1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGObjCRuntime.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/ASTContext.h" 19 #include "clang/AST/Decl.h" 20 #include "clang/Basic/TargetBuiltins.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/CodeGen/CGFunctionInfo.h" 23 #include "llvm/ADT/StringExtras.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/InlineAsm.h" 26 #include "llvm/IR/Intrinsics.h" 27 28 using namespace clang; 29 using namespace CodeGen; 30 using namespace llvm; 31 32 /// getBuiltinLibFunction - Given a builtin id for a function like 33 /// "__builtin_fabsf", return a Function* for "fabsf". 34 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 35 unsigned BuiltinID) { 36 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 37 38 // Get the name, skip over the __builtin_ prefix (if necessary). 39 StringRef Name; 40 GlobalDecl D(FD); 41 42 // If the builtin has been declared explicitly with an assembler label, 43 // use the mangled name. This differs from the plain label on platforms 44 // that prefix labels. 45 if (FD->hasAttr<AsmLabelAttr>()) 46 Name = getMangledName(D); 47 else 48 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 49 50 llvm::FunctionType *Ty = 51 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 52 53 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 54 } 55 56 /// Emit the conversions required to turn the given value into an 57 /// integer of the given size. 58 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 59 QualType T, llvm::IntegerType *IntType) { 60 V = CGF.EmitToMemory(V, T); 61 62 if (V->getType()->isPointerTy()) 63 return CGF.Builder.CreatePtrToInt(V, IntType); 64 65 assert(V->getType() == IntType); 66 return V; 67 } 68 69 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 70 QualType T, llvm::Type *ResultType) { 71 V = CGF.EmitFromMemory(V, T); 72 73 if (ResultType->isPointerTy()) 74 return CGF.Builder.CreateIntToPtr(V, ResultType); 75 76 assert(V->getType() == ResultType); 77 return V; 78 } 79 80 /// Utility to insert an atomic instruction based on Instrinsic::ID 81 /// and the expression node. 82 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 83 llvm::AtomicRMWInst::BinOp Kind, 84 const CallExpr *E) { 85 QualType T = E->getType(); 86 assert(E->getArg(0)->getType()->isPointerType()); 87 assert(CGF.getContext().hasSameUnqualifiedType(T, 88 E->getArg(0)->getType()->getPointeeType())); 89 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 90 91 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 92 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 93 94 llvm::IntegerType *IntType = 95 llvm::IntegerType::get(CGF.getLLVMContext(), 96 CGF.getContext().getTypeSize(T)); 97 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 98 99 llvm::Value *Args[2]; 100 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 101 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 102 llvm::Type *ValueType = Args[1]->getType(); 103 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 104 105 llvm::Value *Result = 106 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 107 llvm::SequentiallyConsistent); 108 Result = EmitFromInt(CGF, Result, T, ValueType); 109 return RValue::get(Result); 110 } 111 112 /// Utility to insert an atomic instruction based Instrinsic::ID and 113 /// the expression node, where the return value is the result of the 114 /// operation. 115 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 116 llvm::AtomicRMWInst::BinOp Kind, 117 const CallExpr *E, 118 Instruction::BinaryOps Op, 119 bool Invert = false) { 120 QualType T = E->getType(); 121 assert(E->getArg(0)->getType()->isPointerType()); 122 assert(CGF.getContext().hasSameUnqualifiedType(T, 123 E->getArg(0)->getType()->getPointeeType())); 124 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 125 126 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 127 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 128 129 llvm::IntegerType *IntType = 130 llvm::IntegerType::get(CGF.getLLVMContext(), 131 CGF.getContext().getTypeSize(T)); 132 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 133 134 llvm::Value *Args[2]; 135 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 136 llvm::Type *ValueType = Args[1]->getType(); 137 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 138 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 139 140 llvm::Value *Result = 141 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 142 llvm::SequentiallyConsistent); 143 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 144 if (Invert) 145 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 146 llvm::ConstantInt::get(IntType, -1)); 147 Result = EmitFromInt(CGF, Result, T, ValueType); 148 return RValue::get(Result); 149 } 150 151 /// EmitFAbs - Emit a call to @llvm.fabs(). 152 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 153 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 154 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 155 Call->setDoesNotAccessMemory(); 156 return Call; 157 } 158 159 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 160 const CallExpr *E, llvm::Value *calleeValue) { 161 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 162 ReturnValueSlot(), Fn); 163 } 164 165 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 166 /// depending on IntrinsicID. 167 /// 168 /// \arg CGF The current codegen function. 169 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 170 /// \arg X The first argument to the llvm.*.with.overflow.*. 171 /// \arg Y The second argument to the llvm.*.with.overflow.*. 172 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 173 /// \returns The result (i.e. sum/product) returned by the intrinsic. 174 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 175 const llvm::Intrinsic::ID IntrinsicID, 176 llvm::Value *X, llvm::Value *Y, 177 llvm::Value *&Carry) { 178 // Make sure we have integers of the same width. 179 assert(X->getType() == Y->getType() && 180 "Arguments must be the same type. (Did you forget to make sure both " 181 "arguments have the same integer width?)"); 182 183 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 184 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 185 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 186 return CGF.Builder.CreateExtractValue(Tmp, 0); 187 } 188 189 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 190 unsigned BuiltinID, const CallExpr *E, 191 ReturnValueSlot ReturnValue) { 192 // See if we can constant fold this builtin. If so, don't emit it at all. 193 Expr::EvalResult Result; 194 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 195 !Result.hasSideEffects()) { 196 if (Result.Val.isInt()) 197 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 198 Result.Val.getInt())); 199 if (Result.Val.isFloat()) 200 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 201 Result.Val.getFloat())); 202 } 203 204 switch (BuiltinID) { 205 default: break; // Handle intrinsics and libm functions below. 206 case Builtin::BI__builtin___CFStringMakeConstantString: 207 case Builtin::BI__builtin___NSStringMakeConstantString: 208 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 209 case Builtin::BI__builtin_stdarg_start: 210 case Builtin::BI__builtin_va_start: 211 case Builtin::BI__va_start: 212 case Builtin::BI__builtin_va_end: { 213 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 214 ? EmitScalarExpr(E->getArg(0)) 215 : EmitVAListRef(E->getArg(0)); 216 llvm::Type *DestType = Int8PtrTy; 217 if (ArgValue->getType() != DestType) 218 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 219 ArgValue->getName().data()); 220 221 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 222 Intrinsic::vaend : Intrinsic::vastart; 223 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 224 } 225 case Builtin::BI__builtin_va_copy: { 226 Value *DstPtr = EmitVAListRef(E->getArg(0)); 227 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 228 229 llvm::Type *Type = Int8PtrTy; 230 231 DstPtr = Builder.CreateBitCast(DstPtr, Type); 232 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 233 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 234 DstPtr, SrcPtr)); 235 } 236 case Builtin::BI__builtin_abs: 237 case Builtin::BI__builtin_labs: 238 case Builtin::BI__builtin_llabs: { 239 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 240 241 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 242 Value *CmpResult = 243 Builder.CreateICmpSGE(ArgValue, 244 llvm::Constant::getNullValue(ArgValue->getType()), 245 "abscond"); 246 Value *Result = 247 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 248 249 return RValue::get(Result); 250 } 251 case Builtin::BI__builtin_fabs: 252 case Builtin::BI__builtin_fabsf: 253 case Builtin::BI__builtin_fabsl: { 254 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 255 Value *Result = EmitFAbs(*this, Arg1); 256 return RValue::get(Result); 257 } 258 case Builtin::BI__builtin_fmod: 259 case Builtin::BI__builtin_fmodf: 260 case Builtin::BI__builtin_fmodl: { 261 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 262 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 263 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 264 return RValue::get(Result); 265 } 266 267 case Builtin::BI__builtin_conj: 268 case Builtin::BI__builtin_conjf: 269 case Builtin::BI__builtin_conjl: { 270 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 271 Value *Real = ComplexVal.first; 272 Value *Imag = ComplexVal.second; 273 Value *Zero = 274 Imag->getType()->isFPOrFPVectorTy() 275 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 276 : llvm::Constant::getNullValue(Imag->getType()); 277 278 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 279 return RValue::getComplex(std::make_pair(Real, Imag)); 280 } 281 case Builtin::BI__builtin_creal: 282 case Builtin::BI__builtin_crealf: 283 case Builtin::BI__builtin_creall: 284 case Builtin::BIcreal: 285 case Builtin::BIcrealf: 286 case Builtin::BIcreall: { 287 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 288 return RValue::get(ComplexVal.first); 289 } 290 291 case Builtin::BI__builtin_cimag: 292 case Builtin::BI__builtin_cimagf: 293 case Builtin::BI__builtin_cimagl: 294 case Builtin::BIcimag: 295 case Builtin::BIcimagf: 296 case Builtin::BIcimagl: { 297 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 298 return RValue::get(ComplexVal.second); 299 } 300 301 case Builtin::BI__builtin_ctzs: 302 case Builtin::BI__builtin_ctz: 303 case Builtin::BI__builtin_ctzl: 304 case Builtin::BI__builtin_ctzll: { 305 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 306 307 llvm::Type *ArgType = ArgValue->getType(); 308 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 309 310 llvm::Type *ResultType = ConvertType(E->getType()); 311 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 312 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 313 if (Result->getType() != ResultType) 314 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 315 "cast"); 316 return RValue::get(Result); 317 } 318 case Builtin::BI__builtin_clzs: 319 case Builtin::BI__builtin_clz: 320 case Builtin::BI__builtin_clzl: 321 case Builtin::BI__builtin_clzll: { 322 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 323 324 llvm::Type *ArgType = ArgValue->getType(); 325 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 326 327 llvm::Type *ResultType = ConvertType(E->getType()); 328 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 329 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 330 if (Result->getType() != ResultType) 331 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 332 "cast"); 333 return RValue::get(Result); 334 } 335 case Builtin::BI__builtin_ffs: 336 case Builtin::BI__builtin_ffsl: 337 case Builtin::BI__builtin_ffsll: { 338 // ffs(x) -> x ? cttz(x) + 1 : 0 339 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 340 341 llvm::Type *ArgType = ArgValue->getType(); 342 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 343 344 llvm::Type *ResultType = ConvertType(E->getType()); 345 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 346 Builder.getTrue()), 347 llvm::ConstantInt::get(ArgType, 1)); 348 Value *Zero = llvm::Constant::getNullValue(ArgType); 349 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 350 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 351 if (Result->getType() != ResultType) 352 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 353 "cast"); 354 return RValue::get(Result); 355 } 356 case Builtin::BI__builtin_parity: 357 case Builtin::BI__builtin_parityl: 358 case Builtin::BI__builtin_parityll: { 359 // parity(x) -> ctpop(x) & 1 360 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 361 362 llvm::Type *ArgType = ArgValue->getType(); 363 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 364 365 llvm::Type *ResultType = ConvertType(E->getType()); 366 Value *Tmp = Builder.CreateCall(F, ArgValue); 367 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 368 if (Result->getType() != ResultType) 369 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 370 "cast"); 371 return RValue::get(Result); 372 } 373 case Builtin::BI__builtin_popcount: 374 case Builtin::BI__builtin_popcountl: 375 case Builtin::BI__builtin_popcountll: { 376 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 377 378 llvm::Type *ArgType = ArgValue->getType(); 379 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 380 381 llvm::Type *ResultType = ConvertType(E->getType()); 382 Value *Result = Builder.CreateCall(F, ArgValue); 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_expect: { 389 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 390 llvm::Type *ArgType = ArgValue->getType(); 391 392 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 393 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 394 395 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 396 "expval"); 397 return RValue::get(Result); 398 } 399 case Builtin::BI__builtin_assume_aligned: { 400 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 401 Value *OffsetValue = 402 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 403 404 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 405 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 406 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 407 408 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 409 return RValue::get(PtrValue); 410 } 411 case Builtin::BI__assume: 412 case Builtin::BI__builtin_assume: { 413 if (E->getArg(0)->HasSideEffects(getContext())) 414 return RValue::get(nullptr); 415 416 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 417 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 418 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 419 } 420 case Builtin::BI__builtin_bswap16: 421 case Builtin::BI__builtin_bswap32: 422 case Builtin::BI__builtin_bswap64: { 423 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 424 llvm::Type *ArgType = ArgValue->getType(); 425 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 426 return RValue::get(Builder.CreateCall(F, ArgValue)); 427 } 428 case Builtin::BI__builtin_object_size: { 429 // We rely on constant folding to deal with expressions with side effects. 430 assert(!E->getArg(0)->HasSideEffects(getContext()) && 431 "should have been constant folded"); 432 433 // We pass this builtin onto the optimizer so that it can 434 // figure out the object size in more complex cases. 435 llvm::Type *ResType = ConvertType(E->getType()); 436 437 // LLVM only supports 0 and 2, make sure that we pass along that 438 // as a boolean. 439 Value *Ty = EmitScalarExpr(E->getArg(1)); 440 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 441 assert(CI); 442 uint64_t val = CI->getZExtValue(); 443 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 444 // FIXME: Get right address space. 445 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 446 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 447 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 448 } 449 case Builtin::BI__builtin_prefetch: { 450 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 451 // FIXME: Technically these constants should of type 'int', yes? 452 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 453 llvm::ConstantInt::get(Int32Ty, 0); 454 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 455 llvm::ConstantInt::get(Int32Ty, 3); 456 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 457 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 458 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 459 } 460 case Builtin::BI__builtin_readcyclecounter: { 461 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 462 return RValue::get(Builder.CreateCall(F)); 463 } 464 case Builtin::BI__builtin___clear_cache: { 465 Value *Begin = EmitScalarExpr(E->getArg(0)); 466 Value *End = EmitScalarExpr(E->getArg(1)); 467 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 468 return RValue::get(Builder.CreateCall2(F, Begin, End)); 469 } 470 case Builtin::BI__builtin_trap: { 471 Value *F = CGM.getIntrinsic(Intrinsic::trap); 472 return RValue::get(Builder.CreateCall(F)); 473 } 474 case Builtin::BI__debugbreak: { 475 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 476 return RValue::get(Builder.CreateCall(F)); 477 } 478 case Builtin::BI__builtin_unreachable: { 479 if (SanOpts.has(SanitizerKind::Unreachable)) { 480 SanitizerScope SanScope(this); 481 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 482 SanitizerKind::Unreachable), 483 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 484 None); 485 } else 486 Builder.CreateUnreachable(); 487 488 // We do need to preserve an insertion point. 489 EmitBlock(createBasicBlock("unreachable.cont")); 490 491 return RValue::get(nullptr); 492 } 493 494 case Builtin::BI__builtin_powi: 495 case Builtin::BI__builtin_powif: 496 case Builtin::BI__builtin_powil: { 497 Value *Base = EmitScalarExpr(E->getArg(0)); 498 Value *Exponent = EmitScalarExpr(E->getArg(1)); 499 llvm::Type *ArgType = Base->getType(); 500 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 501 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 502 } 503 504 case Builtin::BI__builtin_isgreater: 505 case Builtin::BI__builtin_isgreaterequal: 506 case Builtin::BI__builtin_isless: 507 case Builtin::BI__builtin_islessequal: 508 case Builtin::BI__builtin_islessgreater: 509 case Builtin::BI__builtin_isunordered: { 510 // Ordered comparisons: we know the arguments to these are matching scalar 511 // floating point values. 512 Value *LHS = EmitScalarExpr(E->getArg(0)); 513 Value *RHS = EmitScalarExpr(E->getArg(1)); 514 515 switch (BuiltinID) { 516 default: llvm_unreachable("Unknown ordered comparison"); 517 case Builtin::BI__builtin_isgreater: 518 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 519 break; 520 case Builtin::BI__builtin_isgreaterequal: 521 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 522 break; 523 case Builtin::BI__builtin_isless: 524 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 525 break; 526 case Builtin::BI__builtin_islessequal: 527 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 528 break; 529 case Builtin::BI__builtin_islessgreater: 530 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 531 break; 532 case Builtin::BI__builtin_isunordered: 533 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 534 break; 535 } 536 // ZExt bool to int type. 537 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 538 } 539 case Builtin::BI__builtin_isnan: { 540 Value *V = EmitScalarExpr(E->getArg(0)); 541 V = Builder.CreateFCmpUNO(V, V, "cmp"); 542 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 543 } 544 545 case Builtin::BI__builtin_isinf: { 546 // isinf(x) --> fabs(x) == infinity 547 Value *V = EmitScalarExpr(E->getArg(0)); 548 V = EmitFAbs(*this, V); 549 550 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 551 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 552 } 553 554 // TODO: BI__builtin_isinf_sign 555 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 556 557 case Builtin::BI__builtin_isnormal: { 558 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 559 Value *V = EmitScalarExpr(E->getArg(0)); 560 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 561 562 Value *Abs = EmitFAbs(*this, V); 563 Value *IsLessThanInf = 564 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 565 APFloat Smallest = APFloat::getSmallestNormalized( 566 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 567 Value *IsNormal = 568 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 569 "isnormal"); 570 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 571 V = Builder.CreateAnd(V, IsNormal, "and"); 572 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 573 } 574 575 case Builtin::BI__builtin_isfinite: { 576 // isfinite(x) --> x == x && fabs(x) != infinity; 577 Value *V = EmitScalarExpr(E->getArg(0)); 578 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 579 580 Value *Abs = EmitFAbs(*this, V); 581 Value *IsNotInf = 582 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 583 584 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 585 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 586 } 587 588 case Builtin::BI__builtin_fpclassify: { 589 Value *V = EmitScalarExpr(E->getArg(5)); 590 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 591 592 // Create Result 593 BasicBlock *Begin = Builder.GetInsertBlock(); 594 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 595 Builder.SetInsertPoint(End); 596 PHINode *Result = 597 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 598 "fpclassify_result"); 599 600 // if (V==0) return FP_ZERO 601 Builder.SetInsertPoint(Begin); 602 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 603 "iszero"); 604 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 605 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 606 Builder.CreateCondBr(IsZero, End, NotZero); 607 Result->addIncoming(ZeroLiteral, Begin); 608 609 // if (V != V) return FP_NAN 610 Builder.SetInsertPoint(NotZero); 611 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 612 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 613 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 614 Builder.CreateCondBr(IsNan, End, NotNan); 615 Result->addIncoming(NanLiteral, NotZero); 616 617 // if (fabs(V) == infinity) return FP_INFINITY 618 Builder.SetInsertPoint(NotNan); 619 Value *VAbs = EmitFAbs(*this, V); 620 Value *IsInf = 621 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 622 "isinf"); 623 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 624 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 625 Builder.CreateCondBr(IsInf, End, NotInf); 626 Result->addIncoming(InfLiteral, NotNan); 627 628 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 629 Builder.SetInsertPoint(NotInf); 630 APFloat Smallest = APFloat::getSmallestNormalized( 631 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 632 Value *IsNormal = 633 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 634 "isnormal"); 635 Value *NormalResult = 636 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 637 EmitScalarExpr(E->getArg(3))); 638 Builder.CreateBr(End); 639 Result->addIncoming(NormalResult, NotInf); 640 641 // return Result 642 Builder.SetInsertPoint(End); 643 return RValue::get(Result); 644 } 645 646 case Builtin::BIalloca: 647 case Builtin::BI_alloca: 648 case Builtin::BI__builtin_alloca: { 649 Value *Size = EmitScalarExpr(E->getArg(0)); 650 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 651 } 652 case Builtin::BIbzero: 653 case Builtin::BI__builtin_bzero: { 654 std::pair<llvm::Value*, unsigned> Dest = 655 EmitPointerWithAlignment(E->getArg(0)); 656 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 657 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 658 Dest.second, false); 659 return RValue::get(Dest.first); 660 } 661 case Builtin::BImemcpy: 662 case Builtin::BI__builtin_memcpy: { 663 std::pair<llvm::Value*, unsigned> Dest = 664 EmitPointerWithAlignment(E->getArg(0)); 665 std::pair<llvm::Value*, unsigned> Src = 666 EmitPointerWithAlignment(E->getArg(1)); 667 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 668 unsigned Align = std::min(Dest.second, Src.second); 669 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 670 return RValue::get(Dest.first); 671 } 672 673 case Builtin::BI__builtin___memcpy_chk: { 674 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 675 llvm::APSInt Size, DstSize; 676 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 677 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 678 break; 679 if (Size.ugt(DstSize)) 680 break; 681 std::pair<llvm::Value*, unsigned> Dest = 682 EmitPointerWithAlignment(E->getArg(0)); 683 std::pair<llvm::Value*, unsigned> Src = 684 EmitPointerWithAlignment(E->getArg(1)); 685 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 686 unsigned Align = std::min(Dest.second, Src.second); 687 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 688 return RValue::get(Dest.first); 689 } 690 691 case Builtin::BI__builtin_objc_memmove_collectable: { 692 Value *Address = EmitScalarExpr(E->getArg(0)); 693 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 694 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 695 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 696 Address, SrcAddr, SizeVal); 697 return RValue::get(Address); 698 } 699 700 case Builtin::BI__builtin___memmove_chk: { 701 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 702 llvm::APSInt Size, DstSize; 703 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 704 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 705 break; 706 if (Size.ugt(DstSize)) 707 break; 708 std::pair<llvm::Value*, unsigned> Dest = 709 EmitPointerWithAlignment(E->getArg(0)); 710 std::pair<llvm::Value*, unsigned> Src = 711 EmitPointerWithAlignment(E->getArg(1)); 712 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 713 unsigned Align = std::min(Dest.second, Src.second); 714 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 715 return RValue::get(Dest.first); 716 } 717 718 case Builtin::BImemmove: 719 case Builtin::BI__builtin_memmove: { 720 std::pair<llvm::Value*, unsigned> Dest = 721 EmitPointerWithAlignment(E->getArg(0)); 722 std::pair<llvm::Value*, unsigned> Src = 723 EmitPointerWithAlignment(E->getArg(1)); 724 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 725 unsigned Align = std::min(Dest.second, Src.second); 726 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 727 return RValue::get(Dest.first); 728 } 729 case Builtin::BImemset: 730 case Builtin::BI__builtin_memset: { 731 std::pair<llvm::Value*, unsigned> Dest = 732 EmitPointerWithAlignment(E->getArg(0)); 733 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 734 Builder.getInt8Ty()); 735 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 736 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 737 return RValue::get(Dest.first); 738 } 739 case Builtin::BI__builtin___memset_chk: { 740 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 741 llvm::APSInt Size, DstSize; 742 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 743 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 744 break; 745 if (Size.ugt(DstSize)) 746 break; 747 std::pair<llvm::Value*, unsigned> Dest = 748 EmitPointerWithAlignment(E->getArg(0)); 749 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 750 Builder.getInt8Ty()); 751 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 752 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 753 return RValue::get(Dest.first); 754 } 755 case Builtin::BI__builtin_dwarf_cfa: { 756 // The offset in bytes from the first argument to the CFA. 757 // 758 // Why on earth is this in the frontend? Is there any reason at 759 // all that the backend can't reasonably determine this while 760 // lowering llvm.eh.dwarf.cfa()? 761 // 762 // TODO: If there's a satisfactory reason, add a target hook for 763 // this instead of hard-coding 0, which is correct for most targets. 764 int32_t Offset = 0; 765 766 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 767 return RValue::get(Builder.CreateCall(F, 768 llvm::ConstantInt::get(Int32Ty, Offset))); 769 } 770 case Builtin::BI__builtin_return_address: { 771 Value *Depth = EmitScalarExpr(E->getArg(0)); 772 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 773 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 774 return RValue::get(Builder.CreateCall(F, Depth)); 775 } 776 case Builtin::BI__builtin_frame_address: { 777 Value *Depth = EmitScalarExpr(E->getArg(0)); 778 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 779 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 780 return RValue::get(Builder.CreateCall(F, Depth)); 781 } 782 case Builtin::BI__builtin_extract_return_addr: { 783 Value *Address = EmitScalarExpr(E->getArg(0)); 784 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 785 return RValue::get(Result); 786 } 787 case Builtin::BI__builtin_frob_return_addr: { 788 Value *Address = EmitScalarExpr(E->getArg(0)); 789 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 790 return RValue::get(Result); 791 } 792 case Builtin::BI__builtin_dwarf_sp_column: { 793 llvm::IntegerType *Ty 794 = cast<llvm::IntegerType>(ConvertType(E->getType())); 795 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 796 if (Column == -1) { 797 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 798 return RValue::get(llvm::UndefValue::get(Ty)); 799 } 800 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 801 } 802 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 803 Value *Address = EmitScalarExpr(E->getArg(0)); 804 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 805 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 806 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 807 } 808 case Builtin::BI__builtin_eh_return: { 809 Value *Int = EmitScalarExpr(E->getArg(0)); 810 Value *Ptr = EmitScalarExpr(E->getArg(1)); 811 812 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 813 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 814 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 815 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 816 ? Intrinsic::eh_return_i32 817 : Intrinsic::eh_return_i64); 818 Builder.CreateCall2(F, Int, Ptr); 819 Builder.CreateUnreachable(); 820 821 // We do need to preserve an insertion point. 822 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 823 824 return RValue::get(nullptr); 825 } 826 case Builtin::BI__builtin_unwind_init: { 827 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 828 return RValue::get(Builder.CreateCall(F)); 829 } 830 case Builtin::BI__builtin_extend_pointer: { 831 // Extends a pointer to the size of an _Unwind_Word, which is 832 // uint64_t on all platforms. Generally this gets poked into a 833 // register and eventually used as an address, so if the 834 // addressing registers are wider than pointers and the platform 835 // doesn't implicitly ignore high-order bits when doing 836 // addressing, we need to make sure we zext / sext based on 837 // the platform's expectations. 838 // 839 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 840 841 // Cast the pointer to intptr_t. 842 Value *Ptr = EmitScalarExpr(E->getArg(0)); 843 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 844 845 // If that's 64 bits, we're done. 846 if (IntPtrTy->getBitWidth() == 64) 847 return RValue::get(Result); 848 849 // Otherwise, ask the codegen data what to do. 850 if (getTargetHooks().extendPointerWithSExt()) 851 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 852 else 853 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 854 } 855 case Builtin::BI__builtin_setjmp: { 856 // Buffer is a void**. 857 Value *Buf = EmitScalarExpr(E->getArg(0)); 858 859 // Store the frame pointer to the setjmp buffer. 860 Value *FrameAddr = 861 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 862 ConstantInt::get(Int32Ty, 0)); 863 Builder.CreateStore(FrameAddr, Buf); 864 865 // Store the stack pointer to the setjmp buffer. 866 Value *StackAddr = 867 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 868 Value *StackSaveSlot = 869 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 870 Builder.CreateStore(StackAddr, StackSaveSlot); 871 872 // Call LLVM's EH setjmp, which is lightweight. 873 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 874 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 875 return RValue::get(Builder.CreateCall(F, Buf)); 876 } 877 case Builtin::BI__builtin_longjmp: { 878 Value *Buf = EmitScalarExpr(E->getArg(0)); 879 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 880 881 // Call LLVM's EH longjmp, which is lightweight. 882 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 883 884 // longjmp doesn't return; mark this as unreachable. 885 Builder.CreateUnreachable(); 886 887 // We do need to preserve an insertion point. 888 EmitBlock(createBasicBlock("longjmp.cont")); 889 890 return RValue::get(nullptr); 891 } 892 case Builtin::BI__sync_fetch_and_add: 893 case Builtin::BI__sync_fetch_and_sub: 894 case Builtin::BI__sync_fetch_and_or: 895 case Builtin::BI__sync_fetch_and_and: 896 case Builtin::BI__sync_fetch_and_xor: 897 case Builtin::BI__sync_fetch_and_nand: 898 case Builtin::BI__sync_add_and_fetch: 899 case Builtin::BI__sync_sub_and_fetch: 900 case Builtin::BI__sync_and_and_fetch: 901 case Builtin::BI__sync_or_and_fetch: 902 case Builtin::BI__sync_xor_and_fetch: 903 case Builtin::BI__sync_nand_and_fetch: 904 case Builtin::BI__sync_val_compare_and_swap: 905 case Builtin::BI__sync_bool_compare_and_swap: 906 case Builtin::BI__sync_lock_test_and_set: 907 case Builtin::BI__sync_lock_release: 908 case Builtin::BI__sync_swap: 909 llvm_unreachable("Shouldn't make it through sema"); 910 case Builtin::BI__sync_fetch_and_add_1: 911 case Builtin::BI__sync_fetch_and_add_2: 912 case Builtin::BI__sync_fetch_and_add_4: 913 case Builtin::BI__sync_fetch_and_add_8: 914 case Builtin::BI__sync_fetch_and_add_16: 915 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 916 case Builtin::BI__sync_fetch_and_sub_1: 917 case Builtin::BI__sync_fetch_and_sub_2: 918 case Builtin::BI__sync_fetch_and_sub_4: 919 case Builtin::BI__sync_fetch_and_sub_8: 920 case Builtin::BI__sync_fetch_and_sub_16: 921 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 922 case Builtin::BI__sync_fetch_and_or_1: 923 case Builtin::BI__sync_fetch_and_or_2: 924 case Builtin::BI__sync_fetch_and_or_4: 925 case Builtin::BI__sync_fetch_and_or_8: 926 case Builtin::BI__sync_fetch_and_or_16: 927 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 928 case Builtin::BI__sync_fetch_and_and_1: 929 case Builtin::BI__sync_fetch_and_and_2: 930 case Builtin::BI__sync_fetch_and_and_4: 931 case Builtin::BI__sync_fetch_and_and_8: 932 case Builtin::BI__sync_fetch_and_and_16: 933 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 934 case Builtin::BI__sync_fetch_and_xor_1: 935 case Builtin::BI__sync_fetch_and_xor_2: 936 case Builtin::BI__sync_fetch_and_xor_4: 937 case Builtin::BI__sync_fetch_and_xor_8: 938 case Builtin::BI__sync_fetch_and_xor_16: 939 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 940 case Builtin::BI__sync_fetch_and_nand_1: 941 case Builtin::BI__sync_fetch_and_nand_2: 942 case Builtin::BI__sync_fetch_and_nand_4: 943 case Builtin::BI__sync_fetch_and_nand_8: 944 case Builtin::BI__sync_fetch_and_nand_16: 945 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 946 947 // Clang extensions: not overloaded yet. 948 case Builtin::BI__sync_fetch_and_min: 949 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 950 case Builtin::BI__sync_fetch_and_max: 951 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 952 case Builtin::BI__sync_fetch_and_umin: 953 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 954 case Builtin::BI__sync_fetch_and_umax: 955 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 956 957 case Builtin::BI__sync_add_and_fetch_1: 958 case Builtin::BI__sync_add_and_fetch_2: 959 case Builtin::BI__sync_add_and_fetch_4: 960 case Builtin::BI__sync_add_and_fetch_8: 961 case Builtin::BI__sync_add_and_fetch_16: 962 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 963 llvm::Instruction::Add); 964 case Builtin::BI__sync_sub_and_fetch_1: 965 case Builtin::BI__sync_sub_and_fetch_2: 966 case Builtin::BI__sync_sub_and_fetch_4: 967 case Builtin::BI__sync_sub_and_fetch_8: 968 case Builtin::BI__sync_sub_and_fetch_16: 969 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 970 llvm::Instruction::Sub); 971 case Builtin::BI__sync_and_and_fetch_1: 972 case Builtin::BI__sync_and_and_fetch_2: 973 case Builtin::BI__sync_and_and_fetch_4: 974 case Builtin::BI__sync_and_and_fetch_8: 975 case Builtin::BI__sync_and_and_fetch_16: 976 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 977 llvm::Instruction::And); 978 case Builtin::BI__sync_or_and_fetch_1: 979 case Builtin::BI__sync_or_and_fetch_2: 980 case Builtin::BI__sync_or_and_fetch_4: 981 case Builtin::BI__sync_or_and_fetch_8: 982 case Builtin::BI__sync_or_and_fetch_16: 983 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 984 llvm::Instruction::Or); 985 case Builtin::BI__sync_xor_and_fetch_1: 986 case Builtin::BI__sync_xor_and_fetch_2: 987 case Builtin::BI__sync_xor_and_fetch_4: 988 case Builtin::BI__sync_xor_and_fetch_8: 989 case Builtin::BI__sync_xor_and_fetch_16: 990 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 991 llvm::Instruction::Xor); 992 case Builtin::BI__sync_nand_and_fetch_1: 993 case Builtin::BI__sync_nand_and_fetch_2: 994 case Builtin::BI__sync_nand_and_fetch_4: 995 case Builtin::BI__sync_nand_and_fetch_8: 996 case Builtin::BI__sync_nand_and_fetch_16: 997 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 998 llvm::Instruction::And, true); 999 1000 case Builtin::BI__sync_val_compare_and_swap_1: 1001 case Builtin::BI__sync_val_compare_and_swap_2: 1002 case Builtin::BI__sync_val_compare_and_swap_4: 1003 case Builtin::BI__sync_val_compare_and_swap_8: 1004 case Builtin::BI__sync_val_compare_and_swap_16: { 1005 QualType T = E->getType(); 1006 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1007 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1008 1009 llvm::IntegerType *IntType = 1010 llvm::IntegerType::get(getLLVMContext(), 1011 getContext().getTypeSize(T)); 1012 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1013 1014 Value *Args[3]; 1015 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1016 Args[1] = EmitScalarExpr(E->getArg(1)); 1017 llvm::Type *ValueType = Args[1]->getType(); 1018 Args[1] = EmitToInt(*this, Args[1], T, IntType); 1019 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1020 1021 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1022 llvm::SequentiallyConsistent, 1023 llvm::SequentiallyConsistent); 1024 Result = Builder.CreateExtractValue(Result, 0); 1025 Result = EmitFromInt(*this, Result, T, ValueType); 1026 return RValue::get(Result); 1027 } 1028 1029 case Builtin::BI__sync_bool_compare_and_swap_1: 1030 case Builtin::BI__sync_bool_compare_and_swap_2: 1031 case Builtin::BI__sync_bool_compare_and_swap_4: 1032 case Builtin::BI__sync_bool_compare_and_swap_8: 1033 case Builtin::BI__sync_bool_compare_and_swap_16: { 1034 QualType T = E->getArg(1)->getType(); 1035 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1036 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1037 1038 llvm::IntegerType *IntType = 1039 llvm::IntegerType::get(getLLVMContext(), 1040 getContext().getTypeSize(T)); 1041 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1042 1043 Value *Args[3]; 1044 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1045 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1046 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1047 1048 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1049 llvm::SequentiallyConsistent, 1050 llvm::SequentiallyConsistent); 1051 Value *Result = Builder.CreateExtractValue(Pair, 1); 1052 // zext bool to int. 1053 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1054 return RValue::get(Result); 1055 } 1056 1057 case Builtin::BI__sync_swap_1: 1058 case Builtin::BI__sync_swap_2: 1059 case Builtin::BI__sync_swap_4: 1060 case Builtin::BI__sync_swap_8: 1061 case Builtin::BI__sync_swap_16: 1062 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1063 1064 case Builtin::BI__sync_lock_test_and_set_1: 1065 case Builtin::BI__sync_lock_test_and_set_2: 1066 case Builtin::BI__sync_lock_test_and_set_4: 1067 case Builtin::BI__sync_lock_test_and_set_8: 1068 case Builtin::BI__sync_lock_test_and_set_16: 1069 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1070 1071 case Builtin::BI__sync_lock_release_1: 1072 case Builtin::BI__sync_lock_release_2: 1073 case Builtin::BI__sync_lock_release_4: 1074 case Builtin::BI__sync_lock_release_8: 1075 case Builtin::BI__sync_lock_release_16: { 1076 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1077 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1078 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1079 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1080 StoreSize.getQuantity() * 8); 1081 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1082 llvm::StoreInst *Store = 1083 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1084 Store->setAlignment(StoreSize.getQuantity()); 1085 Store->setAtomic(llvm::Release); 1086 return RValue::get(nullptr); 1087 } 1088 1089 case Builtin::BI__sync_synchronize: { 1090 // We assume this is supposed to correspond to a C++0x-style 1091 // sequentially-consistent fence (i.e. this is only usable for 1092 // synchonization, not device I/O or anything like that). This intrinsic 1093 // is really badly designed in the sense that in theory, there isn't 1094 // any way to safely use it... but in practice, it mostly works 1095 // to use it with non-atomic loads and stores to get acquire/release 1096 // semantics. 1097 Builder.CreateFence(llvm::SequentiallyConsistent); 1098 return RValue::get(nullptr); 1099 } 1100 1101 case Builtin::BI__c11_atomic_is_lock_free: 1102 case Builtin::BI__atomic_is_lock_free: { 1103 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1104 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1105 // _Atomic(T) is always properly-aligned. 1106 const char *LibCallName = "__atomic_is_lock_free"; 1107 CallArgList Args; 1108 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1109 getContext().getSizeType()); 1110 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1111 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1112 getContext().VoidPtrTy); 1113 else 1114 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1115 getContext().VoidPtrTy); 1116 const CGFunctionInfo &FuncInfo = 1117 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1118 FunctionType::ExtInfo(), 1119 RequiredArgs::All); 1120 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1121 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1122 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1123 } 1124 1125 case Builtin::BI__atomic_test_and_set: { 1126 // Look at the argument type to determine whether this is a volatile 1127 // operation. The parameter type is always volatile. 1128 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1129 bool Volatile = 1130 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1131 1132 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1133 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1134 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1135 Value *NewVal = Builder.getInt8(1); 1136 Value *Order = EmitScalarExpr(E->getArg(1)); 1137 if (isa<llvm::ConstantInt>(Order)) { 1138 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1139 AtomicRMWInst *Result = nullptr; 1140 switch (ord) { 1141 case 0: // memory_order_relaxed 1142 default: // invalid order 1143 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1144 Ptr, NewVal, 1145 llvm::Monotonic); 1146 break; 1147 case 1: // memory_order_consume 1148 case 2: // memory_order_acquire 1149 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1150 Ptr, NewVal, 1151 llvm::Acquire); 1152 break; 1153 case 3: // memory_order_release 1154 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1155 Ptr, NewVal, 1156 llvm::Release); 1157 break; 1158 case 4: // memory_order_acq_rel 1159 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1160 Ptr, NewVal, 1161 llvm::AcquireRelease); 1162 break; 1163 case 5: // memory_order_seq_cst 1164 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1165 Ptr, NewVal, 1166 llvm::SequentiallyConsistent); 1167 break; 1168 } 1169 Result->setVolatile(Volatile); 1170 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1171 } 1172 1173 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1174 1175 llvm::BasicBlock *BBs[5] = { 1176 createBasicBlock("monotonic", CurFn), 1177 createBasicBlock("acquire", CurFn), 1178 createBasicBlock("release", CurFn), 1179 createBasicBlock("acqrel", CurFn), 1180 createBasicBlock("seqcst", CurFn) 1181 }; 1182 llvm::AtomicOrdering Orders[5] = { 1183 llvm::Monotonic, llvm::Acquire, llvm::Release, 1184 llvm::AcquireRelease, llvm::SequentiallyConsistent 1185 }; 1186 1187 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1188 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1189 1190 Builder.SetInsertPoint(ContBB); 1191 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1192 1193 for (unsigned i = 0; i < 5; ++i) { 1194 Builder.SetInsertPoint(BBs[i]); 1195 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1196 Ptr, NewVal, Orders[i]); 1197 RMW->setVolatile(Volatile); 1198 Result->addIncoming(RMW, BBs[i]); 1199 Builder.CreateBr(ContBB); 1200 } 1201 1202 SI->addCase(Builder.getInt32(0), BBs[0]); 1203 SI->addCase(Builder.getInt32(1), BBs[1]); 1204 SI->addCase(Builder.getInt32(2), BBs[1]); 1205 SI->addCase(Builder.getInt32(3), BBs[2]); 1206 SI->addCase(Builder.getInt32(4), BBs[3]); 1207 SI->addCase(Builder.getInt32(5), BBs[4]); 1208 1209 Builder.SetInsertPoint(ContBB); 1210 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1211 } 1212 1213 case Builtin::BI__atomic_clear: { 1214 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1215 bool Volatile = 1216 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1217 1218 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1219 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1220 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1221 Value *NewVal = Builder.getInt8(0); 1222 Value *Order = EmitScalarExpr(E->getArg(1)); 1223 if (isa<llvm::ConstantInt>(Order)) { 1224 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1225 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1226 Store->setAlignment(1); 1227 switch (ord) { 1228 case 0: // memory_order_relaxed 1229 default: // invalid order 1230 Store->setOrdering(llvm::Monotonic); 1231 break; 1232 case 3: // memory_order_release 1233 Store->setOrdering(llvm::Release); 1234 break; 1235 case 5: // memory_order_seq_cst 1236 Store->setOrdering(llvm::SequentiallyConsistent); 1237 break; 1238 } 1239 return RValue::get(nullptr); 1240 } 1241 1242 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1243 1244 llvm::BasicBlock *BBs[3] = { 1245 createBasicBlock("monotonic", CurFn), 1246 createBasicBlock("release", CurFn), 1247 createBasicBlock("seqcst", CurFn) 1248 }; 1249 llvm::AtomicOrdering Orders[3] = { 1250 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1251 }; 1252 1253 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1254 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1255 1256 for (unsigned i = 0; i < 3; ++i) { 1257 Builder.SetInsertPoint(BBs[i]); 1258 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1259 Store->setAlignment(1); 1260 Store->setOrdering(Orders[i]); 1261 Builder.CreateBr(ContBB); 1262 } 1263 1264 SI->addCase(Builder.getInt32(0), BBs[0]); 1265 SI->addCase(Builder.getInt32(3), BBs[1]); 1266 SI->addCase(Builder.getInt32(5), BBs[2]); 1267 1268 Builder.SetInsertPoint(ContBB); 1269 return RValue::get(nullptr); 1270 } 1271 1272 case Builtin::BI__atomic_thread_fence: 1273 case Builtin::BI__atomic_signal_fence: 1274 case Builtin::BI__c11_atomic_thread_fence: 1275 case Builtin::BI__c11_atomic_signal_fence: { 1276 llvm::SynchronizationScope Scope; 1277 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1278 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1279 Scope = llvm::SingleThread; 1280 else 1281 Scope = llvm::CrossThread; 1282 Value *Order = EmitScalarExpr(E->getArg(0)); 1283 if (isa<llvm::ConstantInt>(Order)) { 1284 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1285 switch (ord) { 1286 case 0: // memory_order_relaxed 1287 default: // invalid order 1288 break; 1289 case 1: // memory_order_consume 1290 case 2: // memory_order_acquire 1291 Builder.CreateFence(llvm::Acquire, Scope); 1292 break; 1293 case 3: // memory_order_release 1294 Builder.CreateFence(llvm::Release, Scope); 1295 break; 1296 case 4: // memory_order_acq_rel 1297 Builder.CreateFence(llvm::AcquireRelease, Scope); 1298 break; 1299 case 5: // memory_order_seq_cst 1300 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1301 break; 1302 } 1303 return RValue::get(nullptr); 1304 } 1305 1306 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1307 AcquireBB = createBasicBlock("acquire", CurFn); 1308 ReleaseBB = createBasicBlock("release", CurFn); 1309 AcqRelBB = createBasicBlock("acqrel", CurFn); 1310 SeqCstBB = createBasicBlock("seqcst", CurFn); 1311 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1312 1313 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1314 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1315 1316 Builder.SetInsertPoint(AcquireBB); 1317 Builder.CreateFence(llvm::Acquire, Scope); 1318 Builder.CreateBr(ContBB); 1319 SI->addCase(Builder.getInt32(1), AcquireBB); 1320 SI->addCase(Builder.getInt32(2), AcquireBB); 1321 1322 Builder.SetInsertPoint(ReleaseBB); 1323 Builder.CreateFence(llvm::Release, Scope); 1324 Builder.CreateBr(ContBB); 1325 SI->addCase(Builder.getInt32(3), ReleaseBB); 1326 1327 Builder.SetInsertPoint(AcqRelBB); 1328 Builder.CreateFence(llvm::AcquireRelease, Scope); 1329 Builder.CreateBr(ContBB); 1330 SI->addCase(Builder.getInt32(4), AcqRelBB); 1331 1332 Builder.SetInsertPoint(SeqCstBB); 1333 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1334 Builder.CreateBr(ContBB); 1335 SI->addCase(Builder.getInt32(5), SeqCstBB); 1336 1337 Builder.SetInsertPoint(ContBB); 1338 return RValue::get(nullptr); 1339 } 1340 1341 // Library functions with special handling. 1342 case Builtin::BIsqrt: 1343 case Builtin::BIsqrtf: 1344 case Builtin::BIsqrtl: { 1345 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1346 // in finite- or unsafe-math mode (the intrinsic has different semantics 1347 // for handling negative numbers compared to the library function, so 1348 // -fmath-errno=0 is not enough). 1349 if (!FD->hasAttr<ConstAttr>()) 1350 break; 1351 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1352 CGM.getCodeGenOpts().NoNaNsFPMath)) 1353 break; 1354 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1355 llvm::Type *ArgType = Arg0->getType(); 1356 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1357 return RValue::get(Builder.CreateCall(F, Arg0)); 1358 } 1359 1360 case Builtin::BIpow: 1361 case Builtin::BIpowf: 1362 case Builtin::BIpowl: { 1363 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1364 if (!FD->hasAttr<ConstAttr>()) 1365 break; 1366 Value *Base = EmitScalarExpr(E->getArg(0)); 1367 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1368 llvm::Type *ArgType = Base->getType(); 1369 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1370 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1371 } 1372 1373 case Builtin::BIfma: 1374 case Builtin::BIfmaf: 1375 case Builtin::BIfmal: 1376 case Builtin::BI__builtin_fma: 1377 case Builtin::BI__builtin_fmaf: 1378 case Builtin::BI__builtin_fmal: { 1379 // Rewrite fma to intrinsic. 1380 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1381 llvm::Type *ArgType = FirstArg->getType(); 1382 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1383 return RValue::get(Builder.CreateCall3(F, FirstArg, 1384 EmitScalarExpr(E->getArg(1)), 1385 EmitScalarExpr(E->getArg(2)))); 1386 } 1387 1388 case Builtin::BI__builtin_signbit: 1389 case Builtin::BI__builtin_signbitf: 1390 case Builtin::BI__builtin_signbitl: { 1391 LLVMContext &C = CGM.getLLVMContext(); 1392 1393 Value *Arg = EmitScalarExpr(E->getArg(0)); 1394 llvm::Type *ArgTy = Arg->getType(); 1395 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1396 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1397 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1398 if (ArgTy->isPPC_FP128Ty()) { 1399 // The higher-order double comes first, and so we need to truncate the 1400 // pair to extract the overall sign. The order of the pair is the same 1401 // in both little- and big-Endian modes. 1402 ArgWidth >>= 1; 1403 ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1404 BCArg = Builder.CreateTrunc(BCArg, ArgIntTy); 1405 } 1406 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1407 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1408 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1409 } 1410 case Builtin::BI__builtin_annotation: { 1411 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1412 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1413 AnnVal->getType()); 1414 1415 // Get the annotation string, go through casts. Sema requires this to be a 1416 // non-wide string literal, potentially casted, so the cast<> is safe. 1417 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1418 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1419 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1420 } 1421 case Builtin::BI__builtin_addcb: 1422 case Builtin::BI__builtin_addcs: 1423 case Builtin::BI__builtin_addc: 1424 case Builtin::BI__builtin_addcl: 1425 case Builtin::BI__builtin_addcll: 1426 case Builtin::BI__builtin_subcb: 1427 case Builtin::BI__builtin_subcs: 1428 case Builtin::BI__builtin_subc: 1429 case Builtin::BI__builtin_subcl: 1430 case Builtin::BI__builtin_subcll: { 1431 1432 // We translate all of these builtins from expressions of the form: 1433 // int x = ..., y = ..., carryin = ..., carryout, result; 1434 // result = __builtin_addc(x, y, carryin, &carryout); 1435 // 1436 // to LLVM IR of the form: 1437 // 1438 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1439 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1440 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1441 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1442 // i32 %carryin) 1443 // %result = extractvalue {i32, i1} %tmp2, 0 1444 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1445 // %tmp3 = or i1 %carry1, %carry2 1446 // %tmp4 = zext i1 %tmp3 to i32 1447 // store i32 %tmp4, i32* %carryout 1448 1449 // Scalarize our inputs. 1450 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1451 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1452 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1453 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1454 EmitPointerWithAlignment(E->getArg(3)); 1455 1456 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1457 llvm::Intrinsic::ID IntrinsicId; 1458 switch (BuiltinID) { 1459 default: llvm_unreachable("Unknown multiprecision builtin id."); 1460 case Builtin::BI__builtin_addcb: 1461 case Builtin::BI__builtin_addcs: 1462 case Builtin::BI__builtin_addc: 1463 case Builtin::BI__builtin_addcl: 1464 case Builtin::BI__builtin_addcll: 1465 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1466 break; 1467 case Builtin::BI__builtin_subcb: 1468 case Builtin::BI__builtin_subcs: 1469 case Builtin::BI__builtin_subc: 1470 case Builtin::BI__builtin_subcl: 1471 case Builtin::BI__builtin_subcll: 1472 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1473 break; 1474 } 1475 1476 // Construct our resulting LLVM IR expression. 1477 llvm::Value *Carry1; 1478 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1479 X, Y, Carry1); 1480 llvm::Value *Carry2; 1481 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1482 Sum1, Carryin, Carry2); 1483 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1484 X->getType()); 1485 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1486 CarryOutPtr.first); 1487 CarryOutStore->setAlignment(CarryOutPtr.second); 1488 return RValue::get(Sum2); 1489 } 1490 case Builtin::BI__builtin_uadd_overflow: 1491 case Builtin::BI__builtin_uaddl_overflow: 1492 case Builtin::BI__builtin_uaddll_overflow: 1493 case Builtin::BI__builtin_usub_overflow: 1494 case Builtin::BI__builtin_usubl_overflow: 1495 case Builtin::BI__builtin_usubll_overflow: 1496 case Builtin::BI__builtin_umul_overflow: 1497 case Builtin::BI__builtin_umull_overflow: 1498 case Builtin::BI__builtin_umulll_overflow: 1499 case Builtin::BI__builtin_sadd_overflow: 1500 case Builtin::BI__builtin_saddl_overflow: 1501 case Builtin::BI__builtin_saddll_overflow: 1502 case Builtin::BI__builtin_ssub_overflow: 1503 case Builtin::BI__builtin_ssubl_overflow: 1504 case Builtin::BI__builtin_ssubll_overflow: 1505 case Builtin::BI__builtin_smul_overflow: 1506 case Builtin::BI__builtin_smull_overflow: 1507 case Builtin::BI__builtin_smulll_overflow: { 1508 1509 // We translate all of these builtins directly to the relevant llvm IR node. 1510 1511 // Scalarize our inputs. 1512 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1513 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1514 std::pair<llvm::Value *, unsigned> SumOutPtr = 1515 EmitPointerWithAlignment(E->getArg(2)); 1516 1517 // Decide which of the overflow intrinsics we are lowering to: 1518 llvm::Intrinsic::ID IntrinsicId; 1519 switch (BuiltinID) { 1520 default: llvm_unreachable("Unknown security overflow builtin id."); 1521 case Builtin::BI__builtin_uadd_overflow: 1522 case Builtin::BI__builtin_uaddl_overflow: 1523 case Builtin::BI__builtin_uaddll_overflow: 1524 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1525 break; 1526 case Builtin::BI__builtin_usub_overflow: 1527 case Builtin::BI__builtin_usubl_overflow: 1528 case Builtin::BI__builtin_usubll_overflow: 1529 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1530 break; 1531 case Builtin::BI__builtin_umul_overflow: 1532 case Builtin::BI__builtin_umull_overflow: 1533 case Builtin::BI__builtin_umulll_overflow: 1534 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1535 break; 1536 case Builtin::BI__builtin_sadd_overflow: 1537 case Builtin::BI__builtin_saddl_overflow: 1538 case Builtin::BI__builtin_saddll_overflow: 1539 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1540 break; 1541 case Builtin::BI__builtin_ssub_overflow: 1542 case Builtin::BI__builtin_ssubl_overflow: 1543 case Builtin::BI__builtin_ssubll_overflow: 1544 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1545 break; 1546 case Builtin::BI__builtin_smul_overflow: 1547 case Builtin::BI__builtin_smull_overflow: 1548 case Builtin::BI__builtin_smulll_overflow: 1549 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1550 break; 1551 } 1552 1553 1554 llvm::Value *Carry; 1555 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1556 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1557 SumOutStore->setAlignment(SumOutPtr.second); 1558 1559 return RValue::get(Carry); 1560 } 1561 case Builtin::BI__builtin_addressof: 1562 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1563 case Builtin::BI__builtin_operator_new: 1564 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1565 E->getArg(0), false); 1566 case Builtin::BI__builtin_operator_delete: 1567 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1568 E->getArg(0), true); 1569 case Builtin::BI__noop: 1570 // __noop always evaluates to an integer literal zero. 1571 return RValue::get(ConstantInt::get(IntTy, 0)); 1572 case Builtin::BI__builtin_call_with_static_chain: { 1573 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1574 const Expr *Chain = E->getArg(1); 1575 return EmitCall(Call->getCallee()->getType(), 1576 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1577 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1578 } 1579 case Builtin::BI_InterlockedExchange: 1580 case Builtin::BI_InterlockedExchangePointer: 1581 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1582 case Builtin::BI_InterlockedCompareExchangePointer: { 1583 llvm::Type *RTy; 1584 llvm::IntegerType *IntType = 1585 IntegerType::get(getLLVMContext(), 1586 getContext().getTypeSize(E->getType())); 1587 llvm::Type *IntPtrType = IntType->getPointerTo(); 1588 1589 llvm::Value *Destination = 1590 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1591 1592 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1593 RTy = Exchange->getType(); 1594 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1595 1596 llvm::Value *Comparand = 1597 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1598 1599 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1600 SequentiallyConsistent, 1601 SequentiallyConsistent); 1602 Result->setVolatile(true); 1603 1604 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1605 0), 1606 RTy)); 1607 } 1608 case Builtin::BI_InterlockedCompareExchange: { 1609 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1610 EmitScalarExpr(E->getArg(0)), 1611 EmitScalarExpr(E->getArg(2)), 1612 EmitScalarExpr(E->getArg(1)), 1613 SequentiallyConsistent, 1614 SequentiallyConsistent); 1615 CXI->setVolatile(true); 1616 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1617 } 1618 case Builtin::BI_InterlockedIncrement: { 1619 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1620 AtomicRMWInst::Add, 1621 EmitScalarExpr(E->getArg(0)), 1622 ConstantInt::get(Int32Ty, 1), 1623 llvm::SequentiallyConsistent); 1624 RMWI->setVolatile(true); 1625 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1626 } 1627 case Builtin::BI_InterlockedDecrement: { 1628 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1629 AtomicRMWInst::Sub, 1630 EmitScalarExpr(E->getArg(0)), 1631 ConstantInt::get(Int32Ty, 1), 1632 llvm::SequentiallyConsistent); 1633 RMWI->setVolatile(true); 1634 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1635 } 1636 case Builtin::BI_InterlockedExchangeAdd: { 1637 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1638 AtomicRMWInst::Add, 1639 EmitScalarExpr(E->getArg(0)), 1640 EmitScalarExpr(E->getArg(1)), 1641 llvm::SequentiallyConsistent); 1642 RMWI->setVolatile(true); 1643 return RValue::get(RMWI); 1644 } 1645 case Builtin::BI__readfsdword: { 1646 Value *IntToPtr = 1647 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1648 llvm::PointerType::get(CGM.Int32Ty, 257)); 1649 LoadInst *Load = 1650 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1651 return RValue::get(Load); 1652 } 1653 } 1654 1655 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1656 // the call using the normal call path, but using the unmangled 1657 // version of the function name. 1658 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1659 return emitLibraryCall(*this, FD, E, 1660 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1661 1662 // If this is a predefined lib function (e.g. malloc), emit the call 1663 // using exactly the normal call path. 1664 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1665 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1666 1667 // See if we have a target specific intrinsic. 1668 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1669 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1670 if (const char *Prefix = 1671 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1672 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1673 // NOTE we dont need to perform a compatibility flag check here since the 1674 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1675 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1676 if (IntrinsicID == Intrinsic::not_intrinsic) 1677 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1678 } 1679 1680 if (IntrinsicID != Intrinsic::not_intrinsic) { 1681 SmallVector<Value*, 16> Args; 1682 1683 // Find out if any arguments are required to be integer constant 1684 // expressions. 1685 unsigned ICEArguments = 0; 1686 ASTContext::GetBuiltinTypeError Error; 1687 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1688 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1689 1690 Function *F = CGM.getIntrinsic(IntrinsicID); 1691 llvm::FunctionType *FTy = F->getFunctionType(); 1692 1693 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1694 Value *ArgValue; 1695 // If this is a normal argument, just emit it as a scalar. 1696 if ((ICEArguments & (1 << i)) == 0) { 1697 ArgValue = EmitScalarExpr(E->getArg(i)); 1698 } else { 1699 // If this is required to be a constant, constant fold it so that we 1700 // know that the generated intrinsic gets a ConstantInt. 1701 llvm::APSInt Result; 1702 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1703 assert(IsConst && "Constant arg isn't actually constant?"); 1704 (void)IsConst; 1705 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1706 } 1707 1708 // If the intrinsic arg type is different from the builtin arg type 1709 // we need to do a bit cast. 1710 llvm::Type *PTy = FTy->getParamType(i); 1711 if (PTy != ArgValue->getType()) { 1712 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1713 "Must be able to losslessly bit cast to param"); 1714 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1715 } 1716 1717 Args.push_back(ArgValue); 1718 } 1719 1720 Value *V = Builder.CreateCall(F, Args); 1721 QualType BuiltinRetType = E->getType(); 1722 1723 llvm::Type *RetTy = VoidTy; 1724 if (!BuiltinRetType->isVoidType()) 1725 RetTy = ConvertType(BuiltinRetType); 1726 1727 if (RetTy != V->getType()) { 1728 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1729 "Must be able to losslessly bit cast result type"); 1730 V = Builder.CreateBitCast(V, RetTy); 1731 } 1732 1733 return RValue::get(V); 1734 } 1735 1736 // See if we have a target specific builtin that needs to be lowered. 1737 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1738 return RValue::get(V); 1739 1740 ErrorUnsupported(E, "builtin function"); 1741 1742 // Unknown builtin, for now just dump it out and return undef. 1743 return GetUndefRValue(E->getType()); 1744 } 1745 1746 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1747 const CallExpr *E) { 1748 switch (getTarget().getTriple().getArch()) { 1749 case llvm::Triple::arm: 1750 case llvm::Triple::armeb: 1751 case llvm::Triple::thumb: 1752 case llvm::Triple::thumbeb: 1753 return EmitARMBuiltinExpr(BuiltinID, E); 1754 case llvm::Triple::aarch64: 1755 case llvm::Triple::aarch64_be: 1756 return EmitAArch64BuiltinExpr(BuiltinID, E); 1757 case llvm::Triple::x86: 1758 case llvm::Triple::x86_64: 1759 return EmitX86BuiltinExpr(BuiltinID, E); 1760 case llvm::Triple::ppc: 1761 case llvm::Triple::ppc64: 1762 case llvm::Triple::ppc64le: 1763 return EmitPPCBuiltinExpr(BuiltinID, E); 1764 case llvm::Triple::r600: 1765 case llvm::Triple::amdgcn: 1766 return EmitR600BuiltinExpr(BuiltinID, E); 1767 default: 1768 return nullptr; 1769 } 1770 } 1771 1772 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1773 NeonTypeFlags TypeFlags, 1774 bool V1Ty=false) { 1775 int IsQuad = TypeFlags.isQuad(); 1776 switch (TypeFlags.getEltType()) { 1777 case NeonTypeFlags::Int8: 1778 case NeonTypeFlags::Poly8: 1779 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1780 case NeonTypeFlags::Int16: 1781 case NeonTypeFlags::Poly16: 1782 case NeonTypeFlags::Float16: 1783 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1784 case NeonTypeFlags::Int32: 1785 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1786 case NeonTypeFlags::Int64: 1787 case NeonTypeFlags::Poly64: 1788 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1789 case NeonTypeFlags::Poly128: 1790 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1791 // There is a lot of i128 and f128 API missing. 1792 // so we use v16i8 to represent poly128 and get pattern matched. 1793 return llvm::VectorType::get(CGF->Int8Ty, 16); 1794 case NeonTypeFlags::Float32: 1795 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1796 case NeonTypeFlags::Float64: 1797 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1798 } 1799 llvm_unreachable("Unknown vector element type!"); 1800 } 1801 1802 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1803 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1804 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1805 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1806 } 1807 1808 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1809 const char *name, 1810 unsigned shift, bool rightshift) { 1811 unsigned j = 0; 1812 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1813 ai != ae; ++ai, ++j) 1814 if (shift > 0 && shift == j) 1815 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1816 else 1817 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1818 1819 return Builder.CreateCall(F, Ops, name); 1820 } 1821 1822 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1823 bool neg) { 1824 int SV = cast<ConstantInt>(V)->getSExtValue(); 1825 1826 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1827 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1828 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1829 } 1830 1831 // \brief Right-shift a vector by a constant. 1832 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1833 llvm::Type *Ty, bool usgn, 1834 const char *name) { 1835 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1836 1837 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1838 int EltSize = VTy->getScalarSizeInBits(); 1839 1840 Vec = Builder.CreateBitCast(Vec, Ty); 1841 1842 // lshr/ashr are undefined when the shift amount is equal to the vector 1843 // element size. 1844 if (ShiftAmt == EltSize) { 1845 if (usgn) { 1846 // Right-shifting an unsigned value by its size yields 0. 1847 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1848 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1849 } else { 1850 // Right-shifting a signed value by its size is equivalent 1851 // to a shift of size-1. 1852 --ShiftAmt; 1853 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1854 } 1855 } 1856 1857 Shift = EmitNeonShiftVector(Shift, Ty, false); 1858 if (usgn) 1859 return Builder.CreateLShr(Vec, Shift, name); 1860 else 1861 return Builder.CreateAShr(Vec, Shift, name); 1862 } 1863 1864 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1865 /// alignment of the type referenced by the pointer. Skip over implicit 1866 /// casts. 1867 std::pair<llvm::Value*, unsigned> 1868 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1869 assert(Addr->getType()->isPointerType()); 1870 Addr = Addr->IgnoreParens(); 1871 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1872 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1873 ICE->getSubExpr()->getType()->isPointerType()) { 1874 std::pair<llvm::Value*, unsigned> Ptr = 1875 EmitPointerWithAlignment(ICE->getSubExpr()); 1876 Ptr.first = Builder.CreateBitCast(Ptr.first, 1877 ConvertType(Addr->getType())); 1878 return Ptr; 1879 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1880 LValue LV = EmitLValue(ICE->getSubExpr()); 1881 unsigned Align = LV.getAlignment().getQuantity(); 1882 if (!Align) { 1883 // FIXME: Once LValues are fixed to always set alignment, 1884 // zap this code. 1885 QualType PtTy = ICE->getSubExpr()->getType(); 1886 if (!PtTy->isIncompleteType()) 1887 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1888 else 1889 Align = 1; 1890 } 1891 return std::make_pair(LV.getAddress(), Align); 1892 } 1893 } 1894 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1895 if (UO->getOpcode() == UO_AddrOf) { 1896 LValue LV = EmitLValue(UO->getSubExpr()); 1897 unsigned Align = LV.getAlignment().getQuantity(); 1898 if (!Align) { 1899 // FIXME: Once LValues are fixed to always set alignment, 1900 // zap this code. 1901 QualType PtTy = UO->getSubExpr()->getType(); 1902 if (!PtTy->isIncompleteType()) 1903 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1904 else 1905 Align = 1; 1906 } 1907 return std::make_pair(LV.getAddress(), Align); 1908 } 1909 } 1910 1911 unsigned Align = 1; 1912 QualType PtTy = Addr->getType()->getPointeeType(); 1913 if (!PtTy->isIncompleteType()) 1914 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1915 1916 return std::make_pair(EmitScalarExpr(Addr), Align); 1917 } 1918 1919 enum { 1920 AddRetType = (1 << 0), 1921 Add1ArgType = (1 << 1), 1922 Add2ArgTypes = (1 << 2), 1923 1924 VectorizeRetType = (1 << 3), 1925 VectorizeArgTypes = (1 << 4), 1926 1927 InventFloatType = (1 << 5), 1928 UnsignedAlts = (1 << 6), 1929 1930 Use64BitVectors = (1 << 7), 1931 Use128BitVectors = (1 << 8), 1932 1933 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 1934 VectorRet = AddRetType | VectorizeRetType, 1935 VectorRetGetArgs01 = 1936 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 1937 FpCmpzModifiers = 1938 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 1939 }; 1940 1941 struct NeonIntrinsicInfo { 1942 unsigned BuiltinID; 1943 unsigned LLVMIntrinsic; 1944 unsigned AltLLVMIntrinsic; 1945 const char *NameHint; 1946 unsigned TypeModifier; 1947 1948 bool operator<(unsigned RHSBuiltinID) const { 1949 return BuiltinID < RHSBuiltinID; 1950 } 1951 }; 1952 1953 #define NEONMAP0(NameBase) \ 1954 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 1955 1956 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 1957 { NEON:: BI__builtin_neon_ ## NameBase, \ 1958 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 1959 1960 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 1961 { NEON:: BI__builtin_neon_ ## NameBase, \ 1962 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 1963 #NameBase, TypeModifier } 1964 1965 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 1966 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1967 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1968 NEONMAP1(vabs_v, arm_neon_vabs, 0), 1969 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 1970 NEONMAP0(vaddhn_v), 1971 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 1972 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 1973 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 1974 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 1975 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 1976 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 1977 NEONMAP1(vcage_v, arm_neon_vacge, 0), 1978 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 1979 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 1980 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 1981 NEONMAP1(vcale_v, arm_neon_vacge, 0), 1982 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 1983 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 1984 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 1985 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 1986 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 1987 NEONMAP1(vclz_v, ctlz, Add1ArgType), 1988 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 1989 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 1990 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 1991 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 1992 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 1993 NEONMAP0(vcvt_f32_v), 1994 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1995 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1996 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1997 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1998 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1999 NEONMAP0(vcvt_s32_v), 2000 NEONMAP0(vcvt_s64_v), 2001 NEONMAP0(vcvt_u32_v), 2002 NEONMAP0(vcvt_u64_v), 2003 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2004 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2005 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2006 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2007 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2008 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2009 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2010 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2011 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2012 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2013 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2014 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2015 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2016 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2017 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2018 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2019 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2020 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2021 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2022 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2023 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2024 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2025 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2026 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2027 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2028 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2029 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2030 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2031 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2032 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2033 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2034 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2035 NEONMAP0(vcvtq_f32_v), 2036 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2037 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2038 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2039 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2040 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2041 NEONMAP0(vcvtq_s32_v), 2042 NEONMAP0(vcvtq_s64_v), 2043 NEONMAP0(vcvtq_u32_v), 2044 NEONMAP0(vcvtq_u64_v), 2045 NEONMAP0(vext_v), 2046 NEONMAP0(vextq_v), 2047 NEONMAP0(vfma_v), 2048 NEONMAP0(vfmaq_v), 2049 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2050 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2051 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2052 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2053 NEONMAP0(vld1_dup_v), 2054 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2055 NEONMAP0(vld1q_dup_v), 2056 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2057 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2058 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2059 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2060 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2061 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2062 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2063 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2064 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2065 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2066 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2067 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2068 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2069 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2070 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2071 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2072 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2073 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2074 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2075 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2076 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2077 NEONMAP0(vmovl_v), 2078 NEONMAP0(vmovn_v), 2079 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2080 NEONMAP0(vmull_v), 2081 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2082 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2083 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2084 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2085 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2086 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2087 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2088 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2089 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2090 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2091 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2092 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2093 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2094 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2095 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2096 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2097 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2098 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2099 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2100 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2101 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2102 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2103 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2104 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2105 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2106 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2107 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2108 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2109 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2110 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2111 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2112 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2113 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2114 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2115 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2116 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2117 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2118 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2119 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2120 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2121 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2122 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2123 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2124 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2125 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2126 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2127 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2128 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2129 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2130 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2131 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2132 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2133 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2134 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2135 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2136 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2137 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2138 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2139 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2140 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2141 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2142 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2143 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2144 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2145 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2146 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2147 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2148 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2149 NEONMAP0(vshl_n_v), 2150 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2151 NEONMAP0(vshll_n_v), 2152 NEONMAP0(vshlq_n_v), 2153 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2154 NEONMAP0(vshr_n_v), 2155 NEONMAP0(vshrn_n_v), 2156 NEONMAP0(vshrq_n_v), 2157 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2158 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2159 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2160 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2161 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2162 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2163 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2164 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2165 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2166 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2167 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2168 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2169 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2170 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2171 NEONMAP0(vsubhn_v), 2172 NEONMAP0(vtrn_v), 2173 NEONMAP0(vtrnq_v), 2174 NEONMAP0(vtst_v), 2175 NEONMAP0(vtstq_v), 2176 NEONMAP0(vuzp_v), 2177 NEONMAP0(vuzpq_v), 2178 NEONMAP0(vzip_v), 2179 NEONMAP0(vzipq_v) 2180 }; 2181 2182 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2183 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2184 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2185 NEONMAP0(vaddhn_v), 2186 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2187 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2188 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2189 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2190 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2191 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2192 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2193 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2194 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2195 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2196 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2197 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2198 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2199 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2200 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2201 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2202 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2203 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2204 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2205 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2206 NEONMAP0(vcvt_f32_v), 2207 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2208 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2209 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2210 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2211 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2212 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2213 NEONMAP0(vcvtq_f32_v), 2214 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2215 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2216 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2217 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2218 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2219 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2220 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2221 NEONMAP0(vext_v), 2222 NEONMAP0(vextq_v), 2223 NEONMAP0(vfma_v), 2224 NEONMAP0(vfmaq_v), 2225 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2226 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2227 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2228 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2229 NEONMAP0(vmovl_v), 2230 NEONMAP0(vmovn_v), 2231 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2232 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2233 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2234 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2235 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2236 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2237 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2238 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2239 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2240 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2241 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2242 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2243 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2244 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2245 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2246 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2247 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2248 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2249 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2250 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2251 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2252 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2253 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2254 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2255 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2256 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2257 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2258 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2259 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2260 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2261 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2262 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2263 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2264 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2265 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2266 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2267 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2268 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2269 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2270 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2271 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2272 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2273 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2274 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2275 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2276 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2277 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2278 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2279 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2280 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2281 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2282 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2283 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2284 NEONMAP0(vshl_n_v), 2285 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2286 NEONMAP0(vshll_n_v), 2287 NEONMAP0(vshlq_n_v), 2288 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2289 NEONMAP0(vshr_n_v), 2290 NEONMAP0(vshrn_n_v), 2291 NEONMAP0(vshrq_n_v), 2292 NEONMAP0(vsubhn_v), 2293 NEONMAP0(vtst_v), 2294 NEONMAP0(vtstq_v), 2295 }; 2296 2297 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2298 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2299 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2300 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2301 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2302 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2303 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2304 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2305 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2306 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2307 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2308 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2309 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2310 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2311 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2312 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2313 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2314 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2315 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2316 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2317 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2318 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2319 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2320 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2321 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2322 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2323 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2324 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2325 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2326 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2327 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2328 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2329 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2330 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2331 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2332 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2333 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2334 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2335 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2336 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2337 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2338 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2339 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2340 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2341 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2342 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2343 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2344 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2345 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2346 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2347 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2348 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2349 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2350 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2351 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2352 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2353 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2354 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2355 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2356 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2357 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2358 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2359 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2360 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2361 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2362 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2363 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2364 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2365 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2366 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2367 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2368 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2369 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2370 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2371 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2372 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2373 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2374 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2375 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2376 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2377 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2378 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2379 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2380 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2381 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2382 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2383 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2384 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2385 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2386 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2387 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2388 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2389 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2390 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2391 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2392 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2393 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2394 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2395 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2396 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2397 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2398 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2399 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2400 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2401 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2402 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2403 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2404 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2405 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2406 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2407 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2408 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2409 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2410 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2411 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2412 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2413 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2414 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2415 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2416 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2417 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2418 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2419 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2420 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2421 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2422 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2423 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2424 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2425 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2426 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2427 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2428 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2429 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2430 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2431 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2432 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2433 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2434 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2435 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2436 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2437 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2438 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2439 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2440 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2441 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2442 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2443 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2444 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2445 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2446 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2447 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2448 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2449 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2450 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2451 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2452 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2453 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2454 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2455 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2456 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2457 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2458 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2459 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2460 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2461 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2462 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2463 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2464 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2465 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2466 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2467 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2468 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2469 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2470 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2471 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2472 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2473 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2474 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2475 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2476 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2477 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2478 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2479 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2480 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2481 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2482 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2483 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2484 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2485 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2486 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2487 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2488 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2489 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2490 }; 2491 2492 #undef NEONMAP0 2493 #undef NEONMAP1 2494 #undef NEONMAP2 2495 2496 static bool NEONSIMDIntrinsicsProvenSorted = false; 2497 2498 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2499 static bool AArch64SISDIntrinsicsProvenSorted = false; 2500 2501 2502 static const NeonIntrinsicInfo * 2503 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2504 unsigned BuiltinID, bool &MapProvenSorted) { 2505 2506 #ifndef NDEBUG 2507 if (!MapProvenSorted) { 2508 // FIXME: use std::is_sorted once C++11 is allowed 2509 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2510 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2511 MapProvenSorted = true; 2512 } 2513 #endif 2514 2515 const NeonIntrinsicInfo *Builtin = 2516 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2517 2518 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2519 return Builtin; 2520 2521 return nullptr; 2522 } 2523 2524 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2525 unsigned Modifier, 2526 llvm::Type *ArgType, 2527 const CallExpr *E) { 2528 int VectorSize = 0; 2529 if (Modifier & Use64BitVectors) 2530 VectorSize = 64; 2531 else if (Modifier & Use128BitVectors) 2532 VectorSize = 128; 2533 2534 // Return type. 2535 SmallVector<llvm::Type *, 3> Tys; 2536 if (Modifier & AddRetType) { 2537 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2538 if (Modifier & VectorizeRetType) 2539 Ty = llvm::VectorType::get( 2540 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2541 2542 Tys.push_back(Ty); 2543 } 2544 2545 // Arguments. 2546 if (Modifier & VectorizeArgTypes) { 2547 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2548 ArgType = llvm::VectorType::get(ArgType, Elts); 2549 } 2550 2551 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2552 Tys.push_back(ArgType); 2553 2554 if (Modifier & Add2ArgTypes) 2555 Tys.push_back(ArgType); 2556 2557 if (Modifier & InventFloatType) 2558 Tys.push_back(FloatTy); 2559 2560 return CGM.getIntrinsic(IntrinsicID, Tys); 2561 } 2562 2563 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2564 const NeonIntrinsicInfo &SISDInfo, 2565 SmallVectorImpl<Value *> &Ops, 2566 const CallExpr *E) { 2567 unsigned BuiltinID = SISDInfo.BuiltinID; 2568 unsigned int Int = SISDInfo.LLVMIntrinsic; 2569 unsigned Modifier = SISDInfo.TypeModifier; 2570 const char *s = SISDInfo.NameHint; 2571 2572 switch (BuiltinID) { 2573 case NEON::BI__builtin_neon_vcled_s64: 2574 case NEON::BI__builtin_neon_vcled_u64: 2575 case NEON::BI__builtin_neon_vcles_f32: 2576 case NEON::BI__builtin_neon_vcled_f64: 2577 case NEON::BI__builtin_neon_vcltd_s64: 2578 case NEON::BI__builtin_neon_vcltd_u64: 2579 case NEON::BI__builtin_neon_vclts_f32: 2580 case NEON::BI__builtin_neon_vcltd_f64: 2581 case NEON::BI__builtin_neon_vcales_f32: 2582 case NEON::BI__builtin_neon_vcaled_f64: 2583 case NEON::BI__builtin_neon_vcalts_f32: 2584 case NEON::BI__builtin_neon_vcaltd_f64: 2585 // Only one direction of comparisons actually exist, cmle is actually a cmge 2586 // with swapped operands. The table gives us the right intrinsic but we 2587 // still need to do the swap. 2588 std::swap(Ops[0], Ops[1]); 2589 break; 2590 } 2591 2592 assert(Int && "Generic code assumes a valid intrinsic"); 2593 2594 // Determine the type(s) of this overloaded AArch64 intrinsic. 2595 const Expr *Arg = E->getArg(0); 2596 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2597 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2598 2599 int j = 0; 2600 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2601 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2602 ai != ae; ++ai, ++j) { 2603 llvm::Type *ArgTy = ai->getType(); 2604 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2605 ArgTy->getPrimitiveSizeInBits()) 2606 continue; 2607 2608 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2609 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2610 // it before inserting. 2611 Ops[j] = 2612 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2613 Ops[j] = 2614 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2615 } 2616 2617 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2618 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2619 if (ResultType->getPrimitiveSizeInBits() < 2620 Result->getType()->getPrimitiveSizeInBits()) 2621 return CGF.Builder.CreateExtractElement(Result, C0); 2622 2623 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2624 } 2625 2626 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2627 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2628 const char *NameHint, unsigned Modifier, const CallExpr *E, 2629 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2630 // Get the last argument, which specifies the vector type. 2631 llvm::APSInt NeonTypeConst; 2632 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2633 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2634 return nullptr; 2635 2636 // Determine the type of this overloaded NEON intrinsic. 2637 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2638 bool Usgn = Type.isUnsigned(); 2639 bool Quad = Type.isQuad(); 2640 2641 llvm::VectorType *VTy = GetNeonType(this, Type); 2642 llvm::Type *Ty = VTy; 2643 if (!Ty) 2644 return nullptr; 2645 2646 unsigned Int = LLVMIntrinsic; 2647 if ((Modifier & UnsignedAlts) && !Usgn) 2648 Int = AltLLVMIntrinsic; 2649 2650 switch (BuiltinID) { 2651 default: break; 2652 case NEON::BI__builtin_neon_vabs_v: 2653 case NEON::BI__builtin_neon_vabsq_v: 2654 if (VTy->getElementType()->isFloatingPointTy()) 2655 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2656 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2657 case NEON::BI__builtin_neon_vaddhn_v: { 2658 llvm::VectorType *SrcTy = 2659 llvm::VectorType::getExtendedElementVectorType(VTy); 2660 2661 // %sum = add <4 x i32> %lhs, %rhs 2662 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2663 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2664 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2665 2666 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2667 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2668 SrcTy->getScalarSizeInBits() / 2); 2669 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2670 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2671 2672 // %res = trunc <4 x i32> %high to <4 x i16> 2673 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2674 } 2675 case NEON::BI__builtin_neon_vcale_v: 2676 case NEON::BI__builtin_neon_vcaleq_v: 2677 case NEON::BI__builtin_neon_vcalt_v: 2678 case NEON::BI__builtin_neon_vcaltq_v: 2679 std::swap(Ops[0], Ops[1]); 2680 case NEON::BI__builtin_neon_vcage_v: 2681 case NEON::BI__builtin_neon_vcageq_v: 2682 case NEON::BI__builtin_neon_vcagt_v: 2683 case NEON::BI__builtin_neon_vcagtq_v: { 2684 llvm::Type *VecFlt = llvm::VectorType::get( 2685 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2686 VTy->getNumElements()); 2687 llvm::Type *Tys[] = { VTy, VecFlt }; 2688 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2689 return EmitNeonCall(F, Ops, NameHint); 2690 } 2691 case NEON::BI__builtin_neon_vclz_v: 2692 case NEON::BI__builtin_neon_vclzq_v: 2693 // We generate target-independent intrinsic, which needs a second argument 2694 // for whether or not clz of zero is undefined; on ARM it isn't. 2695 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2696 break; 2697 case NEON::BI__builtin_neon_vcvt_f32_v: 2698 case NEON::BI__builtin_neon_vcvtq_f32_v: 2699 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2700 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2701 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2702 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2703 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2704 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2705 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2706 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2707 bool Double = 2708 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2709 llvm::Type *FloatTy = 2710 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2711 : NeonTypeFlags::Float32, 2712 false, Quad)); 2713 llvm::Type *Tys[2] = { FloatTy, Ty }; 2714 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2715 Function *F = CGM.getIntrinsic(Int, Tys); 2716 return EmitNeonCall(F, Ops, "vcvt_n"); 2717 } 2718 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2719 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2720 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2721 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2722 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2723 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2724 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2725 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2726 bool Double = 2727 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2728 llvm::Type *FloatTy = 2729 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2730 : NeonTypeFlags::Float32, 2731 false, Quad)); 2732 llvm::Type *Tys[2] = { Ty, FloatTy }; 2733 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2734 return EmitNeonCall(F, Ops, "vcvt_n"); 2735 } 2736 case NEON::BI__builtin_neon_vcvt_s32_v: 2737 case NEON::BI__builtin_neon_vcvt_u32_v: 2738 case NEON::BI__builtin_neon_vcvt_s64_v: 2739 case NEON::BI__builtin_neon_vcvt_u64_v: 2740 case NEON::BI__builtin_neon_vcvtq_s32_v: 2741 case NEON::BI__builtin_neon_vcvtq_u32_v: 2742 case NEON::BI__builtin_neon_vcvtq_s64_v: 2743 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2744 bool Double = 2745 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2746 llvm::Type *FloatTy = 2747 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2748 : NeonTypeFlags::Float32, 2749 false, Quad)); 2750 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2751 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2752 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2753 } 2754 case NEON::BI__builtin_neon_vcvta_s32_v: 2755 case NEON::BI__builtin_neon_vcvta_s64_v: 2756 case NEON::BI__builtin_neon_vcvta_u32_v: 2757 case NEON::BI__builtin_neon_vcvta_u64_v: 2758 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2759 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2760 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2761 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2762 case NEON::BI__builtin_neon_vcvtn_s32_v: 2763 case NEON::BI__builtin_neon_vcvtn_s64_v: 2764 case NEON::BI__builtin_neon_vcvtn_u32_v: 2765 case NEON::BI__builtin_neon_vcvtn_u64_v: 2766 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2767 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2768 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2769 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2770 case NEON::BI__builtin_neon_vcvtp_s32_v: 2771 case NEON::BI__builtin_neon_vcvtp_s64_v: 2772 case NEON::BI__builtin_neon_vcvtp_u32_v: 2773 case NEON::BI__builtin_neon_vcvtp_u64_v: 2774 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2775 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2776 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2777 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2778 case NEON::BI__builtin_neon_vcvtm_s32_v: 2779 case NEON::BI__builtin_neon_vcvtm_s64_v: 2780 case NEON::BI__builtin_neon_vcvtm_u32_v: 2781 case NEON::BI__builtin_neon_vcvtm_u64_v: 2782 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2783 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2784 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2785 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2786 bool Double = 2787 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2788 llvm::Type *InTy = 2789 GetNeonType(this, 2790 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2791 : NeonTypeFlags::Float32, false, Quad)); 2792 llvm::Type *Tys[2] = { Ty, InTy }; 2793 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2794 } 2795 case NEON::BI__builtin_neon_vext_v: 2796 case NEON::BI__builtin_neon_vextq_v: { 2797 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2798 SmallVector<Constant*, 16> Indices; 2799 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2800 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2801 2802 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2803 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2804 Value *SV = llvm::ConstantVector::get(Indices); 2805 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2806 } 2807 case NEON::BI__builtin_neon_vfma_v: 2808 case NEON::BI__builtin_neon_vfmaq_v: { 2809 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2810 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2811 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2812 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2813 2814 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2815 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2816 } 2817 case NEON::BI__builtin_neon_vld1_v: 2818 case NEON::BI__builtin_neon_vld1q_v: 2819 Ops.push_back(Align); 2820 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2821 case NEON::BI__builtin_neon_vld2_v: 2822 case NEON::BI__builtin_neon_vld2q_v: 2823 case NEON::BI__builtin_neon_vld3_v: 2824 case NEON::BI__builtin_neon_vld3q_v: 2825 case NEON::BI__builtin_neon_vld4_v: 2826 case NEON::BI__builtin_neon_vld4q_v: { 2827 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2828 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2829 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2830 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2831 return Builder.CreateStore(Ops[1], Ops[0]); 2832 } 2833 case NEON::BI__builtin_neon_vld1_dup_v: 2834 case NEON::BI__builtin_neon_vld1q_dup_v: { 2835 Value *V = UndefValue::get(Ty); 2836 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2837 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2838 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2839 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2840 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2841 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2842 return EmitNeonSplat(Ops[0], CI); 2843 } 2844 case NEON::BI__builtin_neon_vld2_lane_v: 2845 case NEON::BI__builtin_neon_vld2q_lane_v: 2846 case NEON::BI__builtin_neon_vld3_lane_v: 2847 case NEON::BI__builtin_neon_vld3q_lane_v: 2848 case NEON::BI__builtin_neon_vld4_lane_v: 2849 case NEON::BI__builtin_neon_vld4q_lane_v: { 2850 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2851 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2852 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2853 Ops.push_back(Align); 2854 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2855 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2856 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2857 return Builder.CreateStore(Ops[1], Ops[0]); 2858 } 2859 case NEON::BI__builtin_neon_vmovl_v: { 2860 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2861 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2862 if (Usgn) 2863 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2864 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2865 } 2866 case NEON::BI__builtin_neon_vmovn_v: { 2867 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2868 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2869 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2870 } 2871 case NEON::BI__builtin_neon_vmull_v: 2872 // FIXME: the integer vmull operations could be emitted in terms of pure 2873 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2874 // hoisting the exts outside loops. Until global ISel comes along that can 2875 // see through such movement this leads to bad CodeGen. So we need an 2876 // intrinsic for now. 2877 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2878 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2879 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2880 case NEON::BI__builtin_neon_vpadal_v: 2881 case NEON::BI__builtin_neon_vpadalq_v: { 2882 // The source operand type has twice as many elements of half the size. 2883 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2884 llvm::Type *EltTy = 2885 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2886 llvm::Type *NarrowTy = 2887 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2888 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2889 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2890 } 2891 case NEON::BI__builtin_neon_vpaddl_v: 2892 case NEON::BI__builtin_neon_vpaddlq_v: { 2893 // The source operand type has twice as many elements of half the size. 2894 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2895 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2896 llvm::Type *NarrowTy = 2897 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2898 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2899 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2900 } 2901 case NEON::BI__builtin_neon_vqdmlal_v: 2902 case NEON::BI__builtin_neon_vqdmlsl_v: { 2903 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2904 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2905 MulOps, "vqdmlal"); 2906 2907 SmallVector<Value *, 2> AccumOps; 2908 AccumOps.push_back(Ops[0]); 2909 AccumOps.push_back(Mul); 2910 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2911 AccumOps, NameHint); 2912 } 2913 case NEON::BI__builtin_neon_vqshl_n_v: 2914 case NEON::BI__builtin_neon_vqshlq_n_v: 2915 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2916 1, false); 2917 case NEON::BI__builtin_neon_vqshlu_n_v: 2918 case NEON::BI__builtin_neon_vqshluq_n_v: 2919 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 2920 1, false); 2921 case NEON::BI__builtin_neon_vrecpe_v: 2922 case NEON::BI__builtin_neon_vrecpeq_v: 2923 case NEON::BI__builtin_neon_vrsqrte_v: 2924 case NEON::BI__builtin_neon_vrsqrteq_v: 2925 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2926 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2927 2928 case NEON::BI__builtin_neon_vrshr_n_v: 2929 case NEON::BI__builtin_neon_vrshrq_n_v: 2930 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 2931 1, true); 2932 case NEON::BI__builtin_neon_vshl_n_v: 2933 case NEON::BI__builtin_neon_vshlq_n_v: 2934 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2935 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2936 "vshl_n"); 2937 case NEON::BI__builtin_neon_vshll_n_v: { 2938 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2939 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2940 if (Usgn) 2941 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2942 else 2943 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2944 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2945 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2946 } 2947 case NEON::BI__builtin_neon_vshrn_n_v: { 2948 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2949 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2950 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2951 if (Usgn) 2952 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2953 else 2954 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2955 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2956 } 2957 case NEON::BI__builtin_neon_vshr_n_v: 2958 case NEON::BI__builtin_neon_vshrq_n_v: 2959 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2960 case NEON::BI__builtin_neon_vst1_v: 2961 case NEON::BI__builtin_neon_vst1q_v: 2962 case NEON::BI__builtin_neon_vst2_v: 2963 case NEON::BI__builtin_neon_vst2q_v: 2964 case NEON::BI__builtin_neon_vst3_v: 2965 case NEON::BI__builtin_neon_vst3q_v: 2966 case NEON::BI__builtin_neon_vst4_v: 2967 case NEON::BI__builtin_neon_vst4q_v: 2968 case NEON::BI__builtin_neon_vst2_lane_v: 2969 case NEON::BI__builtin_neon_vst2q_lane_v: 2970 case NEON::BI__builtin_neon_vst3_lane_v: 2971 case NEON::BI__builtin_neon_vst3q_lane_v: 2972 case NEON::BI__builtin_neon_vst4_lane_v: 2973 case NEON::BI__builtin_neon_vst4q_lane_v: 2974 Ops.push_back(Align); 2975 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 2976 case NEON::BI__builtin_neon_vsubhn_v: { 2977 llvm::VectorType *SrcTy = 2978 llvm::VectorType::getExtendedElementVectorType(VTy); 2979 2980 // %sum = add <4 x i32> %lhs, %rhs 2981 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2982 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2983 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2984 2985 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2986 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2987 SrcTy->getScalarSizeInBits() / 2); 2988 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2989 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2990 2991 // %res = trunc <4 x i32> %high to <4 x i16> 2992 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2993 } 2994 case NEON::BI__builtin_neon_vtrn_v: 2995 case NEON::BI__builtin_neon_vtrnq_v: { 2996 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2997 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2998 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2999 Value *SV = nullptr; 3000 3001 for (unsigned vi = 0; vi != 2; ++vi) { 3002 SmallVector<Constant*, 16> Indices; 3003 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3004 Indices.push_back(Builder.getInt32(i+vi)); 3005 Indices.push_back(Builder.getInt32(i+e+vi)); 3006 } 3007 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3008 SV = llvm::ConstantVector::get(Indices); 3009 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3010 SV = Builder.CreateStore(SV, Addr); 3011 } 3012 return SV; 3013 } 3014 case NEON::BI__builtin_neon_vtst_v: 3015 case NEON::BI__builtin_neon_vtstq_v: { 3016 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3017 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3018 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3019 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3020 ConstantAggregateZero::get(Ty)); 3021 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3022 } 3023 case NEON::BI__builtin_neon_vuzp_v: 3024 case NEON::BI__builtin_neon_vuzpq_v: { 3025 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3026 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3027 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3028 Value *SV = nullptr; 3029 3030 for (unsigned vi = 0; vi != 2; ++vi) { 3031 SmallVector<Constant*, 16> Indices; 3032 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3033 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3034 3035 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3036 SV = llvm::ConstantVector::get(Indices); 3037 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3038 SV = Builder.CreateStore(SV, Addr); 3039 } 3040 return SV; 3041 } 3042 case NEON::BI__builtin_neon_vzip_v: 3043 case NEON::BI__builtin_neon_vzipq_v: { 3044 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3045 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3046 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3047 Value *SV = nullptr; 3048 3049 for (unsigned vi = 0; vi != 2; ++vi) { 3050 SmallVector<Constant*, 16> Indices; 3051 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3052 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3053 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3054 } 3055 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3056 SV = llvm::ConstantVector::get(Indices); 3057 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3058 SV = Builder.CreateStore(SV, Addr); 3059 } 3060 return SV; 3061 } 3062 } 3063 3064 assert(Int && "Expected valid intrinsic number"); 3065 3066 // Determine the type(s) of this overloaded AArch64 intrinsic. 3067 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3068 3069 Value *Result = EmitNeonCall(F, Ops, NameHint); 3070 llvm::Type *ResultType = ConvertType(E->getType()); 3071 // AArch64 intrinsic one-element vector type cast to 3072 // scalar type expected by the builtin 3073 return Builder.CreateBitCast(Result, ResultType, NameHint); 3074 } 3075 3076 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3077 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3078 const CmpInst::Predicate Ip, const Twine &Name) { 3079 llvm::Type *OTy = Op->getType(); 3080 3081 // FIXME: this is utterly horrific. We should not be looking at previous 3082 // codegen context to find out what needs doing. Unfortunately TableGen 3083 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3084 // (etc). 3085 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3086 OTy = BI->getOperand(0)->getType(); 3087 3088 Op = Builder.CreateBitCast(Op, OTy); 3089 if (OTy->getScalarType()->isFloatingPointTy()) { 3090 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3091 } else { 3092 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3093 } 3094 return Builder.CreateSExt(Op, Ty, Name); 3095 } 3096 3097 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3098 Value *ExtOp, Value *IndexOp, 3099 llvm::Type *ResTy, unsigned IntID, 3100 const char *Name) { 3101 SmallVector<Value *, 2> TblOps; 3102 if (ExtOp) 3103 TblOps.push_back(ExtOp); 3104 3105 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3106 SmallVector<Constant*, 16> Indices; 3107 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3108 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3109 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3110 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3111 } 3112 Value *SV = llvm::ConstantVector::get(Indices); 3113 3114 int PairPos = 0, End = Ops.size() - 1; 3115 while (PairPos < End) { 3116 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3117 Ops[PairPos+1], SV, Name)); 3118 PairPos += 2; 3119 } 3120 3121 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3122 // of the 128-bit lookup table with zero. 3123 if (PairPos == End) { 3124 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3125 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3126 ZeroTbl, SV, Name)); 3127 } 3128 3129 Function *TblF; 3130 TblOps.push_back(IndexOp); 3131 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3132 3133 return CGF.EmitNeonCall(TblF, TblOps, Name); 3134 } 3135 3136 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3137 switch (BuiltinID) { 3138 default: 3139 return nullptr; 3140 case ARM::BI__builtin_arm_nop: 3141 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3142 llvm::ConstantInt::get(Int32Ty, 0)); 3143 case ARM::BI__builtin_arm_yield: 3144 case ARM::BI__yield: 3145 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3146 llvm::ConstantInt::get(Int32Ty, 1)); 3147 case ARM::BI__builtin_arm_wfe: 3148 case ARM::BI__wfe: 3149 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3150 llvm::ConstantInt::get(Int32Ty, 2)); 3151 case ARM::BI__builtin_arm_wfi: 3152 case ARM::BI__wfi: 3153 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3154 llvm::ConstantInt::get(Int32Ty, 3)); 3155 case ARM::BI__builtin_arm_sev: 3156 case ARM::BI__sev: 3157 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3158 llvm::ConstantInt::get(Int32Ty, 4)); 3159 case ARM::BI__builtin_arm_sevl: 3160 case ARM::BI__sevl: 3161 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3162 llvm::ConstantInt::get(Int32Ty, 5)); 3163 } 3164 } 3165 3166 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3167 const CallExpr *E) { 3168 if (auto Hint = GetValueForARMHint(BuiltinID)) 3169 return Hint; 3170 3171 if (BuiltinID == ARM::BI__emit) { 3172 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3173 llvm::FunctionType *FTy = 3174 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3175 3176 APSInt Value; 3177 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3178 llvm_unreachable("Sema will ensure that the parameter is constant"); 3179 3180 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3181 3182 llvm::InlineAsm *Emit = 3183 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3184 /*SideEffects=*/true) 3185 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3186 /*SideEffects=*/true); 3187 3188 return Builder.CreateCall(Emit); 3189 } 3190 3191 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3192 Value *Option = EmitScalarExpr(E->getArg(0)); 3193 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3194 } 3195 3196 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3197 Value *Address = EmitScalarExpr(E->getArg(0)); 3198 Value *RW = EmitScalarExpr(E->getArg(1)); 3199 Value *IsData = EmitScalarExpr(E->getArg(2)); 3200 3201 // Locality is not supported on ARM target 3202 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3203 3204 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3205 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3206 } 3207 3208 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3209 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3210 EmitScalarExpr(E->getArg(0)), 3211 "rbit"); 3212 } 3213 3214 if (BuiltinID == ARM::BI__clear_cache) { 3215 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3216 const FunctionDecl *FD = E->getDirectCallee(); 3217 SmallVector<Value*, 2> Ops; 3218 for (unsigned i = 0; i < 2; i++) 3219 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3220 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3221 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3222 StringRef Name = FD->getName(); 3223 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3224 } 3225 3226 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3227 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3228 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3229 getContext().getTypeSize(E->getType()) == 64) || 3230 BuiltinID == ARM::BI__ldrexd) { 3231 Function *F; 3232 3233 switch (BuiltinID) { 3234 default: llvm_unreachable("unexpected builtin"); 3235 case ARM::BI__builtin_arm_ldaex: 3236 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3237 break; 3238 case ARM::BI__builtin_arm_ldrexd: 3239 case ARM::BI__builtin_arm_ldrex: 3240 case ARM::BI__ldrexd: 3241 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3242 break; 3243 } 3244 3245 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3246 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3247 "ldrexd"); 3248 3249 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3250 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3251 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3252 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3253 3254 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3255 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3256 Val = Builder.CreateOr(Val, Val1); 3257 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3258 } 3259 3260 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3261 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3262 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3263 3264 QualType Ty = E->getType(); 3265 llvm::Type *RealResTy = ConvertType(Ty); 3266 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3267 getContext().getTypeSize(Ty)); 3268 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3269 3270 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3271 ? Intrinsic::arm_ldaex 3272 : Intrinsic::arm_ldrex, 3273 LoadAddr->getType()); 3274 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3275 3276 if (RealResTy->isPointerTy()) 3277 return Builder.CreateIntToPtr(Val, RealResTy); 3278 else { 3279 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3280 return Builder.CreateBitCast(Val, RealResTy); 3281 } 3282 } 3283 3284 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3285 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3286 BuiltinID == ARM::BI__builtin_arm_strex) && 3287 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3288 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3289 ? Intrinsic::arm_stlexd 3290 : Intrinsic::arm_strexd); 3291 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3292 3293 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3294 Value *Val = EmitScalarExpr(E->getArg(0)); 3295 Builder.CreateStore(Val, Tmp); 3296 3297 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3298 Val = Builder.CreateLoad(LdPtr); 3299 3300 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3301 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3302 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3303 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3304 } 3305 3306 if (BuiltinID == ARM::BI__builtin_arm_strex || 3307 BuiltinID == ARM::BI__builtin_arm_stlex) { 3308 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3309 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3310 3311 QualType Ty = E->getArg(0)->getType(); 3312 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3313 getContext().getTypeSize(Ty)); 3314 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3315 3316 if (StoreVal->getType()->isPointerTy()) 3317 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3318 else { 3319 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3320 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3321 } 3322 3323 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3324 ? Intrinsic::arm_stlex 3325 : Intrinsic::arm_strex, 3326 StoreAddr->getType()); 3327 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3328 } 3329 3330 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3331 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3332 return Builder.CreateCall(F); 3333 } 3334 3335 // CRC32 3336 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3337 switch (BuiltinID) { 3338 case ARM::BI__builtin_arm_crc32b: 3339 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3340 case ARM::BI__builtin_arm_crc32cb: 3341 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3342 case ARM::BI__builtin_arm_crc32h: 3343 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3344 case ARM::BI__builtin_arm_crc32ch: 3345 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3346 case ARM::BI__builtin_arm_crc32w: 3347 case ARM::BI__builtin_arm_crc32d: 3348 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3349 case ARM::BI__builtin_arm_crc32cw: 3350 case ARM::BI__builtin_arm_crc32cd: 3351 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3352 } 3353 3354 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3355 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3356 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3357 3358 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3359 // intrinsics, hence we need different codegen for these cases. 3360 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3361 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3362 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3363 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3364 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3365 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3366 3367 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3368 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3369 return Builder.CreateCall2(F, Res, Arg1b); 3370 } else { 3371 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3372 3373 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3374 return Builder.CreateCall2(F, Arg0, Arg1); 3375 } 3376 } 3377 3378 SmallVector<Value*, 4> Ops; 3379 llvm::Value *Align = nullptr; 3380 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3381 if (i == 0) { 3382 switch (BuiltinID) { 3383 case NEON::BI__builtin_neon_vld1_v: 3384 case NEON::BI__builtin_neon_vld1q_v: 3385 case NEON::BI__builtin_neon_vld1q_lane_v: 3386 case NEON::BI__builtin_neon_vld1_lane_v: 3387 case NEON::BI__builtin_neon_vld1_dup_v: 3388 case NEON::BI__builtin_neon_vld1q_dup_v: 3389 case NEON::BI__builtin_neon_vst1_v: 3390 case NEON::BI__builtin_neon_vst1q_v: 3391 case NEON::BI__builtin_neon_vst1q_lane_v: 3392 case NEON::BI__builtin_neon_vst1_lane_v: 3393 case NEON::BI__builtin_neon_vst2_v: 3394 case NEON::BI__builtin_neon_vst2q_v: 3395 case NEON::BI__builtin_neon_vst2_lane_v: 3396 case NEON::BI__builtin_neon_vst2q_lane_v: 3397 case NEON::BI__builtin_neon_vst3_v: 3398 case NEON::BI__builtin_neon_vst3q_v: 3399 case NEON::BI__builtin_neon_vst3_lane_v: 3400 case NEON::BI__builtin_neon_vst3q_lane_v: 3401 case NEON::BI__builtin_neon_vst4_v: 3402 case NEON::BI__builtin_neon_vst4q_v: 3403 case NEON::BI__builtin_neon_vst4_lane_v: 3404 case NEON::BI__builtin_neon_vst4q_lane_v: 3405 // Get the alignment for the argument in addition to the value; 3406 // we'll use it later. 3407 std::pair<llvm::Value*, unsigned> Src = 3408 EmitPointerWithAlignment(E->getArg(0)); 3409 Ops.push_back(Src.first); 3410 Align = Builder.getInt32(Src.second); 3411 continue; 3412 } 3413 } 3414 if (i == 1) { 3415 switch (BuiltinID) { 3416 case NEON::BI__builtin_neon_vld2_v: 3417 case NEON::BI__builtin_neon_vld2q_v: 3418 case NEON::BI__builtin_neon_vld3_v: 3419 case NEON::BI__builtin_neon_vld3q_v: 3420 case NEON::BI__builtin_neon_vld4_v: 3421 case NEON::BI__builtin_neon_vld4q_v: 3422 case NEON::BI__builtin_neon_vld2_lane_v: 3423 case NEON::BI__builtin_neon_vld2q_lane_v: 3424 case NEON::BI__builtin_neon_vld3_lane_v: 3425 case NEON::BI__builtin_neon_vld3q_lane_v: 3426 case NEON::BI__builtin_neon_vld4_lane_v: 3427 case NEON::BI__builtin_neon_vld4q_lane_v: 3428 case NEON::BI__builtin_neon_vld2_dup_v: 3429 case NEON::BI__builtin_neon_vld3_dup_v: 3430 case NEON::BI__builtin_neon_vld4_dup_v: 3431 // Get the alignment for the argument in addition to the value; 3432 // we'll use it later. 3433 std::pair<llvm::Value*, unsigned> Src = 3434 EmitPointerWithAlignment(E->getArg(1)); 3435 Ops.push_back(Src.first); 3436 Align = Builder.getInt32(Src.second); 3437 continue; 3438 } 3439 } 3440 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3441 } 3442 3443 switch (BuiltinID) { 3444 default: break; 3445 // vget_lane and vset_lane are not overloaded and do not have an extra 3446 // argument that specifies the vector type. 3447 case NEON::BI__builtin_neon_vget_lane_i8: 3448 case NEON::BI__builtin_neon_vget_lane_i16: 3449 case NEON::BI__builtin_neon_vget_lane_i32: 3450 case NEON::BI__builtin_neon_vget_lane_i64: 3451 case NEON::BI__builtin_neon_vget_lane_f32: 3452 case NEON::BI__builtin_neon_vgetq_lane_i8: 3453 case NEON::BI__builtin_neon_vgetq_lane_i16: 3454 case NEON::BI__builtin_neon_vgetq_lane_i32: 3455 case NEON::BI__builtin_neon_vgetq_lane_i64: 3456 case NEON::BI__builtin_neon_vgetq_lane_f32: 3457 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3458 "vget_lane"); 3459 case NEON::BI__builtin_neon_vset_lane_i8: 3460 case NEON::BI__builtin_neon_vset_lane_i16: 3461 case NEON::BI__builtin_neon_vset_lane_i32: 3462 case NEON::BI__builtin_neon_vset_lane_i64: 3463 case NEON::BI__builtin_neon_vset_lane_f32: 3464 case NEON::BI__builtin_neon_vsetq_lane_i8: 3465 case NEON::BI__builtin_neon_vsetq_lane_i16: 3466 case NEON::BI__builtin_neon_vsetq_lane_i32: 3467 case NEON::BI__builtin_neon_vsetq_lane_i64: 3468 case NEON::BI__builtin_neon_vsetq_lane_f32: 3469 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3470 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3471 3472 // Non-polymorphic crypto instructions also not overloaded 3473 case NEON::BI__builtin_neon_vsha1h_u32: 3474 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3475 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3476 "vsha1h"); 3477 case NEON::BI__builtin_neon_vsha1cq_u32: 3478 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3479 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3480 "vsha1h"); 3481 case NEON::BI__builtin_neon_vsha1pq_u32: 3482 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3483 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3484 "vsha1h"); 3485 case NEON::BI__builtin_neon_vsha1mq_u32: 3486 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3487 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3488 "vsha1h"); 3489 } 3490 3491 // Get the last argument, which specifies the vector type. 3492 llvm::APSInt Result; 3493 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3494 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3495 return nullptr; 3496 3497 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3498 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3499 // Determine the overloaded type of this builtin. 3500 llvm::Type *Ty; 3501 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3502 Ty = FloatTy; 3503 else 3504 Ty = DoubleTy; 3505 3506 // Determine whether this is an unsigned conversion or not. 3507 bool usgn = Result.getZExtValue() == 1; 3508 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3509 3510 // Call the appropriate intrinsic. 3511 Function *F = CGM.getIntrinsic(Int, Ty); 3512 return Builder.CreateCall(F, Ops, "vcvtr"); 3513 } 3514 3515 // Determine the type of this overloaded NEON intrinsic. 3516 NeonTypeFlags Type(Result.getZExtValue()); 3517 bool usgn = Type.isUnsigned(); 3518 bool rightShift = false; 3519 3520 llvm::VectorType *VTy = GetNeonType(this, Type); 3521 llvm::Type *Ty = VTy; 3522 if (!Ty) 3523 return nullptr; 3524 3525 // Many NEON builtins have identical semantics and uses in ARM and 3526 // AArch64. Emit these in a single function. 3527 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3528 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3529 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3530 if (Builtin) 3531 return EmitCommonNeonBuiltinExpr( 3532 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3533 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3534 3535 unsigned Int; 3536 switch (BuiltinID) { 3537 default: return nullptr; 3538 case NEON::BI__builtin_neon_vld1q_lane_v: 3539 // Handle 64-bit integer elements as a special case. Use shuffles of 3540 // one-element vectors to avoid poor code for i64 in the backend. 3541 if (VTy->getElementType()->isIntegerTy(64)) { 3542 // Extract the other lane. 3543 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3544 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3545 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3546 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3547 // Load the value as a one-element vector. 3548 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3549 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3550 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3551 // Combine them. 3552 SmallVector<Constant*, 2> Indices; 3553 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3554 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3555 SV = llvm::ConstantVector::get(Indices); 3556 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3557 } 3558 // fall through 3559 case NEON::BI__builtin_neon_vld1_lane_v: { 3560 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3561 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3562 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3563 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3564 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3565 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3566 } 3567 case NEON::BI__builtin_neon_vld2_dup_v: 3568 case NEON::BI__builtin_neon_vld3_dup_v: 3569 case NEON::BI__builtin_neon_vld4_dup_v: { 3570 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3571 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3572 switch (BuiltinID) { 3573 case NEON::BI__builtin_neon_vld2_dup_v: 3574 Int = Intrinsic::arm_neon_vld2; 3575 break; 3576 case NEON::BI__builtin_neon_vld3_dup_v: 3577 Int = Intrinsic::arm_neon_vld3; 3578 break; 3579 case NEON::BI__builtin_neon_vld4_dup_v: 3580 Int = Intrinsic::arm_neon_vld4; 3581 break; 3582 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3583 } 3584 Function *F = CGM.getIntrinsic(Int, Ty); 3585 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3586 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3587 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3588 return Builder.CreateStore(Ops[1], Ops[0]); 3589 } 3590 switch (BuiltinID) { 3591 case NEON::BI__builtin_neon_vld2_dup_v: 3592 Int = Intrinsic::arm_neon_vld2lane; 3593 break; 3594 case NEON::BI__builtin_neon_vld3_dup_v: 3595 Int = Intrinsic::arm_neon_vld3lane; 3596 break; 3597 case NEON::BI__builtin_neon_vld4_dup_v: 3598 Int = Intrinsic::arm_neon_vld4lane; 3599 break; 3600 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3601 } 3602 Function *F = CGM.getIntrinsic(Int, Ty); 3603 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3604 3605 SmallVector<Value*, 6> Args; 3606 Args.push_back(Ops[1]); 3607 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3608 3609 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3610 Args.push_back(CI); 3611 Args.push_back(Align); 3612 3613 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3614 // splat lane 0 to all elts in each vector of the result. 3615 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3616 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3617 Value *Elt = Builder.CreateBitCast(Val, Ty); 3618 Elt = EmitNeonSplat(Elt, CI); 3619 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3620 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3621 } 3622 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3623 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3624 return Builder.CreateStore(Ops[1], Ops[0]); 3625 } 3626 case NEON::BI__builtin_neon_vqrshrn_n_v: 3627 Int = 3628 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3629 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3630 1, true); 3631 case NEON::BI__builtin_neon_vqrshrun_n_v: 3632 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3633 Ops, "vqrshrun_n", 1, true); 3634 case NEON::BI__builtin_neon_vqshrn_n_v: 3635 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3636 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3637 1, true); 3638 case NEON::BI__builtin_neon_vqshrun_n_v: 3639 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3640 Ops, "vqshrun_n", 1, true); 3641 case NEON::BI__builtin_neon_vrecpe_v: 3642 case NEON::BI__builtin_neon_vrecpeq_v: 3643 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3644 Ops, "vrecpe"); 3645 case NEON::BI__builtin_neon_vrshrn_n_v: 3646 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3647 Ops, "vrshrn_n", 1, true); 3648 case NEON::BI__builtin_neon_vrsra_n_v: 3649 case NEON::BI__builtin_neon_vrsraq_n_v: 3650 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3651 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3652 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3653 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3654 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3655 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3656 case NEON::BI__builtin_neon_vsri_n_v: 3657 case NEON::BI__builtin_neon_vsriq_n_v: 3658 rightShift = true; 3659 case NEON::BI__builtin_neon_vsli_n_v: 3660 case NEON::BI__builtin_neon_vsliq_n_v: 3661 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3662 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3663 Ops, "vsli_n"); 3664 case NEON::BI__builtin_neon_vsra_n_v: 3665 case NEON::BI__builtin_neon_vsraq_n_v: 3666 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3667 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3668 return Builder.CreateAdd(Ops[0], Ops[1]); 3669 case NEON::BI__builtin_neon_vst1q_lane_v: 3670 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3671 // a one-element vector and avoid poor code for i64 in the backend. 3672 if (VTy->getElementType()->isIntegerTy(64)) { 3673 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3674 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3675 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3676 Ops[2] = Align; 3677 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3678 Ops[1]->getType()), Ops); 3679 } 3680 // fall through 3681 case NEON::BI__builtin_neon_vst1_lane_v: { 3682 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3683 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3684 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3685 StoreInst *St = Builder.CreateStore(Ops[1], 3686 Builder.CreateBitCast(Ops[0], Ty)); 3687 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3688 return St; 3689 } 3690 case NEON::BI__builtin_neon_vtbl1_v: 3691 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3692 Ops, "vtbl1"); 3693 case NEON::BI__builtin_neon_vtbl2_v: 3694 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3695 Ops, "vtbl2"); 3696 case NEON::BI__builtin_neon_vtbl3_v: 3697 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3698 Ops, "vtbl3"); 3699 case NEON::BI__builtin_neon_vtbl4_v: 3700 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3701 Ops, "vtbl4"); 3702 case NEON::BI__builtin_neon_vtbx1_v: 3703 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3704 Ops, "vtbx1"); 3705 case NEON::BI__builtin_neon_vtbx2_v: 3706 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3707 Ops, "vtbx2"); 3708 case NEON::BI__builtin_neon_vtbx3_v: 3709 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3710 Ops, "vtbx3"); 3711 case NEON::BI__builtin_neon_vtbx4_v: 3712 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3713 Ops, "vtbx4"); 3714 } 3715 } 3716 3717 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3718 const CallExpr *E, 3719 SmallVectorImpl<Value *> &Ops) { 3720 unsigned int Int = 0; 3721 const char *s = nullptr; 3722 3723 switch (BuiltinID) { 3724 default: 3725 return nullptr; 3726 case NEON::BI__builtin_neon_vtbl1_v: 3727 case NEON::BI__builtin_neon_vqtbl1_v: 3728 case NEON::BI__builtin_neon_vqtbl1q_v: 3729 case NEON::BI__builtin_neon_vtbl2_v: 3730 case NEON::BI__builtin_neon_vqtbl2_v: 3731 case NEON::BI__builtin_neon_vqtbl2q_v: 3732 case NEON::BI__builtin_neon_vtbl3_v: 3733 case NEON::BI__builtin_neon_vqtbl3_v: 3734 case NEON::BI__builtin_neon_vqtbl3q_v: 3735 case NEON::BI__builtin_neon_vtbl4_v: 3736 case NEON::BI__builtin_neon_vqtbl4_v: 3737 case NEON::BI__builtin_neon_vqtbl4q_v: 3738 break; 3739 case NEON::BI__builtin_neon_vtbx1_v: 3740 case NEON::BI__builtin_neon_vqtbx1_v: 3741 case NEON::BI__builtin_neon_vqtbx1q_v: 3742 case NEON::BI__builtin_neon_vtbx2_v: 3743 case NEON::BI__builtin_neon_vqtbx2_v: 3744 case NEON::BI__builtin_neon_vqtbx2q_v: 3745 case NEON::BI__builtin_neon_vtbx3_v: 3746 case NEON::BI__builtin_neon_vqtbx3_v: 3747 case NEON::BI__builtin_neon_vqtbx3q_v: 3748 case NEON::BI__builtin_neon_vtbx4_v: 3749 case NEON::BI__builtin_neon_vqtbx4_v: 3750 case NEON::BI__builtin_neon_vqtbx4q_v: 3751 break; 3752 } 3753 3754 assert(E->getNumArgs() >= 3); 3755 3756 // Get the last argument, which specifies the vector type. 3757 llvm::APSInt Result; 3758 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3759 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3760 return nullptr; 3761 3762 // Determine the type of this overloaded NEON intrinsic. 3763 NeonTypeFlags Type(Result.getZExtValue()); 3764 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3765 llvm::Type *Ty = VTy; 3766 if (!Ty) 3767 return nullptr; 3768 3769 unsigned nElts = VTy->getNumElements(); 3770 3771 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3772 3773 // AArch64 scalar builtins are not overloaded, they do not have an extra 3774 // argument that specifies the vector type, need to handle each case. 3775 SmallVector<Value *, 2> TblOps; 3776 switch (BuiltinID) { 3777 case NEON::BI__builtin_neon_vtbl1_v: { 3778 TblOps.push_back(Ops[0]); 3779 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3780 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3781 } 3782 case NEON::BI__builtin_neon_vtbl2_v: { 3783 TblOps.push_back(Ops[0]); 3784 TblOps.push_back(Ops[1]); 3785 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3786 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3787 } 3788 case NEON::BI__builtin_neon_vtbl3_v: { 3789 TblOps.push_back(Ops[0]); 3790 TblOps.push_back(Ops[1]); 3791 TblOps.push_back(Ops[2]); 3792 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3793 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3794 } 3795 case NEON::BI__builtin_neon_vtbl4_v: { 3796 TblOps.push_back(Ops[0]); 3797 TblOps.push_back(Ops[1]); 3798 TblOps.push_back(Ops[2]); 3799 TblOps.push_back(Ops[3]); 3800 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3801 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3802 } 3803 case NEON::BI__builtin_neon_vtbx1_v: { 3804 TblOps.push_back(Ops[1]); 3805 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3806 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3807 3808 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3809 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3810 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3811 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3812 3813 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3814 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3815 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3816 } 3817 case NEON::BI__builtin_neon_vtbx2_v: { 3818 TblOps.push_back(Ops[1]); 3819 TblOps.push_back(Ops[2]); 3820 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3821 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3822 } 3823 case NEON::BI__builtin_neon_vtbx3_v: { 3824 TblOps.push_back(Ops[1]); 3825 TblOps.push_back(Ops[2]); 3826 TblOps.push_back(Ops[3]); 3827 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3828 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3829 3830 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3831 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3832 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3833 TwentyFourV); 3834 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3835 3836 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3837 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3838 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3839 } 3840 case NEON::BI__builtin_neon_vtbx4_v: { 3841 TblOps.push_back(Ops[1]); 3842 TblOps.push_back(Ops[2]); 3843 TblOps.push_back(Ops[3]); 3844 TblOps.push_back(Ops[4]); 3845 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3846 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3847 } 3848 case NEON::BI__builtin_neon_vqtbl1_v: 3849 case NEON::BI__builtin_neon_vqtbl1q_v: 3850 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3851 case NEON::BI__builtin_neon_vqtbl2_v: 3852 case NEON::BI__builtin_neon_vqtbl2q_v: { 3853 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3854 case NEON::BI__builtin_neon_vqtbl3_v: 3855 case NEON::BI__builtin_neon_vqtbl3q_v: 3856 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3857 case NEON::BI__builtin_neon_vqtbl4_v: 3858 case NEON::BI__builtin_neon_vqtbl4q_v: 3859 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3860 case NEON::BI__builtin_neon_vqtbx1_v: 3861 case NEON::BI__builtin_neon_vqtbx1q_v: 3862 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3863 case NEON::BI__builtin_neon_vqtbx2_v: 3864 case NEON::BI__builtin_neon_vqtbx2q_v: 3865 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3866 case NEON::BI__builtin_neon_vqtbx3_v: 3867 case NEON::BI__builtin_neon_vqtbx3q_v: 3868 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3869 case NEON::BI__builtin_neon_vqtbx4_v: 3870 case NEON::BI__builtin_neon_vqtbx4q_v: 3871 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3872 } 3873 } 3874 3875 if (!Int) 3876 return nullptr; 3877 3878 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3879 return CGF.EmitNeonCall(F, Ops, s); 3880 } 3881 3882 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3883 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3884 Op = Builder.CreateBitCast(Op, Int16Ty); 3885 Value *V = UndefValue::get(VTy); 3886 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3887 Op = Builder.CreateInsertElement(V, Op, CI); 3888 return Op; 3889 } 3890 3891 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3892 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3893 Op = Builder.CreateBitCast(Op, Int8Ty); 3894 Value *V = UndefValue::get(VTy); 3895 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3896 Op = Builder.CreateInsertElement(V, Op, CI); 3897 return Op; 3898 } 3899 3900 Value *CodeGenFunction:: 3901 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3902 const char *Name) { 3903 // i8 is not a legal types for AArch64, so we can't just use 3904 // a normal overloaded intrinsic call for these scalar types. Instead 3905 // we'll build 64-bit vectors w/ lane zero being our input values and 3906 // perform the operation on that. The back end can pattern match directly 3907 // to the scalar instruction. 3908 Ops[0] = vectorWrapScalar8(Ops[0]); 3909 Ops[1] = vectorWrapScalar8(Ops[1]); 3910 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3911 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3912 Constant *CI = ConstantInt::get(SizeTy, 0); 3913 return Builder.CreateExtractElement(V, CI, "lane0"); 3914 } 3915 3916 Value *CodeGenFunction:: 3917 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3918 const char *Name) { 3919 // i16 is not a legal types for AArch64, so we can't just use 3920 // a normal overloaded intrinsic call for these scalar types. Instead 3921 // we'll build 64-bit vectors w/ lane zero being our input values and 3922 // perform the operation on that. The back end can pattern match directly 3923 // to the scalar instruction. 3924 Ops[0] = vectorWrapScalar16(Ops[0]); 3925 Ops[1] = vectorWrapScalar16(Ops[1]); 3926 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3927 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3928 Constant *CI = ConstantInt::get(SizeTy, 0); 3929 return Builder.CreateExtractElement(V, CI, "lane0"); 3930 } 3931 3932 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3933 const CallExpr *E) { 3934 unsigned HintID = static_cast<unsigned>(-1); 3935 switch (BuiltinID) { 3936 default: break; 3937 case AArch64::BI__builtin_arm_nop: 3938 HintID = 0; 3939 break; 3940 case AArch64::BI__builtin_arm_yield: 3941 HintID = 1; 3942 break; 3943 case AArch64::BI__builtin_arm_wfe: 3944 HintID = 2; 3945 break; 3946 case AArch64::BI__builtin_arm_wfi: 3947 HintID = 3; 3948 break; 3949 case AArch64::BI__builtin_arm_sev: 3950 HintID = 4; 3951 break; 3952 case AArch64::BI__builtin_arm_sevl: 3953 HintID = 5; 3954 break; 3955 } 3956 3957 if (HintID != static_cast<unsigned>(-1)) { 3958 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 3959 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3960 } 3961 3962 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 3963 Value *Address = EmitScalarExpr(E->getArg(0)); 3964 Value *RW = EmitScalarExpr(E->getArg(1)); 3965 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 3966 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 3967 Value *IsData = EmitScalarExpr(E->getArg(4)); 3968 3969 Value *Locality = nullptr; 3970 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 3971 // Temporal fetch, needs to convert cache level to locality. 3972 Locality = llvm::ConstantInt::get(Int32Ty, 3973 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 3974 } else { 3975 // Streaming fetch. 3976 Locality = llvm::ConstantInt::get(Int32Ty, 0); 3977 } 3978 3979 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 3980 // PLDL3STRM or PLDL2STRM. 3981 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3982 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3983 } 3984 3985 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 3986 assert((getContext().getTypeSize(E->getType()) == 32) && 3987 "rbit of unusual size!"); 3988 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3989 return Builder.CreateCall( 3990 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3991 } 3992 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 3993 assert((getContext().getTypeSize(E->getType()) == 64) && 3994 "rbit of unusual size!"); 3995 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3996 return Builder.CreateCall( 3997 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3998 } 3999 4000 if (BuiltinID == AArch64::BI__clear_cache) { 4001 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4002 const FunctionDecl *FD = E->getDirectCallee(); 4003 SmallVector<Value*, 2> Ops; 4004 for (unsigned i = 0; i < 2; i++) 4005 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4006 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4007 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4008 StringRef Name = FD->getName(); 4009 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4010 } 4011 4012 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4013 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4014 getContext().getTypeSize(E->getType()) == 128) { 4015 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4016 ? Intrinsic::aarch64_ldaxp 4017 : Intrinsic::aarch64_ldxp); 4018 4019 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4020 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4021 "ldxp"); 4022 4023 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4024 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4025 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4026 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4027 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4028 4029 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4030 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4031 Val = Builder.CreateOr(Val, Val1); 4032 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4033 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4034 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4035 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4036 4037 QualType Ty = E->getType(); 4038 llvm::Type *RealResTy = ConvertType(Ty); 4039 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4040 getContext().getTypeSize(Ty)); 4041 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4042 4043 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4044 ? Intrinsic::aarch64_ldaxr 4045 : Intrinsic::aarch64_ldxr, 4046 LoadAddr->getType()); 4047 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4048 4049 if (RealResTy->isPointerTy()) 4050 return Builder.CreateIntToPtr(Val, RealResTy); 4051 4052 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4053 return Builder.CreateBitCast(Val, RealResTy); 4054 } 4055 4056 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4057 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4058 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4059 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4060 ? Intrinsic::aarch64_stlxp 4061 : Intrinsic::aarch64_stxp); 4062 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4063 4064 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4065 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4066 One); 4067 Value *Val = EmitScalarExpr(E->getArg(0)); 4068 Builder.CreateStore(Val, Tmp); 4069 4070 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4071 Val = Builder.CreateLoad(LdPtr); 4072 4073 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4074 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4075 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4076 Int8PtrTy); 4077 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 4078 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 4079 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4080 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4081 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4082 4083 QualType Ty = E->getArg(0)->getType(); 4084 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4085 getContext().getTypeSize(Ty)); 4086 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4087 4088 if (StoreVal->getType()->isPointerTy()) 4089 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4090 else { 4091 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4092 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4093 } 4094 4095 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4096 ? Intrinsic::aarch64_stlxr 4097 : Intrinsic::aarch64_stxr, 4098 StoreAddr->getType()); 4099 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4100 } 4101 4102 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4103 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4104 return Builder.CreateCall(F); 4105 } 4106 4107 // CRC32 4108 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4109 switch (BuiltinID) { 4110 case AArch64::BI__builtin_arm_crc32b: 4111 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4112 case AArch64::BI__builtin_arm_crc32cb: 4113 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4114 case AArch64::BI__builtin_arm_crc32h: 4115 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4116 case AArch64::BI__builtin_arm_crc32ch: 4117 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4118 case AArch64::BI__builtin_arm_crc32w: 4119 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4120 case AArch64::BI__builtin_arm_crc32cw: 4121 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4122 case AArch64::BI__builtin_arm_crc32d: 4123 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4124 case AArch64::BI__builtin_arm_crc32cd: 4125 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4126 } 4127 4128 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4129 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4130 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4131 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4132 4133 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4134 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4135 4136 return Builder.CreateCall2(F, Arg0, Arg1); 4137 } 4138 4139 llvm::SmallVector<Value*, 4> Ops; 4140 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4141 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4142 4143 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4144 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4145 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4146 4147 if (Builtin) { 4148 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4149 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4150 assert(Result && "SISD intrinsic should have been handled"); 4151 return Result; 4152 } 4153 4154 llvm::APSInt Result; 4155 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4156 NeonTypeFlags Type(0); 4157 if (Arg->isIntegerConstantExpr(Result, getContext())) 4158 // Determine the type of this overloaded NEON intrinsic. 4159 Type = NeonTypeFlags(Result.getZExtValue()); 4160 4161 bool usgn = Type.isUnsigned(); 4162 bool quad = Type.isQuad(); 4163 4164 // Handle non-overloaded intrinsics first. 4165 switch (BuiltinID) { 4166 default: break; 4167 case NEON::BI__builtin_neon_vldrq_p128: { 4168 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4169 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4170 return Builder.CreateLoad(Ptr); 4171 } 4172 case NEON::BI__builtin_neon_vstrq_p128: { 4173 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4174 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4175 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4176 } 4177 case NEON::BI__builtin_neon_vcvts_u32_f32: 4178 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4179 usgn = true; 4180 // FALL THROUGH 4181 case NEON::BI__builtin_neon_vcvts_s32_f32: 4182 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4183 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4184 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4185 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4186 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4187 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4188 if (usgn) 4189 return Builder.CreateFPToUI(Ops[0], InTy); 4190 return Builder.CreateFPToSI(Ops[0], InTy); 4191 } 4192 case NEON::BI__builtin_neon_vcvts_f32_u32: 4193 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4194 usgn = true; 4195 // FALL THROUGH 4196 case NEON::BI__builtin_neon_vcvts_f32_s32: 4197 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4198 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4199 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4200 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4201 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4202 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4203 if (usgn) 4204 return Builder.CreateUIToFP(Ops[0], FTy); 4205 return Builder.CreateSIToFP(Ops[0], FTy); 4206 } 4207 case NEON::BI__builtin_neon_vpaddd_s64: { 4208 llvm::Type *Ty = 4209 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4210 Value *Vec = EmitScalarExpr(E->getArg(0)); 4211 // The vector is v2f64, so make sure it's bitcast to that. 4212 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4213 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4214 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4215 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4216 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4217 // Pairwise addition of a v2f64 into a scalar f64. 4218 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4219 } 4220 case NEON::BI__builtin_neon_vpaddd_f64: { 4221 llvm::Type *Ty = 4222 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4223 Value *Vec = EmitScalarExpr(E->getArg(0)); 4224 // The vector is v2f64, so make sure it's bitcast to that. 4225 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4226 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4227 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4228 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4229 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4230 // Pairwise addition of a v2f64 into a scalar f64. 4231 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4232 } 4233 case NEON::BI__builtin_neon_vpadds_f32: { 4234 llvm::Type *Ty = 4235 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4236 Value *Vec = EmitScalarExpr(E->getArg(0)); 4237 // The vector is v2f32, so make sure it's bitcast to that. 4238 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4239 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4240 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4241 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4242 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4243 // Pairwise addition of a v2f32 into a scalar f32. 4244 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4245 } 4246 case NEON::BI__builtin_neon_vceqzd_s64: 4247 case NEON::BI__builtin_neon_vceqzd_f64: 4248 case NEON::BI__builtin_neon_vceqzs_f32: 4249 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4250 return EmitAArch64CompareBuiltinExpr( 4251 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4252 ICmpInst::ICMP_EQ, "vceqz"); 4253 case NEON::BI__builtin_neon_vcgezd_s64: 4254 case NEON::BI__builtin_neon_vcgezd_f64: 4255 case NEON::BI__builtin_neon_vcgezs_f32: 4256 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4257 return EmitAArch64CompareBuiltinExpr( 4258 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4259 ICmpInst::ICMP_SGE, "vcgez"); 4260 case NEON::BI__builtin_neon_vclezd_s64: 4261 case NEON::BI__builtin_neon_vclezd_f64: 4262 case NEON::BI__builtin_neon_vclezs_f32: 4263 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4264 return EmitAArch64CompareBuiltinExpr( 4265 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4266 ICmpInst::ICMP_SLE, "vclez"); 4267 case NEON::BI__builtin_neon_vcgtzd_s64: 4268 case NEON::BI__builtin_neon_vcgtzd_f64: 4269 case NEON::BI__builtin_neon_vcgtzs_f32: 4270 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4271 return EmitAArch64CompareBuiltinExpr( 4272 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4273 ICmpInst::ICMP_SGT, "vcgtz"); 4274 case NEON::BI__builtin_neon_vcltzd_s64: 4275 case NEON::BI__builtin_neon_vcltzd_f64: 4276 case NEON::BI__builtin_neon_vcltzs_f32: 4277 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4278 return EmitAArch64CompareBuiltinExpr( 4279 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4280 ICmpInst::ICMP_SLT, "vcltz"); 4281 4282 case NEON::BI__builtin_neon_vceqzd_u64: { 4283 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4284 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4285 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4286 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4287 llvm::Constant::getNullValue(Ty)); 4288 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4289 } 4290 case NEON::BI__builtin_neon_vceqd_f64: 4291 case NEON::BI__builtin_neon_vcled_f64: 4292 case NEON::BI__builtin_neon_vcltd_f64: 4293 case NEON::BI__builtin_neon_vcged_f64: 4294 case NEON::BI__builtin_neon_vcgtd_f64: { 4295 llvm::CmpInst::Predicate P; 4296 switch (BuiltinID) { 4297 default: llvm_unreachable("missing builtin ID in switch!"); 4298 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4299 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4300 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4301 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4302 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4303 } 4304 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4305 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4306 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4307 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4308 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4309 } 4310 case NEON::BI__builtin_neon_vceqs_f32: 4311 case NEON::BI__builtin_neon_vcles_f32: 4312 case NEON::BI__builtin_neon_vclts_f32: 4313 case NEON::BI__builtin_neon_vcges_f32: 4314 case NEON::BI__builtin_neon_vcgts_f32: { 4315 llvm::CmpInst::Predicate P; 4316 switch (BuiltinID) { 4317 default: llvm_unreachable("missing builtin ID in switch!"); 4318 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4319 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4320 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4321 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4322 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4323 } 4324 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4325 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4326 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4327 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4328 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4329 } 4330 case NEON::BI__builtin_neon_vceqd_s64: 4331 case NEON::BI__builtin_neon_vceqd_u64: 4332 case NEON::BI__builtin_neon_vcgtd_s64: 4333 case NEON::BI__builtin_neon_vcgtd_u64: 4334 case NEON::BI__builtin_neon_vcltd_s64: 4335 case NEON::BI__builtin_neon_vcltd_u64: 4336 case NEON::BI__builtin_neon_vcged_u64: 4337 case NEON::BI__builtin_neon_vcged_s64: 4338 case NEON::BI__builtin_neon_vcled_u64: 4339 case NEON::BI__builtin_neon_vcled_s64: { 4340 llvm::CmpInst::Predicate P; 4341 switch (BuiltinID) { 4342 default: llvm_unreachable("missing builtin ID in switch!"); 4343 case NEON::BI__builtin_neon_vceqd_s64: 4344 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4345 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4346 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4347 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4348 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4349 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4350 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4351 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4352 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4353 } 4354 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4355 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4356 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4357 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4358 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4359 } 4360 case NEON::BI__builtin_neon_vtstd_s64: 4361 case NEON::BI__builtin_neon_vtstd_u64: { 4362 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4363 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4364 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4365 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4366 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4367 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4368 llvm::Constant::getNullValue(Ty)); 4369 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4370 } 4371 case NEON::BI__builtin_neon_vset_lane_i8: 4372 case NEON::BI__builtin_neon_vset_lane_i16: 4373 case NEON::BI__builtin_neon_vset_lane_i32: 4374 case NEON::BI__builtin_neon_vset_lane_i64: 4375 case NEON::BI__builtin_neon_vset_lane_f32: 4376 case NEON::BI__builtin_neon_vsetq_lane_i8: 4377 case NEON::BI__builtin_neon_vsetq_lane_i16: 4378 case NEON::BI__builtin_neon_vsetq_lane_i32: 4379 case NEON::BI__builtin_neon_vsetq_lane_i64: 4380 case NEON::BI__builtin_neon_vsetq_lane_f32: 4381 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4382 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4383 case NEON::BI__builtin_neon_vset_lane_f64: 4384 // The vector type needs a cast for the v1f64 variant. 4385 Ops[1] = Builder.CreateBitCast(Ops[1], 4386 llvm::VectorType::get(DoubleTy, 1)); 4387 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4388 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4389 case NEON::BI__builtin_neon_vsetq_lane_f64: 4390 // The vector type needs a cast for the v2f64 variant. 4391 Ops[1] = Builder.CreateBitCast(Ops[1], 4392 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4393 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4394 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4395 4396 case NEON::BI__builtin_neon_vget_lane_i8: 4397 case NEON::BI__builtin_neon_vdupb_lane_i8: 4398 Ops[0] = Builder.CreateBitCast(Ops[0], 4399 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4400 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4401 "vget_lane"); 4402 case NEON::BI__builtin_neon_vgetq_lane_i8: 4403 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4404 Ops[0] = Builder.CreateBitCast(Ops[0], 4405 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4406 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4407 "vgetq_lane"); 4408 case NEON::BI__builtin_neon_vget_lane_i16: 4409 case NEON::BI__builtin_neon_vduph_lane_i16: 4410 Ops[0] = Builder.CreateBitCast(Ops[0], 4411 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4412 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4413 "vget_lane"); 4414 case NEON::BI__builtin_neon_vgetq_lane_i16: 4415 case NEON::BI__builtin_neon_vduph_laneq_i16: 4416 Ops[0] = Builder.CreateBitCast(Ops[0], 4417 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4418 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4419 "vgetq_lane"); 4420 case NEON::BI__builtin_neon_vget_lane_i32: 4421 case NEON::BI__builtin_neon_vdups_lane_i32: 4422 Ops[0] = Builder.CreateBitCast( 4423 Ops[0], 4424 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4425 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4426 "vget_lane"); 4427 case NEON::BI__builtin_neon_vdups_lane_f32: 4428 Ops[0] = Builder.CreateBitCast(Ops[0], 4429 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4430 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4431 "vdups_lane"); 4432 case NEON::BI__builtin_neon_vgetq_lane_i32: 4433 case NEON::BI__builtin_neon_vdups_laneq_i32: 4434 Ops[0] = Builder.CreateBitCast(Ops[0], 4435 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4436 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4437 "vgetq_lane"); 4438 case NEON::BI__builtin_neon_vget_lane_i64: 4439 case NEON::BI__builtin_neon_vdupd_lane_i64: 4440 Ops[0] = Builder.CreateBitCast(Ops[0], 4441 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4442 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4443 "vget_lane"); 4444 case NEON::BI__builtin_neon_vdupd_lane_f64: 4445 Ops[0] = Builder.CreateBitCast(Ops[0], 4446 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4447 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4448 "vdupd_lane"); 4449 case NEON::BI__builtin_neon_vgetq_lane_i64: 4450 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4451 Ops[0] = Builder.CreateBitCast(Ops[0], 4452 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4453 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4454 "vgetq_lane"); 4455 case NEON::BI__builtin_neon_vget_lane_f32: 4456 Ops[0] = Builder.CreateBitCast(Ops[0], 4457 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4458 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4459 "vget_lane"); 4460 case NEON::BI__builtin_neon_vget_lane_f64: 4461 Ops[0] = Builder.CreateBitCast(Ops[0], 4462 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4463 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4464 "vget_lane"); 4465 case NEON::BI__builtin_neon_vgetq_lane_f32: 4466 case NEON::BI__builtin_neon_vdups_laneq_f32: 4467 Ops[0] = Builder.CreateBitCast(Ops[0], 4468 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4469 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4470 "vgetq_lane"); 4471 case NEON::BI__builtin_neon_vgetq_lane_f64: 4472 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4473 Ops[0] = Builder.CreateBitCast(Ops[0], 4474 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4475 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4476 "vgetq_lane"); 4477 case NEON::BI__builtin_neon_vaddd_s64: 4478 case NEON::BI__builtin_neon_vaddd_u64: 4479 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4480 case NEON::BI__builtin_neon_vsubd_s64: 4481 case NEON::BI__builtin_neon_vsubd_u64: 4482 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4483 case NEON::BI__builtin_neon_vqdmlalh_s16: 4484 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4485 SmallVector<Value *, 2> ProductOps; 4486 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4487 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4488 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4489 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4490 ProductOps, "vqdmlXl"); 4491 Constant *CI = ConstantInt::get(SizeTy, 0); 4492 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4493 4494 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4495 ? Intrinsic::aarch64_neon_sqadd 4496 : Intrinsic::aarch64_neon_sqsub; 4497 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4498 } 4499 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4500 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4501 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4502 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4503 Ops, "vqshlu_n"); 4504 } 4505 case NEON::BI__builtin_neon_vqshld_n_u64: 4506 case NEON::BI__builtin_neon_vqshld_n_s64: { 4507 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4508 ? Intrinsic::aarch64_neon_uqshl 4509 : Intrinsic::aarch64_neon_sqshl; 4510 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4511 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4512 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4513 } 4514 case NEON::BI__builtin_neon_vrshrd_n_u64: 4515 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4516 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4517 ? Intrinsic::aarch64_neon_urshl 4518 : Intrinsic::aarch64_neon_srshl; 4519 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4520 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4521 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4522 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4523 } 4524 case NEON::BI__builtin_neon_vrsrad_n_u64: 4525 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4526 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4527 ? Intrinsic::aarch64_neon_urshl 4528 : Intrinsic::aarch64_neon_srshl; 4529 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4530 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4531 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4532 Builder.CreateSExt(Ops[2], Int64Ty)); 4533 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4534 } 4535 case NEON::BI__builtin_neon_vshld_n_s64: 4536 case NEON::BI__builtin_neon_vshld_n_u64: { 4537 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4538 return Builder.CreateShl( 4539 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4540 } 4541 case NEON::BI__builtin_neon_vshrd_n_s64: { 4542 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4543 return Builder.CreateAShr( 4544 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4545 Amt->getZExtValue())), 4546 "shrd_n"); 4547 } 4548 case NEON::BI__builtin_neon_vshrd_n_u64: { 4549 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4550 uint64_t ShiftAmt = Amt->getZExtValue(); 4551 // Right-shifting an unsigned value by its size yields 0. 4552 if (ShiftAmt == 64) 4553 return ConstantInt::get(Int64Ty, 0); 4554 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4555 "shrd_n"); 4556 } 4557 case NEON::BI__builtin_neon_vsrad_n_s64: { 4558 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4559 Ops[1] = Builder.CreateAShr( 4560 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4561 Amt->getZExtValue())), 4562 "shrd_n"); 4563 return Builder.CreateAdd(Ops[0], Ops[1]); 4564 } 4565 case NEON::BI__builtin_neon_vsrad_n_u64: { 4566 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4567 uint64_t ShiftAmt = Amt->getZExtValue(); 4568 // Right-shifting an unsigned value by its size yields 0. 4569 // As Op + 0 = Op, return Ops[0] directly. 4570 if (ShiftAmt == 64) 4571 return Ops[0]; 4572 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4573 "shrd_n"); 4574 return Builder.CreateAdd(Ops[0], Ops[1]); 4575 } 4576 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4577 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4578 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4579 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4580 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4581 "lane"); 4582 SmallVector<Value *, 2> ProductOps; 4583 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4584 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4585 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4586 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4587 ProductOps, "vqdmlXl"); 4588 Constant *CI = ConstantInt::get(SizeTy, 0); 4589 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4590 Ops.pop_back(); 4591 4592 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4593 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4594 ? Intrinsic::aarch64_neon_sqadd 4595 : Intrinsic::aarch64_neon_sqsub; 4596 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4597 } 4598 case NEON::BI__builtin_neon_vqdmlals_s32: 4599 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4600 SmallVector<Value *, 2> ProductOps; 4601 ProductOps.push_back(Ops[1]); 4602 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4603 Ops[1] = 4604 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4605 ProductOps, "vqdmlXl"); 4606 4607 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4608 ? Intrinsic::aarch64_neon_sqadd 4609 : Intrinsic::aarch64_neon_sqsub; 4610 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4611 } 4612 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4613 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4614 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4615 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4616 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4617 "lane"); 4618 SmallVector<Value *, 2> ProductOps; 4619 ProductOps.push_back(Ops[1]); 4620 ProductOps.push_back(Ops[2]); 4621 Ops[1] = 4622 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4623 ProductOps, "vqdmlXl"); 4624 Ops.pop_back(); 4625 4626 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4627 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4628 ? Intrinsic::aarch64_neon_sqadd 4629 : Intrinsic::aarch64_neon_sqsub; 4630 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4631 } 4632 } 4633 4634 llvm::VectorType *VTy = GetNeonType(this, Type); 4635 llvm::Type *Ty = VTy; 4636 if (!Ty) 4637 return nullptr; 4638 4639 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4640 // defer to common code if it's been added to our special map. 4641 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4642 AArch64SIMDIntrinsicsProvenSorted); 4643 4644 if (Builtin) 4645 return EmitCommonNeonBuiltinExpr( 4646 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4647 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4648 4649 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4650 return V; 4651 4652 unsigned Int; 4653 switch (BuiltinID) { 4654 default: return nullptr; 4655 case NEON::BI__builtin_neon_vbsl_v: 4656 case NEON::BI__builtin_neon_vbslq_v: { 4657 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4658 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4659 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4660 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4661 4662 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4663 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4664 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4665 return Builder.CreateBitCast(Ops[0], Ty); 4666 } 4667 case NEON::BI__builtin_neon_vfma_lane_v: 4668 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4669 // The ARM builtins (and instructions) have the addend as the first 4670 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4671 Value *Addend = Ops[0]; 4672 Value *Multiplicand = Ops[1]; 4673 Value *LaneSource = Ops[2]; 4674 Ops[0] = Multiplicand; 4675 Ops[1] = LaneSource; 4676 Ops[2] = Addend; 4677 4678 // Now adjust things to handle the lane access. 4679 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4680 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4681 VTy; 4682 llvm::Constant *cst = cast<Constant>(Ops[3]); 4683 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4684 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4685 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4686 4687 Ops.pop_back(); 4688 Int = Intrinsic::fma; 4689 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4690 } 4691 case NEON::BI__builtin_neon_vfma_laneq_v: { 4692 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4693 // v1f64 fma should be mapped to Neon scalar f64 fma 4694 if (VTy && VTy->getElementType() == DoubleTy) { 4695 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4696 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4697 llvm::Type *VTy = GetNeonType(this, 4698 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4699 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4700 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4701 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4702 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4703 return Builder.CreateBitCast(Result, Ty); 4704 } 4705 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4706 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4707 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4708 4709 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4710 VTy->getNumElements() * 2); 4711 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4712 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4713 cast<ConstantInt>(Ops[3])); 4714 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4715 4716 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4717 } 4718 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4719 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4720 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4721 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4722 4723 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4724 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4725 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4726 } 4727 case NEON::BI__builtin_neon_vfmas_lane_f32: 4728 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4729 case NEON::BI__builtin_neon_vfmad_lane_f64: 4730 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4731 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4732 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4733 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4734 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4735 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4736 } 4737 case NEON::BI__builtin_neon_vfms_v: 4738 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4739 // FIXME: probably remove when we no longer support aarch64_simd.h 4740 // (arm_neon.h delegates to vfma). 4741 4742 // The ARM builtins (and instructions) have the addend as the first 4743 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4744 Value *Subtrahend = Ops[0]; 4745 Value *Multiplicand = Ops[2]; 4746 Ops[0] = Multiplicand; 4747 Ops[2] = Subtrahend; 4748 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4749 Ops[1] = Builder.CreateFNeg(Ops[1]); 4750 Int = Intrinsic::fma; 4751 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4752 } 4753 case NEON::BI__builtin_neon_vmull_v: 4754 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4755 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4756 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4757 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4758 case NEON::BI__builtin_neon_vmax_v: 4759 case NEON::BI__builtin_neon_vmaxq_v: 4760 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4761 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4762 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4763 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4764 case NEON::BI__builtin_neon_vmin_v: 4765 case NEON::BI__builtin_neon_vminq_v: 4766 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4767 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4768 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4769 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4770 case NEON::BI__builtin_neon_vabd_v: 4771 case NEON::BI__builtin_neon_vabdq_v: 4772 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4773 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4774 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4775 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4776 case NEON::BI__builtin_neon_vpadal_v: 4777 case NEON::BI__builtin_neon_vpadalq_v: { 4778 unsigned ArgElts = VTy->getNumElements(); 4779 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4780 unsigned BitWidth = EltTy->getBitWidth(); 4781 llvm::Type *ArgTy = llvm::VectorType::get( 4782 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4783 llvm::Type* Tys[2] = { VTy, ArgTy }; 4784 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4785 SmallVector<llvm::Value*, 1> TmpOps; 4786 TmpOps.push_back(Ops[1]); 4787 Function *F = CGM.getIntrinsic(Int, Tys); 4788 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4789 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4790 return Builder.CreateAdd(tmp, addend); 4791 } 4792 case NEON::BI__builtin_neon_vpmin_v: 4793 case NEON::BI__builtin_neon_vpminq_v: 4794 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4795 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4796 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4797 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4798 case NEON::BI__builtin_neon_vpmax_v: 4799 case NEON::BI__builtin_neon_vpmaxq_v: 4800 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4801 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4802 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4803 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4804 case NEON::BI__builtin_neon_vminnm_v: 4805 case NEON::BI__builtin_neon_vminnmq_v: 4806 Int = Intrinsic::aarch64_neon_fminnm; 4807 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4808 case NEON::BI__builtin_neon_vmaxnm_v: 4809 case NEON::BI__builtin_neon_vmaxnmq_v: 4810 Int = Intrinsic::aarch64_neon_fmaxnm; 4811 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4812 case NEON::BI__builtin_neon_vrecpss_f32: { 4813 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4814 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4815 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4816 Ops, "vrecps"); 4817 } 4818 case NEON::BI__builtin_neon_vrecpsd_f64: { 4819 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4820 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4821 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4822 Ops, "vrecps"); 4823 } 4824 case NEON::BI__builtin_neon_vqshrun_n_v: 4825 Int = Intrinsic::aarch64_neon_sqshrun; 4826 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4827 case NEON::BI__builtin_neon_vqrshrun_n_v: 4828 Int = Intrinsic::aarch64_neon_sqrshrun; 4829 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4830 case NEON::BI__builtin_neon_vqshrn_n_v: 4831 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4832 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4833 case NEON::BI__builtin_neon_vrshrn_n_v: 4834 Int = Intrinsic::aarch64_neon_rshrn; 4835 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4836 case NEON::BI__builtin_neon_vqrshrn_n_v: 4837 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4838 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4839 case NEON::BI__builtin_neon_vrnda_v: 4840 case NEON::BI__builtin_neon_vrndaq_v: { 4841 Int = Intrinsic::round; 4842 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4843 } 4844 case NEON::BI__builtin_neon_vrndi_v: 4845 case NEON::BI__builtin_neon_vrndiq_v: { 4846 Int = Intrinsic::nearbyint; 4847 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4848 } 4849 case NEON::BI__builtin_neon_vrndm_v: 4850 case NEON::BI__builtin_neon_vrndmq_v: { 4851 Int = Intrinsic::floor; 4852 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4853 } 4854 case NEON::BI__builtin_neon_vrndn_v: 4855 case NEON::BI__builtin_neon_vrndnq_v: { 4856 Int = Intrinsic::aarch64_neon_frintn; 4857 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4858 } 4859 case NEON::BI__builtin_neon_vrndp_v: 4860 case NEON::BI__builtin_neon_vrndpq_v: { 4861 Int = Intrinsic::ceil; 4862 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4863 } 4864 case NEON::BI__builtin_neon_vrndx_v: 4865 case NEON::BI__builtin_neon_vrndxq_v: { 4866 Int = Intrinsic::rint; 4867 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4868 } 4869 case NEON::BI__builtin_neon_vrnd_v: 4870 case NEON::BI__builtin_neon_vrndq_v: { 4871 Int = Intrinsic::trunc; 4872 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4873 } 4874 case NEON::BI__builtin_neon_vceqz_v: 4875 case NEON::BI__builtin_neon_vceqzq_v: 4876 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4877 ICmpInst::ICMP_EQ, "vceqz"); 4878 case NEON::BI__builtin_neon_vcgez_v: 4879 case NEON::BI__builtin_neon_vcgezq_v: 4880 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4881 ICmpInst::ICMP_SGE, "vcgez"); 4882 case NEON::BI__builtin_neon_vclez_v: 4883 case NEON::BI__builtin_neon_vclezq_v: 4884 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4885 ICmpInst::ICMP_SLE, "vclez"); 4886 case NEON::BI__builtin_neon_vcgtz_v: 4887 case NEON::BI__builtin_neon_vcgtzq_v: 4888 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4889 ICmpInst::ICMP_SGT, "vcgtz"); 4890 case NEON::BI__builtin_neon_vcltz_v: 4891 case NEON::BI__builtin_neon_vcltzq_v: 4892 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4893 ICmpInst::ICMP_SLT, "vcltz"); 4894 case NEON::BI__builtin_neon_vcvt_f64_v: 4895 case NEON::BI__builtin_neon_vcvtq_f64_v: 4896 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4897 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4898 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4899 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4900 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4901 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4902 "unexpected vcvt_f64_f32 builtin"); 4903 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4904 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4905 4906 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4907 } 4908 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4909 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4910 "unexpected vcvt_f32_f64 builtin"); 4911 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4912 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4913 4914 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4915 } 4916 case NEON::BI__builtin_neon_vcvt_s32_v: 4917 case NEON::BI__builtin_neon_vcvt_u32_v: 4918 case NEON::BI__builtin_neon_vcvt_s64_v: 4919 case NEON::BI__builtin_neon_vcvt_u64_v: 4920 case NEON::BI__builtin_neon_vcvtq_s32_v: 4921 case NEON::BI__builtin_neon_vcvtq_u32_v: 4922 case NEON::BI__builtin_neon_vcvtq_s64_v: 4923 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4924 bool Double = 4925 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4926 llvm::Type *InTy = 4927 GetNeonType(this, 4928 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4929 : NeonTypeFlags::Float32, false, quad)); 4930 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4931 if (usgn) 4932 return Builder.CreateFPToUI(Ops[0], Ty); 4933 return Builder.CreateFPToSI(Ops[0], Ty); 4934 } 4935 case NEON::BI__builtin_neon_vcvta_s32_v: 4936 case NEON::BI__builtin_neon_vcvtaq_s32_v: 4937 case NEON::BI__builtin_neon_vcvta_u32_v: 4938 case NEON::BI__builtin_neon_vcvtaq_u32_v: 4939 case NEON::BI__builtin_neon_vcvta_s64_v: 4940 case NEON::BI__builtin_neon_vcvtaq_s64_v: 4941 case NEON::BI__builtin_neon_vcvta_u64_v: 4942 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4943 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 4944 bool Double = 4945 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4946 llvm::Type *InTy = 4947 GetNeonType(this, 4948 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4949 : NeonTypeFlags::Float32, false, quad)); 4950 llvm::Type *Tys[2] = { Ty, InTy }; 4951 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 4952 } 4953 case NEON::BI__builtin_neon_vcvtm_s32_v: 4954 case NEON::BI__builtin_neon_vcvtmq_s32_v: 4955 case NEON::BI__builtin_neon_vcvtm_u32_v: 4956 case NEON::BI__builtin_neon_vcvtmq_u32_v: 4957 case NEON::BI__builtin_neon_vcvtm_s64_v: 4958 case NEON::BI__builtin_neon_vcvtmq_s64_v: 4959 case NEON::BI__builtin_neon_vcvtm_u64_v: 4960 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4961 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 4962 bool Double = 4963 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4964 llvm::Type *InTy = 4965 GetNeonType(this, 4966 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4967 : NeonTypeFlags::Float32, false, quad)); 4968 llvm::Type *Tys[2] = { Ty, InTy }; 4969 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 4970 } 4971 case NEON::BI__builtin_neon_vcvtn_s32_v: 4972 case NEON::BI__builtin_neon_vcvtnq_s32_v: 4973 case NEON::BI__builtin_neon_vcvtn_u32_v: 4974 case NEON::BI__builtin_neon_vcvtnq_u32_v: 4975 case NEON::BI__builtin_neon_vcvtn_s64_v: 4976 case NEON::BI__builtin_neon_vcvtnq_s64_v: 4977 case NEON::BI__builtin_neon_vcvtn_u64_v: 4978 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4979 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 4980 bool Double = 4981 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4982 llvm::Type *InTy = 4983 GetNeonType(this, 4984 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4985 : NeonTypeFlags::Float32, false, quad)); 4986 llvm::Type *Tys[2] = { Ty, InTy }; 4987 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 4988 } 4989 case NEON::BI__builtin_neon_vcvtp_s32_v: 4990 case NEON::BI__builtin_neon_vcvtpq_s32_v: 4991 case NEON::BI__builtin_neon_vcvtp_u32_v: 4992 case NEON::BI__builtin_neon_vcvtpq_u32_v: 4993 case NEON::BI__builtin_neon_vcvtp_s64_v: 4994 case NEON::BI__builtin_neon_vcvtpq_s64_v: 4995 case NEON::BI__builtin_neon_vcvtp_u64_v: 4996 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4997 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 4998 bool Double = 4999 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5000 llvm::Type *InTy = 5001 GetNeonType(this, 5002 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5003 : NeonTypeFlags::Float32, false, quad)); 5004 llvm::Type *Tys[2] = { Ty, InTy }; 5005 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5006 } 5007 case NEON::BI__builtin_neon_vmulx_v: 5008 case NEON::BI__builtin_neon_vmulxq_v: { 5009 Int = Intrinsic::aarch64_neon_fmulx; 5010 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5011 } 5012 case NEON::BI__builtin_neon_vmul_lane_v: 5013 case NEON::BI__builtin_neon_vmul_laneq_v: { 5014 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5015 bool Quad = false; 5016 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5017 Quad = true; 5018 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5019 llvm::Type *VTy = GetNeonType(this, 5020 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5021 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5022 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5023 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5024 return Builder.CreateBitCast(Result, Ty); 5025 } 5026 case NEON::BI__builtin_neon_vnegd_s64: 5027 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5028 case NEON::BI__builtin_neon_vpmaxnm_v: 5029 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5030 Int = Intrinsic::aarch64_neon_fmaxnmp; 5031 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5032 } 5033 case NEON::BI__builtin_neon_vpminnm_v: 5034 case NEON::BI__builtin_neon_vpminnmq_v: { 5035 Int = Intrinsic::aarch64_neon_fminnmp; 5036 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5037 } 5038 case NEON::BI__builtin_neon_vsqrt_v: 5039 case NEON::BI__builtin_neon_vsqrtq_v: { 5040 Int = Intrinsic::sqrt; 5041 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5042 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5043 } 5044 case NEON::BI__builtin_neon_vrbit_v: 5045 case NEON::BI__builtin_neon_vrbitq_v: { 5046 Int = Intrinsic::aarch64_neon_rbit; 5047 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5048 } 5049 case NEON::BI__builtin_neon_vaddv_u8: 5050 // FIXME: These are handled by the AArch64 scalar code. 5051 usgn = true; 5052 // FALLTHROUGH 5053 case NEON::BI__builtin_neon_vaddv_s8: { 5054 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5055 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5056 VTy = 5057 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5058 llvm::Type *Tys[2] = { Ty, VTy }; 5059 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5060 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5061 return Builder.CreateTrunc(Ops[0], 5062 llvm::IntegerType::get(getLLVMContext(), 8)); 5063 } 5064 case NEON::BI__builtin_neon_vaddv_u16: 5065 usgn = true; 5066 // FALLTHROUGH 5067 case NEON::BI__builtin_neon_vaddv_s16: { 5068 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5069 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5070 VTy = 5071 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5072 llvm::Type *Tys[2] = { Ty, VTy }; 5073 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5074 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5075 return Builder.CreateTrunc(Ops[0], 5076 llvm::IntegerType::get(getLLVMContext(), 16)); 5077 } 5078 case NEON::BI__builtin_neon_vaddvq_u8: 5079 usgn = true; 5080 // FALLTHROUGH 5081 case NEON::BI__builtin_neon_vaddvq_s8: { 5082 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5083 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5084 VTy = 5085 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5086 llvm::Type *Tys[2] = { Ty, VTy }; 5087 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5088 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5089 return Builder.CreateTrunc(Ops[0], 5090 llvm::IntegerType::get(getLLVMContext(), 8)); 5091 } 5092 case NEON::BI__builtin_neon_vaddvq_u16: 5093 usgn = true; 5094 // FALLTHROUGH 5095 case NEON::BI__builtin_neon_vaddvq_s16: { 5096 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5097 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5098 VTy = 5099 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5100 llvm::Type *Tys[2] = { Ty, VTy }; 5101 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5102 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5103 return Builder.CreateTrunc(Ops[0], 5104 llvm::IntegerType::get(getLLVMContext(), 16)); 5105 } 5106 case NEON::BI__builtin_neon_vmaxv_u8: { 5107 Int = Intrinsic::aarch64_neon_umaxv; 5108 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5109 VTy = 5110 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5111 llvm::Type *Tys[2] = { Ty, VTy }; 5112 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5113 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5114 return Builder.CreateTrunc(Ops[0], 5115 llvm::IntegerType::get(getLLVMContext(), 8)); 5116 } 5117 case NEON::BI__builtin_neon_vmaxv_u16: { 5118 Int = Intrinsic::aarch64_neon_umaxv; 5119 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5120 VTy = 5121 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5122 llvm::Type *Tys[2] = { Ty, VTy }; 5123 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5124 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5125 return Builder.CreateTrunc(Ops[0], 5126 llvm::IntegerType::get(getLLVMContext(), 16)); 5127 } 5128 case NEON::BI__builtin_neon_vmaxvq_u8: { 5129 Int = Intrinsic::aarch64_neon_umaxv; 5130 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5131 VTy = 5132 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5133 llvm::Type *Tys[2] = { Ty, VTy }; 5134 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5135 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5136 return Builder.CreateTrunc(Ops[0], 5137 llvm::IntegerType::get(getLLVMContext(), 8)); 5138 } 5139 case NEON::BI__builtin_neon_vmaxvq_u16: { 5140 Int = Intrinsic::aarch64_neon_umaxv; 5141 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5142 VTy = 5143 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5144 llvm::Type *Tys[2] = { Ty, VTy }; 5145 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5146 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5147 return Builder.CreateTrunc(Ops[0], 5148 llvm::IntegerType::get(getLLVMContext(), 16)); 5149 } 5150 case NEON::BI__builtin_neon_vmaxv_s8: { 5151 Int = Intrinsic::aarch64_neon_smaxv; 5152 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5153 VTy = 5154 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5155 llvm::Type *Tys[2] = { Ty, VTy }; 5156 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5157 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5158 return Builder.CreateTrunc(Ops[0], 5159 llvm::IntegerType::get(getLLVMContext(), 8)); 5160 } 5161 case NEON::BI__builtin_neon_vmaxv_s16: { 5162 Int = Intrinsic::aarch64_neon_smaxv; 5163 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5164 VTy = 5165 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5166 llvm::Type *Tys[2] = { Ty, VTy }; 5167 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5168 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5169 return Builder.CreateTrunc(Ops[0], 5170 llvm::IntegerType::get(getLLVMContext(), 16)); 5171 } 5172 case NEON::BI__builtin_neon_vmaxvq_s8: { 5173 Int = Intrinsic::aarch64_neon_smaxv; 5174 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5175 VTy = 5176 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5177 llvm::Type *Tys[2] = { Ty, VTy }; 5178 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5179 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5180 return Builder.CreateTrunc(Ops[0], 5181 llvm::IntegerType::get(getLLVMContext(), 8)); 5182 } 5183 case NEON::BI__builtin_neon_vmaxvq_s16: { 5184 Int = Intrinsic::aarch64_neon_smaxv; 5185 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5186 VTy = 5187 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5188 llvm::Type *Tys[2] = { Ty, VTy }; 5189 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5190 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5191 return Builder.CreateTrunc(Ops[0], 5192 llvm::IntegerType::get(getLLVMContext(), 16)); 5193 } 5194 case NEON::BI__builtin_neon_vminv_u8: { 5195 Int = Intrinsic::aarch64_neon_uminv; 5196 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5197 VTy = 5198 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5199 llvm::Type *Tys[2] = { Ty, VTy }; 5200 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5201 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5202 return Builder.CreateTrunc(Ops[0], 5203 llvm::IntegerType::get(getLLVMContext(), 8)); 5204 } 5205 case NEON::BI__builtin_neon_vminv_u16: { 5206 Int = Intrinsic::aarch64_neon_uminv; 5207 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5208 VTy = 5209 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5210 llvm::Type *Tys[2] = { Ty, VTy }; 5211 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5212 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5213 return Builder.CreateTrunc(Ops[0], 5214 llvm::IntegerType::get(getLLVMContext(), 16)); 5215 } 5216 case NEON::BI__builtin_neon_vminvq_u8: { 5217 Int = Intrinsic::aarch64_neon_uminv; 5218 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5219 VTy = 5220 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5221 llvm::Type *Tys[2] = { Ty, VTy }; 5222 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5223 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5224 return Builder.CreateTrunc(Ops[0], 5225 llvm::IntegerType::get(getLLVMContext(), 8)); 5226 } 5227 case NEON::BI__builtin_neon_vminvq_u16: { 5228 Int = Intrinsic::aarch64_neon_uminv; 5229 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5230 VTy = 5231 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5232 llvm::Type *Tys[2] = { Ty, VTy }; 5233 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5234 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5235 return Builder.CreateTrunc(Ops[0], 5236 llvm::IntegerType::get(getLLVMContext(), 16)); 5237 } 5238 case NEON::BI__builtin_neon_vminv_s8: { 5239 Int = Intrinsic::aarch64_neon_sminv; 5240 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5241 VTy = 5242 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5243 llvm::Type *Tys[2] = { Ty, VTy }; 5244 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5245 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5246 return Builder.CreateTrunc(Ops[0], 5247 llvm::IntegerType::get(getLLVMContext(), 8)); 5248 } 5249 case NEON::BI__builtin_neon_vminv_s16: { 5250 Int = Intrinsic::aarch64_neon_sminv; 5251 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5252 VTy = 5253 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5254 llvm::Type *Tys[2] = { Ty, VTy }; 5255 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5256 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5257 return Builder.CreateTrunc(Ops[0], 5258 llvm::IntegerType::get(getLLVMContext(), 16)); 5259 } 5260 case NEON::BI__builtin_neon_vminvq_s8: { 5261 Int = Intrinsic::aarch64_neon_sminv; 5262 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5263 VTy = 5264 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5265 llvm::Type *Tys[2] = { Ty, VTy }; 5266 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5267 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5268 return Builder.CreateTrunc(Ops[0], 5269 llvm::IntegerType::get(getLLVMContext(), 8)); 5270 } 5271 case NEON::BI__builtin_neon_vminvq_s16: { 5272 Int = Intrinsic::aarch64_neon_sminv; 5273 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5274 VTy = 5275 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5276 llvm::Type *Tys[2] = { Ty, VTy }; 5277 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5278 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5279 return Builder.CreateTrunc(Ops[0], 5280 llvm::IntegerType::get(getLLVMContext(), 16)); 5281 } 5282 case NEON::BI__builtin_neon_vmul_n_f64: { 5283 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5284 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5285 return Builder.CreateFMul(Ops[0], RHS); 5286 } 5287 case NEON::BI__builtin_neon_vaddlv_u8: { 5288 Int = Intrinsic::aarch64_neon_uaddlv; 5289 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5290 VTy = 5291 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5292 llvm::Type *Tys[2] = { Ty, VTy }; 5293 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5294 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5295 return Builder.CreateTrunc(Ops[0], 5296 llvm::IntegerType::get(getLLVMContext(), 16)); 5297 } 5298 case NEON::BI__builtin_neon_vaddlv_u16: { 5299 Int = Intrinsic::aarch64_neon_uaddlv; 5300 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5301 VTy = 5302 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5303 llvm::Type *Tys[2] = { Ty, VTy }; 5304 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5305 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5306 } 5307 case NEON::BI__builtin_neon_vaddlvq_u8: { 5308 Int = Intrinsic::aarch64_neon_uaddlv; 5309 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5310 VTy = 5311 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5312 llvm::Type *Tys[2] = { Ty, VTy }; 5313 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5314 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5315 return Builder.CreateTrunc(Ops[0], 5316 llvm::IntegerType::get(getLLVMContext(), 16)); 5317 } 5318 case NEON::BI__builtin_neon_vaddlvq_u16: { 5319 Int = Intrinsic::aarch64_neon_uaddlv; 5320 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5321 VTy = 5322 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5323 llvm::Type *Tys[2] = { Ty, VTy }; 5324 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5325 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5326 } 5327 case NEON::BI__builtin_neon_vaddlv_s8: { 5328 Int = Intrinsic::aarch64_neon_saddlv; 5329 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5330 VTy = 5331 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5332 llvm::Type *Tys[2] = { Ty, VTy }; 5333 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5334 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5335 return Builder.CreateTrunc(Ops[0], 5336 llvm::IntegerType::get(getLLVMContext(), 16)); 5337 } 5338 case NEON::BI__builtin_neon_vaddlv_s16: { 5339 Int = Intrinsic::aarch64_neon_saddlv; 5340 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5341 VTy = 5342 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5343 llvm::Type *Tys[2] = { Ty, VTy }; 5344 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5345 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5346 } 5347 case NEON::BI__builtin_neon_vaddlvq_s8: { 5348 Int = Intrinsic::aarch64_neon_saddlv; 5349 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5350 VTy = 5351 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5352 llvm::Type *Tys[2] = { Ty, VTy }; 5353 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5354 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5355 return Builder.CreateTrunc(Ops[0], 5356 llvm::IntegerType::get(getLLVMContext(), 16)); 5357 } 5358 case NEON::BI__builtin_neon_vaddlvq_s16: { 5359 Int = Intrinsic::aarch64_neon_saddlv; 5360 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5361 VTy = 5362 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5363 llvm::Type *Tys[2] = { Ty, VTy }; 5364 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5365 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5366 } 5367 case NEON::BI__builtin_neon_vsri_n_v: 5368 case NEON::BI__builtin_neon_vsriq_n_v: { 5369 Int = Intrinsic::aarch64_neon_vsri; 5370 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5371 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5372 } 5373 case NEON::BI__builtin_neon_vsli_n_v: 5374 case NEON::BI__builtin_neon_vsliq_n_v: { 5375 Int = Intrinsic::aarch64_neon_vsli; 5376 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5377 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5378 } 5379 case NEON::BI__builtin_neon_vsra_n_v: 5380 case NEON::BI__builtin_neon_vsraq_n_v: 5381 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5382 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5383 return Builder.CreateAdd(Ops[0], Ops[1]); 5384 case NEON::BI__builtin_neon_vrsra_n_v: 5385 case NEON::BI__builtin_neon_vrsraq_n_v: { 5386 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5387 SmallVector<llvm::Value*,2> TmpOps; 5388 TmpOps.push_back(Ops[1]); 5389 TmpOps.push_back(Ops[2]); 5390 Function* F = CGM.getIntrinsic(Int, Ty); 5391 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5392 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5393 return Builder.CreateAdd(Ops[0], tmp); 5394 } 5395 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5396 // of an Align parameter here. 5397 case NEON::BI__builtin_neon_vld1_x2_v: 5398 case NEON::BI__builtin_neon_vld1q_x2_v: 5399 case NEON::BI__builtin_neon_vld1_x3_v: 5400 case NEON::BI__builtin_neon_vld1q_x3_v: 5401 case NEON::BI__builtin_neon_vld1_x4_v: 5402 case NEON::BI__builtin_neon_vld1q_x4_v: { 5403 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5404 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5405 llvm::Type *Tys[2] = { VTy, PTy }; 5406 unsigned Int; 5407 switch (BuiltinID) { 5408 case NEON::BI__builtin_neon_vld1_x2_v: 5409 case NEON::BI__builtin_neon_vld1q_x2_v: 5410 Int = Intrinsic::aarch64_neon_ld1x2; 5411 break; 5412 case NEON::BI__builtin_neon_vld1_x3_v: 5413 case NEON::BI__builtin_neon_vld1q_x3_v: 5414 Int = Intrinsic::aarch64_neon_ld1x3; 5415 break; 5416 case NEON::BI__builtin_neon_vld1_x4_v: 5417 case NEON::BI__builtin_neon_vld1q_x4_v: 5418 Int = Intrinsic::aarch64_neon_ld1x4; 5419 break; 5420 } 5421 Function *F = CGM.getIntrinsic(Int, Tys); 5422 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5423 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5424 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5425 return Builder.CreateStore(Ops[1], Ops[0]); 5426 } 5427 case NEON::BI__builtin_neon_vst1_x2_v: 5428 case NEON::BI__builtin_neon_vst1q_x2_v: 5429 case NEON::BI__builtin_neon_vst1_x3_v: 5430 case NEON::BI__builtin_neon_vst1q_x3_v: 5431 case NEON::BI__builtin_neon_vst1_x4_v: 5432 case NEON::BI__builtin_neon_vst1q_x4_v: { 5433 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5434 llvm::Type *Tys[2] = { VTy, PTy }; 5435 unsigned Int; 5436 switch (BuiltinID) { 5437 case NEON::BI__builtin_neon_vst1_x2_v: 5438 case NEON::BI__builtin_neon_vst1q_x2_v: 5439 Int = Intrinsic::aarch64_neon_st1x2; 5440 break; 5441 case NEON::BI__builtin_neon_vst1_x3_v: 5442 case NEON::BI__builtin_neon_vst1q_x3_v: 5443 Int = Intrinsic::aarch64_neon_st1x3; 5444 break; 5445 case NEON::BI__builtin_neon_vst1_x4_v: 5446 case NEON::BI__builtin_neon_vst1q_x4_v: 5447 Int = Intrinsic::aarch64_neon_st1x4; 5448 break; 5449 } 5450 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5451 IntOps.push_back(Ops[0]); 5452 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5453 } 5454 case NEON::BI__builtin_neon_vld1_v: 5455 case NEON::BI__builtin_neon_vld1q_v: 5456 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5457 return Builder.CreateLoad(Ops[0]); 5458 case NEON::BI__builtin_neon_vst1_v: 5459 case NEON::BI__builtin_neon_vst1q_v: 5460 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5461 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5462 return Builder.CreateStore(Ops[1], Ops[0]); 5463 case NEON::BI__builtin_neon_vld1_lane_v: 5464 case NEON::BI__builtin_neon_vld1q_lane_v: 5465 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5466 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5467 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5468 Ops[0] = Builder.CreateLoad(Ops[0]); 5469 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5470 case NEON::BI__builtin_neon_vld1_dup_v: 5471 case NEON::BI__builtin_neon_vld1q_dup_v: { 5472 Value *V = UndefValue::get(Ty); 5473 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5474 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5475 Ops[0] = Builder.CreateLoad(Ops[0]); 5476 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5477 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5478 return EmitNeonSplat(Ops[0], CI); 5479 } 5480 case NEON::BI__builtin_neon_vst1_lane_v: 5481 case NEON::BI__builtin_neon_vst1q_lane_v: 5482 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5483 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5484 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5485 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5486 case NEON::BI__builtin_neon_vld2_v: 5487 case NEON::BI__builtin_neon_vld2q_v: { 5488 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5489 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5490 llvm::Type *Tys[2] = { VTy, PTy }; 5491 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5492 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5493 Ops[0] = Builder.CreateBitCast(Ops[0], 5494 llvm::PointerType::getUnqual(Ops[1]->getType())); 5495 return Builder.CreateStore(Ops[1], Ops[0]); 5496 } 5497 case NEON::BI__builtin_neon_vld3_v: 5498 case NEON::BI__builtin_neon_vld3q_v: { 5499 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5500 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5501 llvm::Type *Tys[2] = { VTy, PTy }; 5502 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5503 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5504 Ops[0] = Builder.CreateBitCast(Ops[0], 5505 llvm::PointerType::getUnqual(Ops[1]->getType())); 5506 return Builder.CreateStore(Ops[1], Ops[0]); 5507 } 5508 case NEON::BI__builtin_neon_vld4_v: 5509 case NEON::BI__builtin_neon_vld4q_v: { 5510 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5511 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5512 llvm::Type *Tys[2] = { VTy, PTy }; 5513 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5514 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5515 Ops[0] = Builder.CreateBitCast(Ops[0], 5516 llvm::PointerType::getUnqual(Ops[1]->getType())); 5517 return Builder.CreateStore(Ops[1], Ops[0]); 5518 } 5519 case NEON::BI__builtin_neon_vld2_dup_v: 5520 case NEON::BI__builtin_neon_vld2q_dup_v: { 5521 llvm::Type *PTy = 5522 llvm::PointerType::getUnqual(VTy->getElementType()); 5523 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5524 llvm::Type *Tys[2] = { VTy, PTy }; 5525 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5526 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5527 Ops[0] = Builder.CreateBitCast(Ops[0], 5528 llvm::PointerType::getUnqual(Ops[1]->getType())); 5529 return Builder.CreateStore(Ops[1], Ops[0]); 5530 } 5531 case NEON::BI__builtin_neon_vld3_dup_v: 5532 case NEON::BI__builtin_neon_vld3q_dup_v: { 5533 llvm::Type *PTy = 5534 llvm::PointerType::getUnqual(VTy->getElementType()); 5535 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5536 llvm::Type *Tys[2] = { VTy, PTy }; 5537 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5538 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5539 Ops[0] = Builder.CreateBitCast(Ops[0], 5540 llvm::PointerType::getUnqual(Ops[1]->getType())); 5541 return Builder.CreateStore(Ops[1], Ops[0]); 5542 } 5543 case NEON::BI__builtin_neon_vld4_dup_v: 5544 case NEON::BI__builtin_neon_vld4q_dup_v: { 5545 llvm::Type *PTy = 5546 llvm::PointerType::getUnqual(VTy->getElementType()); 5547 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5548 llvm::Type *Tys[2] = { VTy, PTy }; 5549 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5550 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5551 Ops[0] = Builder.CreateBitCast(Ops[0], 5552 llvm::PointerType::getUnqual(Ops[1]->getType())); 5553 return Builder.CreateStore(Ops[1], Ops[0]); 5554 } 5555 case NEON::BI__builtin_neon_vld2_lane_v: 5556 case NEON::BI__builtin_neon_vld2q_lane_v: { 5557 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5558 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5559 Ops.push_back(Ops[1]); 5560 Ops.erase(Ops.begin()+1); 5561 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5562 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5563 Ops[3] = Builder.CreateZExt(Ops[3], 5564 llvm::IntegerType::get(getLLVMContext(), 64)); 5565 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5566 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5567 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5568 return Builder.CreateStore(Ops[1], Ops[0]); 5569 } 5570 case NEON::BI__builtin_neon_vld3_lane_v: 5571 case NEON::BI__builtin_neon_vld3q_lane_v: { 5572 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5573 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5574 Ops.push_back(Ops[1]); 5575 Ops.erase(Ops.begin()+1); 5576 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5577 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5578 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5579 Ops[4] = Builder.CreateZExt(Ops[4], 5580 llvm::IntegerType::get(getLLVMContext(), 64)); 5581 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5582 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5583 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5584 return Builder.CreateStore(Ops[1], Ops[0]); 5585 } 5586 case NEON::BI__builtin_neon_vld4_lane_v: 5587 case NEON::BI__builtin_neon_vld4q_lane_v: { 5588 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5589 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5590 Ops.push_back(Ops[1]); 5591 Ops.erase(Ops.begin()+1); 5592 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5593 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5594 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5595 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5596 Ops[5] = Builder.CreateZExt(Ops[5], 5597 llvm::IntegerType::get(getLLVMContext(), 64)); 5598 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5599 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5600 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5601 return Builder.CreateStore(Ops[1], Ops[0]); 5602 } 5603 case NEON::BI__builtin_neon_vst2_v: 5604 case NEON::BI__builtin_neon_vst2q_v: { 5605 Ops.push_back(Ops[0]); 5606 Ops.erase(Ops.begin()); 5607 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5608 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5609 Ops, ""); 5610 } 5611 case NEON::BI__builtin_neon_vst2_lane_v: 5612 case NEON::BI__builtin_neon_vst2q_lane_v: { 5613 Ops.push_back(Ops[0]); 5614 Ops.erase(Ops.begin()); 5615 Ops[2] = Builder.CreateZExt(Ops[2], 5616 llvm::IntegerType::get(getLLVMContext(), 64)); 5617 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5618 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5619 Ops, ""); 5620 } 5621 case NEON::BI__builtin_neon_vst3_v: 5622 case NEON::BI__builtin_neon_vst3q_v: { 5623 Ops.push_back(Ops[0]); 5624 Ops.erase(Ops.begin()); 5625 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5626 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5627 Ops, ""); 5628 } 5629 case NEON::BI__builtin_neon_vst3_lane_v: 5630 case NEON::BI__builtin_neon_vst3q_lane_v: { 5631 Ops.push_back(Ops[0]); 5632 Ops.erase(Ops.begin()); 5633 Ops[3] = Builder.CreateZExt(Ops[3], 5634 llvm::IntegerType::get(getLLVMContext(), 64)); 5635 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5636 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5637 Ops, ""); 5638 } 5639 case NEON::BI__builtin_neon_vst4_v: 5640 case NEON::BI__builtin_neon_vst4q_v: { 5641 Ops.push_back(Ops[0]); 5642 Ops.erase(Ops.begin()); 5643 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5644 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5645 Ops, ""); 5646 } 5647 case NEON::BI__builtin_neon_vst4_lane_v: 5648 case NEON::BI__builtin_neon_vst4q_lane_v: { 5649 Ops.push_back(Ops[0]); 5650 Ops.erase(Ops.begin()); 5651 Ops[4] = Builder.CreateZExt(Ops[4], 5652 llvm::IntegerType::get(getLLVMContext(), 64)); 5653 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5654 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5655 Ops, ""); 5656 } 5657 case NEON::BI__builtin_neon_vtrn_v: 5658 case NEON::BI__builtin_neon_vtrnq_v: { 5659 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5660 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5661 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5662 Value *SV = nullptr; 5663 5664 for (unsigned vi = 0; vi != 2; ++vi) { 5665 SmallVector<Constant*, 16> Indices; 5666 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5667 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5668 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5669 } 5670 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5671 SV = llvm::ConstantVector::get(Indices); 5672 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5673 SV = Builder.CreateStore(SV, Addr); 5674 } 5675 return SV; 5676 } 5677 case NEON::BI__builtin_neon_vuzp_v: 5678 case NEON::BI__builtin_neon_vuzpq_v: { 5679 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5680 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5681 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5682 Value *SV = nullptr; 5683 5684 for (unsigned vi = 0; vi != 2; ++vi) { 5685 SmallVector<Constant*, 16> Indices; 5686 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5687 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5688 5689 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5690 SV = llvm::ConstantVector::get(Indices); 5691 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5692 SV = Builder.CreateStore(SV, Addr); 5693 } 5694 return SV; 5695 } 5696 case NEON::BI__builtin_neon_vzip_v: 5697 case NEON::BI__builtin_neon_vzipq_v: { 5698 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5699 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5700 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5701 Value *SV = nullptr; 5702 5703 for (unsigned vi = 0; vi != 2; ++vi) { 5704 SmallVector<Constant*, 16> Indices; 5705 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5706 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5707 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5708 } 5709 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5710 SV = llvm::ConstantVector::get(Indices); 5711 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5712 SV = Builder.CreateStore(SV, Addr); 5713 } 5714 return SV; 5715 } 5716 case NEON::BI__builtin_neon_vqtbl1q_v: { 5717 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5718 Ops, "vtbl1"); 5719 } 5720 case NEON::BI__builtin_neon_vqtbl2q_v: { 5721 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5722 Ops, "vtbl2"); 5723 } 5724 case NEON::BI__builtin_neon_vqtbl3q_v: { 5725 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5726 Ops, "vtbl3"); 5727 } 5728 case NEON::BI__builtin_neon_vqtbl4q_v: { 5729 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5730 Ops, "vtbl4"); 5731 } 5732 case NEON::BI__builtin_neon_vqtbx1q_v: { 5733 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5734 Ops, "vtbx1"); 5735 } 5736 case NEON::BI__builtin_neon_vqtbx2q_v: { 5737 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5738 Ops, "vtbx2"); 5739 } 5740 case NEON::BI__builtin_neon_vqtbx3q_v: { 5741 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5742 Ops, "vtbx3"); 5743 } 5744 case NEON::BI__builtin_neon_vqtbx4q_v: { 5745 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5746 Ops, "vtbx4"); 5747 } 5748 case NEON::BI__builtin_neon_vsqadd_v: 5749 case NEON::BI__builtin_neon_vsqaddq_v: { 5750 Int = Intrinsic::aarch64_neon_usqadd; 5751 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5752 } 5753 case NEON::BI__builtin_neon_vuqadd_v: 5754 case NEON::BI__builtin_neon_vuqaddq_v: { 5755 Int = Intrinsic::aarch64_neon_suqadd; 5756 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5757 } 5758 } 5759 } 5760 5761 llvm::Value *CodeGenFunction:: 5762 BuildVector(ArrayRef<llvm::Value*> Ops) { 5763 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5764 "Not a power-of-two sized vector!"); 5765 bool AllConstants = true; 5766 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5767 AllConstants &= isa<Constant>(Ops[i]); 5768 5769 // If this is a constant vector, create a ConstantVector. 5770 if (AllConstants) { 5771 SmallVector<llvm::Constant*, 16> CstOps; 5772 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5773 CstOps.push_back(cast<Constant>(Ops[i])); 5774 return llvm::ConstantVector::get(CstOps); 5775 } 5776 5777 // Otherwise, insertelement the values to build the vector. 5778 Value *Result = 5779 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5780 5781 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5782 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5783 5784 return Result; 5785 } 5786 5787 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5788 const CallExpr *E) { 5789 SmallVector<Value*, 4> Ops; 5790 5791 // Find out if any arguments are required to be integer constant expressions. 5792 unsigned ICEArguments = 0; 5793 ASTContext::GetBuiltinTypeError Error; 5794 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5795 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5796 5797 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5798 // If this is a normal argument, just emit it as a scalar. 5799 if ((ICEArguments & (1 << i)) == 0) { 5800 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5801 continue; 5802 } 5803 5804 // If this is required to be a constant, constant fold it so that we know 5805 // that the generated intrinsic gets a ConstantInt. 5806 llvm::APSInt Result; 5807 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5808 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5809 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5810 } 5811 5812 switch (BuiltinID) { 5813 default: return nullptr; 5814 case X86::BI_mm_prefetch: { 5815 Value *Address = EmitScalarExpr(E->getArg(0)); 5816 Value *RW = ConstantInt::get(Int32Ty, 0); 5817 Value *Locality = EmitScalarExpr(E->getArg(1)); 5818 Value *Data = ConstantInt::get(Int32Ty, 1); 5819 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5820 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5821 } 5822 case X86::BI__builtin_ia32_vec_init_v8qi: 5823 case X86::BI__builtin_ia32_vec_init_v4hi: 5824 case X86::BI__builtin_ia32_vec_init_v2si: 5825 return Builder.CreateBitCast(BuildVector(Ops), 5826 llvm::Type::getX86_MMXTy(getLLVMContext())); 5827 case X86::BI__builtin_ia32_vec_ext_v2si: 5828 return Builder.CreateExtractElement(Ops[0], 5829 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5830 case X86::BI__builtin_ia32_ldmxcsr: { 5831 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5832 Builder.CreateStore(Ops[0], Tmp); 5833 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5834 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5835 } 5836 case X86::BI__builtin_ia32_stmxcsr: { 5837 Value *Tmp = CreateMemTemp(E->getType()); 5838 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5839 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5840 return Builder.CreateLoad(Tmp, "stmxcsr"); 5841 } 5842 case X86::BI__builtin_ia32_storehps: 5843 case X86::BI__builtin_ia32_storelps: { 5844 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5845 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5846 5847 // cast val v2i64 5848 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5849 5850 // extract (0, 1) 5851 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5852 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5853 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5854 5855 // cast pointer to i64 & store 5856 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5857 return Builder.CreateStore(Ops[1], Ops[0]); 5858 } 5859 case X86::BI__builtin_ia32_palignr: { 5860 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5861 5862 // If palignr is shifting the pair of input vectors less than 9 bytes, 5863 // emit a shuffle instruction. 5864 if (shiftVal <= 8) { 5865 SmallVector<llvm::Constant*, 8> Indices; 5866 for (unsigned i = 0; i != 8; ++i) 5867 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5868 5869 Value* SV = llvm::ConstantVector::get(Indices); 5870 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5871 } 5872 5873 // If palignr is shifting the pair of input vectors more than 8 but less 5874 // than 16 bytes, emit a logical right shift of the destination. 5875 if (shiftVal < 16) { 5876 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 5877 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 5878 5879 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5880 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 5881 5882 // create i32 constant 5883 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 5884 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5885 } 5886 5887 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 5888 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5889 } 5890 case X86::BI__builtin_ia32_palignr128: { 5891 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5892 5893 // If palignr is shifting the pair of input vectors less than 17 bytes, 5894 // emit a shuffle instruction. 5895 if (shiftVal <= 16) { 5896 SmallVector<llvm::Constant*, 16> Indices; 5897 for (unsigned i = 0; i != 16; ++i) 5898 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5899 5900 Value* SV = llvm::ConstantVector::get(Indices); 5901 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5902 } 5903 5904 // If palignr is shifting the pair of input vectors more than 16 but less 5905 // than 32 bytes, emit a logical right shift of the destination. 5906 if (shiftVal < 32) { 5907 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5908 5909 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5910 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5911 5912 // create i32 constant 5913 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 5914 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5915 } 5916 5917 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5918 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5919 } 5920 case X86::BI__builtin_ia32_palignr256: { 5921 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5922 5923 // If palignr is shifting the pair of input vectors less than 17 bytes, 5924 // emit a shuffle instruction. 5925 if (shiftVal <= 16) { 5926 SmallVector<llvm::Constant*, 32> Indices; 5927 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5928 for (unsigned l = 0; l != 2; ++l) { 5929 unsigned LaneStart = l * 16; 5930 unsigned LaneEnd = (l+1) * 16; 5931 for (unsigned i = 0; i != 16; ++i) { 5932 unsigned Idx = shiftVal + i + LaneStart; 5933 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 5934 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 5935 } 5936 } 5937 5938 Value* SV = llvm::ConstantVector::get(Indices); 5939 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5940 } 5941 5942 // If palignr is shifting the pair of input vectors more than 16 but less 5943 // than 32 bytes, emit a logical right shift of the destination. 5944 if (shiftVal < 32) { 5945 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 5946 5947 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5948 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5949 5950 // create i32 constant 5951 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 5952 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5953 } 5954 5955 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5956 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5957 } 5958 case X86::BI__builtin_ia32_movntps: 5959 case X86::BI__builtin_ia32_movntps256: 5960 case X86::BI__builtin_ia32_movntpd: 5961 case X86::BI__builtin_ia32_movntpd256: 5962 case X86::BI__builtin_ia32_movntdq: 5963 case X86::BI__builtin_ia32_movntdq256: 5964 case X86::BI__builtin_ia32_movnti: 5965 case X86::BI__builtin_ia32_movnti64: { 5966 llvm::MDNode *Node = llvm::MDNode::get( 5967 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 5968 5969 // Convert the type of the pointer to a pointer to the stored type. 5970 Value *BC = Builder.CreateBitCast(Ops[0], 5971 llvm::PointerType::getUnqual(Ops[1]->getType()), 5972 "cast"); 5973 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 5974 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 5975 5976 // If the operand is an integer, we can't assume alignment. Otherwise, 5977 // assume natural alignment. 5978 QualType ArgTy = E->getArg(1)->getType(); 5979 unsigned Align; 5980 if (ArgTy->isIntegerType()) 5981 Align = 1; 5982 else 5983 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 5984 SI->setAlignment(Align); 5985 return SI; 5986 } 5987 // 3DNow! 5988 case X86::BI__builtin_ia32_pswapdsf: 5989 case X86::BI__builtin_ia32_pswapdsi: { 5990 const char *name; 5991 Intrinsic::ID ID; 5992 switch(BuiltinID) { 5993 default: llvm_unreachable("Unsupported intrinsic!"); 5994 case X86::BI__builtin_ia32_pswapdsf: 5995 case X86::BI__builtin_ia32_pswapdsi: 5996 name = "pswapd"; 5997 ID = Intrinsic::x86_3dnowa_pswapd; 5998 break; 5999 } 6000 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6001 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6002 llvm::Function *F = CGM.getIntrinsic(ID); 6003 return Builder.CreateCall(F, Ops, name); 6004 } 6005 case X86::BI__builtin_ia32_rdrand16_step: 6006 case X86::BI__builtin_ia32_rdrand32_step: 6007 case X86::BI__builtin_ia32_rdrand64_step: 6008 case X86::BI__builtin_ia32_rdseed16_step: 6009 case X86::BI__builtin_ia32_rdseed32_step: 6010 case X86::BI__builtin_ia32_rdseed64_step: { 6011 Intrinsic::ID ID; 6012 switch (BuiltinID) { 6013 default: llvm_unreachable("Unsupported intrinsic!"); 6014 case X86::BI__builtin_ia32_rdrand16_step: 6015 ID = Intrinsic::x86_rdrand_16; 6016 break; 6017 case X86::BI__builtin_ia32_rdrand32_step: 6018 ID = Intrinsic::x86_rdrand_32; 6019 break; 6020 case X86::BI__builtin_ia32_rdrand64_step: 6021 ID = Intrinsic::x86_rdrand_64; 6022 break; 6023 case X86::BI__builtin_ia32_rdseed16_step: 6024 ID = Intrinsic::x86_rdseed_16; 6025 break; 6026 case X86::BI__builtin_ia32_rdseed32_step: 6027 ID = Intrinsic::x86_rdseed_32; 6028 break; 6029 case X86::BI__builtin_ia32_rdseed64_step: 6030 ID = Intrinsic::x86_rdseed_64; 6031 break; 6032 } 6033 6034 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6035 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6036 return Builder.CreateExtractValue(Call, 1); 6037 } 6038 // AVX2 broadcast 6039 case X86::BI__builtin_ia32_vbroadcastsi256: { 6040 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 6041 Builder.CreateStore(Ops[0], VecTmp); 6042 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 6043 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 6044 } 6045 // SSE comparison intrisics 6046 case X86::BI__builtin_ia32_cmpeqps: 6047 case X86::BI__builtin_ia32_cmpltps: 6048 case X86::BI__builtin_ia32_cmpleps: 6049 case X86::BI__builtin_ia32_cmpunordps: 6050 case X86::BI__builtin_ia32_cmpneqps: 6051 case X86::BI__builtin_ia32_cmpnltps: 6052 case X86::BI__builtin_ia32_cmpnleps: 6053 case X86::BI__builtin_ia32_cmpordps: 6054 case X86::BI__builtin_ia32_cmpeqss: 6055 case X86::BI__builtin_ia32_cmpltss: 6056 case X86::BI__builtin_ia32_cmpless: 6057 case X86::BI__builtin_ia32_cmpunordss: 6058 case X86::BI__builtin_ia32_cmpneqss: 6059 case X86::BI__builtin_ia32_cmpnltss: 6060 case X86::BI__builtin_ia32_cmpnless: 6061 case X86::BI__builtin_ia32_cmpordss: 6062 case X86::BI__builtin_ia32_cmpeqpd: 6063 case X86::BI__builtin_ia32_cmpltpd: 6064 case X86::BI__builtin_ia32_cmplepd: 6065 case X86::BI__builtin_ia32_cmpunordpd: 6066 case X86::BI__builtin_ia32_cmpneqpd: 6067 case X86::BI__builtin_ia32_cmpnltpd: 6068 case X86::BI__builtin_ia32_cmpnlepd: 6069 case X86::BI__builtin_ia32_cmpordpd: 6070 case X86::BI__builtin_ia32_cmpeqsd: 6071 case X86::BI__builtin_ia32_cmpltsd: 6072 case X86::BI__builtin_ia32_cmplesd: 6073 case X86::BI__builtin_ia32_cmpunordsd: 6074 case X86::BI__builtin_ia32_cmpneqsd: 6075 case X86::BI__builtin_ia32_cmpnltsd: 6076 case X86::BI__builtin_ia32_cmpnlesd: 6077 case X86::BI__builtin_ia32_cmpordsd: 6078 // These exist so that the builtin that takes an immediate can be bounds 6079 // checked by clang to avoid passing bad immediates to the backend. Since 6080 // AVX has a larger immediate than SSE we would need separate builtins to 6081 // do the different bounds checking. Rather than create a clang specific 6082 // SSE only builtin, this implements eight separate builtins to match gcc 6083 // implementation. 6084 6085 // Choose the immediate. 6086 unsigned Imm; 6087 switch (BuiltinID) { 6088 default: llvm_unreachable("Unsupported intrinsic!"); 6089 case X86::BI__builtin_ia32_cmpeqps: 6090 case X86::BI__builtin_ia32_cmpeqss: 6091 case X86::BI__builtin_ia32_cmpeqpd: 6092 case X86::BI__builtin_ia32_cmpeqsd: 6093 Imm = 0; 6094 break; 6095 case X86::BI__builtin_ia32_cmpltps: 6096 case X86::BI__builtin_ia32_cmpltss: 6097 case X86::BI__builtin_ia32_cmpltpd: 6098 case X86::BI__builtin_ia32_cmpltsd: 6099 Imm = 1; 6100 break; 6101 case X86::BI__builtin_ia32_cmpleps: 6102 case X86::BI__builtin_ia32_cmpless: 6103 case X86::BI__builtin_ia32_cmplepd: 6104 case X86::BI__builtin_ia32_cmplesd: 6105 Imm = 2; 6106 break; 6107 case X86::BI__builtin_ia32_cmpunordps: 6108 case X86::BI__builtin_ia32_cmpunordss: 6109 case X86::BI__builtin_ia32_cmpunordpd: 6110 case X86::BI__builtin_ia32_cmpunordsd: 6111 Imm = 3; 6112 break; 6113 case X86::BI__builtin_ia32_cmpneqps: 6114 case X86::BI__builtin_ia32_cmpneqss: 6115 case X86::BI__builtin_ia32_cmpneqpd: 6116 case X86::BI__builtin_ia32_cmpneqsd: 6117 Imm = 4; 6118 break; 6119 case X86::BI__builtin_ia32_cmpnltps: 6120 case X86::BI__builtin_ia32_cmpnltss: 6121 case X86::BI__builtin_ia32_cmpnltpd: 6122 case X86::BI__builtin_ia32_cmpnltsd: 6123 Imm = 5; 6124 break; 6125 case X86::BI__builtin_ia32_cmpnleps: 6126 case X86::BI__builtin_ia32_cmpnless: 6127 case X86::BI__builtin_ia32_cmpnlepd: 6128 case X86::BI__builtin_ia32_cmpnlesd: 6129 Imm = 6; 6130 break; 6131 case X86::BI__builtin_ia32_cmpordps: 6132 case X86::BI__builtin_ia32_cmpordss: 6133 case X86::BI__builtin_ia32_cmpordpd: 6134 case X86::BI__builtin_ia32_cmpordsd: 6135 Imm = 7; 6136 break; 6137 } 6138 6139 // Choose the intrinsic ID. 6140 const char *name; 6141 Intrinsic::ID ID; 6142 switch (BuiltinID) { 6143 default: llvm_unreachable("Unsupported intrinsic!"); 6144 case X86::BI__builtin_ia32_cmpeqps: 6145 case X86::BI__builtin_ia32_cmpltps: 6146 case X86::BI__builtin_ia32_cmpleps: 6147 case X86::BI__builtin_ia32_cmpunordps: 6148 case X86::BI__builtin_ia32_cmpneqps: 6149 case X86::BI__builtin_ia32_cmpnltps: 6150 case X86::BI__builtin_ia32_cmpnleps: 6151 case X86::BI__builtin_ia32_cmpordps: 6152 name = "cmpps"; 6153 ID = Intrinsic::x86_sse_cmp_ps; 6154 break; 6155 case X86::BI__builtin_ia32_cmpeqss: 6156 case X86::BI__builtin_ia32_cmpltss: 6157 case X86::BI__builtin_ia32_cmpless: 6158 case X86::BI__builtin_ia32_cmpunordss: 6159 case X86::BI__builtin_ia32_cmpneqss: 6160 case X86::BI__builtin_ia32_cmpnltss: 6161 case X86::BI__builtin_ia32_cmpnless: 6162 case X86::BI__builtin_ia32_cmpordss: 6163 name = "cmpss"; 6164 ID = Intrinsic::x86_sse_cmp_ss; 6165 break; 6166 case X86::BI__builtin_ia32_cmpeqpd: 6167 case X86::BI__builtin_ia32_cmpltpd: 6168 case X86::BI__builtin_ia32_cmplepd: 6169 case X86::BI__builtin_ia32_cmpunordpd: 6170 case X86::BI__builtin_ia32_cmpneqpd: 6171 case X86::BI__builtin_ia32_cmpnltpd: 6172 case X86::BI__builtin_ia32_cmpnlepd: 6173 case X86::BI__builtin_ia32_cmpordpd: 6174 name = "cmppd"; 6175 ID = Intrinsic::x86_sse2_cmp_pd; 6176 break; 6177 case X86::BI__builtin_ia32_cmpeqsd: 6178 case X86::BI__builtin_ia32_cmpltsd: 6179 case X86::BI__builtin_ia32_cmplesd: 6180 case X86::BI__builtin_ia32_cmpunordsd: 6181 case X86::BI__builtin_ia32_cmpneqsd: 6182 case X86::BI__builtin_ia32_cmpnltsd: 6183 case X86::BI__builtin_ia32_cmpnlesd: 6184 case X86::BI__builtin_ia32_cmpordsd: 6185 name = "cmpsd"; 6186 ID = Intrinsic::x86_sse2_cmp_sd; 6187 break; 6188 } 6189 6190 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6191 llvm::Function *F = CGM.getIntrinsic(ID); 6192 return Builder.CreateCall(F, Ops, name); 6193 } 6194 } 6195 6196 6197 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6198 const CallExpr *E) { 6199 SmallVector<Value*, 4> Ops; 6200 6201 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6202 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6203 6204 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6205 6206 switch (BuiltinID) { 6207 default: return nullptr; 6208 6209 // vec_ld, vec_lvsl, vec_lvsr 6210 case PPC::BI__builtin_altivec_lvx: 6211 case PPC::BI__builtin_altivec_lvxl: 6212 case PPC::BI__builtin_altivec_lvebx: 6213 case PPC::BI__builtin_altivec_lvehx: 6214 case PPC::BI__builtin_altivec_lvewx: 6215 case PPC::BI__builtin_altivec_lvsl: 6216 case PPC::BI__builtin_altivec_lvsr: 6217 case PPC::BI__builtin_vsx_lxvd2x: 6218 case PPC::BI__builtin_vsx_lxvw4x: 6219 { 6220 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6221 6222 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6223 Ops.pop_back(); 6224 6225 switch (BuiltinID) { 6226 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6227 case PPC::BI__builtin_altivec_lvx: 6228 ID = Intrinsic::ppc_altivec_lvx; 6229 break; 6230 case PPC::BI__builtin_altivec_lvxl: 6231 ID = Intrinsic::ppc_altivec_lvxl; 6232 break; 6233 case PPC::BI__builtin_altivec_lvebx: 6234 ID = Intrinsic::ppc_altivec_lvebx; 6235 break; 6236 case PPC::BI__builtin_altivec_lvehx: 6237 ID = Intrinsic::ppc_altivec_lvehx; 6238 break; 6239 case PPC::BI__builtin_altivec_lvewx: 6240 ID = Intrinsic::ppc_altivec_lvewx; 6241 break; 6242 case PPC::BI__builtin_altivec_lvsl: 6243 ID = Intrinsic::ppc_altivec_lvsl; 6244 break; 6245 case PPC::BI__builtin_altivec_lvsr: 6246 ID = Intrinsic::ppc_altivec_lvsr; 6247 break; 6248 case PPC::BI__builtin_vsx_lxvd2x: 6249 ID = Intrinsic::ppc_vsx_lxvd2x; 6250 break; 6251 case PPC::BI__builtin_vsx_lxvw4x: 6252 ID = Intrinsic::ppc_vsx_lxvw4x; 6253 break; 6254 } 6255 llvm::Function *F = CGM.getIntrinsic(ID); 6256 return Builder.CreateCall(F, Ops, ""); 6257 } 6258 6259 // vec_st 6260 case PPC::BI__builtin_altivec_stvx: 6261 case PPC::BI__builtin_altivec_stvxl: 6262 case PPC::BI__builtin_altivec_stvebx: 6263 case PPC::BI__builtin_altivec_stvehx: 6264 case PPC::BI__builtin_altivec_stvewx: 6265 case PPC::BI__builtin_vsx_stxvd2x: 6266 case PPC::BI__builtin_vsx_stxvw4x: 6267 { 6268 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6269 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6270 Ops.pop_back(); 6271 6272 switch (BuiltinID) { 6273 default: llvm_unreachable("Unsupported st intrinsic!"); 6274 case PPC::BI__builtin_altivec_stvx: 6275 ID = Intrinsic::ppc_altivec_stvx; 6276 break; 6277 case PPC::BI__builtin_altivec_stvxl: 6278 ID = Intrinsic::ppc_altivec_stvxl; 6279 break; 6280 case PPC::BI__builtin_altivec_stvebx: 6281 ID = Intrinsic::ppc_altivec_stvebx; 6282 break; 6283 case PPC::BI__builtin_altivec_stvehx: 6284 ID = Intrinsic::ppc_altivec_stvehx; 6285 break; 6286 case PPC::BI__builtin_altivec_stvewx: 6287 ID = Intrinsic::ppc_altivec_stvewx; 6288 break; 6289 case PPC::BI__builtin_vsx_stxvd2x: 6290 ID = Intrinsic::ppc_vsx_stxvd2x; 6291 break; 6292 case PPC::BI__builtin_vsx_stxvw4x: 6293 ID = Intrinsic::ppc_vsx_stxvw4x; 6294 break; 6295 } 6296 llvm::Function *F = CGM.getIntrinsic(ID); 6297 return Builder.CreateCall(F, Ops, ""); 6298 } 6299 } 6300 } 6301 6302 // Emit an intrinsic that has 1 float or double. 6303 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6304 const CallExpr *E, 6305 unsigned IntrinsicID) { 6306 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6307 6308 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6309 return CGF.Builder.CreateCall(F, Src0); 6310 } 6311 6312 // Emit an intrinsic that has 3 float or double operands. 6313 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6314 const CallExpr *E, 6315 unsigned IntrinsicID) { 6316 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6317 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6318 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6319 6320 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6321 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6322 } 6323 6324 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6325 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6326 const CallExpr *E, 6327 unsigned IntrinsicID) { 6328 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6329 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6330 6331 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6332 return CGF.Builder.CreateCall2(F, Src0, Src1); 6333 } 6334 6335 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6336 const CallExpr *E) { 6337 switch (BuiltinID) { 6338 case R600::BI__builtin_amdgpu_div_scale: 6339 case R600::BI__builtin_amdgpu_div_scalef: { 6340 // Translate from the intrinsics's struct return to the builtin's out 6341 // argument. 6342 6343 std::pair<llvm::Value *, unsigned> FlagOutPtr 6344 = EmitPointerWithAlignment(E->getArg(3)); 6345 6346 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6347 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6348 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6349 6350 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6351 X->getType()); 6352 6353 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6354 6355 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6356 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6357 6358 llvm::Type *RealFlagType 6359 = FlagOutPtr.first->getType()->getPointerElementType(); 6360 6361 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6362 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6363 FlagStore->setAlignment(FlagOutPtr.second); 6364 return Result; 6365 } 6366 case R600::BI__builtin_amdgpu_div_fmas: 6367 case R600::BI__builtin_amdgpu_div_fmasf: { 6368 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6369 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6370 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6371 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6372 6373 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6374 Src0->getType()); 6375 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6376 return Builder.CreateCall4(F, Src0, Src1, Src2, Src3ToBool); 6377 } 6378 case R600::BI__builtin_amdgpu_div_fixup: 6379 case R600::BI__builtin_amdgpu_div_fixupf: 6380 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6381 case R600::BI__builtin_amdgpu_trig_preop: 6382 case R600::BI__builtin_amdgpu_trig_preopf: 6383 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6384 case R600::BI__builtin_amdgpu_rcp: 6385 case R600::BI__builtin_amdgpu_rcpf: 6386 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6387 case R600::BI__builtin_amdgpu_rsq: 6388 case R600::BI__builtin_amdgpu_rsqf: 6389 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6390 case R600::BI__builtin_amdgpu_rsq_clamped: 6391 case R600::BI__builtin_amdgpu_rsq_clampedf: 6392 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6393 case R600::BI__builtin_amdgpu_ldexp: 6394 case R600::BI__builtin_amdgpu_ldexpf: 6395 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6396 case R600::BI__builtin_amdgpu_class: 6397 case R600::BI__builtin_amdgpu_classf: 6398 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6399 default: 6400 return nullptr; 6401 } 6402 } 6403