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 return EmitR600BuiltinExpr(BuiltinID, E); 1766 default: 1767 return nullptr; 1768 } 1769 } 1770 1771 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1772 NeonTypeFlags TypeFlags, 1773 bool V1Ty=false) { 1774 int IsQuad = TypeFlags.isQuad(); 1775 switch (TypeFlags.getEltType()) { 1776 case NeonTypeFlags::Int8: 1777 case NeonTypeFlags::Poly8: 1778 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1779 case NeonTypeFlags::Int16: 1780 case NeonTypeFlags::Poly16: 1781 case NeonTypeFlags::Float16: 1782 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1783 case NeonTypeFlags::Int32: 1784 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1785 case NeonTypeFlags::Int64: 1786 case NeonTypeFlags::Poly64: 1787 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1788 case NeonTypeFlags::Poly128: 1789 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1790 // There is a lot of i128 and f128 API missing. 1791 // so we use v16i8 to represent poly128 and get pattern matched. 1792 return llvm::VectorType::get(CGF->Int8Ty, 16); 1793 case NeonTypeFlags::Float32: 1794 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1795 case NeonTypeFlags::Float64: 1796 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1797 } 1798 llvm_unreachable("Unknown vector element type!"); 1799 } 1800 1801 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1802 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1803 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1804 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1805 } 1806 1807 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1808 const char *name, 1809 unsigned shift, bool rightshift) { 1810 unsigned j = 0; 1811 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1812 ai != ae; ++ai, ++j) 1813 if (shift > 0 && shift == j) 1814 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1815 else 1816 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1817 1818 return Builder.CreateCall(F, Ops, name); 1819 } 1820 1821 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1822 bool neg) { 1823 int SV = cast<ConstantInt>(V)->getSExtValue(); 1824 1825 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1826 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1827 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1828 } 1829 1830 // \brief Right-shift a vector by a constant. 1831 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1832 llvm::Type *Ty, bool usgn, 1833 const char *name) { 1834 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1835 1836 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1837 int EltSize = VTy->getScalarSizeInBits(); 1838 1839 Vec = Builder.CreateBitCast(Vec, Ty); 1840 1841 // lshr/ashr are undefined when the shift amount is equal to the vector 1842 // element size. 1843 if (ShiftAmt == EltSize) { 1844 if (usgn) { 1845 // Right-shifting an unsigned value by its size yields 0. 1846 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1847 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1848 } else { 1849 // Right-shifting a signed value by its size is equivalent 1850 // to a shift of size-1. 1851 --ShiftAmt; 1852 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1853 } 1854 } 1855 1856 Shift = EmitNeonShiftVector(Shift, Ty, false); 1857 if (usgn) 1858 return Builder.CreateLShr(Vec, Shift, name); 1859 else 1860 return Builder.CreateAShr(Vec, Shift, name); 1861 } 1862 1863 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1864 /// alignment of the type referenced by the pointer. Skip over implicit 1865 /// casts. 1866 std::pair<llvm::Value*, unsigned> 1867 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1868 assert(Addr->getType()->isPointerType()); 1869 Addr = Addr->IgnoreParens(); 1870 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1871 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1872 ICE->getSubExpr()->getType()->isPointerType()) { 1873 std::pair<llvm::Value*, unsigned> Ptr = 1874 EmitPointerWithAlignment(ICE->getSubExpr()); 1875 Ptr.first = Builder.CreateBitCast(Ptr.first, 1876 ConvertType(Addr->getType())); 1877 return Ptr; 1878 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1879 LValue LV = EmitLValue(ICE->getSubExpr()); 1880 unsigned Align = LV.getAlignment().getQuantity(); 1881 if (!Align) { 1882 // FIXME: Once LValues are fixed to always set alignment, 1883 // zap this code. 1884 QualType PtTy = ICE->getSubExpr()->getType(); 1885 if (!PtTy->isIncompleteType()) 1886 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1887 else 1888 Align = 1; 1889 } 1890 return std::make_pair(LV.getAddress(), Align); 1891 } 1892 } 1893 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1894 if (UO->getOpcode() == UO_AddrOf) { 1895 LValue LV = EmitLValue(UO->getSubExpr()); 1896 unsigned Align = LV.getAlignment().getQuantity(); 1897 if (!Align) { 1898 // FIXME: Once LValues are fixed to always set alignment, 1899 // zap this code. 1900 QualType PtTy = UO->getSubExpr()->getType(); 1901 if (!PtTy->isIncompleteType()) 1902 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1903 else 1904 Align = 1; 1905 } 1906 return std::make_pair(LV.getAddress(), Align); 1907 } 1908 } 1909 1910 unsigned Align = 1; 1911 QualType PtTy = Addr->getType()->getPointeeType(); 1912 if (!PtTy->isIncompleteType()) 1913 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1914 1915 return std::make_pair(EmitScalarExpr(Addr), Align); 1916 } 1917 1918 enum { 1919 AddRetType = (1 << 0), 1920 Add1ArgType = (1 << 1), 1921 Add2ArgTypes = (1 << 2), 1922 1923 VectorizeRetType = (1 << 3), 1924 VectorizeArgTypes = (1 << 4), 1925 1926 InventFloatType = (1 << 5), 1927 UnsignedAlts = (1 << 6), 1928 1929 Use64BitVectors = (1 << 7), 1930 Use128BitVectors = (1 << 8), 1931 1932 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 1933 VectorRet = AddRetType | VectorizeRetType, 1934 VectorRetGetArgs01 = 1935 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 1936 FpCmpzModifiers = 1937 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 1938 }; 1939 1940 struct NeonIntrinsicInfo { 1941 unsigned BuiltinID; 1942 unsigned LLVMIntrinsic; 1943 unsigned AltLLVMIntrinsic; 1944 const char *NameHint; 1945 unsigned TypeModifier; 1946 1947 bool operator<(unsigned RHSBuiltinID) const { 1948 return BuiltinID < RHSBuiltinID; 1949 } 1950 }; 1951 1952 #define NEONMAP0(NameBase) \ 1953 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 1954 1955 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 1956 { NEON:: BI__builtin_neon_ ## NameBase, \ 1957 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 1958 1959 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 1960 { NEON:: BI__builtin_neon_ ## NameBase, \ 1961 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 1962 #NameBase, TypeModifier } 1963 1964 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 1965 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1966 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 1967 NEONMAP1(vabs_v, arm_neon_vabs, 0), 1968 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 1969 NEONMAP0(vaddhn_v), 1970 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 1971 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 1972 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 1973 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 1974 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 1975 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 1976 NEONMAP1(vcage_v, arm_neon_vacge, 0), 1977 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 1978 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 1979 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 1980 NEONMAP1(vcale_v, arm_neon_vacge, 0), 1981 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 1982 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 1983 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 1984 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 1985 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 1986 NEONMAP1(vclz_v, ctlz, Add1ArgType), 1987 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 1988 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 1989 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 1990 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 1991 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 1992 NEONMAP0(vcvt_f32_v), 1993 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 1994 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 1995 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 1996 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 1997 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 1998 NEONMAP0(vcvt_s32_v), 1999 NEONMAP0(vcvt_s64_v), 2000 NEONMAP0(vcvt_u32_v), 2001 NEONMAP0(vcvt_u64_v), 2002 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2003 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2004 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2005 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2006 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2007 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2008 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2009 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2010 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2011 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2012 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2013 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2014 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2015 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2016 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2017 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2018 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2019 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2020 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2021 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2022 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2023 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2024 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2025 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2026 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2027 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2028 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2029 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2030 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2031 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2032 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2033 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2034 NEONMAP0(vcvtq_f32_v), 2035 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2036 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2037 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2038 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2039 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2040 NEONMAP0(vcvtq_s32_v), 2041 NEONMAP0(vcvtq_s64_v), 2042 NEONMAP0(vcvtq_u32_v), 2043 NEONMAP0(vcvtq_u64_v), 2044 NEONMAP0(vext_v), 2045 NEONMAP0(vextq_v), 2046 NEONMAP0(vfma_v), 2047 NEONMAP0(vfmaq_v), 2048 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2049 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2050 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2051 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2052 NEONMAP0(vld1_dup_v), 2053 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2054 NEONMAP0(vld1q_dup_v), 2055 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2056 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2057 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2058 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2059 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2060 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2061 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2062 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2063 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2064 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2065 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2066 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2067 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2068 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2069 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2070 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2071 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2072 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2073 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2074 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2075 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2076 NEONMAP0(vmovl_v), 2077 NEONMAP0(vmovn_v), 2078 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2079 NEONMAP0(vmull_v), 2080 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2081 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2082 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2083 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2084 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2085 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2086 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2087 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2088 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2089 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2090 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2091 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2092 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2093 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2094 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2095 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2096 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2097 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2098 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2099 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2100 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2101 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2102 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2103 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2104 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2105 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2106 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2107 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2108 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2109 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2110 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2111 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2112 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2113 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2114 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2115 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2116 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2117 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2118 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2119 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2120 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2121 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2122 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2123 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2124 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2125 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2126 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2127 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2128 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2129 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2130 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2131 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2132 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2133 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2134 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2135 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2136 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2137 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2138 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2139 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2140 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2141 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2142 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2143 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2144 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2145 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2146 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2147 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2148 NEONMAP0(vshl_n_v), 2149 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2150 NEONMAP0(vshll_n_v), 2151 NEONMAP0(vshlq_n_v), 2152 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2153 NEONMAP0(vshr_n_v), 2154 NEONMAP0(vshrn_n_v), 2155 NEONMAP0(vshrq_n_v), 2156 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2157 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2158 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2159 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2160 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2161 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2162 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2163 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2164 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2165 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2166 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2167 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2168 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2169 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2170 NEONMAP0(vsubhn_v), 2171 NEONMAP0(vtrn_v), 2172 NEONMAP0(vtrnq_v), 2173 NEONMAP0(vtst_v), 2174 NEONMAP0(vtstq_v), 2175 NEONMAP0(vuzp_v), 2176 NEONMAP0(vuzpq_v), 2177 NEONMAP0(vzip_v), 2178 NEONMAP0(vzipq_v) 2179 }; 2180 2181 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2182 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2183 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2184 NEONMAP0(vaddhn_v), 2185 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2186 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2187 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2188 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2189 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2190 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2191 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2192 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2193 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2194 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2195 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2196 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2197 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2198 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2199 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2200 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2201 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2202 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2203 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2204 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2205 NEONMAP0(vcvt_f32_v), 2206 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2207 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2208 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2209 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2210 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2211 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2212 NEONMAP0(vcvtq_f32_v), 2213 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2214 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2215 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2216 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2217 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2218 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2219 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2220 NEONMAP0(vext_v), 2221 NEONMAP0(vextq_v), 2222 NEONMAP0(vfma_v), 2223 NEONMAP0(vfmaq_v), 2224 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2225 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2226 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2227 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2228 NEONMAP0(vmovl_v), 2229 NEONMAP0(vmovn_v), 2230 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2231 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2232 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2233 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2234 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2235 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2236 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2237 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2238 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2239 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2240 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2241 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2242 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2243 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2244 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2245 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2246 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2247 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2248 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2249 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2250 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2251 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2252 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2253 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2254 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2255 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2256 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2257 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2258 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2259 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2260 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2261 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2262 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2263 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2264 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2265 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2266 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2267 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2268 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2269 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2270 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2271 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2272 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2273 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2274 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2275 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2276 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2277 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2278 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2279 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2280 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2281 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2282 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2283 NEONMAP0(vshl_n_v), 2284 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2285 NEONMAP0(vshll_n_v), 2286 NEONMAP0(vshlq_n_v), 2287 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2288 NEONMAP0(vshr_n_v), 2289 NEONMAP0(vshrn_n_v), 2290 NEONMAP0(vshrq_n_v), 2291 NEONMAP0(vsubhn_v), 2292 NEONMAP0(vtst_v), 2293 NEONMAP0(vtstq_v), 2294 }; 2295 2296 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2297 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2298 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2299 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2300 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2301 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2302 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2303 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2304 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2305 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2306 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2307 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2308 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2309 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2310 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2311 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2312 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2313 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2314 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2315 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2316 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2317 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2318 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2319 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2320 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2321 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2322 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2323 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2324 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2325 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2326 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2327 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2328 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2329 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2330 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2331 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2332 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2333 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2334 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2335 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2336 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2337 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2338 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2339 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2340 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2341 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2342 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2343 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2344 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2345 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2346 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2347 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2348 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2349 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2350 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2351 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2352 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2353 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2354 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2355 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2356 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2357 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2358 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2359 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2360 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2361 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2362 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2363 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2364 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2365 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2366 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2367 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2368 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2369 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2370 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2371 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2372 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2373 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2374 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2375 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2376 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2377 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2378 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2379 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2380 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2381 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2382 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2383 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2384 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2385 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2386 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2387 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2388 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2389 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2390 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2391 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2392 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2393 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2394 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2395 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2396 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2397 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2398 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2399 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2400 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2401 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2402 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2403 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2404 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2405 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2406 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2407 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2408 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2409 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2410 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2411 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2412 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2413 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2414 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2415 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2416 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2417 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2418 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2419 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2420 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2421 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2422 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2423 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2424 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2425 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2426 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2427 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2428 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2429 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2430 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2431 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2432 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2433 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2434 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2435 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2436 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2437 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2438 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2439 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2440 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2441 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2442 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2443 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2444 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2445 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2446 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2447 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2448 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2449 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2450 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2451 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2452 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2453 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2454 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2455 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2456 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2457 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2458 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2459 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2460 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2461 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2462 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2463 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2464 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2465 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2466 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2467 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2468 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2469 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2470 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2471 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2472 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2473 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2474 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2475 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2476 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2477 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2478 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2479 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2480 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2481 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2482 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2483 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2484 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2485 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2486 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2487 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2488 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2489 }; 2490 2491 #undef NEONMAP0 2492 #undef NEONMAP1 2493 #undef NEONMAP2 2494 2495 static bool NEONSIMDIntrinsicsProvenSorted = false; 2496 2497 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2498 static bool AArch64SISDIntrinsicsProvenSorted = false; 2499 2500 2501 static const NeonIntrinsicInfo * 2502 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2503 unsigned BuiltinID, bool &MapProvenSorted) { 2504 2505 #ifndef NDEBUG 2506 if (!MapProvenSorted) { 2507 // FIXME: use std::is_sorted once C++11 is allowed 2508 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2509 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2510 MapProvenSorted = true; 2511 } 2512 #endif 2513 2514 const NeonIntrinsicInfo *Builtin = 2515 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2516 2517 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2518 return Builtin; 2519 2520 return nullptr; 2521 } 2522 2523 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2524 unsigned Modifier, 2525 llvm::Type *ArgType, 2526 const CallExpr *E) { 2527 int VectorSize = 0; 2528 if (Modifier & Use64BitVectors) 2529 VectorSize = 64; 2530 else if (Modifier & Use128BitVectors) 2531 VectorSize = 128; 2532 2533 // Return type. 2534 SmallVector<llvm::Type *, 3> Tys; 2535 if (Modifier & AddRetType) { 2536 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2537 if (Modifier & VectorizeRetType) 2538 Ty = llvm::VectorType::get( 2539 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2540 2541 Tys.push_back(Ty); 2542 } 2543 2544 // Arguments. 2545 if (Modifier & VectorizeArgTypes) { 2546 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2547 ArgType = llvm::VectorType::get(ArgType, Elts); 2548 } 2549 2550 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2551 Tys.push_back(ArgType); 2552 2553 if (Modifier & Add2ArgTypes) 2554 Tys.push_back(ArgType); 2555 2556 if (Modifier & InventFloatType) 2557 Tys.push_back(FloatTy); 2558 2559 return CGM.getIntrinsic(IntrinsicID, Tys); 2560 } 2561 2562 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2563 const NeonIntrinsicInfo &SISDInfo, 2564 SmallVectorImpl<Value *> &Ops, 2565 const CallExpr *E) { 2566 unsigned BuiltinID = SISDInfo.BuiltinID; 2567 unsigned int Int = SISDInfo.LLVMIntrinsic; 2568 unsigned Modifier = SISDInfo.TypeModifier; 2569 const char *s = SISDInfo.NameHint; 2570 2571 switch (BuiltinID) { 2572 case NEON::BI__builtin_neon_vcled_s64: 2573 case NEON::BI__builtin_neon_vcled_u64: 2574 case NEON::BI__builtin_neon_vcles_f32: 2575 case NEON::BI__builtin_neon_vcled_f64: 2576 case NEON::BI__builtin_neon_vcltd_s64: 2577 case NEON::BI__builtin_neon_vcltd_u64: 2578 case NEON::BI__builtin_neon_vclts_f32: 2579 case NEON::BI__builtin_neon_vcltd_f64: 2580 case NEON::BI__builtin_neon_vcales_f32: 2581 case NEON::BI__builtin_neon_vcaled_f64: 2582 case NEON::BI__builtin_neon_vcalts_f32: 2583 case NEON::BI__builtin_neon_vcaltd_f64: 2584 // Only one direction of comparisons actually exist, cmle is actually a cmge 2585 // with swapped operands. The table gives us the right intrinsic but we 2586 // still need to do the swap. 2587 std::swap(Ops[0], Ops[1]); 2588 break; 2589 } 2590 2591 assert(Int && "Generic code assumes a valid intrinsic"); 2592 2593 // Determine the type(s) of this overloaded AArch64 intrinsic. 2594 const Expr *Arg = E->getArg(0); 2595 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2596 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2597 2598 int j = 0; 2599 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2600 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2601 ai != ae; ++ai, ++j) { 2602 llvm::Type *ArgTy = ai->getType(); 2603 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2604 ArgTy->getPrimitiveSizeInBits()) 2605 continue; 2606 2607 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2608 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2609 // it before inserting. 2610 Ops[j] = 2611 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2612 Ops[j] = 2613 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2614 } 2615 2616 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2617 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2618 if (ResultType->getPrimitiveSizeInBits() < 2619 Result->getType()->getPrimitiveSizeInBits()) 2620 return CGF.Builder.CreateExtractElement(Result, C0); 2621 2622 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2623 } 2624 2625 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2626 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2627 const char *NameHint, unsigned Modifier, const CallExpr *E, 2628 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2629 // Get the last argument, which specifies the vector type. 2630 llvm::APSInt NeonTypeConst; 2631 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2632 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2633 return nullptr; 2634 2635 // Determine the type of this overloaded NEON intrinsic. 2636 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2637 bool Usgn = Type.isUnsigned(); 2638 bool Quad = Type.isQuad(); 2639 2640 llvm::VectorType *VTy = GetNeonType(this, Type); 2641 llvm::Type *Ty = VTy; 2642 if (!Ty) 2643 return nullptr; 2644 2645 unsigned Int = LLVMIntrinsic; 2646 if ((Modifier & UnsignedAlts) && !Usgn) 2647 Int = AltLLVMIntrinsic; 2648 2649 switch (BuiltinID) { 2650 default: break; 2651 case NEON::BI__builtin_neon_vabs_v: 2652 case NEON::BI__builtin_neon_vabsq_v: 2653 if (VTy->getElementType()->isFloatingPointTy()) 2654 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2655 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2656 case NEON::BI__builtin_neon_vaddhn_v: { 2657 llvm::VectorType *SrcTy = 2658 llvm::VectorType::getExtendedElementVectorType(VTy); 2659 2660 // %sum = add <4 x i32> %lhs, %rhs 2661 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2662 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2663 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2664 2665 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2666 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2667 SrcTy->getScalarSizeInBits() / 2); 2668 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2669 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2670 2671 // %res = trunc <4 x i32> %high to <4 x i16> 2672 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2673 } 2674 case NEON::BI__builtin_neon_vcale_v: 2675 case NEON::BI__builtin_neon_vcaleq_v: 2676 case NEON::BI__builtin_neon_vcalt_v: 2677 case NEON::BI__builtin_neon_vcaltq_v: 2678 std::swap(Ops[0], Ops[1]); 2679 case NEON::BI__builtin_neon_vcage_v: 2680 case NEON::BI__builtin_neon_vcageq_v: 2681 case NEON::BI__builtin_neon_vcagt_v: 2682 case NEON::BI__builtin_neon_vcagtq_v: { 2683 llvm::Type *VecFlt = llvm::VectorType::get( 2684 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2685 VTy->getNumElements()); 2686 llvm::Type *Tys[] = { VTy, VecFlt }; 2687 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2688 return EmitNeonCall(F, Ops, NameHint); 2689 } 2690 case NEON::BI__builtin_neon_vclz_v: 2691 case NEON::BI__builtin_neon_vclzq_v: 2692 // We generate target-independent intrinsic, which needs a second argument 2693 // for whether or not clz of zero is undefined; on ARM it isn't. 2694 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2695 break; 2696 case NEON::BI__builtin_neon_vcvt_f32_v: 2697 case NEON::BI__builtin_neon_vcvtq_f32_v: 2698 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2699 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2700 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2701 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2702 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2703 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2704 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2705 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2706 bool Double = 2707 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2708 llvm::Type *FloatTy = 2709 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2710 : NeonTypeFlags::Float32, 2711 false, Quad)); 2712 llvm::Type *Tys[2] = { FloatTy, Ty }; 2713 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2714 Function *F = CGM.getIntrinsic(Int, Tys); 2715 return EmitNeonCall(F, Ops, "vcvt_n"); 2716 } 2717 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2718 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2719 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2720 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2721 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2722 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2723 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2724 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2725 bool Double = 2726 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2727 llvm::Type *FloatTy = 2728 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2729 : NeonTypeFlags::Float32, 2730 false, Quad)); 2731 llvm::Type *Tys[2] = { Ty, FloatTy }; 2732 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2733 return EmitNeonCall(F, Ops, "vcvt_n"); 2734 } 2735 case NEON::BI__builtin_neon_vcvt_s32_v: 2736 case NEON::BI__builtin_neon_vcvt_u32_v: 2737 case NEON::BI__builtin_neon_vcvt_s64_v: 2738 case NEON::BI__builtin_neon_vcvt_u64_v: 2739 case NEON::BI__builtin_neon_vcvtq_s32_v: 2740 case NEON::BI__builtin_neon_vcvtq_u32_v: 2741 case NEON::BI__builtin_neon_vcvtq_s64_v: 2742 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2743 bool Double = 2744 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2745 llvm::Type *FloatTy = 2746 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2747 : NeonTypeFlags::Float32, 2748 false, Quad)); 2749 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2750 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2751 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2752 } 2753 case NEON::BI__builtin_neon_vcvta_s32_v: 2754 case NEON::BI__builtin_neon_vcvta_s64_v: 2755 case NEON::BI__builtin_neon_vcvta_u32_v: 2756 case NEON::BI__builtin_neon_vcvta_u64_v: 2757 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2758 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2759 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2760 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2761 case NEON::BI__builtin_neon_vcvtn_s32_v: 2762 case NEON::BI__builtin_neon_vcvtn_s64_v: 2763 case NEON::BI__builtin_neon_vcvtn_u32_v: 2764 case NEON::BI__builtin_neon_vcvtn_u64_v: 2765 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2766 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2767 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2768 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2769 case NEON::BI__builtin_neon_vcvtp_s32_v: 2770 case NEON::BI__builtin_neon_vcvtp_s64_v: 2771 case NEON::BI__builtin_neon_vcvtp_u32_v: 2772 case NEON::BI__builtin_neon_vcvtp_u64_v: 2773 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2774 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2775 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2776 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2777 case NEON::BI__builtin_neon_vcvtm_s32_v: 2778 case NEON::BI__builtin_neon_vcvtm_s64_v: 2779 case NEON::BI__builtin_neon_vcvtm_u32_v: 2780 case NEON::BI__builtin_neon_vcvtm_u64_v: 2781 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2782 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2783 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2784 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2785 bool Double = 2786 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2787 llvm::Type *InTy = 2788 GetNeonType(this, 2789 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2790 : NeonTypeFlags::Float32, false, Quad)); 2791 llvm::Type *Tys[2] = { Ty, InTy }; 2792 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2793 } 2794 case NEON::BI__builtin_neon_vext_v: 2795 case NEON::BI__builtin_neon_vextq_v: { 2796 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2797 SmallVector<Constant*, 16> Indices; 2798 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2799 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2800 2801 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2802 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2803 Value *SV = llvm::ConstantVector::get(Indices); 2804 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2805 } 2806 case NEON::BI__builtin_neon_vfma_v: 2807 case NEON::BI__builtin_neon_vfmaq_v: { 2808 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2809 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2810 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2811 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2812 2813 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2814 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2815 } 2816 case NEON::BI__builtin_neon_vld1_v: 2817 case NEON::BI__builtin_neon_vld1q_v: 2818 Ops.push_back(Align); 2819 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2820 case NEON::BI__builtin_neon_vld2_v: 2821 case NEON::BI__builtin_neon_vld2q_v: 2822 case NEON::BI__builtin_neon_vld3_v: 2823 case NEON::BI__builtin_neon_vld3q_v: 2824 case NEON::BI__builtin_neon_vld4_v: 2825 case NEON::BI__builtin_neon_vld4q_v: { 2826 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2827 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2828 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2829 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2830 return Builder.CreateStore(Ops[1], Ops[0]); 2831 } 2832 case NEON::BI__builtin_neon_vld1_dup_v: 2833 case NEON::BI__builtin_neon_vld1q_dup_v: { 2834 Value *V = UndefValue::get(Ty); 2835 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2836 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2837 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2838 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2839 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2840 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2841 return EmitNeonSplat(Ops[0], CI); 2842 } 2843 case NEON::BI__builtin_neon_vld2_lane_v: 2844 case NEON::BI__builtin_neon_vld2q_lane_v: 2845 case NEON::BI__builtin_neon_vld3_lane_v: 2846 case NEON::BI__builtin_neon_vld3q_lane_v: 2847 case NEON::BI__builtin_neon_vld4_lane_v: 2848 case NEON::BI__builtin_neon_vld4q_lane_v: { 2849 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2850 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2851 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2852 Ops.push_back(Align); 2853 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2854 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2855 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2856 return Builder.CreateStore(Ops[1], Ops[0]); 2857 } 2858 case NEON::BI__builtin_neon_vmovl_v: { 2859 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2860 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2861 if (Usgn) 2862 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2863 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2864 } 2865 case NEON::BI__builtin_neon_vmovn_v: { 2866 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2867 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2868 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2869 } 2870 case NEON::BI__builtin_neon_vmull_v: 2871 // FIXME: the integer vmull operations could be emitted in terms of pure 2872 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2873 // hoisting the exts outside loops. Until global ISel comes along that can 2874 // see through such movement this leads to bad CodeGen. So we need an 2875 // intrinsic for now. 2876 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2877 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2878 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2879 case NEON::BI__builtin_neon_vpadal_v: 2880 case NEON::BI__builtin_neon_vpadalq_v: { 2881 // The source operand type has twice as many elements of half the size. 2882 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2883 llvm::Type *EltTy = 2884 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2885 llvm::Type *NarrowTy = 2886 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2887 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2888 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2889 } 2890 case NEON::BI__builtin_neon_vpaddl_v: 2891 case NEON::BI__builtin_neon_vpaddlq_v: { 2892 // The source operand type has twice as many elements of half the size. 2893 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2894 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2895 llvm::Type *NarrowTy = 2896 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2897 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2898 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2899 } 2900 case NEON::BI__builtin_neon_vqdmlal_v: 2901 case NEON::BI__builtin_neon_vqdmlsl_v: { 2902 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2903 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2904 MulOps, "vqdmlal"); 2905 2906 SmallVector<Value *, 2> AccumOps; 2907 AccumOps.push_back(Ops[0]); 2908 AccumOps.push_back(Mul); 2909 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2910 AccumOps, NameHint); 2911 } 2912 case NEON::BI__builtin_neon_vqshl_n_v: 2913 case NEON::BI__builtin_neon_vqshlq_n_v: 2914 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2915 1, false); 2916 case NEON::BI__builtin_neon_vqshlu_n_v: 2917 case NEON::BI__builtin_neon_vqshluq_n_v: 2918 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 2919 1, false); 2920 case NEON::BI__builtin_neon_vrecpe_v: 2921 case NEON::BI__builtin_neon_vrecpeq_v: 2922 case NEON::BI__builtin_neon_vrsqrte_v: 2923 case NEON::BI__builtin_neon_vrsqrteq_v: 2924 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2925 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2926 2927 case NEON::BI__builtin_neon_vrshr_n_v: 2928 case NEON::BI__builtin_neon_vrshrq_n_v: 2929 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 2930 1, true); 2931 case NEON::BI__builtin_neon_vshl_n_v: 2932 case NEON::BI__builtin_neon_vshlq_n_v: 2933 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2934 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2935 "vshl_n"); 2936 case NEON::BI__builtin_neon_vshll_n_v: { 2937 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2938 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2939 if (Usgn) 2940 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2941 else 2942 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2943 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2944 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2945 } 2946 case NEON::BI__builtin_neon_vshrn_n_v: { 2947 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2948 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2949 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2950 if (Usgn) 2951 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2952 else 2953 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2954 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2955 } 2956 case NEON::BI__builtin_neon_vshr_n_v: 2957 case NEON::BI__builtin_neon_vshrq_n_v: 2958 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2959 case NEON::BI__builtin_neon_vst1_v: 2960 case NEON::BI__builtin_neon_vst1q_v: 2961 case NEON::BI__builtin_neon_vst2_v: 2962 case NEON::BI__builtin_neon_vst2q_v: 2963 case NEON::BI__builtin_neon_vst3_v: 2964 case NEON::BI__builtin_neon_vst3q_v: 2965 case NEON::BI__builtin_neon_vst4_v: 2966 case NEON::BI__builtin_neon_vst4q_v: 2967 case NEON::BI__builtin_neon_vst2_lane_v: 2968 case NEON::BI__builtin_neon_vst2q_lane_v: 2969 case NEON::BI__builtin_neon_vst3_lane_v: 2970 case NEON::BI__builtin_neon_vst3q_lane_v: 2971 case NEON::BI__builtin_neon_vst4_lane_v: 2972 case NEON::BI__builtin_neon_vst4q_lane_v: 2973 Ops.push_back(Align); 2974 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 2975 case NEON::BI__builtin_neon_vsubhn_v: { 2976 llvm::VectorType *SrcTy = 2977 llvm::VectorType::getExtendedElementVectorType(VTy); 2978 2979 // %sum = add <4 x i32> %lhs, %rhs 2980 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2981 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2982 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2983 2984 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2985 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2986 SrcTy->getScalarSizeInBits() / 2); 2987 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2988 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2989 2990 // %res = trunc <4 x i32> %high to <4 x i16> 2991 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2992 } 2993 case NEON::BI__builtin_neon_vtrn_v: 2994 case NEON::BI__builtin_neon_vtrnq_v: { 2995 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2996 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2997 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2998 Value *SV = nullptr; 2999 3000 for (unsigned vi = 0; vi != 2; ++vi) { 3001 SmallVector<Constant*, 16> Indices; 3002 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3003 Indices.push_back(Builder.getInt32(i+vi)); 3004 Indices.push_back(Builder.getInt32(i+e+vi)); 3005 } 3006 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3007 SV = llvm::ConstantVector::get(Indices); 3008 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3009 SV = Builder.CreateStore(SV, Addr); 3010 } 3011 return SV; 3012 } 3013 case NEON::BI__builtin_neon_vtst_v: 3014 case NEON::BI__builtin_neon_vtstq_v: { 3015 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3016 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3017 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3018 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3019 ConstantAggregateZero::get(Ty)); 3020 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3021 } 3022 case NEON::BI__builtin_neon_vuzp_v: 3023 case NEON::BI__builtin_neon_vuzpq_v: { 3024 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3025 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3026 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3027 Value *SV = nullptr; 3028 3029 for (unsigned vi = 0; vi != 2; ++vi) { 3030 SmallVector<Constant*, 16> Indices; 3031 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3032 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3033 3034 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3035 SV = llvm::ConstantVector::get(Indices); 3036 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3037 SV = Builder.CreateStore(SV, Addr); 3038 } 3039 return SV; 3040 } 3041 case NEON::BI__builtin_neon_vzip_v: 3042 case NEON::BI__builtin_neon_vzipq_v: { 3043 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3044 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3045 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3046 Value *SV = nullptr; 3047 3048 for (unsigned vi = 0; vi != 2; ++vi) { 3049 SmallVector<Constant*, 16> Indices; 3050 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3051 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3052 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3053 } 3054 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3055 SV = llvm::ConstantVector::get(Indices); 3056 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3057 SV = Builder.CreateStore(SV, Addr); 3058 } 3059 return SV; 3060 } 3061 } 3062 3063 assert(Int && "Expected valid intrinsic number"); 3064 3065 // Determine the type(s) of this overloaded AArch64 intrinsic. 3066 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3067 3068 Value *Result = EmitNeonCall(F, Ops, NameHint); 3069 llvm::Type *ResultType = ConvertType(E->getType()); 3070 // AArch64 intrinsic one-element vector type cast to 3071 // scalar type expected by the builtin 3072 return Builder.CreateBitCast(Result, ResultType, NameHint); 3073 } 3074 3075 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3076 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3077 const CmpInst::Predicate Ip, const Twine &Name) { 3078 llvm::Type *OTy = Op->getType(); 3079 3080 // FIXME: this is utterly horrific. We should not be looking at previous 3081 // codegen context to find out what needs doing. Unfortunately TableGen 3082 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3083 // (etc). 3084 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3085 OTy = BI->getOperand(0)->getType(); 3086 3087 Op = Builder.CreateBitCast(Op, OTy); 3088 if (OTy->getScalarType()->isFloatingPointTy()) { 3089 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3090 } else { 3091 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3092 } 3093 return Builder.CreateSExt(Op, Ty, Name); 3094 } 3095 3096 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3097 Value *ExtOp, Value *IndexOp, 3098 llvm::Type *ResTy, unsigned IntID, 3099 const char *Name) { 3100 SmallVector<Value *, 2> TblOps; 3101 if (ExtOp) 3102 TblOps.push_back(ExtOp); 3103 3104 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3105 SmallVector<Constant*, 16> Indices; 3106 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3107 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3108 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3109 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3110 } 3111 Value *SV = llvm::ConstantVector::get(Indices); 3112 3113 int PairPos = 0, End = Ops.size() - 1; 3114 while (PairPos < End) { 3115 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3116 Ops[PairPos+1], SV, Name)); 3117 PairPos += 2; 3118 } 3119 3120 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3121 // of the 128-bit lookup table with zero. 3122 if (PairPos == End) { 3123 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3124 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3125 ZeroTbl, SV, Name)); 3126 } 3127 3128 Function *TblF; 3129 TblOps.push_back(IndexOp); 3130 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3131 3132 return CGF.EmitNeonCall(TblF, TblOps, Name); 3133 } 3134 3135 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3136 switch (BuiltinID) { 3137 default: 3138 return nullptr; 3139 case ARM::BI__builtin_arm_nop: 3140 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3141 llvm::ConstantInt::get(Int32Ty, 0)); 3142 case ARM::BI__builtin_arm_yield: 3143 case ARM::BI__yield: 3144 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3145 llvm::ConstantInt::get(Int32Ty, 1)); 3146 case ARM::BI__builtin_arm_wfe: 3147 case ARM::BI__wfe: 3148 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3149 llvm::ConstantInt::get(Int32Ty, 2)); 3150 case ARM::BI__builtin_arm_wfi: 3151 case ARM::BI__wfi: 3152 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3153 llvm::ConstantInt::get(Int32Ty, 3)); 3154 case ARM::BI__builtin_arm_sev: 3155 case ARM::BI__sev: 3156 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3157 llvm::ConstantInt::get(Int32Ty, 4)); 3158 case ARM::BI__builtin_arm_sevl: 3159 case ARM::BI__sevl: 3160 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3161 llvm::ConstantInt::get(Int32Ty, 5)); 3162 } 3163 } 3164 3165 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3166 const CallExpr *E) { 3167 if (auto Hint = GetValueForARMHint(BuiltinID)) 3168 return Hint; 3169 3170 if (BuiltinID == ARM::BI__emit) { 3171 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3172 llvm::FunctionType *FTy = 3173 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3174 3175 APSInt Value; 3176 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3177 llvm_unreachable("Sema will ensure that the parameter is constant"); 3178 3179 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3180 3181 llvm::InlineAsm *Emit = 3182 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3183 /*SideEffects=*/true) 3184 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3185 /*SideEffects=*/true); 3186 3187 return Builder.CreateCall(Emit); 3188 } 3189 3190 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3191 Value *Option = EmitScalarExpr(E->getArg(0)); 3192 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3193 } 3194 3195 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3196 Value *Address = EmitScalarExpr(E->getArg(0)); 3197 Value *RW = EmitScalarExpr(E->getArg(1)); 3198 Value *IsData = EmitScalarExpr(E->getArg(2)); 3199 3200 // Locality is not supported on ARM target 3201 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3202 3203 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3204 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3205 } 3206 3207 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3208 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3209 EmitScalarExpr(E->getArg(0)), 3210 "rbit"); 3211 } 3212 3213 if (BuiltinID == ARM::BI__clear_cache) { 3214 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3215 const FunctionDecl *FD = E->getDirectCallee(); 3216 SmallVector<Value*, 2> Ops; 3217 for (unsigned i = 0; i < 2; i++) 3218 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3219 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3220 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3221 StringRef Name = FD->getName(); 3222 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3223 } 3224 3225 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3226 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3227 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3228 getContext().getTypeSize(E->getType()) == 64) || 3229 BuiltinID == ARM::BI__ldrexd) { 3230 Function *F; 3231 3232 switch (BuiltinID) { 3233 default: llvm_unreachable("unexpected builtin"); 3234 case ARM::BI__builtin_arm_ldaex: 3235 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3236 break; 3237 case ARM::BI__builtin_arm_ldrexd: 3238 case ARM::BI__builtin_arm_ldrex: 3239 case ARM::BI__ldrexd: 3240 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3241 break; 3242 } 3243 3244 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3245 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3246 "ldrexd"); 3247 3248 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3249 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3250 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3251 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3252 3253 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3254 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3255 Val = Builder.CreateOr(Val, Val1); 3256 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3257 } 3258 3259 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3260 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3261 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3262 3263 QualType Ty = E->getType(); 3264 llvm::Type *RealResTy = ConvertType(Ty); 3265 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3266 getContext().getTypeSize(Ty)); 3267 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3268 3269 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3270 ? Intrinsic::arm_ldaex 3271 : Intrinsic::arm_ldrex, 3272 LoadAddr->getType()); 3273 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3274 3275 if (RealResTy->isPointerTy()) 3276 return Builder.CreateIntToPtr(Val, RealResTy); 3277 else { 3278 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3279 return Builder.CreateBitCast(Val, RealResTy); 3280 } 3281 } 3282 3283 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3284 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3285 BuiltinID == ARM::BI__builtin_arm_strex) && 3286 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3287 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3288 ? Intrinsic::arm_stlexd 3289 : Intrinsic::arm_strexd); 3290 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3291 3292 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3293 Value *Val = EmitScalarExpr(E->getArg(0)); 3294 Builder.CreateStore(Val, Tmp); 3295 3296 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3297 Val = Builder.CreateLoad(LdPtr); 3298 3299 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3300 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3301 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3302 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3303 } 3304 3305 if (BuiltinID == ARM::BI__builtin_arm_strex || 3306 BuiltinID == ARM::BI__builtin_arm_stlex) { 3307 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3308 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3309 3310 QualType Ty = E->getArg(0)->getType(); 3311 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3312 getContext().getTypeSize(Ty)); 3313 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3314 3315 if (StoreVal->getType()->isPointerTy()) 3316 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3317 else { 3318 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3319 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3320 } 3321 3322 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3323 ? Intrinsic::arm_stlex 3324 : Intrinsic::arm_strex, 3325 StoreAddr->getType()); 3326 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3327 } 3328 3329 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3330 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3331 return Builder.CreateCall(F); 3332 } 3333 3334 // CRC32 3335 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3336 switch (BuiltinID) { 3337 case ARM::BI__builtin_arm_crc32b: 3338 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3339 case ARM::BI__builtin_arm_crc32cb: 3340 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3341 case ARM::BI__builtin_arm_crc32h: 3342 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3343 case ARM::BI__builtin_arm_crc32ch: 3344 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3345 case ARM::BI__builtin_arm_crc32w: 3346 case ARM::BI__builtin_arm_crc32d: 3347 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3348 case ARM::BI__builtin_arm_crc32cw: 3349 case ARM::BI__builtin_arm_crc32cd: 3350 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3351 } 3352 3353 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3354 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3355 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3356 3357 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3358 // intrinsics, hence we need different codegen for these cases. 3359 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3360 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3361 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3362 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3363 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3364 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3365 3366 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3367 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3368 return Builder.CreateCall2(F, Res, Arg1b); 3369 } else { 3370 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3371 3372 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3373 return Builder.CreateCall2(F, Arg0, Arg1); 3374 } 3375 } 3376 3377 SmallVector<Value*, 4> Ops; 3378 llvm::Value *Align = nullptr; 3379 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3380 if (i == 0) { 3381 switch (BuiltinID) { 3382 case NEON::BI__builtin_neon_vld1_v: 3383 case NEON::BI__builtin_neon_vld1q_v: 3384 case NEON::BI__builtin_neon_vld1q_lane_v: 3385 case NEON::BI__builtin_neon_vld1_lane_v: 3386 case NEON::BI__builtin_neon_vld1_dup_v: 3387 case NEON::BI__builtin_neon_vld1q_dup_v: 3388 case NEON::BI__builtin_neon_vst1_v: 3389 case NEON::BI__builtin_neon_vst1q_v: 3390 case NEON::BI__builtin_neon_vst1q_lane_v: 3391 case NEON::BI__builtin_neon_vst1_lane_v: 3392 case NEON::BI__builtin_neon_vst2_v: 3393 case NEON::BI__builtin_neon_vst2q_v: 3394 case NEON::BI__builtin_neon_vst2_lane_v: 3395 case NEON::BI__builtin_neon_vst2q_lane_v: 3396 case NEON::BI__builtin_neon_vst3_v: 3397 case NEON::BI__builtin_neon_vst3q_v: 3398 case NEON::BI__builtin_neon_vst3_lane_v: 3399 case NEON::BI__builtin_neon_vst3q_lane_v: 3400 case NEON::BI__builtin_neon_vst4_v: 3401 case NEON::BI__builtin_neon_vst4q_v: 3402 case NEON::BI__builtin_neon_vst4_lane_v: 3403 case NEON::BI__builtin_neon_vst4q_lane_v: 3404 // Get the alignment for the argument in addition to the value; 3405 // we'll use it later. 3406 std::pair<llvm::Value*, unsigned> Src = 3407 EmitPointerWithAlignment(E->getArg(0)); 3408 Ops.push_back(Src.first); 3409 Align = Builder.getInt32(Src.second); 3410 continue; 3411 } 3412 } 3413 if (i == 1) { 3414 switch (BuiltinID) { 3415 case NEON::BI__builtin_neon_vld2_v: 3416 case NEON::BI__builtin_neon_vld2q_v: 3417 case NEON::BI__builtin_neon_vld3_v: 3418 case NEON::BI__builtin_neon_vld3q_v: 3419 case NEON::BI__builtin_neon_vld4_v: 3420 case NEON::BI__builtin_neon_vld4q_v: 3421 case NEON::BI__builtin_neon_vld2_lane_v: 3422 case NEON::BI__builtin_neon_vld2q_lane_v: 3423 case NEON::BI__builtin_neon_vld3_lane_v: 3424 case NEON::BI__builtin_neon_vld3q_lane_v: 3425 case NEON::BI__builtin_neon_vld4_lane_v: 3426 case NEON::BI__builtin_neon_vld4q_lane_v: 3427 case NEON::BI__builtin_neon_vld2_dup_v: 3428 case NEON::BI__builtin_neon_vld3_dup_v: 3429 case NEON::BI__builtin_neon_vld4_dup_v: 3430 // Get the alignment for the argument in addition to the value; 3431 // we'll use it later. 3432 std::pair<llvm::Value*, unsigned> Src = 3433 EmitPointerWithAlignment(E->getArg(1)); 3434 Ops.push_back(Src.first); 3435 Align = Builder.getInt32(Src.second); 3436 continue; 3437 } 3438 } 3439 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3440 } 3441 3442 switch (BuiltinID) { 3443 default: break; 3444 // vget_lane and vset_lane are not overloaded and do not have an extra 3445 // argument that specifies the vector type. 3446 case NEON::BI__builtin_neon_vget_lane_i8: 3447 case NEON::BI__builtin_neon_vget_lane_i16: 3448 case NEON::BI__builtin_neon_vget_lane_i32: 3449 case NEON::BI__builtin_neon_vget_lane_i64: 3450 case NEON::BI__builtin_neon_vget_lane_f32: 3451 case NEON::BI__builtin_neon_vgetq_lane_i8: 3452 case NEON::BI__builtin_neon_vgetq_lane_i16: 3453 case NEON::BI__builtin_neon_vgetq_lane_i32: 3454 case NEON::BI__builtin_neon_vgetq_lane_i64: 3455 case NEON::BI__builtin_neon_vgetq_lane_f32: 3456 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3457 "vget_lane"); 3458 case NEON::BI__builtin_neon_vset_lane_i8: 3459 case NEON::BI__builtin_neon_vset_lane_i16: 3460 case NEON::BI__builtin_neon_vset_lane_i32: 3461 case NEON::BI__builtin_neon_vset_lane_i64: 3462 case NEON::BI__builtin_neon_vset_lane_f32: 3463 case NEON::BI__builtin_neon_vsetq_lane_i8: 3464 case NEON::BI__builtin_neon_vsetq_lane_i16: 3465 case NEON::BI__builtin_neon_vsetq_lane_i32: 3466 case NEON::BI__builtin_neon_vsetq_lane_i64: 3467 case NEON::BI__builtin_neon_vsetq_lane_f32: 3468 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3469 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3470 3471 // Non-polymorphic crypto instructions also not overloaded 3472 case NEON::BI__builtin_neon_vsha1h_u32: 3473 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3474 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3475 "vsha1h"); 3476 case NEON::BI__builtin_neon_vsha1cq_u32: 3477 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3478 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3479 "vsha1h"); 3480 case NEON::BI__builtin_neon_vsha1pq_u32: 3481 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3482 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3483 "vsha1h"); 3484 case NEON::BI__builtin_neon_vsha1mq_u32: 3485 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3486 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3487 "vsha1h"); 3488 } 3489 3490 // Get the last argument, which specifies the vector type. 3491 llvm::APSInt Result; 3492 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3493 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3494 return nullptr; 3495 3496 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3497 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3498 // Determine the overloaded type of this builtin. 3499 llvm::Type *Ty; 3500 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3501 Ty = FloatTy; 3502 else 3503 Ty = DoubleTy; 3504 3505 // Determine whether this is an unsigned conversion or not. 3506 bool usgn = Result.getZExtValue() == 1; 3507 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3508 3509 // Call the appropriate intrinsic. 3510 Function *F = CGM.getIntrinsic(Int, Ty); 3511 return Builder.CreateCall(F, Ops, "vcvtr"); 3512 } 3513 3514 // Determine the type of this overloaded NEON intrinsic. 3515 NeonTypeFlags Type(Result.getZExtValue()); 3516 bool usgn = Type.isUnsigned(); 3517 bool rightShift = false; 3518 3519 llvm::VectorType *VTy = GetNeonType(this, Type); 3520 llvm::Type *Ty = VTy; 3521 if (!Ty) 3522 return nullptr; 3523 3524 // Many NEON builtins have identical semantics and uses in ARM and 3525 // AArch64. Emit these in a single function. 3526 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3527 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3528 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3529 if (Builtin) 3530 return EmitCommonNeonBuiltinExpr( 3531 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3532 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3533 3534 unsigned Int; 3535 switch (BuiltinID) { 3536 default: return nullptr; 3537 case NEON::BI__builtin_neon_vld1q_lane_v: 3538 // Handle 64-bit integer elements as a special case. Use shuffles of 3539 // one-element vectors to avoid poor code for i64 in the backend. 3540 if (VTy->getElementType()->isIntegerTy(64)) { 3541 // Extract the other lane. 3542 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3543 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3544 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3545 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3546 // Load the value as a one-element vector. 3547 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3548 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3549 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3550 // Combine them. 3551 SmallVector<Constant*, 2> Indices; 3552 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3553 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3554 SV = llvm::ConstantVector::get(Indices); 3555 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3556 } 3557 // fall through 3558 case NEON::BI__builtin_neon_vld1_lane_v: { 3559 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3560 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3561 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3562 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3563 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3564 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3565 } 3566 case NEON::BI__builtin_neon_vld2_dup_v: 3567 case NEON::BI__builtin_neon_vld3_dup_v: 3568 case NEON::BI__builtin_neon_vld4_dup_v: { 3569 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3570 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3571 switch (BuiltinID) { 3572 case NEON::BI__builtin_neon_vld2_dup_v: 3573 Int = Intrinsic::arm_neon_vld2; 3574 break; 3575 case NEON::BI__builtin_neon_vld3_dup_v: 3576 Int = Intrinsic::arm_neon_vld3; 3577 break; 3578 case NEON::BI__builtin_neon_vld4_dup_v: 3579 Int = Intrinsic::arm_neon_vld4; 3580 break; 3581 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3582 } 3583 Function *F = CGM.getIntrinsic(Int, Ty); 3584 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3585 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3586 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3587 return Builder.CreateStore(Ops[1], Ops[0]); 3588 } 3589 switch (BuiltinID) { 3590 case NEON::BI__builtin_neon_vld2_dup_v: 3591 Int = Intrinsic::arm_neon_vld2lane; 3592 break; 3593 case NEON::BI__builtin_neon_vld3_dup_v: 3594 Int = Intrinsic::arm_neon_vld3lane; 3595 break; 3596 case NEON::BI__builtin_neon_vld4_dup_v: 3597 Int = Intrinsic::arm_neon_vld4lane; 3598 break; 3599 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3600 } 3601 Function *F = CGM.getIntrinsic(Int, Ty); 3602 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3603 3604 SmallVector<Value*, 6> Args; 3605 Args.push_back(Ops[1]); 3606 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3607 3608 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3609 Args.push_back(CI); 3610 Args.push_back(Align); 3611 3612 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3613 // splat lane 0 to all elts in each vector of the result. 3614 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3615 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3616 Value *Elt = Builder.CreateBitCast(Val, Ty); 3617 Elt = EmitNeonSplat(Elt, CI); 3618 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3619 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3620 } 3621 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3622 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3623 return Builder.CreateStore(Ops[1], Ops[0]); 3624 } 3625 case NEON::BI__builtin_neon_vqrshrn_n_v: 3626 Int = 3627 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3628 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3629 1, true); 3630 case NEON::BI__builtin_neon_vqrshrun_n_v: 3631 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3632 Ops, "vqrshrun_n", 1, true); 3633 case NEON::BI__builtin_neon_vqshrn_n_v: 3634 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3635 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3636 1, true); 3637 case NEON::BI__builtin_neon_vqshrun_n_v: 3638 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3639 Ops, "vqshrun_n", 1, true); 3640 case NEON::BI__builtin_neon_vrecpe_v: 3641 case NEON::BI__builtin_neon_vrecpeq_v: 3642 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3643 Ops, "vrecpe"); 3644 case NEON::BI__builtin_neon_vrshrn_n_v: 3645 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3646 Ops, "vrshrn_n", 1, true); 3647 case NEON::BI__builtin_neon_vrsra_n_v: 3648 case NEON::BI__builtin_neon_vrsraq_n_v: 3649 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3650 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3651 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3652 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3653 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3654 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3655 case NEON::BI__builtin_neon_vsri_n_v: 3656 case NEON::BI__builtin_neon_vsriq_n_v: 3657 rightShift = true; 3658 case NEON::BI__builtin_neon_vsli_n_v: 3659 case NEON::BI__builtin_neon_vsliq_n_v: 3660 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3661 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3662 Ops, "vsli_n"); 3663 case NEON::BI__builtin_neon_vsra_n_v: 3664 case NEON::BI__builtin_neon_vsraq_n_v: 3665 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3666 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3667 return Builder.CreateAdd(Ops[0], Ops[1]); 3668 case NEON::BI__builtin_neon_vst1q_lane_v: 3669 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3670 // a one-element vector and avoid poor code for i64 in the backend. 3671 if (VTy->getElementType()->isIntegerTy(64)) { 3672 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3673 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3674 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3675 Ops[2] = Align; 3676 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3677 Ops[1]->getType()), Ops); 3678 } 3679 // fall through 3680 case NEON::BI__builtin_neon_vst1_lane_v: { 3681 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3682 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3683 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3684 StoreInst *St = Builder.CreateStore(Ops[1], 3685 Builder.CreateBitCast(Ops[0], Ty)); 3686 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3687 return St; 3688 } 3689 case NEON::BI__builtin_neon_vtbl1_v: 3690 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3691 Ops, "vtbl1"); 3692 case NEON::BI__builtin_neon_vtbl2_v: 3693 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3694 Ops, "vtbl2"); 3695 case NEON::BI__builtin_neon_vtbl3_v: 3696 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3697 Ops, "vtbl3"); 3698 case NEON::BI__builtin_neon_vtbl4_v: 3699 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3700 Ops, "vtbl4"); 3701 case NEON::BI__builtin_neon_vtbx1_v: 3702 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3703 Ops, "vtbx1"); 3704 case NEON::BI__builtin_neon_vtbx2_v: 3705 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3706 Ops, "vtbx2"); 3707 case NEON::BI__builtin_neon_vtbx3_v: 3708 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3709 Ops, "vtbx3"); 3710 case NEON::BI__builtin_neon_vtbx4_v: 3711 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3712 Ops, "vtbx4"); 3713 } 3714 } 3715 3716 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3717 const CallExpr *E, 3718 SmallVectorImpl<Value *> &Ops) { 3719 unsigned int Int = 0; 3720 const char *s = nullptr; 3721 3722 switch (BuiltinID) { 3723 default: 3724 return nullptr; 3725 case NEON::BI__builtin_neon_vtbl1_v: 3726 case NEON::BI__builtin_neon_vqtbl1_v: 3727 case NEON::BI__builtin_neon_vqtbl1q_v: 3728 case NEON::BI__builtin_neon_vtbl2_v: 3729 case NEON::BI__builtin_neon_vqtbl2_v: 3730 case NEON::BI__builtin_neon_vqtbl2q_v: 3731 case NEON::BI__builtin_neon_vtbl3_v: 3732 case NEON::BI__builtin_neon_vqtbl3_v: 3733 case NEON::BI__builtin_neon_vqtbl3q_v: 3734 case NEON::BI__builtin_neon_vtbl4_v: 3735 case NEON::BI__builtin_neon_vqtbl4_v: 3736 case NEON::BI__builtin_neon_vqtbl4q_v: 3737 break; 3738 case NEON::BI__builtin_neon_vtbx1_v: 3739 case NEON::BI__builtin_neon_vqtbx1_v: 3740 case NEON::BI__builtin_neon_vqtbx1q_v: 3741 case NEON::BI__builtin_neon_vtbx2_v: 3742 case NEON::BI__builtin_neon_vqtbx2_v: 3743 case NEON::BI__builtin_neon_vqtbx2q_v: 3744 case NEON::BI__builtin_neon_vtbx3_v: 3745 case NEON::BI__builtin_neon_vqtbx3_v: 3746 case NEON::BI__builtin_neon_vqtbx3q_v: 3747 case NEON::BI__builtin_neon_vtbx4_v: 3748 case NEON::BI__builtin_neon_vqtbx4_v: 3749 case NEON::BI__builtin_neon_vqtbx4q_v: 3750 break; 3751 } 3752 3753 assert(E->getNumArgs() >= 3); 3754 3755 // Get the last argument, which specifies the vector type. 3756 llvm::APSInt Result; 3757 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3758 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3759 return nullptr; 3760 3761 // Determine the type of this overloaded NEON intrinsic. 3762 NeonTypeFlags Type(Result.getZExtValue()); 3763 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3764 llvm::Type *Ty = VTy; 3765 if (!Ty) 3766 return nullptr; 3767 3768 unsigned nElts = VTy->getNumElements(); 3769 3770 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3771 3772 // AArch64 scalar builtins are not overloaded, they do not have an extra 3773 // argument that specifies the vector type, need to handle each case. 3774 SmallVector<Value *, 2> TblOps; 3775 switch (BuiltinID) { 3776 case NEON::BI__builtin_neon_vtbl1_v: { 3777 TblOps.push_back(Ops[0]); 3778 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3779 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3780 } 3781 case NEON::BI__builtin_neon_vtbl2_v: { 3782 TblOps.push_back(Ops[0]); 3783 TblOps.push_back(Ops[1]); 3784 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3785 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3786 } 3787 case NEON::BI__builtin_neon_vtbl3_v: { 3788 TblOps.push_back(Ops[0]); 3789 TblOps.push_back(Ops[1]); 3790 TblOps.push_back(Ops[2]); 3791 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3792 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3793 } 3794 case NEON::BI__builtin_neon_vtbl4_v: { 3795 TblOps.push_back(Ops[0]); 3796 TblOps.push_back(Ops[1]); 3797 TblOps.push_back(Ops[2]); 3798 TblOps.push_back(Ops[3]); 3799 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3800 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3801 } 3802 case NEON::BI__builtin_neon_vtbx1_v: { 3803 TblOps.push_back(Ops[1]); 3804 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3805 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3806 3807 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3808 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3809 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3810 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3811 3812 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3813 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3814 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3815 } 3816 case NEON::BI__builtin_neon_vtbx2_v: { 3817 TblOps.push_back(Ops[1]); 3818 TblOps.push_back(Ops[2]); 3819 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3820 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3821 } 3822 case NEON::BI__builtin_neon_vtbx3_v: { 3823 TblOps.push_back(Ops[1]); 3824 TblOps.push_back(Ops[2]); 3825 TblOps.push_back(Ops[3]); 3826 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3827 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3828 3829 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3830 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3831 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3832 TwentyFourV); 3833 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3834 3835 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3836 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3837 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3838 } 3839 case NEON::BI__builtin_neon_vtbx4_v: { 3840 TblOps.push_back(Ops[1]); 3841 TblOps.push_back(Ops[2]); 3842 TblOps.push_back(Ops[3]); 3843 TblOps.push_back(Ops[4]); 3844 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3845 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3846 } 3847 case NEON::BI__builtin_neon_vqtbl1_v: 3848 case NEON::BI__builtin_neon_vqtbl1q_v: 3849 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3850 case NEON::BI__builtin_neon_vqtbl2_v: 3851 case NEON::BI__builtin_neon_vqtbl2q_v: { 3852 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3853 case NEON::BI__builtin_neon_vqtbl3_v: 3854 case NEON::BI__builtin_neon_vqtbl3q_v: 3855 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3856 case NEON::BI__builtin_neon_vqtbl4_v: 3857 case NEON::BI__builtin_neon_vqtbl4q_v: 3858 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3859 case NEON::BI__builtin_neon_vqtbx1_v: 3860 case NEON::BI__builtin_neon_vqtbx1q_v: 3861 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3862 case NEON::BI__builtin_neon_vqtbx2_v: 3863 case NEON::BI__builtin_neon_vqtbx2q_v: 3864 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3865 case NEON::BI__builtin_neon_vqtbx3_v: 3866 case NEON::BI__builtin_neon_vqtbx3q_v: 3867 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3868 case NEON::BI__builtin_neon_vqtbx4_v: 3869 case NEON::BI__builtin_neon_vqtbx4q_v: 3870 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3871 } 3872 } 3873 3874 if (!Int) 3875 return nullptr; 3876 3877 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3878 return CGF.EmitNeonCall(F, Ops, s); 3879 } 3880 3881 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3882 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3883 Op = Builder.CreateBitCast(Op, Int16Ty); 3884 Value *V = UndefValue::get(VTy); 3885 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3886 Op = Builder.CreateInsertElement(V, Op, CI); 3887 return Op; 3888 } 3889 3890 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3891 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3892 Op = Builder.CreateBitCast(Op, Int8Ty); 3893 Value *V = UndefValue::get(VTy); 3894 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3895 Op = Builder.CreateInsertElement(V, Op, CI); 3896 return Op; 3897 } 3898 3899 Value *CodeGenFunction:: 3900 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3901 const char *Name) { 3902 // i8 is not a legal types for AArch64, so we can't just use 3903 // a normal overloaded intrinsic call for these scalar types. Instead 3904 // we'll build 64-bit vectors w/ lane zero being our input values and 3905 // perform the operation on that. The back end can pattern match directly 3906 // to the scalar instruction. 3907 Ops[0] = vectorWrapScalar8(Ops[0]); 3908 Ops[1] = vectorWrapScalar8(Ops[1]); 3909 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3910 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3911 Constant *CI = ConstantInt::get(SizeTy, 0); 3912 return Builder.CreateExtractElement(V, CI, "lane0"); 3913 } 3914 3915 Value *CodeGenFunction:: 3916 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3917 const char *Name) { 3918 // i16 is not a legal types for AArch64, so we can't just use 3919 // a normal overloaded intrinsic call for these scalar types. Instead 3920 // we'll build 64-bit vectors w/ lane zero being our input values and 3921 // perform the operation on that. The back end can pattern match directly 3922 // to the scalar instruction. 3923 Ops[0] = vectorWrapScalar16(Ops[0]); 3924 Ops[1] = vectorWrapScalar16(Ops[1]); 3925 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3926 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3927 Constant *CI = ConstantInt::get(SizeTy, 0); 3928 return Builder.CreateExtractElement(V, CI, "lane0"); 3929 } 3930 3931 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3932 const CallExpr *E) { 3933 unsigned HintID = static_cast<unsigned>(-1); 3934 switch (BuiltinID) { 3935 default: break; 3936 case AArch64::BI__builtin_arm_nop: 3937 HintID = 0; 3938 break; 3939 case AArch64::BI__builtin_arm_yield: 3940 HintID = 1; 3941 break; 3942 case AArch64::BI__builtin_arm_wfe: 3943 HintID = 2; 3944 break; 3945 case AArch64::BI__builtin_arm_wfi: 3946 HintID = 3; 3947 break; 3948 case AArch64::BI__builtin_arm_sev: 3949 HintID = 4; 3950 break; 3951 case AArch64::BI__builtin_arm_sevl: 3952 HintID = 5; 3953 break; 3954 } 3955 3956 if (HintID != static_cast<unsigned>(-1)) { 3957 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 3958 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 3959 } 3960 3961 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 3962 Value *Address = EmitScalarExpr(E->getArg(0)); 3963 Value *RW = EmitScalarExpr(E->getArg(1)); 3964 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 3965 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 3966 Value *IsData = EmitScalarExpr(E->getArg(4)); 3967 3968 Value *Locality = nullptr; 3969 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 3970 // Temporal fetch, needs to convert cache level to locality. 3971 Locality = llvm::ConstantInt::get(Int32Ty, 3972 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 3973 } else { 3974 // Streaming fetch. 3975 Locality = llvm::ConstantInt::get(Int32Ty, 0); 3976 } 3977 3978 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 3979 // PLDL3STRM or PLDL2STRM. 3980 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3981 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3982 } 3983 3984 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 3985 assert((getContext().getTypeSize(E->getType()) == 32) && 3986 "rbit of unusual size!"); 3987 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3988 return Builder.CreateCall( 3989 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3990 } 3991 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 3992 assert((getContext().getTypeSize(E->getType()) == 64) && 3993 "rbit of unusual size!"); 3994 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 3995 return Builder.CreateCall( 3996 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 3997 } 3998 3999 if (BuiltinID == AArch64::BI__clear_cache) { 4000 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4001 const FunctionDecl *FD = E->getDirectCallee(); 4002 SmallVector<Value*, 2> Ops; 4003 for (unsigned i = 0; i < 2; i++) 4004 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4005 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4006 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4007 StringRef Name = FD->getName(); 4008 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4009 } 4010 4011 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4012 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4013 getContext().getTypeSize(E->getType()) == 128) { 4014 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4015 ? Intrinsic::aarch64_ldaxp 4016 : Intrinsic::aarch64_ldxp); 4017 4018 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4019 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4020 "ldxp"); 4021 4022 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4023 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4024 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4025 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4026 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4027 4028 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4029 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4030 Val = Builder.CreateOr(Val, Val1); 4031 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4032 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4033 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4034 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4035 4036 QualType Ty = E->getType(); 4037 llvm::Type *RealResTy = ConvertType(Ty); 4038 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4039 getContext().getTypeSize(Ty)); 4040 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4041 4042 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4043 ? Intrinsic::aarch64_ldaxr 4044 : Intrinsic::aarch64_ldxr, 4045 LoadAddr->getType()); 4046 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4047 4048 if (RealResTy->isPointerTy()) 4049 return Builder.CreateIntToPtr(Val, RealResTy); 4050 4051 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4052 return Builder.CreateBitCast(Val, RealResTy); 4053 } 4054 4055 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4056 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4057 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4058 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4059 ? Intrinsic::aarch64_stlxp 4060 : Intrinsic::aarch64_stxp); 4061 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4062 4063 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4064 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4065 One); 4066 Value *Val = EmitScalarExpr(E->getArg(0)); 4067 Builder.CreateStore(Val, Tmp); 4068 4069 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4070 Val = Builder.CreateLoad(LdPtr); 4071 4072 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4073 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4074 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4075 Int8PtrTy); 4076 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 4077 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 4078 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4079 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4080 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4081 4082 QualType Ty = E->getArg(0)->getType(); 4083 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4084 getContext().getTypeSize(Ty)); 4085 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4086 4087 if (StoreVal->getType()->isPointerTy()) 4088 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4089 else { 4090 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4091 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4092 } 4093 4094 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4095 ? Intrinsic::aarch64_stlxr 4096 : Intrinsic::aarch64_stxr, 4097 StoreAddr->getType()); 4098 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4099 } 4100 4101 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4102 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4103 return Builder.CreateCall(F); 4104 } 4105 4106 // CRC32 4107 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4108 switch (BuiltinID) { 4109 case AArch64::BI__builtin_arm_crc32b: 4110 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4111 case AArch64::BI__builtin_arm_crc32cb: 4112 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4113 case AArch64::BI__builtin_arm_crc32h: 4114 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4115 case AArch64::BI__builtin_arm_crc32ch: 4116 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4117 case AArch64::BI__builtin_arm_crc32w: 4118 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4119 case AArch64::BI__builtin_arm_crc32cw: 4120 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4121 case AArch64::BI__builtin_arm_crc32d: 4122 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4123 case AArch64::BI__builtin_arm_crc32cd: 4124 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4125 } 4126 4127 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4128 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4129 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4130 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4131 4132 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4133 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4134 4135 return Builder.CreateCall2(F, Arg0, Arg1); 4136 } 4137 4138 llvm::SmallVector<Value*, 4> Ops; 4139 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4140 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4141 4142 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4143 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4144 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4145 4146 if (Builtin) { 4147 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4148 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4149 assert(Result && "SISD intrinsic should have been handled"); 4150 return Result; 4151 } 4152 4153 llvm::APSInt Result; 4154 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4155 NeonTypeFlags Type(0); 4156 if (Arg->isIntegerConstantExpr(Result, getContext())) 4157 // Determine the type of this overloaded NEON intrinsic. 4158 Type = NeonTypeFlags(Result.getZExtValue()); 4159 4160 bool usgn = Type.isUnsigned(); 4161 bool quad = Type.isQuad(); 4162 4163 // Handle non-overloaded intrinsics first. 4164 switch (BuiltinID) { 4165 default: break; 4166 case NEON::BI__builtin_neon_vldrq_p128: { 4167 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4168 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4169 return Builder.CreateLoad(Ptr); 4170 } 4171 case NEON::BI__builtin_neon_vstrq_p128: { 4172 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4173 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4174 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4175 } 4176 case NEON::BI__builtin_neon_vcvts_u32_f32: 4177 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4178 usgn = true; 4179 // FALL THROUGH 4180 case NEON::BI__builtin_neon_vcvts_s32_f32: 4181 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4182 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4183 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4184 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4185 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4186 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4187 if (usgn) 4188 return Builder.CreateFPToUI(Ops[0], InTy); 4189 return Builder.CreateFPToSI(Ops[0], InTy); 4190 } 4191 case NEON::BI__builtin_neon_vcvts_f32_u32: 4192 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4193 usgn = true; 4194 // FALL THROUGH 4195 case NEON::BI__builtin_neon_vcvts_f32_s32: 4196 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4197 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4198 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4199 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4200 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4201 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4202 if (usgn) 4203 return Builder.CreateUIToFP(Ops[0], FTy); 4204 return Builder.CreateSIToFP(Ops[0], FTy); 4205 } 4206 case NEON::BI__builtin_neon_vpaddd_s64: { 4207 llvm::Type *Ty = 4208 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4209 Value *Vec = EmitScalarExpr(E->getArg(0)); 4210 // The vector is v2f64, so make sure it's bitcast to that. 4211 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4212 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4213 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4214 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4215 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4216 // Pairwise addition of a v2f64 into a scalar f64. 4217 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4218 } 4219 case NEON::BI__builtin_neon_vpaddd_f64: { 4220 llvm::Type *Ty = 4221 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4222 Value *Vec = EmitScalarExpr(E->getArg(0)); 4223 // The vector is v2f64, so make sure it's bitcast to that. 4224 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4225 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4226 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4227 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4228 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4229 // Pairwise addition of a v2f64 into a scalar f64. 4230 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4231 } 4232 case NEON::BI__builtin_neon_vpadds_f32: { 4233 llvm::Type *Ty = 4234 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4235 Value *Vec = EmitScalarExpr(E->getArg(0)); 4236 // The vector is v2f32, so make sure it's bitcast to that. 4237 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4238 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4239 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4240 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4241 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4242 // Pairwise addition of a v2f32 into a scalar f32. 4243 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4244 } 4245 case NEON::BI__builtin_neon_vceqzd_s64: 4246 case NEON::BI__builtin_neon_vceqzd_f64: 4247 case NEON::BI__builtin_neon_vceqzs_f32: 4248 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4249 return EmitAArch64CompareBuiltinExpr( 4250 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4251 ICmpInst::ICMP_EQ, "vceqz"); 4252 case NEON::BI__builtin_neon_vcgezd_s64: 4253 case NEON::BI__builtin_neon_vcgezd_f64: 4254 case NEON::BI__builtin_neon_vcgezs_f32: 4255 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4256 return EmitAArch64CompareBuiltinExpr( 4257 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4258 ICmpInst::ICMP_SGE, "vcgez"); 4259 case NEON::BI__builtin_neon_vclezd_s64: 4260 case NEON::BI__builtin_neon_vclezd_f64: 4261 case NEON::BI__builtin_neon_vclezs_f32: 4262 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4263 return EmitAArch64CompareBuiltinExpr( 4264 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4265 ICmpInst::ICMP_SLE, "vclez"); 4266 case NEON::BI__builtin_neon_vcgtzd_s64: 4267 case NEON::BI__builtin_neon_vcgtzd_f64: 4268 case NEON::BI__builtin_neon_vcgtzs_f32: 4269 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4270 return EmitAArch64CompareBuiltinExpr( 4271 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4272 ICmpInst::ICMP_SGT, "vcgtz"); 4273 case NEON::BI__builtin_neon_vcltzd_s64: 4274 case NEON::BI__builtin_neon_vcltzd_f64: 4275 case NEON::BI__builtin_neon_vcltzs_f32: 4276 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4277 return EmitAArch64CompareBuiltinExpr( 4278 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4279 ICmpInst::ICMP_SLT, "vcltz"); 4280 4281 case NEON::BI__builtin_neon_vceqzd_u64: { 4282 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4283 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4284 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4285 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4286 llvm::Constant::getNullValue(Ty)); 4287 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4288 } 4289 case NEON::BI__builtin_neon_vceqd_f64: 4290 case NEON::BI__builtin_neon_vcled_f64: 4291 case NEON::BI__builtin_neon_vcltd_f64: 4292 case NEON::BI__builtin_neon_vcged_f64: 4293 case NEON::BI__builtin_neon_vcgtd_f64: { 4294 llvm::CmpInst::Predicate P; 4295 switch (BuiltinID) { 4296 default: llvm_unreachable("missing builtin ID in switch!"); 4297 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4298 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4299 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4300 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4301 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4302 } 4303 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4304 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4305 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4306 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4307 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4308 } 4309 case NEON::BI__builtin_neon_vceqs_f32: 4310 case NEON::BI__builtin_neon_vcles_f32: 4311 case NEON::BI__builtin_neon_vclts_f32: 4312 case NEON::BI__builtin_neon_vcges_f32: 4313 case NEON::BI__builtin_neon_vcgts_f32: { 4314 llvm::CmpInst::Predicate P; 4315 switch (BuiltinID) { 4316 default: llvm_unreachable("missing builtin ID in switch!"); 4317 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4318 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4319 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4320 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4321 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4322 } 4323 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4324 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4325 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4326 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4327 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4328 } 4329 case NEON::BI__builtin_neon_vceqd_s64: 4330 case NEON::BI__builtin_neon_vceqd_u64: 4331 case NEON::BI__builtin_neon_vcgtd_s64: 4332 case NEON::BI__builtin_neon_vcgtd_u64: 4333 case NEON::BI__builtin_neon_vcltd_s64: 4334 case NEON::BI__builtin_neon_vcltd_u64: 4335 case NEON::BI__builtin_neon_vcged_u64: 4336 case NEON::BI__builtin_neon_vcged_s64: 4337 case NEON::BI__builtin_neon_vcled_u64: 4338 case NEON::BI__builtin_neon_vcled_s64: { 4339 llvm::CmpInst::Predicate P; 4340 switch (BuiltinID) { 4341 default: llvm_unreachable("missing builtin ID in switch!"); 4342 case NEON::BI__builtin_neon_vceqd_s64: 4343 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4344 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4345 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4346 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4347 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4348 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4349 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4350 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4351 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4352 } 4353 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4354 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4355 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4356 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4357 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4358 } 4359 case NEON::BI__builtin_neon_vtstd_s64: 4360 case NEON::BI__builtin_neon_vtstd_u64: { 4361 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4362 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4363 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4364 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4365 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4366 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4367 llvm::Constant::getNullValue(Ty)); 4368 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4369 } 4370 case NEON::BI__builtin_neon_vset_lane_i8: 4371 case NEON::BI__builtin_neon_vset_lane_i16: 4372 case NEON::BI__builtin_neon_vset_lane_i32: 4373 case NEON::BI__builtin_neon_vset_lane_i64: 4374 case NEON::BI__builtin_neon_vset_lane_f32: 4375 case NEON::BI__builtin_neon_vsetq_lane_i8: 4376 case NEON::BI__builtin_neon_vsetq_lane_i16: 4377 case NEON::BI__builtin_neon_vsetq_lane_i32: 4378 case NEON::BI__builtin_neon_vsetq_lane_i64: 4379 case NEON::BI__builtin_neon_vsetq_lane_f32: 4380 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4381 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4382 case NEON::BI__builtin_neon_vset_lane_f64: 4383 // The vector type needs a cast for the v1f64 variant. 4384 Ops[1] = Builder.CreateBitCast(Ops[1], 4385 llvm::VectorType::get(DoubleTy, 1)); 4386 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4387 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4388 case NEON::BI__builtin_neon_vsetq_lane_f64: 4389 // The vector type needs a cast for the v2f64 variant. 4390 Ops[1] = Builder.CreateBitCast(Ops[1], 4391 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4392 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4393 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4394 4395 case NEON::BI__builtin_neon_vget_lane_i8: 4396 case NEON::BI__builtin_neon_vdupb_lane_i8: 4397 Ops[0] = Builder.CreateBitCast(Ops[0], 4398 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4399 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4400 "vget_lane"); 4401 case NEON::BI__builtin_neon_vgetq_lane_i8: 4402 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4403 Ops[0] = Builder.CreateBitCast(Ops[0], 4404 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4405 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4406 "vgetq_lane"); 4407 case NEON::BI__builtin_neon_vget_lane_i16: 4408 case NEON::BI__builtin_neon_vduph_lane_i16: 4409 Ops[0] = Builder.CreateBitCast(Ops[0], 4410 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4411 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4412 "vget_lane"); 4413 case NEON::BI__builtin_neon_vgetq_lane_i16: 4414 case NEON::BI__builtin_neon_vduph_laneq_i16: 4415 Ops[0] = Builder.CreateBitCast(Ops[0], 4416 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4417 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4418 "vgetq_lane"); 4419 case NEON::BI__builtin_neon_vget_lane_i32: 4420 case NEON::BI__builtin_neon_vdups_lane_i32: 4421 Ops[0] = Builder.CreateBitCast( 4422 Ops[0], 4423 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4424 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4425 "vget_lane"); 4426 case NEON::BI__builtin_neon_vdups_lane_f32: 4427 Ops[0] = Builder.CreateBitCast(Ops[0], 4428 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4429 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4430 "vdups_lane"); 4431 case NEON::BI__builtin_neon_vgetq_lane_i32: 4432 case NEON::BI__builtin_neon_vdups_laneq_i32: 4433 Ops[0] = Builder.CreateBitCast(Ops[0], 4434 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4435 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4436 "vgetq_lane"); 4437 case NEON::BI__builtin_neon_vget_lane_i64: 4438 case NEON::BI__builtin_neon_vdupd_lane_i64: 4439 Ops[0] = Builder.CreateBitCast(Ops[0], 4440 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4441 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4442 "vget_lane"); 4443 case NEON::BI__builtin_neon_vdupd_lane_f64: 4444 Ops[0] = Builder.CreateBitCast(Ops[0], 4445 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4446 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4447 "vdupd_lane"); 4448 case NEON::BI__builtin_neon_vgetq_lane_i64: 4449 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4450 Ops[0] = Builder.CreateBitCast(Ops[0], 4451 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4452 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4453 "vgetq_lane"); 4454 case NEON::BI__builtin_neon_vget_lane_f32: 4455 Ops[0] = Builder.CreateBitCast(Ops[0], 4456 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4457 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4458 "vget_lane"); 4459 case NEON::BI__builtin_neon_vget_lane_f64: 4460 Ops[0] = Builder.CreateBitCast(Ops[0], 4461 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4462 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4463 "vget_lane"); 4464 case NEON::BI__builtin_neon_vgetq_lane_f32: 4465 case NEON::BI__builtin_neon_vdups_laneq_f32: 4466 Ops[0] = Builder.CreateBitCast(Ops[0], 4467 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4468 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4469 "vgetq_lane"); 4470 case NEON::BI__builtin_neon_vgetq_lane_f64: 4471 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4472 Ops[0] = Builder.CreateBitCast(Ops[0], 4473 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4474 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4475 "vgetq_lane"); 4476 case NEON::BI__builtin_neon_vaddd_s64: 4477 case NEON::BI__builtin_neon_vaddd_u64: 4478 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4479 case NEON::BI__builtin_neon_vsubd_s64: 4480 case NEON::BI__builtin_neon_vsubd_u64: 4481 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4482 case NEON::BI__builtin_neon_vqdmlalh_s16: 4483 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4484 SmallVector<Value *, 2> ProductOps; 4485 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4486 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4487 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4488 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4489 ProductOps, "vqdmlXl"); 4490 Constant *CI = ConstantInt::get(SizeTy, 0); 4491 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4492 4493 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4494 ? Intrinsic::aarch64_neon_sqadd 4495 : Intrinsic::aarch64_neon_sqsub; 4496 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4497 } 4498 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4499 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4500 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4501 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4502 Ops, "vqshlu_n"); 4503 } 4504 case NEON::BI__builtin_neon_vqshld_n_u64: 4505 case NEON::BI__builtin_neon_vqshld_n_s64: { 4506 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4507 ? Intrinsic::aarch64_neon_uqshl 4508 : Intrinsic::aarch64_neon_sqshl; 4509 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4510 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4511 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4512 } 4513 case NEON::BI__builtin_neon_vrshrd_n_u64: 4514 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4515 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4516 ? Intrinsic::aarch64_neon_urshl 4517 : Intrinsic::aarch64_neon_srshl; 4518 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4519 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4520 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4521 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4522 } 4523 case NEON::BI__builtin_neon_vrsrad_n_u64: 4524 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4525 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4526 ? Intrinsic::aarch64_neon_urshl 4527 : Intrinsic::aarch64_neon_srshl; 4528 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4529 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4530 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4531 Builder.CreateSExt(Ops[2], Int64Ty)); 4532 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4533 } 4534 case NEON::BI__builtin_neon_vshld_n_s64: 4535 case NEON::BI__builtin_neon_vshld_n_u64: { 4536 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4537 return Builder.CreateShl( 4538 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4539 } 4540 case NEON::BI__builtin_neon_vshrd_n_s64: { 4541 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4542 return Builder.CreateAShr( 4543 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4544 Amt->getZExtValue())), 4545 "shrd_n"); 4546 } 4547 case NEON::BI__builtin_neon_vshrd_n_u64: { 4548 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4549 uint64_t ShiftAmt = Amt->getZExtValue(); 4550 // Right-shifting an unsigned value by its size yields 0. 4551 if (ShiftAmt == 64) 4552 return ConstantInt::get(Int64Ty, 0); 4553 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4554 "shrd_n"); 4555 } 4556 case NEON::BI__builtin_neon_vsrad_n_s64: { 4557 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4558 Ops[1] = Builder.CreateAShr( 4559 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4560 Amt->getZExtValue())), 4561 "shrd_n"); 4562 return Builder.CreateAdd(Ops[0], Ops[1]); 4563 } 4564 case NEON::BI__builtin_neon_vsrad_n_u64: { 4565 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4566 uint64_t ShiftAmt = Amt->getZExtValue(); 4567 // Right-shifting an unsigned value by its size yields 0. 4568 // As Op + 0 = Op, return Ops[0] directly. 4569 if (ShiftAmt == 64) 4570 return Ops[0]; 4571 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4572 "shrd_n"); 4573 return Builder.CreateAdd(Ops[0], Ops[1]); 4574 } 4575 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4576 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4577 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4578 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4579 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4580 "lane"); 4581 SmallVector<Value *, 2> ProductOps; 4582 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4583 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4584 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4585 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4586 ProductOps, "vqdmlXl"); 4587 Constant *CI = ConstantInt::get(SizeTy, 0); 4588 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4589 Ops.pop_back(); 4590 4591 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4592 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4593 ? Intrinsic::aarch64_neon_sqadd 4594 : Intrinsic::aarch64_neon_sqsub; 4595 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4596 } 4597 case NEON::BI__builtin_neon_vqdmlals_s32: 4598 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4599 SmallVector<Value *, 2> ProductOps; 4600 ProductOps.push_back(Ops[1]); 4601 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4602 Ops[1] = 4603 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4604 ProductOps, "vqdmlXl"); 4605 4606 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4607 ? Intrinsic::aarch64_neon_sqadd 4608 : Intrinsic::aarch64_neon_sqsub; 4609 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4610 } 4611 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4612 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4613 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4614 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4615 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4616 "lane"); 4617 SmallVector<Value *, 2> ProductOps; 4618 ProductOps.push_back(Ops[1]); 4619 ProductOps.push_back(Ops[2]); 4620 Ops[1] = 4621 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4622 ProductOps, "vqdmlXl"); 4623 Ops.pop_back(); 4624 4625 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4626 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4627 ? Intrinsic::aarch64_neon_sqadd 4628 : Intrinsic::aarch64_neon_sqsub; 4629 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4630 } 4631 } 4632 4633 llvm::VectorType *VTy = GetNeonType(this, Type); 4634 llvm::Type *Ty = VTy; 4635 if (!Ty) 4636 return nullptr; 4637 4638 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4639 // defer to common code if it's been added to our special map. 4640 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4641 AArch64SIMDIntrinsicsProvenSorted); 4642 4643 if (Builtin) 4644 return EmitCommonNeonBuiltinExpr( 4645 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4646 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4647 4648 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4649 return V; 4650 4651 unsigned Int; 4652 switch (BuiltinID) { 4653 default: return nullptr; 4654 case NEON::BI__builtin_neon_vbsl_v: 4655 case NEON::BI__builtin_neon_vbslq_v: { 4656 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4657 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4658 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4659 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4660 4661 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4662 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4663 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4664 return Builder.CreateBitCast(Ops[0], Ty); 4665 } 4666 case NEON::BI__builtin_neon_vfma_lane_v: 4667 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4668 // The ARM builtins (and instructions) have the addend as the first 4669 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4670 Value *Addend = Ops[0]; 4671 Value *Multiplicand = Ops[1]; 4672 Value *LaneSource = Ops[2]; 4673 Ops[0] = Multiplicand; 4674 Ops[1] = LaneSource; 4675 Ops[2] = Addend; 4676 4677 // Now adjust things to handle the lane access. 4678 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4679 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4680 VTy; 4681 llvm::Constant *cst = cast<Constant>(Ops[3]); 4682 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4683 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4684 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4685 4686 Ops.pop_back(); 4687 Int = Intrinsic::fma; 4688 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4689 } 4690 case NEON::BI__builtin_neon_vfma_laneq_v: { 4691 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4692 // v1f64 fma should be mapped to Neon scalar f64 fma 4693 if (VTy && VTy->getElementType() == DoubleTy) { 4694 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4695 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4696 llvm::Type *VTy = GetNeonType(this, 4697 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4698 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4699 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4700 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4701 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4702 return Builder.CreateBitCast(Result, Ty); 4703 } 4704 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4705 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4706 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4707 4708 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4709 VTy->getNumElements() * 2); 4710 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4711 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4712 cast<ConstantInt>(Ops[3])); 4713 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4714 4715 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4716 } 4717 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4718 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4719 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4720 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4721 4722 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4723 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4724 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4725 } 4726 case NEON::BI__builtin_neon_vfmas_lane_f32: 4727 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4728 case NEON::BI__builtin_neon_vfmad_lane_f64: 4729 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4730 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4731 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4732 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4733 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4734 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4735 } 4736 case NEON::BI__builtin_neon_vfms_v: 4737 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4738 // FIXME: probably remove when we no longer support aarch64_simd.h 4739 // (arm_neon.h delegates to vfma). 4740 4741 // The ARM builtins (and instructions) have the addend as the first 4742 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4743 Value *Subtrahend = Ops[0]; 4744 Value *Multiplicand = Ops[2]; 4745 Ops[0] = Multiplicand; 4746 Ops[2] = Subtrahend; 4747 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4748 Ops[1] = Builder.CreateFNeg(Ops[1]); 4749 Int = Intrinsic::fma; 4750 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4751 } 4752 case NEON::BI__builtin_neon_vmull_v: 4753 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4754 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4755 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4756 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4757 case NEON::BI__builtin_neon_vmax_v: 4758 case NEON::BI__builtin_neon_vmaxq_v: 4759 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4760 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4761 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4762 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4763 case NEON::BI__builtin_neon_vmin_v: 4764 case NEON::BI__builtin_neon_vminq_v: 4765 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4766 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4767 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4768 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4769 case NEON::BI__builtin_neon_vabd_v: 4770 case NEON::BI__builtin_neon_vabdq_v: 4771 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4772 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4773 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4774 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4775 case NEON::BI__builtin_neon_vpadal_v: 4776 case NEON::BI__builtin_neon_vpadalq_v: { 4777 unsigned ArgElts = VTy->getNumElements(); 4778 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4779 unsigned BitWidth = EltTy->getBitWidth(); 4780 llvm::Type *ArgTy = llvm::VectorType::get( 4781 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4782 llvm::Type* Tys[2] = { VTy, ArgTy }; 4783 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4784 SmallVector<llvm::Value*, 1> TmpOps; 4785 TmpOps.push_back(Ops[1]); 4786 Function *F = CGM.getIntrinsic(Int, Tys); 4787 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4788 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4789 return Builder.CreateAdd(tmp, addend); 4790 } 4791 case NEON::BI__builtin_neon_vpmin_v: 4792 case NEON::BI__builtin_neon_vpminq_v: 4793 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4794 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4795 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4796 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4797 case NEON::BI__builtin_neon_vpmax_v: 4798 case NEON::BI__builtin_neon_vpmaxq_v: 4799 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4800 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4801 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4802 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4803 case NEON::BI__builtin_neon_vminnm_v: 4804 case NEON::BI__builtin_neon_vminnmq_v: 4805 Int = Intrinsic::aarch64_neon_fminnm; 4806 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4807 case NEON::BI__builtin_neon_vmaxnm_v: 4808 case NEON::BI__builtin_neon_vmaxnmq_v: 4809 Int = Intrinsic::aarch64_neon_fmaxnm; 4810 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4811 case NEON::BI__builtin_neon_vrecpss_f32: { 4812 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4813 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4814 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4815 Ops, "vrecps"); 4816 } 4817 case NEON::BI__builtin_neon_vrecpsd_f64: { 4818 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4819 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4820 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4821 Ops, "vrecps"); 4822 } 4823 case NEON::BI__builtin_neon_vqshrun_n_v: 4824 Int = Intrinsic::aarch64_neon_sqshrun; 4825 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4826 case NEON::BI__builtin_neon_vqrshrun_n_v: 4827 Int = Intrinsic::aarch64_neon_sqrshrun; 4828 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4829 case NEON::BI__builtin_neon_vqshrn_n_v: 4830 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4831 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4832 case NEON::BI__builtin_neon_vrshrn_n_v: 4833 Int = Intrinsic::aarch64_neon_rshrn; 4834 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4835 case NEON::BI__builtin_neon_vqrshrn_n_v: 4836 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4837 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4838 case NEON::BI__builtin_neon_vrnda_v: 4839 case NEON::BI__builtin_neon_vrndaq_v: { 4840 Int = Intrinsic::round; 4841 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4842 } 4843 case NEON::BI__builtin_neon_vrndi_v: 4844 case NEON::BI__builtin_neon_vrndiq_v: { 4845 Int = Intrinsic::nearbyint; 4846 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4847 } 4848 case NEON::BI__builtin_neon_vrndm_v: 4849 case NEON::BI__builtin_neon_vrndmq_v: { 4850 Int = Intrinsic::floor; 4851 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4852 } 4853 case NEON::BI__builtin_neon_vrndn_v: 4854 case NEON::BI__builtin_neon_vrndnq_v: { 4855 Int = Intrinsic::aarch64_neon_frintn; 4856 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4857 } 4858 case NEON::BI__builtin_neon_vrndp_v: 4859 case NEON::BI__builtin_neon_vrndpq_v: { 4860 Int = Intrinsic::ceil; 4861 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4862 } 4863 case NEON::BI__builtin_neon_vrndx_v: 4864 case NEON::BI__builtin_neon_vrndxq_v: { 4865 Int = Intrinsic::rint; 4866 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4867 } 4868 case NEON::BI__builtin_neon_vrnd_v: 4869 case NEON::BI__builtin_neon_vrndq_v: { 4870 Int = Intrinsic::trunc; 4871 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4872 } 4873 case NEON::BI__builtin_neon_vceqz_v: 4874 case NEON::BI__builtin_neon_vceqzq_v: 4875 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4876 ICmpInst::ICMP_EQ, "vceqz"); 4877 case NEON::BI__builtin_neon_vcgez_v: 4878 case NEON::BI__builtin_neon_vcgezq_v: 4879 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4880 ICmpInst::ICMP_SGE, "vcgez"); 4881 case NEON::BI__builtin_neon_vclez_v: 4882 case NEON::BI__builtin_neon_vclezq_v: 4883 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4884 ICmpInst::ICMP_SLE, "vclez"); 4885 case NEON::BI__builtin_neon_vcgtz_v: 4886 case NEON::BI__builtin_neon_vcgtzq_v: 4887 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4888 ICmpInst::ICMP_SGT, "vcgtz"); 4889 case NEON::BI__builtin_neon_vcltz_v: 4890 case NEON::BI__builtin_neon_vcltzq_v: 4891 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4892 ICmpInst::ICMP_SLT, "vcltz"); 4893 case NEON::BI__builtin_neon_vcvt_f64_v: 4894 case NEON::BI__builtin_neon_vcvtq_f64_v: 4895 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4896 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4897 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4898 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4899 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4900 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4901 "unexpected vcvt_f64_f32 builtin"); 4902 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4903 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4904 4905 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4906 } 4907 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4908 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4909 "unexpected vcvt_f32_f64 builtin"); 4910 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4911 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4912 4913 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4914 } 4915 case NEON::BI__builtin_neon_vcvt_s32_v: 4916 case NEON::BI__builtin_neon_vcvt_u32_v: 4917 case NEON::BI__builtin_neon_vcvt_s64_v: 4918 case NEON::BI__builtin_neon_vcvt_u64_v: 4919 case NEON::BI__builtin_neon_vcvtq_s32_v: 4920 case NEON::BI__builtin_neon_vcvtq_u32_v: 4921 case NEON::BI__builtin_neon_vcvtq_s64_v: 4922 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4923 bool Double = 4924 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4925 llvm::Type *InTy = 4926 GetNeonType(this, 4927 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4928 : NeonTypeFlags::Float32, false, quad)); 4929 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4930 if (usgn) 4931 return Builder.CreateFPToUI(Ops[0], Ty); 4932 return Builder.CreateFPToSI(Ops[0], Ty); 4933 } 4934 case NEON::BI__builtin_neon_vcvta_s32_v: 4935 case NEON::BI__builtin_neon_vcvtaq_s32_v: 4936 case NEON::BI__builtin_neon_vcvta_u32_v: 4937 case NEON::BI__builtin_neon_vcvtaq_u32_v: 4938 case NEON::BI__builtin_neon_vcvta_s64_v: 4939 case NEON::BI__builtin_neon_vcvtaq_s64_v: 4940 case NEON::BI__builtin_neon_vcvta_u64_v: 4941 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4942 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 4943 bool Double = 4944 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4945 llvm::Type *InTy = 4946 GetNeonType(this, 4947 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4948 : NeonTypeFlags::Float32, false, quad)); 4949 llvm::Type *Tys[2] = { Ty, InTy }; 4950 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 4951 } 4952 case NEON::BI__builtin_neon_vcvtm_s32_v: 4953 case NEON::BI__builtin_neon_vcvtmq_s32_v: 4954 case NEON::BI__builtin_neon_vcvtm_u32_v: 4955 case NEON::BI__builtin_neon_vcvtmq_u32_v: 4956 case NEON::BI__builtin_neon_vcvtm_s64_v: 4957 case NEON::BI__builtin_neon_vcvtmq_s64_v: 4958 case NEON::BI__builtin_neon_vcvtm_u64_v: 4959 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4960 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 4961 bool Double = 4962 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4963 llvm::Type *InTy = 4964 GetNeonType(this, 4965 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4966 : NeonTypeFlags::Float32, false, quad)); 4967 llvm::Type *Tys[2] = { Ty, InTy }; 4968 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 4969 } 4970 case NEON::BI__builtin_neon_vcvtn_s32_v: 4971 case NEON::BI__builtin_neon_vcvtnq_s32_v: 4972 case NEON::BI__builtin_neon_vcvtn_u32_v: 4973 case NEON::BI__builtin_neon_vcvtnq_u32_v: 4974 case NEON::BI__builtin_neon_vcvtn_s64_v: 4975 case NEON::BI__builtin_neon_vcvtnq_s64_v: 4976 case NEON::BI__builtin_neon_vcvtn_u64_v: 4977 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4978 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 4979 bool Double = 4980 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4981 llvm::Type *InTy = 4982 GetNeonType(this, 4983 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4984 : NeonTypeFlags::Float32, false, quad)); 4985 llvm::Type *Tys[2] = { Ty, InTy }; 4986 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 4987 } 4988 case NEON::BI__builtin_neon_vcvtp_s32_v: 4989 case NEON::BI__builtin_neon_vcvtpq_s32_v: 4990 case NEON::BI__builtin_neon_vcvtp_u32_v: 4991 case NEON::BI__builtin_neon_vcvtpq_u32_v: 4992 case NEON::BI__builtin_neon_vcvtp_s64_v: 4993 case NEON::BI__builtin_neon_vcvtpq_s64_v: 4994 case NEON::BI__builtin_neon_vcvtp_u64_v: 4995 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4996 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 4997 bool Double = 4998 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4999 llvm::Type *InTy = 5000 GetNeonType(this, 5001 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5002 : NeonTypeFlags::Float32, false, quad)); 5003 llvm::Type *Tys[2] = { Ty, InTy }; 5004 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5005 } 5006 case NEON::BI__builtin_neon_vmulx_v: 5007 case NEON::BI__builtin_neon_vmulxq_v: { 5008 Int = Intrinsic::aarch64_neon_fmulx; 5009 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5010 } 5011 case NEON::BI__builtin_neon_vmul_lane_v: 5012 case NEON::BI__builtin_neon_vmul_laneq_v: { 5013 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5014 bool Quad = false; 5015 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5016 Quad = true; 5017 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5018 llvm::Type *VTy = GetNeonType(this, 5019 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5020 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5021 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5022 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5023 return Builder.CreateBitCast(Result, Ty); 5024 } 5025 case NEON::BI__builtin_neon_vnegd_s64: 5026 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5027 case NEON::BI__builtin_neon_vpmaxnm_v: 5028 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5029 Int = Intrinsic::aarch64_neon_fmaxnmp; 5030 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5031 } 5032 case NEON::BI__builtin_neon_vpminnm_v: 5033 case NEON::BI__builtin_neon_vpminnmq_v: { 5034 Int = Intrinsic::aarch64_neon_fminnmp; 5035 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5036 } 5037 case NEON::BI__builtin_neon_vsqrt_v: 5038 case NEON::BI__builtin_neon_vsqrtq_v: { 5039 Int = Intrinsic::sqrt; 5040 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5041 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5042 } 5043 case NEON::BI__builtin_neon_vrbit_v: 5044 case NEON::BI__builtin_neon_vrbitq_v: { 5045 Int = Intrinsic::aarch64_neon_rbit; 5046 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5047 } 5048 case NEON::BI__builtin_neon_vaddv_u8: 5049 // FIXME: These are handled by the AArch64 scalar code. 5050 usgn = true; 5051 // FALLTHROUGH 5052 case NEON::BI__builtin_neon_vaddv_s8: { 5053 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5054 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5055 VTy = 5056 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5057 llvm::Type *Tys[2] = { Ty, VTy }; 5058 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5059 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5060 return Builder.CreateTrunc(Ops[0], 5061 llvm::IntegerType::get(getLLVMContext(), 8)); 5062 } 5063 case NEON::BI__builtin_neon_vaddv_u16: 5064 usgn = true; 5065 // FALLTHROUGH 5066 case NEON::BI__builtin_neon_vaddv_s16: { 5067 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5068 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5069 VTy = 5070 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5071 llvm::Type *Tys[2] = { Ty, VTy }; 5072 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5073 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5074 return Builder.CreateTrunc(Ops[0], 5075 llvm::IntegerType::get(getLLVMContext(), 16)); 5076 } 5077 case NEON::BI__builtin_neon_vaddvq_u8: 5078 usgn = true; 5079 // FALLTHROUGH 5080 case NEON::BI__builtin_neon_vaddvq_s8: { 5081 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5082 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5083 VTy = 5084 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5085 llvm::Type *Tys[2] = { Ty, VTy }; 5086 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5087 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5088 return Builder.CreateTrunc(Ops[0], 5089 llvm::IntegerType::get(getLLVMContext(), 8)); 5090 } 5091 case NEON::BI__builtin_neon_vaddvq_u16: 5092 usgn = true; 5093 // FALLTHROUGH 5094 case NEON::BI__builtin_neon_vaddvq_s16: { 5095 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5096 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5097 VTy = 5098 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5099 llvm::Type *Tys[2] = { Ty, VTy }; 5100 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5101 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5102 return Builder.CreateTrunc(Ops[0], 5103 llvm::IntegerType::get(getLLVMContext(), 16)); 5104 } 5105 case NEON::BI__builtin_neon_vmaxv_u8: { 5106 Int = Intrinsic::aarch64_neon_umaxv; 5107 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5108 VTy = 5109 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5110 llvm::Type *Tys[2] = { Ty, VTy }; 5111 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5112 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5113 return Builder.CreateTrunc(Ops[0], 5114 llvm::IntegerType::get(getLLVMContext(), 8)); 5115 } 5116 case NEON::BI__builtin_neon_vmaxv_u16: { 5117 Int = Intrinsic::aarch64_neon_umaxv; 5118 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5119 VTy = 5120 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5121 llvm::Type *Tys[2] = { Ty, VTy }; 5122 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5123 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5124 return Builder.CreateTrunc(Ops[0], 5125 llvm::IntegerType::get(getLLVMContext(), 16)); 5126 } 5127 case NEON::BI__builtin_neon_vmaxvq_u8: { 5128 Int = Intrinsic::aarch64_neon_umaxv; 5129 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5130 VTy = 5131 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5132 llvm::Type *Tys[2] = { Ty, VTy }; 5133 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5134 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5135 return Builder.CreateTrunc(Ops[0], 5136 llvm::IntegerType::get(getLLVMContext(), 8)); 5137 } 5138 case NEON::BI__builtin_neon_vmaxvq_u16: { 5139 Int = Intrinsic::aarch64_neon_umaxv; 5140 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5141 VTy = 5142 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5143 llvm::Type *Tys[2] = { Ty, VTy }; 5144 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5145 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5146 return Builder.CreateTrunc(Ops[0], 5147 llvm::IntegerType::get(getLLVMContext(), 16)); 5148 } 5149 case NEON::BI__builtin_neon_vmaxv_s8: { 5150 Int = Intrinsic::aarch64_neon_smaxv; 5151 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5152 VTy = 5153 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5154 llvm::Type *Tys[2] = { Ty, VTy }; 5155 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5156 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5157 return Builder.CreateTrunc(Ops[0], 5158 llvm::IntegerType::get(getLLVMContext(), 8)); 5159 } 5160 case NEON::BI__builtin_neon_vmaxv_s16: { 5161 Int = Intrinsic::aarch64_neon_smaxv; 5162 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5163 VTy = 5164 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5165 llvm::Type *Tys[2] = { Ty, VTy }; 5166 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5167 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5168 return Builder.CreateTrunc(Ops[0], 5169 llvm::IntegerType::get(getLLVMContext(), 16)); 5170 } 5171 case NEON::BI__builtin_neon_vmaxvq_s8: { 5172 Int = Intrinsic::aarch64_neon_smaxv; 5173 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5174 VTy = 5175 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5176 llvm::Type *Tys[2] = { Ty, VTy }; 5177 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5178 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5179 return Builder.CreateTrunc(Ops[0], 5180 llvm::IntegerType::get(getLLVMContext(), 8)); 5181 } 5182 case NEON::BI__builtin_neon_vmaxvq_s16: { 5183 Int = Intrinsic::aarch64_neon_smaxv; 5184 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5185 VTy = 5186 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5187 llvm::Type *Tys[2] = { Ty, VTy }; 5188 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5189 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5190 return Builder.CreateTrunc(Ops[0], 5191 llvm::IntegerType::get(getLLVMContext(), 16)); 5192 } 5193 case NEON::BI__builtin_neon_vminv_u8: { 5194 Int = Intrinsic::aarch64_neon_uminv; 5195 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5196 VTy = 5197 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5198 llvm::Type *Tys[2] = { Ty, VTy }; 5199 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5200 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5201 return Builder.CreateTrunc(Ops[0], 5202 llvm::IntegerType::get(getLLVMContext(), 8)); 5203 } 5204 case NEON::BI__builtin_neon_vminv_u16: { 5205 Int = Intrinsic::aarch64_neon_uminv; 5206 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5207 VTy = 5208 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5209 llvm::Type *Tys[2] = { Ty, VTy }; 5210 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5211 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5212 return Builder.CreateTrunc(Ops[0], 5213 llvm::IntegerType::get(getLLVMContext(), 16)); 5214 } 5215 case NEON::BI__builtin_neon_vminvq_u8: { 5216 Int = Intrinsic::aarch64_neon_uminv; 5217 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5218 VTy = 5219 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5220 llvm::Type *Tys[2] = { Ty, VTy }; 5221 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5222 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5223 return Builder.CreateTrunc(Ops[0], 5224 llvm::IntegerType::get(getLLVMContext(), 8)); 5225 } 5226 case NEON::BI__builtin_neon_vminvq_u16: { 5227 Int = Intrinsic::aarch64_neon_uminv; 5228 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5229 VTy = 5230 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5231 llvm::Type *Tys[2] = { Ty, VTy }; 5232 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5233 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5234 return Builder.CreateTrunc(Ops[0], 5235 llvm::IntegerType::get(getLLVMContext(), 16)); 5236 } 5237 case NEON::BI__builtin_neon_vminv_s8: { 5238 Int = Intrinsic::aarch64_neon_sminv; 5239 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5240 VTy = 5241 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5242 llvm::Type *Tys[2] = { Ty, VTy }; 5243 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5244 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5245 return Builder.CreateTrunc(Ops[0], 5246 llvm::IntegerType::get(getLLVMContext(), 8)); 5247 } 5248 case NEON::BI__builtin_neon_vminv_s16: { 5249 Int = Intrinsic::aarch64_neon_sminv; 5250 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5251 VTy = 5252 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5253 llvm::Type *Tys[2] = { Ty, VTy }; 5254 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5255 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5256 return Builder.CreateTrunc(Ops[0], 5257 llvm::IntegerType::get(getLLVMContext(), 16)); 5258 } 5259 case NEON::BI__builtin_neon_vminvq_s8: { 5260 Int = Intrinsic::aarch64_neon_sminv; 5261 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5262 VTy = 5263 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5264 llvm::Type *Tys[2] = { Ty, VTy }; 5265 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5266 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5267 return Builder.CreateTrunc(Ops[0], 5268 llvm::IntegerType::get(getLLVMContext(), 8)); 5269 } 5270 case NEON::BI__builtin_neon_vminvq_s16: { 5271 Int = Intrinsic::aarch64_neon_sminv; 5272 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5273 VTy = 5274 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5275 llvm::Type *Tys[2] = { Ty, VTy }; 5276 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5277 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5278 return Builder.CreateTrunc(Ops[0], 5279 llvm::IntegerType::get(getLLVMContext(), 16)); 5280 } 5281 case NEON::BI__builtin_neon_vmul_n_f64: { 5282 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5283 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5284 return Builder.CreateFMul(Ops[0], RHS); 5285 } 5286 case NEON::BI__builtin_neon_vaddlv_u8: { 5287 Int = Intrinsic::aarch64_neon_uaddlv; 5288 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5289 VTy = 5290 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5291 llvm::Type *Tys[2] = { Ty, VTy }; 5292 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5293 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5294 return Builder.CreateTrunc(Ops[0], 5295 llvm::IntegerType::get(getLLVMContext(), 16)); 5296 } 5297 case NEON::BI__builtin_neon_vaddlv_u16: { 5298 Int = Intrinsic::aarch64_neon_uaddlv; 5299 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5300 VTy = 5301 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5302 llvm::Type *Tys[2] = { Ty, VTy }; 5303 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5304 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5305 } 5306 case NEON::BI__builtin_neon_vaddlvq_u8: { 5307 Int = Intrinsic::aarch64_neon_uaddlv; 5308 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5309 VTy = 5310 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5311 llvm::Type *Tys[2] = { Ty, VTy }; 5312 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5313 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5314 return Builder.CreateTrunc(Ops[0], 5315 llvm::IntegerType::get(getLLVMContext(), 16)); 5316 } 5317 case NEON::BI__builtin_neon_vaddlvq_u16: { 5318 Int = Intrinsic::aarch64_neon_uaddlv; 5319 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5320 VTy = 5321 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5322 llvm::Type *Tys[2] = { Ty, VTy }; 5323 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5324 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5325 } 5326 case NEON::BI__builtin_neon_vaddlv_s8: { 5327 Int = Intrinsic::aarch64_neon_saddlv; 5328 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5329 VTy = 5330 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5331 llvm::Type *Tys[2] = { Ty, VTy }; 5332 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5333 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5334 return Builder.CreateTrunc(Ops[0], 5335 llvm::IntegerType::get(getLLVMContext(), 16)); 5336 } 5337 case NEON::BI__builtin_neon_vaddlv_s16: { 5338 Int = Intrinsic::aarch64_neon_saddlv; 5339 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5340 VTy = 5341 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5342 llvm::Type *Tys[2] = { Ty, VTy }; 5343 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5344 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5345 } 5346 case NEON::BI__builtin_neon_vaddlvq_s8: { 5347 Int = Intrinsic::aarch64_neon_saddlv; 5348 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5349 VTy = 5350 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5351 llvm::Type *Tys[2] = { Ty, VTy }; 5352 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5353 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5354 return Builder.CreateTrunc(Ops[0], 5355 llvm::IntegerType::get(getLLVMContext(), 16)); 5356 } 5357 case NEON::BI__builtin_neon_vaddlvq_s16: { 5358 Int = Intrinsic::aarch64_neon_saddlv; 5359 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5360 VTy = 5361 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5362 llvm::Type *Tys[2] = { Ty, VTy }; 5363 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5364 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5365 } 5366 case NEON::BI__builtin_neon_vsri_n_v: 5367 case NEON::BI__builtin_neon_vsriq_n_v: { 5368 Int = Intrinsic::aarch64_neon_vsri; 5369 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5370 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5371 } 5372 case NEON::BI__builtin_neon_vsli_n_v: 5373 case NEON::BI__builtin_neon_vsliq_n_v: { 5374 Int = Intrinsic::aarch64_neon_vsli; 5375 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5376 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5377 } 5378 case NEON::BI__builtin_neon_vsra_n_v: 5379 case NEON::BI__builtin_neon_vsraq_n_v: 5380 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5381 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5382 return Builder.CreateAdd(Ops[0], Ops[1]); 5383 case NEON::BI__builtin_neon_vrsra_n_v: 5384 case NEON::BI__builtin_neon_vrsraq_n_v: { 5385 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5386 SmallVector<llvm::Value*,2> TmpOps; 5387 TmpOps.push_back(Ops[1]); 5388 TmpOps.push_back(Ops[2]); 5389 Function* F = CGM.getIntrinsic(Int, Ty); 5390 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5391 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5392 return Builder.CreateAdd(Ops[0], tmp); 5393 } 5394 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5395 // of an Align parameter here. 5396 case NEON::BI__builtin_neon_vld1_x2_v: 5397 case NEON::BI__builtin_neon_vld1q_x2_v: 5398 case NEON::BI__builtin_neon_vld1_x3_v: 5399 case NEON::BI__builtin_neon_vld1q_x3_v: 5400 case NEON::BI__builtin_neon_vld1_x4_v: 5401 case NEON::BI__builtin_neon_vld1q_x4_v: { 5402 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5403 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5404 llvm::Type *Tys[2] = { VTy, PTy }; 5405 unsigned Int; 5406 switch (BuiltinID) { 5407 case NEON::BI__builtin_neon_vld1_x2_v: 5408 case NEON::BI__builtin_neon_vld1q_x2_v: 5409 Int = Intrinsic::aarch64_neon_ld1x2; 5410 break; 5411 case NEON::BI__builtin_neon_vld1_x3_v: 5412 case NEON::BI__builtin_neon_vld1q_x3_v: 5413 Int = Intrinsic::aarch64_neon_ld1x3; 5414 break; 5415 case NEON::BI__builtin_neon_vld1_x4_v: 5416 case NEON::BI__builtin_neon_vld1q_x4_v: 5417 Int = Intrinsic::aarch64_neon_ld1x4; 5418 break; 5419 } 5420 Function *F = CGM.getIntrinsic(Int, Tys); 5421 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5422 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5423 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5424 return Builder.CreateStore(Ops[1], Ops[0]); 5425 } 5426 case NEON::BI__builtin_neon_vst1_x2_v: 5427 case NEON::BI__builtin_neon_vst1q_x2_v: 5428 case NEON::BI__builtin_neon_vst1_x3_v: 5429 case NEON::BI__builtin_neon_vst1q_x3_v: 5430 case NEON::BI__builtin_neon_vst1_x4_v: 5431 case NEON::BI__builtin_neon_vst1q_x4_v: { 5432 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5433 llvm::Type *Tys[2] = { VTy, PTy }; 5434 unsigned Int; 5435 switch (BuiltinID) { 5436 case NEON::BI__builtin_neon_vst1_x2_v: 5437 case NEON::BI__builtin_neon_vst1q_x2_v: 5438 Int = Intrinsic::aarch64_neon_st1x2; 5439 break; 5440 case NEON::BI__builtin_neon_vst1_x3_v: 5441 case NEON::BI__builtin_neon_vst1q_x3_v: 5442 Int = Intrinsic::aarch64_neon_st1x3; 5443 break; 5444 case NEON::BI__builtin_neon_vst1_x4_v: 5445 case NEON::BI__builtin_neon_vst1q_x4_v: 5446 Int = Intrinsic::aarch64_neon_st1x4; 5447 break; 5448 } 5449 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5450 IntOps.push_back(Ops[0]); 5451 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5452 } 5453 case NEON::BI__builtin_neon_vld1_v: 5454 case NEON::BI__builtin_neon_vld1q_v: 5455 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5456 return Builder.CreateLoad(Ops[0]); 5457 case NEON::BI__builtin_neon_vst1_v: 5458 case NEON::BI__builtin_neon_vst1q_v: 5459 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5460 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5461 return Builder.CreateStore(Ops[1], Ops[0]); 5462 case NEON::BI__builtin_neon_vld1_lane_v: 5463 case NEON::BI__builtin_neon_vld1q_lane_v: 5464 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5465 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5466 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5467 Ops[0] = Builder.CreateLoad(Ops[0]); 5468 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5469 case NEON::BI__builtin_neon_vld1_dup_v: 5470 case NEON::BI__builtin_neon_vld1q_dup_v: { 5471 Value *V = UndefValue::get(Ty); 5472 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5473 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5474 Ops[0] = Builder.CreateLoad(Ops[0]); 5475 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5476 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5477 return EmitNeonSplat(Ops[0], CI); 5478 } 5479 case NEON::BI__builtin_neon_vst1_lane_v: 5480 case NEON::BI__builtin_neon_vst1q_lane_v: 5481 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5482 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5483 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5484 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5485 case NEON::BI__builtin_neon_vld2_v: 5486 case NEON::BI__builtin_neon_vld2q_v: { 5487 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5488 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5489 llvm::Type *Tys[2] = { VTy, PTy }; 5490 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5491 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5492 Ops[0] = Builder.CreateBitCast(Ops[0], 5493 llvm::PointerType::getUnqual(Ops[1]->getType())); 5494 return Builder.CreateStore(Ops[1], Ops[0]); 5495 } 5496 case NEON::BI__builtin_neon_vld3_v: 5497 case NEON::BI__builtin_neon_vld3q_v: { 5498 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5499 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5500 llvm::Type *Tys[2] = { VTy, PTy }; 5501 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5502 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5503 Ops[0] = Builder.CreateBitCast(Ops[0], 5504 llvm::PointerType::getUnqual(Ops[1]->getType())); 5505 return Builder.CreateStore(Ops[1], Ops[0]); 5506 } 5507 case NEON::BI__builtin_neon_vld4_v: 5508 case NEON::BI__builtin_neon_vld4q_v: { 5509 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5510 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5511 llvm::Type *Tys[2] = { VTy, PTy }; 5512 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5513 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5514 Ops[0] = Builder.CreateBitCast(Ops[0], 5515 llvm::PointerType::getUnqual(Ops[1]->getType())); 5516 return Builder.CreateStore(Ops[1], Ops[0]); 5517 } 5518 case NEON::BI__builtin_neon_vld2_dup_v: 5519 case NEON::BI__builtin_neon_vld2q_dup_v: { 5520 llvm::Type *PTy = 5521 llvm::PointerType::getUnqual(VTy->getElementType()); 5522 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5523 llvm::Type *Tys[2] = { VTy, PTy }; 5524 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5525 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5526 Ops[0] = Builder.CreateBitCast(Ops[0], 5527 llvm::PointerType::getUnqual(Ops[1]->getType())); 5528 return Builder.CreateStore(Ops[1], Ops[0]); 5529 } 5530 case NEON::BI__builtin_neon_vld3_dup_v: 5531 case NEON::BI__builtin_neon_vld3q_dup_v: { 5532 llvm::Type *PTy = 5533 llvm::PointerType::getUnqual(VTy->getElementType()); 5534 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5535 llvm::Type *Tys[2] = { VTy, PTy }; 5536 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5537 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5538 Ops[0] = Builder.CreateBitCast(Ops[0], 5539 llvm::PointerType::getUnqual(Ops[1]->getType())); 5540 return Builder.CreateStore(Ops[1], Ops[0]); 5541 } 5542 case NEON::BI__builtin_neon_vld4_dup_v: 5543 case NEON::BI__builtin_neon_vld4q_dup_v: { 5544 llvm::Type *PTy = 5545 llvm::PointerType::getUnqual(VTy->getElementType()); 5546 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5547 llvm::Type *Tys[2] = { VTy, PTy }; 5548 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5549 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5550 Ops[0] = Builder.CreateBitCast(Ops[0], 5551 llvm::PointerType::getUnqual(Ops[1]->getType())); 5552 return Builder.CreateStore(Ops[1], Ops[0]); 5553 } 5554 case NEON::BI__builtin_neon_vld2_lane_v: 5555 case NEON::BI__builtin_neon_vld2q_lane_v: { 5556 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5557 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5558 Ops.push_back(Ops[1]); 5559 Ops.erase(Ops.begin()+1); 5560 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5561 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5562 Ops[3] = Builder.CreateZExt(Ops[3], 5563 llvm::IntegerType::get(getLLVMContext(), 64)); 5564 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5565 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5566 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5567 return Builder.CreateStore(Ops[1], Ops[0]); 5568 } 5569 case NEON::BI__builtin_neon_vld3_lane_v: 5570 case NEON::BI__builtin_neon_vld3q_lane_v: { 5571 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5572 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5573 Ops.push_back(Ops[1]); 5574 Ops.erase(Ops.begin()+1); 5575 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5576 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5577 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5578 Ops[4] = Builder.CreateZExt(Ops[4], 5579 llvm::IntegerType::get(getLLVMContext(), 64)); 5580 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5581 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5582 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5583 return Builder.CreateStore(Ops[1], Ops[0]); 5584 } 5585 case NEON::BI__builtin_neon_vld4_lane_v: 5586 case NEON::BI__builtin_neon_vld4q_lane_v: { 5587 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5588 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5589 Ops.push_back(Ops[1]); 5590 Ops.erase(Ops.begin()+1); 5591 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5592 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5593 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5594 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5595 Ops[5] = Builder.CreateZExt(Ops[5], 5596 llvm::IntegerType::get(getLLVMContext(), 64)); 5597 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5598 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5599 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5600 return Builder.CreateStore(Ops[1], Ops[0]); 5601 } 5602 case NEON::BI__builtin_neon_vst2_v: 5603 case NEON::BI__builtin_neon_vst2q_v: { 5604 Ops.push_back(Ops[0]); 5605 Ops.erase(Ops.begin()); 5606 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5607 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5608 Ops, ""); 5609 } 5610 case NEON::BI__builtin_neon_vst2_lane_v: 5611 case NEON::BI__builtin_neon_vst2q_lane_v: { 5612 Ops.push_back(Ops[0]); 5613 Ops.erase(Ops.begin()); 5614 Ops[2] = Builder.CreateZExt(Ops[2], 5615 llvm::IntegerType::get(getLLVMContext(), 64)); 5616 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5617 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5618 Ops, ""); 5619 } 5620 case NEON::BI__builtin_neon_vst3_v: 5621 case NEON::BI__builtin_neon_vst3q_v: { 5622 Ops.push_back(Ops[0]); 5623 Ops.erase(Ops.begin()); 5624 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5625 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5626 Ops, ""); 5627 } 5628 case NEON::BI__builtin_neon_vst3_lane_v: 5629 case NEON::BI__builtin_neon_vst3q_lane_v: { 5630 Ops.push_back(Ops[0]); 5631 Ops.erase(Ops.begin()); 5632 Ops[3] = Builder.CreateZExt(Ops[3], 5633 llvm::IntegerType::get(getLLVMContext(), 64)); 5634 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5635 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5636 Ops, ""); 5637 } 5638 case NEON::BI__builtin_neon_vst4_v: 5639 case NEON::BI__builtin_neon_vst4q_v: { 5640 Ops.push_back(Ops[0]); 5641 Ops.erase(Ops.begin()); 5642 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5643 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5644 Ops, ""); 5645 } 5646 case NEON::BI__builtin_neon_vst4_lane_v: 5647 case NEON::BI__builtin_neon_vst4q_lane_v: { 5648 Ops.push_back(Ops[0]); 5649 Ops.erase(Ops.begin()); 5650 Ops[4] = Builder.CreateZExt(Ops[4], 5651 llvm::IntegerType::get(getLLVMContext(), 64)); 5652 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5653 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5654 Ops, ""); 5655 } 5656 case NEON::BI__builtin_neon_vtrn_v: 5657 case NEON::BI__builtin_neon_vtrnq_v: { 5658 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5659 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5660 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5661 Value *SV = nullptr; 5662 5663 for (unsigned vi = 0; vi != 2; ++vi) { 5664 SmallVector<Constant*, 16> Indices; 5665 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5666 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5667 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5668 } 5669 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5670 SV = llvm::ConstantVector::get(Indices); 5671 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5672 SV = Builder.CreateStore(SV, Addr); 5673 } 5674 return SV; 5675 } 5676 case NEON::BI__builtin_neon_vuzp_v: 5677 case NEON::BI__builtin_neon_vuzpq_v: { 5678 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5679 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5680 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5681 Value *SV = nullptr; 5682 5683 for (unsigned vi = 0; vi != 2; ++vi) { 5684 SmallVector<Constant*, 16> Indices; 5685 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5686 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5687 5688 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5689 SV = llvm::ConstantVector::get(Indices); 5690 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5691 SV = Builder.CreateStore(SV, Addr); 5692 } 5693 return SV; 5694 } 5695 case NEON::BI__builtin_neon_vzip_v: 5696 case NEON::BI__builtin_neon_vzipq_v: { 5697 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5698 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5699 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5700 Value *SV = nullptr; 5701 5702 for (unsigned vi = 0; vi != 2; ++vi) { 5703 SmallVector<Constant*, 16> Indices; 5704 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5705 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5706 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5707 } 5708 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5709 SV = llvm::ConstantVector::get(Indices); 5710 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5711 SV = Builder.CreateStore(SV, Addr); 5712 } 5713 return SV; 5714 } 5715 case NEON::BI__builtin_neon_vqtbl1q_v: { 5716 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5717 Ops, "vtbl1"); 5718 } 5719 case NEON::BI__builtin_neon_vqtbl2q_v: { 5720 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5721 Ops, "vtbl2"); 5722 } 5723 case NEON::BI__builtin_neon_vqtbl3q_v: { 5724 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5725 Ops, "vtbl3"); 5726 } 5727 case NEON::BI__builtin_neon_vqtbl4q_v: { 5728 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5729 Ops, "vtbl4"); 5730 } 5731 case NEON::BI__builtin_neon_vqtbx1q_v: { 5732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5733 Ops, "vtbx1"); 5734 } 5735 case NEON::BI__builtin_neon_vqtbx2q_v: { 5736 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5737 Ops, "vtbx2"); 5738 } 5739 case NEON::BI__builtin_neon_vqtbx3q_v: { 5740 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5741 Ops, "vtbx3"); 5742 } 5743 case NEON::BI__builtin_neon_vqtbx4q_v: { 5744 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5745 Ops, "vtbx4"); 5746 } 5747 case NEON::BI__builtin_neon_vsqadd_v: 5748 case NEON::BI__builtin_neon_vsqaddq_v: { 5749 Int = Intrinsic::aarch64_neon_usqadd; 5750 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5751 } 5752 case NEON::BI__builtin_neon_vuqadd_v: 5753 case NEON::BI__builtin_neon_vuqaddq_v: { 5754 Int = Intrinsic::aarch64_neon_suqadd; 5755 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5756 } 5757 } 5758 } 5759 5760 llvm::Value *CodeGenFunction:: 5761 BuildVector(ArrayRef<llvm::Value*> Ops) { 5762 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5763 "Not a power-of-two sized vector!"); 5764 bool AllConstants = true; 5765 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5766 AllConstants &= isa<Constant>(Ops[i]); 5767 5768 // If this is a constant vector, create a ConstantVector. 5769 if (AllConstants) { 5770 SmallVector<llvm::Constant*, 16> CstOps; 5771 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5772 CstOps.push_back(cast<Constant>(Ops[i])); 5773 return llvm::ConstantVector::get(CstOps); 5774 } 5775 5776 // Otherwise, insertelement the values to build the vector. 5777 Value *Result = 5778 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5779 5780 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5781 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5782 5783 return Result; 5784 } 5785 5786 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5787 const CallExpr *E) { 5788 SmallVector<Value*, 4> Ops; 5789 5790 // Find out if any arguments are required to be integer constant expressions. 5791 unsigned ICEArguments = 0; 5792 ASTContext::GetBuiltinTypeError Error; 5793 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5794 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5795 5796 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5797 // If this is a normal argument, just emit it as a scalar. 5798 if ((ICEArguments & (1 << i)) == 0) { 5799 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5800 continue; 5801 } 5802 5803 // If this is required to be a constant, constant fold it so that we know 5804 // that the generated intrinsic gets a ConstantInt. 5805 llvm::APSInt Result; 5806 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5807 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5808 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5809 } 5810 5811 switch (BuiltinID) { 5812 default: return nullptr; 5813 case X86::BI_mm_prefetch: { 5814 Value *Address = EmitScalarExpr(E->getArg(0)); 5815 Value *RW = ConstantInt::get(Int32Ty, 0); 5816 Value *Locality = EmitScalarExpr(E->getArg(1)); 5817 Value *Data = ConstantInt::get(Int32Ty, 1); 5818 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5819 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5820 } 5821 case X86::BI__builtin_ia32_vec_init_v8qi: 5822 case X86::BI__builtin_ia32_vec_init_v4hi: 5823 case X86::BI__builtin_ia32_vec_init_v2si: 5824 return Builder.CreateBitCast(BuildVector(Ops), 5825 llvm::Type::getX86_MMXTy(getLLVMContext())); 5826 case X86::BI__builtin_ia32_vec_ext_v2si: 5827 return Builder.CreateExtractElement(Ops[0], 5828 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5829 case X86::BI__builtin_ia32_ldmxcsr: { 5830 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5831 Builder.CreateStore(Ops[0], Tmp); 5832 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5833 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5834 } 5835 case X86::BI__builtin_ia32_stmxcsr: { 5836 Value *Tmp = CreateMemTemp(E->getType()); 5837 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5838 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5839 return Builder.CreateLoad(Tmp, "stmxcsr"); 5840 } 5841 case X86::BI__builtin_ia32_storehps: 5842 case X86::BI__builtin_ia32_storelps: { 5843 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5844 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5845 5846 // cast val v2i64 5847 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5848 5849 // extract (0, 1) 5850 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5851 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5852 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5853 5854 // cast pointer to i64 & store 5855 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5856 return Builder.CreateStore(Ops[1], Ops[0]); 5857 } 5858 case X86::BI__builtin_ia32_palignr: { 5859 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5860 5861 // If palignr is shifting the pair of input vectors less than 9 bytes, 5862 // emit a shuffle instruction. 5863 if (shiftVal <= 8) { 5864 SmallVector<llvm::Constant*, 8> Indices; 5865 for (unsigned i = 0; i != 8; ++i) 5866 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5867 5868 Value* SV = llvm::ConstantVector::get(Indices); 5869 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5870 } 5871 5872 // If palignr is shifting the pair of input vectors more than 8 but less 5873 // than 16 bytes, emit a logical right shift of the destination. 5874 if (shiftVal < 16) { 5875 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 5876 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 5877 5878 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5879 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 5880 5881 // create i32 constant 5882 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 5883 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5884 } 5885 5886 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 5887 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5888 } 5889 case X86::BI__builtin_ia32_palignr128: { 5890 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5891 5892 // If palignr is shifting the pair of input vectors less than 17 bytes, 5893 // emit a shuffle instruction. 5894 if (shiftVal <= 16) { 5895 SmallVector<llvm::Constant*, 16> Indices; 5896 for (unsigned i = 0; i != 16; ++i) 5897 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 5898 5899 Value* SV = llvm::ConstantVector::get(Indices); 5900 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5901 } 5902 5903 // If palignr is shifting the pair of input vectors more than 16 but less 5904 // than 32 bytes, emit a logical right shift of the destination. 5905 if (shiftVal < 32) { 5906 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5907 5908 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5909 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5910 5911 // create i32 constant 5912 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 5913 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5914 } 5915 5916 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5917 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5918 } 5919 case X86::BI__builtin_ia32_palignr256: { 5920 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5921 5922 // If palignr is shifting the pair of input vectors less than 17 bytes, 5923 // emit a shuffle instruction. 5924 if (shiftVal <= 16) { 5925 SmallVector<llvm::Constant*, 32> Indices; 5926 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5927 for (unsigned l = 0; l != 2; ++l) { 5928 unsigned LaneStart = l * 16; 5929 unsigned LaneEnd = (l+1) * 16; 5930 for (unsigned i = 0; i != 16; ++i) { 5931 unsigned Idx = shiftVal + i + LaneStart; 5932 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 5933 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 5934 } 5935 } 5936 5937 Value* SV = llvm::ConstantVector::get(Indices); 5938 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5939 } 5940 5941 // If palignr is shifting the pair of input vectors more than 16 but less 5942 // than 32 bytes, emit a logical right shift of the destination. 5943 if (shiftVal < 32) { 5944 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 5945 5946 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5947 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 5948 5949 // create i32 constant 5950 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 5951 return Builder.CreateCall(F, makeArrayRef(Ops.data(), 2), "palignr"); 5952 } 5953 5954 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 5955 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5956 } 5957 case X86::BI__builtin_ia32_movntps: 5958 case X86::BI__builtin_ia32_movntps256: 5959 case X86::BI__builtin_ia32_movntpd: 5960 case X86::BI__builtin_ia32_movntpd256: 5961 case X86::BI__builtin_ia32_movntdq: 5962 case X86::BI__builtin_ia32_movntdq256: 5963 case X86::BI__builtin_ia32_movnti: 5964 case X86::BI__builtin_ia32_movnti64: { 5965 llvm::MDNode *Node = llvm::MDNode::get( 5966 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 5967 5968 // Convert the type of the pointer to a pointer to the stored type. 5969 Value *BC = Builder.CreateBitCast(Ops[0], 5970 llvm::PointerType::getUnqual(Ops[1]->getType()), 5971 "cast"); 5972 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 5973 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 5974 5975 // If the operand is an integer, we can't assume alignment. Otherwise, 5976 // assume natural alignment. 5977 QualType ArgTy = E->getArg(1)->getType(); 5978 unsigned Align; 5979 if (ArgTy->isIntegerType()) 5980 Align = 1; 5981 else 5982 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 5983 SI->setAlignment(Align); 5984 return SI; 5985 } 5986 // 3DNow! 5987 case X86::BI__builtin_ia32_pswapdsf: 5988 case X86::BI__builtin_ia32_pswapdsi: { 5989 const char *name = nullptr; 5990 Intrinsic::ID ID = Intrinsic::not_intrinsic; 5991 switch(BuiltinID) { 5992 default: llvm_unreachable("Unsupported intrinsic!"); 5993 case X86::BI__builtin_ia32_pswapdsf: 5994 case X86::BI__builtin_ia32_pswapdsi: 5995 name = "pswapd"; 5996 ID = Intrinsic::x86_3dnowa_pswapd; 5997 break; 5998 } 5999 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6000 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6001 llvm::Function *F = CGM.getIntrinsic(ID); 6002 return Builder.CreateCall(F, Ops, name); 6003 } 6004 case X86::BI__builtin_ia32_rdrand16_step: 6005 case X86::BI__builtin_ia32_rdrand32_step: 6006 case X86::BI__builtin_ia32_rdrand64_step: 6007 case X86::BI__builtin_ia32_rdseed16_step: 6008 case X86::BI__builtin_ia32_rdseed32_step: 6009 case X86::BI__builtin_ia32_rdseed64_step: { 6010 Intrinsic::ID ID; 6011 switch (BuiltinID) { 6012 default: llvm_unreachable("Unsupported intrinsic!"); 6013 case X86::BI__builtin_ia32_rdrand16_step: 6014 ID = Intrinsic::x86_rdrand_16; 6015 break; 6016 case X86::BI__builtin_ia32_rdrand32_step: 6017 ID = Intrinsic::x86_rdrand_32; 6018 break; 6019 case X86::BI__builtin_ia32_rdrand64_step: 6020 ID = Intrinsic::x86_rdrand_64; 6021 break; 6022 case X86::BI__builtin_ia32_rdseed16_step: 6023 ID = Intrinsic::x86_rdseed_16; 6024 break; 6025 case X86::BI__builtin_ia32_rdseed32_step: 6026 ID = Intrinsic::x86_rdseed_32; 6027 break; 6028 case X86::BI__builtin_ia32_rdseed64_step: 6029 ID = Intrinsic::x86_rdseed_64; 6030 break; 6031 } 6032 6033 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6034 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6035 return Builder.CreateExtractValue(Call, 1); 6036 } 6037 // AVX2 broadcast 6038 case X86::BI__builtin_ia32_vbroadcastsi256: { 6039 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 6040 Builder.CreateStore(Ops[0], VecTmp); 6041 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 6042 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 6043 } 6044 } 6045 } 6046 6047 6048 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6049 const CallExpr *E) { 6050 SmallVector<Value*, 4> Ops; 6051 6052 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6053 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6054 6055 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6056 6057 switch (BuiltinID) { 6058 default: return nullptr; 6059 6060 // vec_ld, vec_lvsl, vec_lvsr 6061 case PPC::BI__builtin_altivec_lvx: 6062 case PPC::BI__builtin_altivec_lvxl: 6063 case PPC::BI__builtin_altivec_lvebx: 6064 case PPC::BI__builtin_altivec_lvehx: 6065 case PPC::BI__builtin_altivec_lvewx: 6066 case PPC::BI__builtin_altivec_lvsl: 6067 case PPC::BI__builtin_altivec_lvsr: 6068 case PPC::BI__builtin_vsx_lxvd2x: 6069 case PPC::BI__builtin_vsx_lxvw4x: 6070 { 6071 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6072 6073 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6074 Ops.pop_back(); 6075 6076 switch (BuiltinID) { 6077 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6078 case PPC::BI__builtin_altivec_lvx: 6079 ID = Intrinsic::ppc_altivec_lvx; 6080 break; 6081 case PPC::BI__builtin_altivec_lvxl: 6082 ID = Intrinsic::ppc_altivec_lvxl; 6083 break; 6084 case PPC::BI__builtin_altivec_lvebx: 6085 ID = Intrinsic::ppc_altivec_lvebx; 6086 break; 6087 case PPC::BI__builtin_altivec_lvehx: 6088 ID = Intrinsic::ppc_altivec_lvehx; 6089 break; 6090 case PPC::BI__builtin_altivec_lvewx: 6091 ID = Intrinsic::ppc_altivec_lvewx; 6092 break; 6093 case PPC::BI__builtin_altivec_lvsl: 6094 ID = Intrinsic::ppc_altivec_lvsl; 6095 break; 6096 case PPC::BI__builtin_altivec_lvsr: 6097 ID = Intrinsic::ppc_altivec_lvsr; 6098 break; 6099 case PPC::BI__builtin_vsx_lxvd2x: 6100 ID = Intrinsic::ppc_vsx_lxvd2x; 6101 break; 6102 case PPC::BI__builtin_vsx_lxvw4x: 6103 ID = Intrinsic::ppc_vsx_lxvw4x; 6104 break; 6105 } 6106 llvm::Function *F = CGM.getIntrinsic(ID); 6107 return Builder.CreateCall(F, Ops, ""); 6108 } 6109 6110 // vec_st 6111 case PPC::BI__builtin_altivec_stvx: 6112 case PPC::BI__builtin_altivec_stvxl: 6113 case PPC::BI__builtin_altivec_stvebx: 6114 case PPC::BI__builtin_altivec_stvehx: 6115 case PPC::BI__builtin_altivec_stvewx: 6116 case PPC::BI__builtin_vsx_stxvd2x: 6117 case PPC::BI__builtin_vsx_stxvw4x: 6118 { 6119 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6120 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6121 Ops.pop_back(); 6122 6123 switch (BuiltinID) { 6124 default: llvm_unreachable("Unsupported st intrinsic!"); 6125 case PPC::BI__builtin_altivec_stvx: 6126 ID = Intrinsic::ppc_altivec_stvx; 6127 break; 6128 case PPC::BI__builtin_altivec_stvxl: 6129 ID = Intrinsic::ppc_altivec_stvxl; 6130 break; 6131 case PPC::BI__builtin_altivec_stvebx: 6132 ID = Intrinsic::ppc_altivec_stvebx; 6133 break; 6134 case PPC::BI__builtin_altivec_stvehx: 6135 ID = Intrinsic::ppc_altivec_stvehx; 6136 break; 6137 case PPC::BI__builtin_altivec_stvewx: 6138 ID = Intrinsic::ppc_altivec_stvewx; 6139 break; 6140 case PPC::BI__builtin_vsx_stxvd2x: 6141 ID = Intrinsic::ppc_vsx_stxvd2x; 6142 break; 6143 case PPC::BI__builtin_vsx_stxvw4x: 6144 ID = Intrinsic::ppc_vsx_stxvw4x; 6145 break; 6146 } 6147 llvm::Function *F = CGM.getIntrinsic(ID); 6148 return Builder.CreateCall(F, Ops, ""); 6149 } 6150 } 6151 } 6152 6153 // Emit an intrinsic that has 1 float or double. 6154 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6155 const CallExpr *E, 6156 unsigned IntrinsicID) { 6157 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6158 6159 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6160 return CGF.Builder.CreateCall(F, Src0); 6161 } 6162 6163 // Emit an intrinsic that has 3 float or double operands. 6164 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6165 const CallExpr *E, 6166 unsigned IntrinsicID) { 6167 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6168 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6169 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6170 6171 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6172 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6173 } 6174 6175 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6176 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6177 const CallExpr *E, 6178 unsigned IntrinsicID) { 6179 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6180 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6181 6182 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6183 return CGF.Builder.CreateCall2(F, Src0, Src1); 6184 } 6185 6186 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6187 const CallExpr *E) { 6188 switch (BuiltinID) { 6189 case R600::BI__builtin_amdgpu_div_scale: 6190 case R600::BI__builtin_amdgpu_div_scalef: { 6191 // Translate from the intrinsics's struct return to the builtin's out 6192 // argument. 6193 6194 std::pair<llvm::Value *, unsigned> FlagOutPtr 6195 = EmitPointerWithAlignment(E->getArg(3)); 6196 6197 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6198 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6199 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6200 6201 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6202 X->getType()); 6203 6204 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6205 6206 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6207 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6208 6209 llvm::Type *RealFlagType 6210 = FlagOutPtr.first->getType()->getPointerElementType(); 6211 6212 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6213 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6214 FlagStore->setAlignment(FlagOutPtr.second); 6215 return Result; 6216 } 6217 case R600::BI__builtin_amdgpu_div_fmas: 6218 case R600::BI__builtin_amdgpu_div_fmasf: { 6219 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6220 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6221 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6222 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6223 6224 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6225 Src0->getType()); 6226 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6227 return Builder.CreateCall4(F, Src0, Src1, Src2, Src3ToBool); 6228 } 6229 case R600::BI__builtin_amdgpu_div_fixup: 6230 case R600::BI__builtin_amdgpu_div_fixupf: 6231 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6232 case R600::BI__builtin_amdgpu_trig_preop: 6233 case R600::BI__builtin_amdgpu_trig_preopf: 6234 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6235 case R600::BI__builtin_amdgpu_rcp: 6236 case R600::BI__builtin_amdgpu_rcpf: 6237 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6238 case R600::BI__builtin_amdgpu_rsq: 6239 case R600::BI__builtin_amdgpu_rsqf: 6240 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6241 case R600::BI__builtin_amdgpu_rsq_clamped: 6242 case R600::BI__builtin_amdgpu_rsq_clampedf: 6243 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6244 case R600::BI__builtin_amdgpu_ldexp: 6245 case R600::BI__builtin_amdgpu_ldexpf: 6246 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6247 default: 6248 return nullptr; 6249 } 6250 } 6251