI__builtin_ia32_cmpeqss: 6364 case X86::BI__builtin_ia32_cmpltss: 6365 case X86::BI__builtin_ia32_cmpless: 6366 case X86::BI__builtin_ia32_cmpunordss: 6367 case X86::BI__builtin_ia32_cmpneqss: 6368 case X86::BI__builtin_ia32_cmpnltss: 6369 case X86::BI__builtin_ia32_cmpnless: 6370 case X86::BI__builtin_ia32_cmpordss: 6371 name = "cmpss"; 6372 ID = Intrinsic::x86_sse_cmp_ss; 6373 break; 6374 case X86::BI__builtin_ia32_cmpeqpd: 6375 case X86::BI__builtin_ia32_cmpltpd: 6376 case X86::BI__builtin_ia32_cmplepd: 6377 case X86::BI__builtin_ia32_cmpunordpd: 6378 case X86::BI__builtin_ia32_cmpneqpd: 6379 case X86::BI__builtin_ia32_cmpnltpd: 6380 case X86::BI__builtin_ia32_cmpnlepd: 6381 case X86::BI__builtin_ia32_cmpordpd: 6382 name = "cmppd"; 6383 ID = Intrinsic::x86_sse2_cmp_pd; 6384 break; 6385 case X86::BI__builtin_ia32_cmpeqsd: 6386 case X86::BI__builtin_ia32_cmpltsd: 6387 case X86::BI__builtin_ia32_cmplesd: 6388 case X86::BI__builtin_ia32_cmpunordsd: 6389 case X86::BI__builtin_ia32_cmpneqsd: 6390 case X86::BI__builtin_ia32_cmpnltsd: 6391 case X86::BI__builtin_ia32_cmpnlesd: 6392 case X86::BI__builtin_ia32_cmpordsd: 6393 name = "cmpsd"; 6394 ID = Intrinsic::x86_sse2_cmp_sd; 6395 break; 6396 } 6397 6398 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6399 llvm::Function *F = CGM.getIntrinsic(ID); 6400 return Builder.CreateCall(F, Ops, name); 6401 } 6402 } 6403 6404 6405 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6406 const CallExpr *E) { 6407 SmallVector<Value*, 4> Ops; 6408 6409 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6410 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6411 6412 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6413 6414 switch (BuiltinID) { 6415 default: return nullptr; 6416 6417 // vec_ld, vec_lvsl, vec_lvsr 6418 case PPC::BI__builtin_altivec_lvx: 6419 case PPC::BI__builtin_altivec_lvxl: 6420 case PPC::BI__builtin_altivec_lvebx: 6421 case PPC::BI__builtin_altivec_lvehx: 6422 case PPC::BI__builtin_altivec_lvewx: 6423 case PPC::BI__builtin_altivec_lvsl: 6424 case PPC::BI__builtin_altivec_lvsr: 6425 case PPC::BI__builtin_vsx_lxvd2x: 6426 case PPC::BI__builtin_vsx_lxvw4x: 6427 { 6428 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6429 6430 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6431 Ops.pop_back(); 6432 6433 switch (BuiltinID) { 6434 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6435 case PPC::BI__builtin_altivec_lvx: 6436 ID = Intrinsic::ppc_altivec_lvx; 6437 break; 6438 case PPC::BI__builtin_altivec_lvxl: 6439 ID = Intrinsic::ppc_altivec_lvxl; 6440 break; 6441 case PPC::BI__builtin_altivec_lvebx: 6442 ID = Intrinsic::ppc_altivec_lvebx; 6443 break; 6444 case PPC::BI__builtin_altivec_lvehx: 6445 ID = Intrinsic::ppc_altivec_lvehx; 6446 break; 6447 case PPC::BI__builtin_altivec_lvewx: 6448 ID = Intrinsic::ppc_altivec_lvewx; 6449 break; 6450 case PPC::BI__builtin_altivec_lvsl: 6451 ID = Intrinsic::ppc_altivec_lvsl; 6452 break; 6453 case PPC::BI__builtin_altivec_lvsr: 6454 ID = Intrinsic::ppc_altivec_lvsr; 6455 break; 6456 case PPC::BI__builtin_vsx_lxvd2x: 6457 ID = Intrinsic::ppc_vsx_lxvd2x; 6458 break; 6459 case PPC::BI__builtin_vsx_lxvw4x: 6460 ID = Intrinsic::ppc_vsx_lxvw4x; 6461 break; 6462 } 6463 llvm::Function *F = CGM.getIntrinsic(ID); 6464 return Builder.CreateCall(F, Ops, ""); 6465 } 6466 6467 // vec_st 6468 case PPC::BI__builtin_altivec_stvx: 6469 case PPC::BI__builtin_altivec_stvxl: 6470 case PPC::BI__builtin_altivec_stvebx: 6471 case PPC::BI__builtin_altivec_stvehx: 6472 case PPC::BI__builtin_altivec_stvewx: 6473 case PPC::BI__builtin_vsx_stxvd2x: 6474 case PPC::BI__builtin_vsx_stxvw4x: 6475 { 6476 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6477 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6478 Ops.pop_back(); 6479 6480 switch (BuiltinID) { 6481 default: llvm_unreachable("Unsupported st intrinsic!"); 6482 case PPC::BI__builtin_altivec_stvx: 6483 ID = Intrinsic::ppc_altivec_stvx; 6484 break; 6485 case PPC::BI__builtin_altivec_stvxl: 6486 ID = Intrinsic::ppc_altivec_stvxl; 6487 break; 6488 case PPC::BI__builtin_altivec_stvebx: 6489 ID = Intrinsic::ppc_altivec_stvebx; 6490 break; 6491 case PPC::BI__builtin_altivec_stvehx: 6492 ID = Intrinsic::ppc_altivec_stvehx; 6493 break; 6494 case PPC::BI__builtin_altivec_stvewx: 6495 ID = Intrinsic::ppc_altivec_stvewx; 6496 break; 6497 case PPC::BI__builtin_vsx_stxvd2x: 6498 ID = Intrinsic::ppc_vsx_stxvd2x; 6499 break; 6500 case PPC::BI__builtin_vsx_stxvw4x: 6501 ID = Intrinsic::ppc_vsx_stxvw4x; 6502 break; 6503 } 6504 llvm::Function *F = CGM.getIntrinsic(ID); 6505 return Builder.CreateCall(F, Ops, ""); 6506 } 6507 // Square root 6508 case PPC::BI__builtin_vsx_xvsqrtsp: 6509 case PPC::BI__builtin_vsx_xvsqrtdp: { 6510 llvm::Type *ResultType = ConvertType(E->getType()); 6511 Value *X = EmitScalarExpr(E->getArg(0)); 6512 ID = Intrinsic::sqrt; 6513 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6514 return Builder.CreateCall(F, X); 6515 } 6516 // Count leading zeros 6517 case PPC::BI__builtin_altivec_vclzb: 6518 case PPC::BI__builtin_altivec_vclzh: 6519 case PPC::BI__builtin_altivec_vclzw: 6520 case PPC::BI__builtin_altivec_vclzd: { 6521 llvm::Type *ResultType = ConvertType(E->getType()); 6522 Value *X = EmitScalarExpr(E->getArg(0)); 6523 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6524 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6525 return Builder.CreateCall(F, {X, Undef}); 6526 } 6527 // Copy sign 6528 case PPC::BI__builtin_vsx_xvcpsgnsp: 6529 case PPC::BI__builtin_vsx_xvcpsgndp: { 6530 llvm::Type *ResultType = ConvertType(E->getType()); 6531 Value *X = EmitScalarExpr(E->getArg(0)); 6532 Value *Y = EmitScalarExpr(E->getArg(1)); 6533 ID = Intrinsic::copysign; 6534 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6535 return Builder.CreateCall(F, {X, Y}); 6536 } 6537 // Rounding/truncation 6538 case PPC::BI__builtin_vsx_xvrspip: 6539 case PPC::BI__builtin_vsx_xvrdpip: 6540 case PPC::BI__builtin_vsx_xvrdpim: 6541 case PPC::BI__builtin_vsx_xvrspim: 6542 case PPC::BI__builtin_vsx_xvrdpi: 6543 case PPC::BI__builtin_vsx_xvrspi: 6544 case PPC::BI__builtin_vsx_xvrdpic: 6545 case PPC::BI__builtin_vsx_xvrspic: 6546 case PPC::BI__builtin_vsx_xvrdpiz: 6547 case PPC::BI__builtin_vsx_xvrspiz: { 6548 llvm::Type *ResultType = ConvertType(E->getType()); 6549 Value *X = EmitScalarExpr(E->getArg(0)); 6550 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6551 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6552 ID = Intrinsic::floor; 6553 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6554 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6555 ID = Intrinsic::round; 6556 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6557 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6558 ID = Intrinsic::nearbyint; 6559 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6560 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6561 ID = Intrinsic::ceil; 6562 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6563 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6564 ID = Intrinsic::trunc; 6565 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6566 return Builder.CreateCall(F, X); 6567 } 6568 // FMA variations 6569 case PPC::BI__builtin_vsx_xvmaddadp: 6570 case PPC::BI__builtin_vsx_xvmaddasp: 6571 case PPC::BI__builtin_vsx_xvnmaddadp: 6572 case PPC::BI__builtin_vsx_xvnmaddasp: 6573 case PPC::BI__builtin_vsx_xvmsubadp: 6574 case PPC::BI__builtin_vsx_xvmsubasp: 6575 case PPC::BI__builtin_vsx_xvnmsubadp: 6576 case PPC::BI__builtin_vsx_xvnmsubasp: { 6577 llvm::Type *ResultType = ConvertType(E->getType()); 6578 Value *X = EmitScalarExpr(E->getArg(0)); 6579 Value *Y = EmitScalarExpr(E->getArg(1)); 6580 Value *Z = EmitScalarExpr(E->getArg(2)); 6581 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6582 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6583 switch (BuiltinID) { 6584 case PPC::BI__builtin_vsx_xvmaddadp: 6585 case PPC::BI__builtin_vsx_xvmaddasp: 6586 return Builder.CreateCall(F, {X, Y, Z}); 6587 case PPC::BI__builtin_vsx_xvnmaddadp: 6588 case PPC::BI__builtin_vsx_xvnmaddasp: 6589 return Builder.CreateFSub(Zero, 6590 Builder.CreateCall(F, {X, Y, Z}), "sub"); 6591 case PPC::BI__builtin_vsx_xvmsubadp: 6592 case PPC::BI__builtin_vsx_xvmsubasp: 6593 return Builder.CreateCall(F, 6594 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6595 case PPC::BI__builtin_vsx_xvnmsubadp: 6596 case PPC::BI__builtin_vsx_xvnmsubasp: 6597 Value *FsubRes = 6598 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6599 return Builder.CreateFSub(Zero, FsubRes, "sub"); 6600 } 6601 llvm_unreachable("Unknown FMA operation"); 6602 return nullptr; // Suppress no-return warning 6603 } 6604 } 6605 } 6606 6607 // Emit an intrinsic that has 1 float or double. 6608 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6609 const CallExpr *E, 6610 unsigned IntrinsicID) { 6611 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6612 6613 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6614 return CGF.Builder.CreateCall(F, Src0); 6615 } 6616 6617 // Emit an intrinsic that has 3 float or double operands. 6618 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6619 const CallExpr *E, 6620 unsigned IntrinsicID) { 6621 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6622 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6623 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6624 6625 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6626 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6627 } 6628 6629 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6630 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6631 const CallExpr *E, 6632 unsigned IntrinsicID) { 6633 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6634 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6635 6636 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6637 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6638 } 6639 6640 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 6641 const CallExpr *E) { 6642 switch (BuiltinID) { 6643 case AMDGPU::BI__builtin_amdgpu_div_scale: 6644 case AMDGPU::BI__builtin_amdgpu_div_scalef: { 6645 // Translate from the intrinsics's struct return to the builtin's out 6646 // argument. 6647 6648 std::pair<llvm::Value *, unsigned> FlagOutPtr 6649 = EmitPointerWithAlignment(E->getArg(3)); 6650 6651 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6652 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6653 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6654 6655 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6656 X->getType()); 6657 6658 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6659 6660 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6661 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6662 6663 llvm::Type *RealFlagType 6664 = FlagOutPtr.first->getType()->getPointerElementType(); 6665 6666 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6667 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6668 FlagStore->setAlignment(FlagOutPtr.second); 6669 return Result; 6670 } 6671 case AMDGPU::BI__builtin_amdgpu_div_fmas: 6672 case AMDGPU::BI__builtin_amdgpu_div_fmasf: { 6673 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6674 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6675 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6676 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6677 6678 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6679 Src0->getType()); 6680 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6681 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6682 } 6683 case AMDGPU::BI__builtin_amdgpu_div_fixup: 6684 case AMDGPU::BI__builtin_amdgpu_div_fixupf: 6685 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6686 case AMDGPU::BI__builtin_amdgpu_trig_preop: 6687 case AMDGPU::BI__builtin_amdgpu_trig_preopf: 6688 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6689 case AMDGPU::BI__builtin_amdgpu_rcp: 6690 case AMDGPU::BI__builtin_amdgpu_rcpf: 6691 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6692 case AMDGPU::BI__builtin_amdgpu_rsq: 6693 case AMDGPU::BI__builtin_amdgpu_rsqf: 6694 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6695 case AMDGPU::BI__builtin_amdgpu_rsq_clamped: 6696 case AMDGPU::BI__builtin_amdgpu_rsq_clampedf: 6697 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6698 case AMDGPU::BI__builtin_amdgpu_ldexp: 6699 case AMDGPU::BI__builtin_amdgpu_ldexpf: 6700 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6701 case AMDGPU::BI__builtin_amdgpu_class: 6702 case AMDGPU::BI__builtin_amdgpu_classf: 6703 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6704 default: 6705 return nullptr; 6706 } 6707 } 6708 6709 /// Handle a SystemZ function in which the final argument is a pointer 6710 /// to an int that receives the post-instruction CC value. At the LLVM level 6711 /// this is represented as a function that returns a {result, cc} pair. 6712 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6713 unsigned IntrinsicID, 6714 const CallExpr *E) { 6715 unsigned NumArgs = E->getNumArgs() - 1; 6716 SmallVector<Value *, 8> Args(NumArgs); 6717 for (unsigned I = 0; I < NumArgs; ++I) 6718 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6719 Value *CCPtr = CGF.EmitScalarExpr(E->getArg(NumArgs)); 6720 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6721 Value *Call = CGF.Builder.CreateCall(F, Args); 6722 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6723 CGF.Builder.CreateStore(CC, CCPtr); 6724 return CGF.Builder.CreateExtractValue(Call, 0); 6725 } 6726 6727 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6728 const CallExpr *E) { 6729 switch (BuiltinID) { 6730 case SystemZ::BI__builtin_tbegin: { 6731 Value *TDB = EmitScalarExpr(E->getArg(0)); 6732 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6733 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6734 return Builder.CreateCall(F, {TDB, Control}); 6735 } 6736 case SystemZ::BI__builtin_tbegin_nofloat: { 6737 Value *TDB = EmitScalarExpr(E->getArg(0)); 6738 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6739 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6740 return Builder.CreateCall(F, {TDB, Control}); 6741 } 6742 case SystemZ::BI__builtin_tbeginc: { 6743 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6744 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6745 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6746 return Builder.CreateCall(F, {TDB, Control}); 6747 } 6748 case SystemZ::BI__builtin_tabort: { 6749 Value *Data = EmitScalarExpr(E->getArg(0)); 6750 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6751 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6752 } 6753 case SystemZ::BI__builtin_non_tx_store: { 6754 Value *Address = EmitScalarExpr(E->getArg(0)); 6755 Value *Data = EmitScalarExpr(E->getArg(1)); 6756 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6757 return Builder.CreateCall(F, {Data, Address}); 6758 } 6759 6760 // Vector builtins. Note that most vector builtins are mapped automatically 6761 // to target-specific LLVM intrinsics. The ones handled specially here can 6762 // be represented via standard LLVM IR, which is preferable to enable common 6763 // LLVM optimizations. 6764 6765 case SystemZ::BI__builtin_s390_vpopctb: 6766 case SystemZ::BI__builtin_s390_vpopcth: 6767 case SystemZ::BI__builtin_s390_vpopctf: 6768 case SystemZ::BI__builtin_s390_vpopctg: { 6769 llvm::Type *ResultType = ConvertType(E->getType()); 6770 Value *X = EmitScalarExpr(E->getArg(0)); 6771 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6772 return Builder.CreateCall(F, X); 6773 } 6774 6775 case SystemZ::BI__builtin_s390_vclzb: 6776 case SystemZ::BI__builtin_s390_vclzh: 6777 case SystemZ::BI__builtin_s390_vclzf: 6778 case SystemZ::BI__builtin_s390_vclzg: { 6779 llvm::Type *ResultType = ConvertType(E->getType()); 6780 Value *X = EmitScalarExpr(E->getArg(0)); 6781 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6782 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6783 return Builder.CreateCall(F, {X, Undef}); 6784 } 6785 6786 case SystemZ::BI__builtin_s390_vctzb: 6787 case SystemZ::BI__builtin_s390_vctzh: 6788 case SystemZ::BI__builtin_s390_vctzf: 6789 case SystemZ::BI__builtin_s390_vctzg: { 6790 llvm::Type *ResultType = ConvertType(E->getType()); 6791 Value *X = EmitScalarExpr(E->getArg(0)); 6792 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6793 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6794 return Builder.CreateCall(F, {X, Undef}); 6795 } 6796 6797 case SystemZ::BI__builtin_s390_vfsqdb: { 6798 llvm::Type *ResultType = ConvertType(E->getType()); 6799 Value *X = EmitScalarExpr(E->getArg(0)); 6800 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6801 return Builder.CreateCall(F, X); 6802 } 6803 case SystemZ::BI__builtin_s390_vfmadb: { 6804 llvm::Type *ResultType = ConvertType(E->getType()); 6805 Value *X = EmitScalarExpr(E->getArg(0)); 6806 Value *Y = EmitScalarExpr(E->getArg(1)); 6807 Value *Z = EmitScalarExpr(E->getArg(2)); 6808 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6809 return Builder.CreateCall(F, {X, Y, Z}); 6810 } 6811 case SystemZ::BI__builtin_s390_vfmsdb: { 6812 llvm::Type *ResultType = ConvertType(E->getType()); 6813 Value *X = EmitScalarExpr(E->getArg(0)); 6814 Value *Y = EmitScalarExpr(E->getArg(1)); 6815 Value *Z = EmitScalarExpr(E->getArg(2)); 6816 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6817 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6818 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6819 } 6820 case SystemZ::BI__builtin_s390_vflpdb: { 6821 llvm::Type *ResultType = ConvertType(E->getType()); 6822 Value *X = EmitScalarExpr(E->getArg(0)); 6823 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6824 return Builder.CreateCall(F, X); 6825 } 6826 case SystemZ::BI__builtin_s390_vflndb: { 6827 llvm::Type *ResultType = ConvertType(E->getType()); 6828 Value *X = EmitScalarExpr(E->getArg(0)); 6829 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6830 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6831 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 6832 } 6833 case SystemZ::BI__builtin_s390_vfidb: { 6834 llvm::Type *ResultType = ConvertType(E->getType()); 6835 Value *X = EmitScalarExpr(E->getArg(0)); 6836 // Constant-fold the M4 and M5 mask arguments. 6837 llvm::APSInt M4, M5; 6838 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 6839 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 6840 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 6841 (void)IsConstM4; (void)IsConstM5; 6842 // Check whether this instance of vfidb can be represented via a LLVM 6843 // standard intrinsic. We only support some combinations of M4 and M5. 6844 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6845 switch (M4.getZExtValue()) { 6846 default: break; 6847 case 0: // IEEE-inexact exception allowed 6848 switch (M5.getZExtValue()) { 6849 default: break; 6850 case 0: ID = Intrinsic::rint; break; 6851 } 6852 break; 6853 case 4: // IEEE-inexact exception suppressed 6854 switch (M5.getZExtValue()) { 6855 default: break; 6856 case 0: ID = Intrinsic::nearbyint; break; 6857 case 1: ID = Intrinsic::round; break; 6858 case 5: ID = Intrinsic::trunc; break; 6859 case 6: ID = Intrinsic::ceil; break; 6860 case 7: ID = Intrinsic::floor; break; 6861 } 6862 break; 6863 } 6864 if (ID != Intrinsic::not_intrinsic) { 6865 Function *F = CGM.getIntrinsic(ID, ResultType); 6866 return Builder.CreateCall(F, X); 6867 } 6868 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 6869 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 6870 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 6871 return Builder.CreateCall(F, {X, M4Value, M5Value}); 6872 } 6873 6874 // Vector intrisincs that output the post-instruction CC value. 6875 6876 #define INTRINSIC_WITH_CC(NAME) \ 6877 case SystemZ::BI__builtin_##NAME: \ 6878 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 6879 6880 INTRINSIC_WITH_CC(s390_vpkshs); 6881 INTRINSIC_WITH_CC(s390_vpksfs); 6882 INTRINSIC_WITH_CC(s390_vpksgs); 6883 6884 INTRINSIC_WITH_CC(s390_vpklshs); 6885 INTRINSIC_WITH_CC(s390_vpklsfs); 6886 INTRINSIC_WITH_CC(s390_vpklsgs); 6887 6888 INTRINSIC_WITH_CC(s390_vceqbs); 6889 INTRINSIC_WITH_CC(s390_vceqhs); 6890 INTRINSIC_WITH_CC(s390_vceqfs); 6891 INTRINSIC_WITH_CC(s390_vceqgs); 6892 6893 INTRINSIC_WITH_CC(s390_vchbs); 6894 INTRINSIC_WITH_CC(s390_vchhs); 6895 INTRINSIC_WITH_CC(s390_vchfs); 6896 INTRINSIC_WITH_CC(s390_vchgs); 6897 6898 INTRINSIC_WITH_CC(s390_vchlbs); 6899 INTRINSIC_WITH_CC(s390_vchlhs); 6900 INTRINSIC_WITH_CC(s390_vchlfs); 6901 INTRINSIC_WITH_CC(s390_vchlgs); 6902 6903 INTRINSIC_WITH_CC(s390_vfaebs); 6904 INTRINSIC_WITH_CC(s390_vfaehs); 6905 INTRINSIC_WITH_CC(s390_vfaefs); 6906 6907 INTRINSIC_WITH_CC(s390_vfaezbs); 6908 INTRINSIC_WITH_CC(s390_vfaezhs); 6909 INTRINSIC_WITH_CC(s390_vfaezfs); 6910 6911 INTRINSIC_WITH_CC(s390_vfeebs); 6912 INTRINSIC_WITH_CC(s390_vfeehs); 6913 INTRINSIC_WITH_CC(s390_vfeefs); 6914 6915 INTRINSIC_WITH_CC(s390_vfeezbs); 6916 INTRINSIC_WITH_CC(s390_vfeezhs); 6917 INTRINSIC_WITH_CC(s390_vfeezfs); 6918 6919 INTRINSIC_WITH_CC(s390_vfenebs); 6920 INTRINSIC_WITH_CC(s390_vfenehs); 6921 INTRINSIC_WITH_CC(s390_vfenefs); 6922 6923 INTRINSIC_WITH_CC(s390_vfenezbs); 6924 INTRINSIC_WITH_CC(s390_vfenezhs); 6925 INTRINSIC_WITH_CC(s390_vfenezfs); 6926 6927 INTRINSIC_WITH_CC(s390_vistrbs); 6928 INTRINSIC_WITH_CC(s390_vistrhs); 6929 INTRINSIC_WITH_CC(s390_vistrfs); 6930 6931 INTRINSIC_WITH_CC(s390_vstrcbs); 6932 INTRINSIC_WITH_CC(s390_vstrchs); 6933 INTRINSIC_WITH_CC(s390_vstrcfs); 6934 6935 INTRINSIC_WITH_CC(s390_vstrczbs); 6936 INTRINSIC_WITH_CC(s390_vstrczhs); 6937 INTRINSIC_WITH_CC(s390_vstrczfs); 6938 6939 INTRINSIC_WITH_CC(s390_vfcedbs); 6940 INTRINSIC_WITH_CC(s390_vfchdbs); 6941 INTRINSIC_WITH_CC(s390_vfchedbs); 6942 6943 INTRINSIC_WITH_CC(s390_vftcidb); 6944 6945 #undef INTRINSIC_WITH_CC 6946 6947 default: 6948 return nullptr; 6949 } 6950 } 6951 6952 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 6953 const CallExpr *E) { 6954 switch (BuiltinID) { 6955 case NVPTX::BI__nvvm_atom_add_gen_i: 6956 case NVPTX::BI__nvvm_atom_add_gen_l: 6957 case NVPTX::BI__nvvm_atom_add_gen_ll: 6958 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 6959 6960 case NVPTX::BI__nvvm_atom_sub_gen_i: 6961 case NVPTX::BI__nvvm_atom_sub_gen_l: 6962 case NVPTX::BI__nvvm_atom_sub_gen_ll: 6963 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 6964 6965 case NVPTX::BI__nvvm_atom_and_gen_i: 6966 case NVPTX::BI__nvvm_atom_and_gen_l: 6967 case NVPTX::BI__nvvm_atom_and_gen_ll: 6968 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 6969 6970 case NVPTX::BI__nvvm_atom_or_gen_i: 6971 case NVPTX::BI__nvvm_atom_or_gen_l: 6972 case NVPTX::BI__nvvm_atom_or_gen_ll: 6973 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 6974 6975 case NVPTX::BI__nvvm_atom_xor_gen_i: 6976 case NVPTX::BI__nvvm_atom_xor_gen_l: 6977 case NVPTX::BI__nvvm_atom_xor_gen_ll: 6978 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 6979 6980 case NVPTX::BI__nvvm_atom_xchg_gen_i: 6981 case NVPTX::BI__nvvm_atom_xchg_gen_l: 6982 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 6983 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 6984 6985 case NVPTX::BI__nvvm_atom_max_gen_i: 6986 case NVPTX::BI__nvvm_atom_max_gen_l: 6987 case NVPTX::BI__nvvm_atom_max_gen_ll: 6988 case NVPTX::BI__nvvm_atom_max_gen_ui: 6989 case NVPTX::BI__nvvm_atom_max_gen_ul: 6990 case NVPTX::BI__nvvm_atom_max_gen_ull: 6991 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 6992 6993 case NVPTX::BI__nvvm_atom_min_gen_i: 6994 case NVPTX::BI__nvvm_atom_min_gen_l: 6995 case NVPTX::BI__nvvm_atom_min_gen_ll: 6996 case NVPTX::BI__nvvm_atom_min_gen_ui: 6997 case NVPTX::BI__nvvm_atom_min_gen_ul: 6998 case NVPTX::BI__nvvm_atom_min_gen_ull: 6999 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7000 7001 case NVPTX::BI__nvvm_atom_cas_gen_i: 7002 case NVPTX::BI__nvvm_atom_cas_gen_l: 7003 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7004 return MakeAtomicCmpXchgValue(*this, E, true); 7005 7006 case NVPTX::BI__nvvm_atom_add_gen_f: { 7007 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7008 Value *Val = EmitScalarExpr(E->getArg(1)); 7009 // atomicrmw only deals with integer arguments so we need to use 7010 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7011 Value *FnALAF32 = 7012 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7013 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7014 } 7015 7016 default: 7017 return nullptr; 7018 } 7019 } 7020