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/CallSite.h" 25 #include "llvm/IR/DataLayout.h" 26 #include "llvm/IR/InlineAsm.h" 27 #include "llvm/IR/Intrinsics.h" 28 29 using namespace clang; 30 using namespace CodeGen; 31 using namespace llvm; 32 33 /// getBuiltinLibFunction - Given a builtin id for a function like 34 /// "__builtin_fabsf", return a Function* for "fabsf". 35 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 36 unsigned BuiltinID) { 37 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 38 39 // Get the name, skip over the __builtin_ prefix (if necessary). 40 StringRef Name; 41 GlobalDecl D(FD); 42 43 // If the builtin has been declared explicitly with an assembler label, 44 // use the mangled name. This differs from the plain label on platforms 45 // that prefix labels. 46 if (FD->hasAttr<AsmLabelAttr>()) 47 Name = getMangledName(D); 48 else 49 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 50 51 llvm::FunctionType *Ty = 52 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 53 54 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 55 } 56 57 /// Emit the conversions required to turn the given value into an 58 /// integer of the given size. 59 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 60 QualType T, llvm::IntegerType *IntType) { 61 V = CGF.EmitToMemory(V, T); 62 63 if (V->getType()->isPointerTy()) 64 return CGF.Builder.CreatePtrToInt(V, IntType); 65 66 assert(V->getType() == IntType); 67 return V; 68 } 69 70 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 71 QualType T, llvm::Type *ResultType) { 72 V = CGF.EmitFromMemory(V, T); 73 74 if (ResultType->isPointerTy()) 75 return CGF.Builder.CreateIntToPtr(V, ResultType); 76 77 assert(V->getType() == ResultType); 78 return V; 79 } 80 81 /// Utility to insert an atomic instruction based on Instrinsic::ID 82 /// and the expression node. 83 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 84 llvm::AtomicRMWInst::BinOp Kind, 85 const CallExpr *E) { 86 QualType T = E->getType(); 87 assert(E->getArg(0)->getType()->isPointerType()); 88 assert(CGF.getContext().hasSameUnqualifiedType(T, 89 E->getArg(0)->getType()->getPointeeType())); 90 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 91 92 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 93 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 94 95 llvm::IntegerType *IntType = 96 llvm::IntegerType::get(CGF.getLLVMContext(), 97 CGF.getContext().getTypeSize(T)); 98 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 99 100 llvm::Value *Args[2]; 101 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 102 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 103 llvm::Type *ValueType = Args[1]->getType(); 104 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 105 106 llvm::Value *Result = 107 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 108 llvm::SequentiallyConsistent); 109 Result = EmitFromInt(CGF, Result, T, ValueType); 110 return RValue::get(Result); 111 } 112 113 /// Utility to insert an atomic instruction based Instrinsic::ID and 114 /// the expression node, where the return value is the result of the 115 /// operation. 116 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 117 llvm::AtomicRMWInst::BinOp Kind, 118 const CallExpr *E, 119 Instruction::BinaryOps Op, 120 bool Invert = false) { 121 QualType T = E->getType(); 122 assert(E->getArg(0)->getType()->isPointerType()); 123 assert(CGF.getContext().hasSameUnqualifiedType(T, 124 E->getArg(0)->getType()->getPointeeType())); 125 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 126 127 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 128 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 129 130 llvm::IntegerType *IntType = 131 llvm::IntegerType::get(CGF.getLLVMContext(), 132 CGF.getContext().getTypeSize(T)); 133 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 134 135 llvm::Value *Args[2]; 136 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 137 llvm::Type *ValueType = Args[1]->getType(); 138 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 139 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 140 141 llvm::Value *Result = 142 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 143 llvm::SequentiallyConsistent); 144 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 145 if (Invert) 146 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 147 llvm::ConstantInt::get(IntType, -1)); 148 Result = EmitFromInt(CGF, Result, T, ValueType); 149 return RValue::get(Result); 150 } 151 152 /// EmitFAbs - Emit a call to @llvm.fabs(). 153 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 154 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 155 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 156 Call->setDoesNotAccessMemory(); 157 return Call; 158 } 159 160 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 161 const CallExpr *E, llvm::Value *calleeValue) { 162 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 163 ReturnValueSlot(), Fn); 164 } 165 166 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 167 /// depending on IntrinsicID. 168 /// 169 /// \arg CGF The current codegen function. 170 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 171 /// \arg X The first argument to the llvm.*.with.overflow.*. 172 /// \arg Y The second argument to the llvm.*.with.overflow.*. 173 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 174 /// \returns The result (i.e. sum/product) returned by the intrinsic. 175 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 176 const llvm::Intrinsic::ID IntrinsicID, 177 llvm::Value *X, llvm::Value *Y, 178 llvm::Value *&Carry) { 179 // Make sure we have integers of the same width. 180 assert(X->getType() == Y->getType() && 181 "Arguments must be the same type. (Did you forget to make sure both " 182 "arguments have the same integer width?)"); 183 184 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 185 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 186 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 187 return CGF.Builder.CreateExtractValue(Tmp, 0); 188 } 189 190 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 191 unsigned BuiltinID, const CallExpr *E, 192 ReturnValueSlot ReturnValue) { 193 // See if we can constant fold this builtin. If so, don't emit it at all. 194 Expr::EvalResult Result; 195 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 196 !Result.hasSideEffects()) { 197 if (Result.Val.isInt()) 198 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 199 Result.Val.getInt())); 200 if (Result.Val.isFloat()) 201 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 202 Result.Val.getFloat())); 203 } 204 205 switch (BuiltinID) { 206 default: break; // Handle intrinsics and libm functions below. 207 case Builtin::BI__builtin___CFStringMakeConstantString: 208 case Builtin::BI__builtin___NSStringMakeConstantString: 209 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 210 case Builtin::BI__builtin_stdarg_start: 211 case Builtin::BI__builtin_va_start: 212 case Builtin::BI__va_start: 213 case Builtin::BI__builtin_va_end: { 214 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 215 ? EmitScalarExpr(E->getArg(0)) 216 : EmitVAListRef(E->getArg(0)); 217 llvm::Type *DestType = Int8PtrTy; 218 if (ArgValue->getType() != DestType) 219 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 220 ArgValue->getName().data()); 221 222 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 223 Intrinsic::vaend : Intrinsic::vastart; 224 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 225 } 226 case Builtin::BI__builtin_va_copy: { 227 Value *DstPtr = EmitVAListRef(E->getArg(0)); 228 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 229 230 llvm::Type *Type = Int8PtrTy; 231 232 DstPtr = Builder.CreateBitCast(DstPtr, Type); 233 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 234 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 235 DstPtr, SrcPtr)); 236 } 237 case Builtin::BI__builtin_abs: 238 case Builtin::BI__builtin_labs: 239 case Builtin::BI__builtin_llabs: { 240 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 241 242 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 243 Value *CmpResult = 244 Builder.CreateICmpSGE(ArgValue, 245 llvm::Constant::getNullValue(ArgValue->getType()), 246 "abscond"); 247 Value *Result = 248 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 249 250 return RValue::get(Result); 251 } 252 case Builtin::BI__builtin_fabs: 253 case Builtin::BI__builtin_fabsf: 254 case Builtin::BI__builtin_fabsl: { 255 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 256 Value *Result = EmitFAbs(*this, Arg1); 257 return RValue::get(Result); 258 } 259 case Builtin::BI__builtin_fmod: 260 case Builtin::BI__builtin_fmodf: 261 case Builtin::BI__builtin_fmodl: { 262 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 263 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 264 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 265 return RValue::get(Result); 266 } 267 268 case Builtin::BI__builtin_conj: 269 case Builtin::BI__builtin_conjf: 270 case Builtin::BI__builtin_conjl: { 271 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 272 Value *Real = ComplexVal.first; 273 Value *Imag = ComplexVal.second; 274 Value *Zero = 275 Imag->getType()->isFPOrFPVectorTy() 276 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 277 : llvm::Constant::getNullValue(Imag->getType()); 278 279 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 280 return RValue::getComplex(std::make_pair(Real, Imag)); 281 } 282 case Builtin::BI__builtin_creal: 283 case Builtin::BI__builtin_crealf: 284 case Builtin::BI__builtin_creall: 285 case Builtin::BIcreal: 286 case Builtin::BIcrealf: 287 case Builtin::BIcreall: { 288 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 289 return RValue::get(ComplexVal.first); 290 } 291 292 case Builtin::BI__builtin_cimag: 293 case Builtin::BI__builtin_cimagf: 294 case Builtin::BI__builtin_cimagl: 295 case Builtin::BIcimag: 296 case Builtin::BIcimagf: 297 case Builtin::BIcimagl: { 298 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 299 return RValue::get(ComplexVal.second); 300 } 301 302 case Builtin::BI__builtin_ctzs: 303 case Builtin::BI__builtin_ctz: 304 case Builtin::BI__builtin_ctzl: 305 case Builtin::BI__builtin_ctzll: { 306 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 307 308 llvm::Type *ArgType = ArgValue->getType(); 309 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 310 311 llvm::Type *ResultType = ConvertType(E->getType()); 312 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 313 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 314 if (Result->getType() != ResultType) 315 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 316 "cast"); 317 return RValue::get(Result); 318 } 319 case Builtin::BI__builtin_clzs: 320 case Builtin::BI__builtin_clz: 321 case Builtin::BI__builtin_clzl: 322 case Builtin::BI__builtin_clzll: { 323 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 324 325 llvm::Type *ArgType = ArgValue->getType(); 326 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 327 328 llvm::Type *ResultType = ConvertType(E->getType()); 329 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 330 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 331 if (Result->getType() != ResultType) 332 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 333 "cast"); 334 return RValue::get(Result); 335 } 336 case Builtin::BI__builtin_ffs: 337 case Builtin::BI__builtin_ffsl: 338 case Builtin::BI__builtin_ffsll: { 339 // ffs(x) -> x ? cttz(x) + 1 : 0 340 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 341 342 llvm::Type *ArgType = ArgValue->getType(); 343 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 344 345 llvm::Type *ResultType = ConvertType(E->getType()); 346 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 347 Builder.getTrue()), 348 llvm::ConstantInt::get(ArgType, 1)); 349 Value *Zero = llvm::Constant::getNullValue(ArgType); 350 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 351 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 352 if (Result->getType() != ResultType) 353 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 354 "cast"); 355 return RValue::get(Result); 356 } 357 case Builtin::BI__builtin_parity: 358 case Builtin::BI__builtin_parityl: 359 case Builtin::BI__builtin_parityll: { 360 // parity(x) -> ctpop(x) & 1 361 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 362 363 llvm::Type *ArgType = ArgValue->getType(); 364 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 365 366 llvm::Type *ResultType = ConvertType(E->getType()); 367 Value *Tmp = Builder.CreateCall(F, ArgValue); 368 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 369 if (Result->getType() != ResultType) 370 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 371 "cast"); 372 return RValue::get(Result); 373 } 374 case Builtin::BI__builtin_popcount: 375 case Builtin::BI__builtin_popcountl: 376 case Builtin::BI__builtin_popcountll: { 377 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 378 379 llvm::Type *ArgType = ArgValue->getType(); 380 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 381 382 llvm::Type *ResultType = ConvertType(E->getType()); 383 Value *Result = Builder.CreateCall(F, ArgValue); 384 if (Result->getType() != ResultType) 385 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 386 "cast"); 387 return RValue::get(Result); 388 } 389 case Builtin::BI__builtin_expect: { 390 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 391 llvm::Type *ArgType = ArgValue->getType(); 392 393 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 394 // Don't generate llvm.expect on -O0 as the backend won't use it for 395 // anything. 396 // Note, we still IRGen ExpectedValue because it could have side-effects. 397 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 398 return RValue::get(ArgValue); 399 400 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 401 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 402 "expval"); 403 return RValue::get(Result); 404 } 405 case Builtin::BI__builtin_assume_aligned: { 406 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 407 Value *OffsetValue = 408 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 409 410 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 411 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 412 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 413 414 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 415 return RValue::get(PtrValue); 416 } 417 case Builtin::BI__assume: 418 case Builtin::BI__builtin_assume: { 419 if (E->getArg(0)->HasSideEffects(getContext())) 420 return RValue::get(nullptr); 421 422 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 423 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 424 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 425 } 426 case Builtin::BI__builtin_bswap16: 427 case Builtin::BI__builtin_bswap32: 428 case Builtin::BI__builtin_bswap64: { 429 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 430 llvm::Type *ArgType = ArgValue->getType(); 431 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 432 return RValue::get(Builder.CreateCall(F, ArgValue)); 433 } 434 case Builtin::BI__builtin_object_size: { 435 // We rely on constant folding to deal with expressions with side effects. 436 assert(!E->getArg(0)->HasSideEffects(getContext()) && 437 "should have been constant folded"); 438 439 // We pass this builtin onto the optimizer so that it can 440 // figure out the object size in more complex cases. 441 llvm::Type *ResType = ConvertType(E->getType()); 442 443 // LLVM only supports 0 and 2, make sure that we pass along that 444 // as a boolean. 445 Value *Ty = EmitScalarExpr(E->getArg(1)); 446 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 447 assert(CI); 448 uint64_t val = CI->getZExtValue(); 449 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 450 // FIXME: Get right address space. 451 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 452 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 453 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 454 } 455 case Builtin::BI__builtin_prefetch: { 456 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 457 // FIXME: Technically these constants should of type 'int', yes? 458 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 459 llvm::ConstantInt::get(Int32Ty, 0); 460 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 461 llvm::ConstantInt::get(Int32Ty, 3); 462 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 463 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 464 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 465 } 466 case Builtin::BI__builtin_readcyclecounter: { 467 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 468 return RValue::get(Builder.CreateCall(F)); 469 } 470 case Builtin::BI__builtin___clear_cache: { 471 Value *Begin = EmitScalarExpr(E->getArg(0)); 472 Value *End = EmitScalarExpr(E->getArg(1)); 473 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 474 return RValue::get(Builder.CreateCall2(F, Begin, End)); 475 } 476 case Builtin::BI__builtin_trap: { 477 Value *F = CGM.getIntrinsic(Intrinsic::trap); 478 return RValue::get(Builder.CreateCall(F)); 479 } 480 case Builtin::BI__debugbreak: { 481 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 482 return RValue::get(Builder.CreateCall(F)); 483 } 484 case Builtin::BI__builtin_unreachable: { 485 if (SanOpts.has(SanitizerKind::Unreachable)) { 486 SanitizerScope SanScope(this); 487 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 488 SanitizerKind::Unreachable), 489 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 490 None); 491 } else 492 Builder.CreateUnreachable(); 493 494 // We do need to preserve an insertion point. 495 EmitBlock(createBasicBlock("unreachable.cont")); 496 497 return RValue::get(nullptr); 498 } 499 500 case Builtin::BI__builtin_powi: 501 case Builtin::BI__builtin_powif: 502 case Builtin::BI__builtin_powil: { 503 Value *Base = EmitScalarExpr(E->getArg(0)); 504 Value *Exponent = EmitScalarExpr(E->getArg(1)); 505 llvm::Type *ArgType = Base->getType(); 506 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 507 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 508 } 509 510 case Builtin::BI__builtin_isgreater: 511 case Builtin::BI__builtin_isgreaterequal: 512 case Builtin::BI__builtin_isless: 513 case Builtin::BI__builtin_islessequal: 514 case Builtin::BI__builtin_islessgreater: 515 case Builtin::BI__builtin_isunordered: { 516 // Ordered comparisons: we know the arguments to these are matching scalar 517 // floating point values. 518 Value *LHS = EmitScalarExpr(E->getArg(0)); 519 Value *RHS = EmitScalarExpr(E->getArg(1)); 520 521 switch (BuiltinID) { 522 default: llvm_unreachable("Unknown ordered comparison"); 523 case Builtin::BI__builtin_isgreater: 524 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 525 break; 526 case Builtin::BI__builtin_isgreaterequal: 527 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 528 break; 529 case Builtin::BI__builtin_isless: 530 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 531 break; 532 case Builtin::BI__builtin_islessequal: 533 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 534 break; 535 case Builtin::BI__builtin_islessgreater: 536 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 537 break; 538 case Builtin::BI__builtin_isunordered: 539 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 540 break; 541 } 542 // ZExt bool to int type. 543 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 544 } 545 case Builtin::BI__builtin_isnan: { 546 Value *V = EmitScalarExpr(E->getArg(0)); 547 V = Builder.CreateFCmpUNO(V, V, "cmp"); 548 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 549 } 550 551 case Builtin::BI__builtin_isinf: { 552 // isinf(x) --> fabs(x) == infinity 553 Value *V = EmitScalarExpr(E->getArg(0)); 554 V = EmitFAbs(*this, V); 555 556 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 557 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 558 } 559 560 // TODO: BI__builtin_isinf_sign 561 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 562 563 case Builtin::BI__builtin_isnormal: { 564 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 565 Value *V = EmitScalarExpr(E->getArg(0)); 566 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 567 568 Value *Abs = EmitFAbs(*this, V); 569 Value *IsLessThanInf = 570 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 571 APFloat Smallest = APFloat::getSmallestNormalized( 572 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 573 Value *IsNormal = 574 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 575 "isnormal"); 576 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 577 V = Builder.CreateAnd(V, IsNormal, "and"); 578 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 579 } 580 581 case Builtin::BI__builtin_isfinite: { 582 // isfinite(x) --> x == x && fabs(x) != infinity; 583 Value *V = EmitScalarExpr(E->getArg(0)); 584 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 585 586 Value *Abs = EmitFAbs(*this, V); 587 Value *IsNotInf = 588 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 589 590 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 591 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 592 } 593 594 case Builtin::BI__builtin_fpclassify: { 595 Value *V = EmitScalarExpr(E->getArg(5)); 596 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 597 598 // Create Result 599 BasicBlock *Begin = Builder.GetInsertBlock(); 600 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 601 Builder.SetInsertPoint(End); 602 PHINode *Result = 603 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 604 "fpclassify_result"); 605 606 // if (V==0) return FP_ZERO 607 Builder.SetInsertPoint(Begin); 608 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 609 "iszero"); 610 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 611 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 612 Builder.CreateCondBr(IsZero, End, NotZero); 613 Result->addIncoming(ZeroLiteral, Begin); 614 615 // if (V != V) return FP_NAN 616 Builder.SetInsertPoint(NotZero); 617 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 618 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 619 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 620 Builder.CreateCondBr(IsNan, End, NotNan); 621 Result->addIncoming(NanLiteral, NotZero); 622 623 // if (fabs(V) == infinity) return FP_INFINITY 624 Builder.SetInsertPoint(NotNan); 625 Value *VAbs = EmitFAbs(*this, V); 626 Value *IsInf = 627 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 628 "isinf"); 629 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 630 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 631 Builder.CreateCondBr(IsInf, End, NotInf); 632 Result->addIncoming(InfLiteral, NotNan); 633 634 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 635 Builder.SetInsertPoint(NotInf); 636 APFloat Smallest = APFloat::getSmallestNormalized( 637 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 638 Value *IsNormal = 639 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 640 "isnormal"); 641 Value *NormalResult = 642 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 643 EmitScalarExpr(E->getArg(3))); 644 Builder.CreateBr(End); 645 Result->addIncoming(NormalResult, NotInf); 646 647 // return Result 648 Builder.SetInsertPoint(End); 649 return RValue::get(Result); 650 } 651 652 case Builtin::BIalloca: 653 case Builtin::BI_alloca: 654 case Builtin::BI__builtin_alloca: { 655 Value *Size = EmitScalarExpr(E->getArg(0)); 656 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 657 } 658 case Builtin::BIbzero: 659 case Builtin::BI__builtin_bzero: { 660 std::pair<llvm::Value*, unsigned> Dest = 661 EmitPointerWithAlignment(E->getArg(0)); 662 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 663 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 664 Dest.second, false); 665 return RValue::get(Dest.first); 666 } 667 case Builtin::BImemcpy: 668 case Builtin::BI__builtin_memcpy: { 669 std::pair<llvm::Value*, unsigned> Dest = 670 EmitPointerWithAlignment(E->getArg(0)); 671 std::pair<llvm::Value*, unsigned> Src = 672 EmitPointerWithAlignment(E->getArg(1)); 673 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 674 unsigned Align = std::min(Dest.second, Src.second); 675 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 676 return RValue::get(Dest.first); 677 } 678 679 case Builtin::BI__builtin___memcpy_chk: { 680 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 681 llvm::APSInt Size, DstSize; 682 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 683 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 684 break; 685 if (Size.ugt(DstSize)) 686 break; 687 std::pair<llvm::Value*, unsigned> Dest = 688 EmitPointerWithAlignment(E->getArg(0)); 689 std::pair<llvm::Value*, unsigned> Src = 690 EmitPointerWithAlignment(E->getArg(1)); 691 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 692 unsigned Align = std::min(Dest.second, Src.second); 693 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 694 return RValue::get(Dest.first); 695 } 696 697 case Builtin::BI__builtin_objc_memmove_collectable: { 698 Value *Address = EmitScalarExpr(E->getArg(0)); 699 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 700 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 701 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 702 Address, SrcAddr, SizeVal); 703 return RValue::get(Address); 704 } 705 706 case Builtin::BI__builtin___memmove_chk: { 707 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 708 llvm::APSInt Size, DstSize; 709 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 710 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 711 break; 712 if (Size.ugt(DstSize)) 713 break; 714 std::pair<llvm::Value*, unsigned> Dest = 715 EmitPointerWithAlignment(E->getArg(0)); 716 std::pair<llvm::Value*, unsigned> Src = 717 EmitPointerWithAlignment(E->getArg(1)); 718 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 719 unsigned Align = std::min(Dest.second, Src.second); 720 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 721 return RValue::get(Dest.first); 722 } 723 724 case Builtin::BImemmove: 725 case Builtin::BI__builtin_memmove: { 726 std::pair<llvm::Value*, unsigned> Dest = 727 EmitPointerWithAlignment(E->getArg(0)); 728 std::pair<llvm::Value*, unsigned> Src = 729 EmitPointerWithAlignment(E->getArg(1)); 730 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 731 unsigned Align = std::min(Dest.second, Src.second); 732 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 733 return RValue::get(Dest.first); 734 } 735 case Builtin::BImemset: 736 case Builtin::BI__builtin_memset: { 737 std::pair<llvm::Value*, unsigned> Dest = 738 EmitPointerWithAlignment(E->getArg(0)); 739 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 740 Builder.getInt8Ty()); 741 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 742 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 743 return RValue::get(Dest.first); 744 } 745 case Builtin::BI__builtin___memset_chk: { 746 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 747 llvm::APSInt Size, DstSize; 748 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 749 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 750 break; 751 if (Size.ugt(DstSize)) 752 break; 753 std::pair<llvm::Value*, unsigned> Dest = 754 EmitPointerWithAlignment(E->getArg(0)); 755 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 756 Builder.getInt8Ty()); 757 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 758 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 759 return RValue::get(Dest.first); 760 } 761 case Builtin::BI__builtin_dwarf_cfa: { 762 // The offset in bytes from the first argument to the CFA. 763 // 764 // Why on earth is this in the frontend? Is there any reason at 765 // all that the backend can't reasonably determine this while 766 // lowering llvm.eh.dwarf.cfa()? 767 // 768 // TODO: If there's a satisfactory reason, add a target hook for 769 // this instead of hard-coding 0, which is correct for most targets. 770 int32_t Offset = 0; 771 772 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 773 return RValue::get(Builder.CreateCall(F, 774 llvm::ConstantInt::get(Int32Ty, Offset))); 775 } 776 case Builtin::BI__builtin_return_address: { 777 Value *Depth = EmitScalarExpr(E->getArg(0)); 778 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 779 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 780 return RValue::get(Builder.CreateCall(F, Depth)); 781 } 782 case Builtin::BI__builtin_frame_address: { 783 Value *Depth = EmitScalarExpr(E->getArg(0)); 784 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 785 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 786 return RValue::get(Builder.CreateCall(F, Depth)); 787 } 788 case Builtin::BI__builtin_extract_return_addr: { 789 Value *Address = EmitScalarExpr(E->getArg(0)); 790 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 791 return RValue::get(Result); 792 } 793 case Builtin::BI__builtin_frob_return_addr: { 794 Value *Address = EmitScalarExpr(E->getArg(0)); 795 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 796 return RValue::get(Result); 797 } 798 case Builtin::BI__builtin_dwarf_sp_column: { 799 llvm::IntegerType *Ty 800 = cast<llvm::IntegerType>(ConvertType(E->getType())); 801 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 802 if (Column == -1) { 803 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 804 return RValue::get(llvm::UndefValue::get(Ty)); 805 } 806 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 807 } 808 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 809 Value *Address = EmitScalarExpr(E->getArg(0)); 810 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 811 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 812 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 813 } 814 case Builtin::BI__builtin_eh_return: { 815 Value *Int = EmitScalarExpr(E->getArg(0)); 816 Value *Ptr = EmitScalarExpr(E->getArg(1)); 817 818 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 819 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 820 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 821 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 822 ? Intrinsic::eh_return_i32 823 : Intrinsic::eh_return_i64); 824 Builder.CreateCall2(F, Int, Ptr); 825 Builder.CreateUnreachable(); 826 827 // We do need to preserve an insertion point. 828 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 829 830 return RValue::get(nullptr); 831 } 832 case Builtin::BI__builtin_unwind_init: { 833 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 834 return RValue::get(Builder.CreateCall(F)); 835 } 836 case Builtin::BI__builtin_extend_pointer: { 837 // Extends a pointer to the size of an _Unwind_Word, which is 838 // uint64_t on all platforms. Generally this gets poked into a 839 // register and eventually used as an address, so if the 840 // addressing registers are wider than pointers and the platform 841 // doesn't implicitly ignore high-order bits when doing 842 // addressing, we need to make sure we zext / sext based on 843 // the platform's expectations. 844 // 845 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 846 847 // Cast the pointer to intptr_t. 848 Value *Ptr = EmitScalarExpr(E->getArg(0)); 849 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 850 851 // If that's 64 bits, we're done. 852 if (IntPtrTy->getBitWidth() == 64) 853 return RValue::get(Result); 854 855 // Otherwise, ask the codegen data what to do. 856 if (getTargetHooks().extendPointerWithSExt()) 857 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 858 else 859 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 860 } 861 case Builtin::BI__builtin_setjmp: { 862 // Buffer is a void**. 863 Value *Buf = EmitScalarExpr(E->getArg(0)); 864 865 // Store the frame pointer to the setjmp buffer. 866 Value *FrameAddr = 867 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 868 ConstantInt::get(Int32Ty, 0)); 869 Builder.CreateStore(FrameAddr, Buf); 870 871 // Store the stack pointer to the setjmp buffer. 872 Value *StackAddr = 873 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 874 Value *StackSaveSlot = 875 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 876 Builder.CreateStore(StackAddr, StackSaveSlot); 877 878 // Call LLVM's EH setjmp, which is lightweight. 879 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 880 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 881 return RValue::get(Builder.CreateCall(F, Buf)); 882 } 883 case Builtin::BI__builtin_longjmp: { 884 Value *Buf = EmitScalarExpr(E->getArg(0)); 885 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 886 887 // Call LLVM's EH longjmp, which is lightweight. 888 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 889 890 // longjmp doesn't return; mark this as unreachable. 891 Builder.CreateUnreachable(); 892 893 // We do need to preserve an insertion point. 894 EmitBlock(createBasicBlock("longjmp.cont")); 895 896 return RValue::get(nullptr); 897 } 898 case Builtin::BI__sync_fetch_and_add: 899 case Builtin::BI__sync_fetch_and_sub: 900 case Builtin::BI__sync_fetch_and_or: 901 case Builtin::BI__sync_fetch_and_and: 902 case Builtin::BI__sync_fetch_and_xor: 903 case Builtin::BI__sync_fetch_and_nand: 904 case Builtin::BI__sync_add_and_fetch: 905 case Builtin::BI__sync_sub_and_fetch: 906 case Builtin::BI__sync_and_and_fetch: 907 case Builtin::BI__sync_or_and_fetch: 908 case Builtin::BI__sync_xor_and_fetch: 909 case Builtin::BI__sync_nand_and_fetch: 910 case Builtin::BI__sync_val_compare_and_swap: 911 case Builtin::BI__sync_bool_compare_and_swap: 912 case Builtin::BI__sync_lock_test_and_set: 913 case Builtin::BI__sync_lock_release: 914 case Builtin::BI__sync_swap: 915 llvm_unreachable("Shouldn't make it through sema"); 916 case Builtin::BI__sync_fetch_and_add_1: 917 case Builtin::BI__sync_fetch_and_add_2: 918 case Builtin::BI__sync_fetch_and_add_4: 919 case Builtin::BI__sync_fetch_and_add_8: 920 case Builtin::BI__sync_fetch_and_add_16: 921 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 922 case Builtin::BI__sync_fetch_and_sub_1: 923 case Builtin::BI__sync_fetch_and_sub_2: 924 case Builtin::BI__sync_fetch_and_sub_4: 925 case Builtin::BI__sync_fetch_and_sub_8: 926 case Builtin::BI__sync_fetch_and_sub_16: 927 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 928 case Builtin::BI__sync_fetch_and_or_1: 929 case Builtin::BI__sync_fetch_and_or_2: 930 case Builtin::BI__sync_fetch_and_or_4: 931 case Builtin::BI__sync_fetch_and_or_8: 932 case Builtin::BI__sync_fetch_and_or_16: 933 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 934 case Builtin::BI__sync_fetch_and_and_1: 935 case Builtin::BI__sync_fetch_and_and_2: 936 case Builtin::BI__sync_fetch_and_and_4: 937 case Builtin::BI__sync_fetch_and_and_8: 938 case Builtin::BI__sync_fetch_and_and_16: 939 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 940 case Builtin::BI__sync_fetch_and_xor_1: 941 case Builtin::BI__sync_fetch_and_xor_2: 942 case Builtin::BI__sync_fetch_and_xor_4: 943 case Builtin::BI__sync_fetch_and_xor_8: 944 case Builtin::BI__sync_fetch_and_xor_16: 945 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 946 case Builtin::BI__sync_fetch_and_nand_1: 947 case Builtin::BI__sync_fetch_and_nand_2: 948 case Builtin::BI__sync_fetch_and_nand_4: 949 case Builtin::BI__sync_fetch_and_nand_8: 950 case Builtin::BI__sync_fetch_and_nand_16: 951 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 952 953 // Clang extensions: not overloaded yet. 954 case Builtin::BI__sync_fetch_and_min: 955 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 956 case Builtin::BI__sync_fetch_and_max: 957 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 958 case Builtin::BI__sync_fetch_and_umin: 959 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 960 case Builtin::BI__sync_fetch_and_umax: 961 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 962 963 case Builtin::BI__sync_add_and_fetch_1: 964 case Builtin::BI__sync_add_and_fetch_2: 965 case Builtin::BI__sync_add_and_fetch_4: 966 case Builtin::BI__sync_add_and_fetch_8: 967 case Builtin::BI__sync_add_and_fetch_16: 968 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 969 llvm::Instruction::Add); 970 case Builtin::BI__sync_sub_and_fetch_1: 971 case Builtin::BI__sync_sub_and_fetch_2: 972 case Builtin::BI__sync_sub_and_fetch_4: 973 case Builtin::BI__sync_sub_and_fetch_8: 974 case Builtin::BI__sync_sub_and_fetch_16: 975 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 976 llvm::Instruction::Sub); 977 case Builtin::BI__sync_and_and_fetch_1: 978 case Builtin::BI__sync_and_and_fetch_2: 979 case Builtin::BI__sync_and_and_fetch_4: 980 case Builtin::BI__sync_and_and_fetch_8: 981 case Builtin::BI__sync_and_and_fetch_16: 982 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 983 llvm::Instruction::And); 984 case Builtin::BI__sync_or_and_fetch_1: 985 case Builtin::BI__sync_or_and_fetch_2: 986 case Builtin::BI__sync_or_and_fetch_4: 987 case Builtin::BI__sync_or_and_fetch_8: 988 case Builtin::BI__sync_or_and_fetch_16: 989 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 990 llvm::Instruction::Or); 991 case Builtin::BI__sync_xor_and_fetch_1: 992 case Builtin::BI__sync_xor_and_fetch_2: 993 case Builtin::BI__sync_xor_and_fetch_4: 994 case Builtin::BI__sync_xor_and_fetch_8: 995 case Builtin::BI__sync_xor_and_fetch_16: 996 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 997 llvm::Instruction::Xor); 998 case Builtin::BI__sync_nand_and_fetch_1: 999 case Builtin::BI__sync_nand_and_fetch_2: 1000 case Builtin::BI__sync_nand_and_fetch_4: 1001 case Builtin::BI__sync_nand_and_fetch_8: 1002 case Builtin::BI__sync_nand_and_fetch_16: 1003 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1004 llvm::Instruction::And, true); 1005 1006 case Builtin::BI__sync_val_compare_and_swap_1: 1007 case Builtin::BI__sync_val_compare_and_swap_2: 1008 case Builtin::BI__sync_val_compare_and_swap_4: 1009 case Builtin::BI__sync_val_compare_and_swap_8: 1010 case Builtin::BI__sync_val_compare_and_swap_16: { 1011 QualType T = E->getType(); 1012 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1013 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1014 1015 llvm::IntegerType *IntType = 1016 llvm::IntegerType::get(getLLVMContext(), 1017 getContext().getTypeSize(T)); 1018 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1019 1020 Value *Args[3]; 1021 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1022 Args[1] = EmitScalarExpr(E->getArg(1)); 1023 llvm::Type *ValueType = Args[1]->getType(); 1024 Args[1] = EmitToInt(*this, Args[1], T, IntType); 1025 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1026 1027 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1028 llvm::SequentiallyConsistent, 1029 llvm::SequentiallyConsistent); 1030 Result = Builder.CreateExtractValue(Result, 0); 1031 Result = EmitFromInt(*this, Result, T, ValueType); 1032 return RValue::get(Result); 1033 } 1034 1035 case Builtin::BI__sync_bool_compare_and_swap_1: 1036 case Builtin::BI__sync_bool_compare_and_swap_2: 1037 case Builtin::BI__sync_bool_compare_and_swap_4: 1038 case Builtin::BI__sync_bool_compare_and_swap_8: 1039 case Builtin::BI__sync_bool_compare_and_swap_16: { 1040 QualType T = E->getArg(1)->getType(); 1041 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 1042 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 1043 1044 llvm::IntegerType *IntType = 1045 llvm::IntegerType::get(getLLVMContext(), 1046 getContext().getTypeSize(T)); 1047 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 1048 1049 Value *Args[3]; 1050 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 1051 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 1052 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 1053 1054 Value *Pair = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 1055 llvm::SequentiallyConsistent, 1056 llvm::SequentiallyConsistent); 1057 Value *Result = Builder.CreateExtractValue(Pair, 1); 1058 // zext bool to int. 1059 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 1060 return RValue::get(Result); 1061 } 1062 1063 case Builtin::BI__sync_swap_1: 1064 case Builtin::BI__sync_swap_2: 1065 case Builtin::BI__sync_swap_4: 1066 case Builtin::BI__sync_swap_8: 1067 case Builtin::BI__sync_swap_16: 1068 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1069 1070 case Builtin::BI__sync_lock_test_and_set_1: 1071 case Builtin::BI__sync_lock_test_and_set_2: 1072 case Builtin::BI__sync_lock_test_and_set_4: 1073 case Builtin::BI__sync_lock_test_and_set_8: 1074 case Builtin::BI__sync_lock_test_and_set_16: 1075 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1076 1077 case Builtin::BI__sync_lock_release_1: 1078 case Builtin::BI__sync_lock_release_2: 1079 case Builtin::BI__sync_lock_release_4: 1080 case Builtin::BI__sync_lock_release_8: 1081 case Builtin::BI__sync_lock_release_16: { 1082 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1083 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1084 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1085 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1086 StoreSize.getQuantity() * 8); 1087 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1088 llvm::StoreInst *Store = 1089 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1090 Store->setAlignment(StoreSize.getQuantity()); 1091 Store->setAtomic(llvm::Release); 1092 return RValue::get(nullptr); 1093 } 1094 1095 case Builtin::BI__sync_synchronize: { 1096 // We assume this is supposed to correspond to a C++0x-style 1097 // sequentially-consistent fence (i.e. this is only usable for 1098 // synchonization, not device I/O or anything like that). This intrinsic 1099 // is really badly designed in the sense that in theory, there isn't 1100 // any way to safely use it... but in practice, it mostly works 1101 // to use it with non-atomic loads and stores to get acquire/release 1102 // semantics. 1103 Builder.CreateFence(llvm::SequentiallyConsistent); 1104 return RValue::get(nullptr); 1105 } 1106 1107 case Builtin::BI__c11_atomic_is_lock_free: 1108 case Builtin::BI__atomic_is_lock_free: { 1109 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1110 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1111 // _Atomic(T) is always properly-aligned. 1112 const char *LibCallName = "__atomic_is_lock_free"; 1113 CallArgList Args; 1114 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1115 getContext().getSizeType()); 1116 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1117 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1118 getContext().VoidPtrTy); 1119 else 1120 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1121 getContext().VoidPtrTy); 1122 const CGFunctionInfo &FuncInfo = 1123 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1124 FunctionType::ExtInfo(), 1125 RequiredArgs::All); 1126 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1127 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1128 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1129 } 1130 1131 case Builtin::BI__atomic_test_and_set: { 1132 // Look at the argument type to determine whether this is a volatile 1133 // operation. The parameter type is always volatile. 1134 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1135 bool Volatile = 1136 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1137 1138 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1139 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1140 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1141 Value *NewVal = Builder.getInt8(1); 1142 Value *Order = EmitScalarExpr(E->getArg(1)); 1143 if (isa<llvm::ConstantInt>(Order)) { 1144 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1145 AtomicRMWInst *Result = nullptr; 1146 switch (ord) { 1147 case 0: // memory_order_relaxed 1148 default: // invalid order 1149 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1150 Ptr, NewVal, 1151 llvm::Monotonic); 1152 break; 1153 case 1: // memory_order_consume 1154 case 2: // memory_order_acquire 1155 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1156 Ptr, NewVal, 1157 llvm::Acquire); 1158 break; 1159 case 3: // memory_order_release 1160 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1161 Ptr, NewVal, 1162 llvm::Release); 1163 break; 1164 case 4: // memory_order_acq_rel 1165 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1166 Ptr, NewVal, 1167 llvm::AcquireRelease); 1168 break; 1169 case 5: // memory_order_seq_cst 1170 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1171 Ptr, NewVal, 1172 llvm::SequentiallyConsistent); 1173 break; 1174 } 1175 Result->setVolatile(Volatile); 1176 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1177 } 1178 1179 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1180 1181 llvm::BasicBlock *BBs[5] = { 1182 createBasicBlock("monotonic", CurFn), 1183 createBasicBlock("acquire", CurFn), 1184 createBasicBlock("release", CurFn), 1185 createBasicBlock("acqrel", CurFn), 1186 createBasicBlock("seqcst", CurFn) 1187 }; 1188 llvm::AtomicOrdering Orders[5] = { 1189 llvm::Monotonic, llvm::Acquire, llvm::Release, 1190 llvm::AcquireRelease, llvm::SequentiallyConsistent 1191 }; 1192 1193 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1194 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1195 1196 Builder.SetInsertPoint(ContBB); 1197 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1198 1199 for (unsigned i = 0; i < 5; ++i) { 1200 Builder.SetInsertPoint(BBs[i]); 1201 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1202 Ptr, NewVal, Orders[i]); 1203 RMW->setVolatile(Volatile); 1204 Result->addIncoming(RMW, BBs[i]); 1205 Builder.CreateBr(ContBB); 1206 } 1207 1208 SI->addCase(Builder.getInt32(0), BBs[0]); 1209 SI->addCase(Builder.getInt32(1), BBs[1]); 1210 SI->addCase(Builder.getInt32(2), BBs[1]); 1211 SI->addCase(Builder.getInt32(3), BBs[2]); 1212 SI->addCase(Builder.getInt32(4), BBs[3]); 1213 SI->addCase(Builder.getInt32(5), BBs[4]); 1214 1215 Builder.SetInsertPoint(ContBB); 1216 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1217 } 1218 1219 case Builtin::BI__atomic_clear: { 1220 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1221 bool Volatile = 1222 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1223 1224 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1225 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1226 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1227 Value *NewVal = Builder.getInt8(0); 1228 Value *Order = EmitScalarExpr(E->getArg(1)); 1229 if (isa<llvm::ConstantInt>(Order)) { 1230 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1231 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1232 Store->setAlignment(1); 1233 switch (ord) { 1234 case 0: // memory_order_relaxed 1235 default: // invalid order 1236 Store->setOrdering(llvm::Monotonic); 1237 break; 1238 case 3: // memory_order_release 1239 Store->setOrdering(llvm::Release); 1240 break; 1241 case 5: // memory_order_seq_cst 1242 Store->setOrdering(llvm::SequentiallyConsistent); 1243 break; 1244 } 1245 return RValue::get(nullptr); 1246 } 1247 1248 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1249 1250 llvm::BasicBlock *BBs[3] = { 1251 createBasicBlock("monotonic", CurFn), 1252 createBasicBlock("release", CurFn), 1253 createBasicBlock("seqcst", CurFn) 1254 }; 1255 llvm::AtomicOrdering Orders[3] = { 1256 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1257 }; 1258 1259 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1260 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1261 1262 for (unsigned i = 0; i < 3; ++i) { 1263 Builder.SetInsertPoint(BBs[i]); 1264 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1265 Store->setAlignment(1); 1266 Store->setOrdering(Orders[i]); 1267 Builder.CreateBr(ContBB); 1268 } 1269 1270 SI->addCase(Builder.getInt32(0), BBs[0]); 1271 SI->addCase(Builder.getInt32(3), BBs[1]); 1272 SI->addCase(Builder.getInt32(5), BBs[2]); 1273 1274 Builder.SetInsertPoint(ContBB); 1275 return RValue::get(nullptr); 1276 } 1277 1278 case Builtin::BI__atomic_thread_fence: 1279 case Builtin::BI__atomic_signal_fence: 1280 case Builtin::BI__c11_atomic_thread_fence: 1281 case Builtin::BI__c11_atomic_signal_fence: { 1282 llvm::SynchronizationScope Scope; 1283 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1284 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1285 Scope = llvm::SingleThread; 1286 else 1287 Scope = llvm::CrossThread; 1288 Value *Order = EmitScalarExpr(E->getArg(0)); 1289 if (isa<llvm::ConstantInt>(Order)) { 1290 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1291 switch (ord) { 1292 case 0: // memory_order_relaxed 1293 default: // invalid order 1294 break; 1295 case 1: // memory_order_consume 1296 case 2: // memory_order_acquire 1297 Builder.CreateFence(llvm::Acquire, Scope); 1298 break; 1299 case 3: // memory_order_release 1300 Builder.CreateFence(llvm::Release, Scope); 1301 break; 1302 case 4: // memory_order_acq_rel 1303 Builder.CreateFence(llvm::AcquireRelease, Scope); 1304 break; 1305 case 5: // memory_order_seq_cst 1306 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1307 break; 1308 } 1309 return RValue::get(nullptr); 1310 } 1311 1312 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1313 AcquireBB = createBasicBlock("acquire", CurFn); 1314 ReleaseBB = createBasicBlock("release", CurFn); 1315 AcqRelBB = createBasicBlock("acqrel", CurFn); 1316 SeqCstBB = createBasicBlock("seqcst", CurFn); 1317 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1318 1319 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1320 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1321 1322 Builder.SetInsertPoint(AcquireBB); 1323 Builder.CreateFence(llvm::Acquire, Scope); 1324 Builder.CreateBr(ContBB); 1325 SI->addCase(Builder.getInt32(1), AcquireBB); 1326 SI->addCase(Builder.getInt32(2), AcquireBB); 1327 1328 Builder.SetInsertPoint(ReleaseBB); 1329 Builder.CreateFence(llvm::Release, Scope); 1330 Builder.CreateBr(ContBB); 1331 SI->addCase(Builder.getInt32(3), ReleaseBB); 1332 1333 Builder.SetInsertPoint(AcqRelBB); 1334 Builder.CreateFence(llvm::AcquireRelease, Scope); 1335 Builder.CreateBr(ContBB); 1336 SI->addCase(Builder.getInt32(4), AcqRelBB); 1337 1338 Builder.SetInsertPoint(SeqCstBB); 1339 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1340 Builder.CreateBr(ContBB); 1341 SI->addCase(Builder.getInt32(5), SeqCstBB); 1342 1343 Builder.SetInsertPoint(ContBB); 1344 return RValue::get(nullptr); 1345 } 1346 1347 // Library functions with special handling. 1348 case Builtin::BIsqrt: 1349 case Builtin::BIsqrtf: 1350 case Builtin::BIsqrtl: { 1351 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1352 // in finite- or unsafe-math mode (the intrinsic has different semantics 1353 // for handling negative numbers compared to the library function, so 1354 // -fmath-errno=0 is not enough). 1355 if (!FD->hasAttr<ConstAttr>()) 1356 break; 1357 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1358 CGM.getCodeGenOpts().NoNaNsFPMath)) 1359 break; 1360 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1361 llvm::Type *ArgType = Arg0->getType(); 1362 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1363 return RValue::get(Builder.CreateCall(F, Arg0)); 1364 } 1365 1366 case Builtin::BI__builtin_pow: 1367 case Builtin::BI__builtin_powf: 1368 case Builtin::BI__builtin_powl: 1369 case Builtin::BIpow: 1370 case Builtin::BIpowf: 1371 case Builtin::BIpowl: { 1372 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1373 if (!FD->hasAttr<ConstAttr>()) 1374 break; 1375 Value *Base = EmitScalarExpr(E->getArg(0)); 1376 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1377 llvm::Type *ArgType = Base->getType(); 1378 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1379 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1380 } 1381 1382 case Builtin::BIfma: 1383 case Builtin::BIfmaf: 1384 case Builtin::BIfmal: 1385 case Builtin::BI__builtin_fma: 1386 case Builtin::BI__builtin_fmaf: 1387 case Builtin::BI__builtin_fmal: { 1388 // Rewrite fma to intrinsic. 1389 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1390 llvm::Type *ArgType = FirstArg->getType(); 1391 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1392 return RValue::get(Builder.CreateCall3(F, FirstArg, 1393 EmitScalarExpr(E->getArg(1)), 1394 EmitScalarExpr(E->getArg(2)))); 1395 } 1396 1397 case Builtin::BI__builtin_signbit: 1398 case Builtin::BI__builtin_signbitf: 1399 case Builtin::BI__builtin_signbitl: { 1400 LLVMContext &C = CGM.getLLVMContext(); 1401 1402 Value *Arg = EmitScalarExpr(E->getArg(0)); 1403 llvm::Type *ArgTy = Arg->getType(); 1404 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1405 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1406 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1407 if (ArgTy->isPPC_FP128Ty()) { 1408 // The higher-order double comes first, and so we need to truncate the 1409 // pair to extract the overall sign. The order of the pair is the same 1410 // in both little- and big-Endian modes. 1411 ArgWidth >>= 1; 1412 ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1413 BCArg = Builder.CreateTrunc(BCArg, ArgIntTy); 1414 } 1415 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1416 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1417 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1418 } 1419 case Builtin::BI__builtin_annotation: { 1420 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1421 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1422 AnnVal->getType()); 1423 1424 // Get the annotation string, go through casts. Sema requires this to be a 1425 // non-wide string literal, potentially casted, so the cast<> is safe. 1426 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1427 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1428 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1429 } 1430 case Builtin::BI__builtin_addcb: 1431 case Builtin::BI__builtin_addcs: 1432 case Builtin::BI__builtin_addc: 1433 case Builtin::BI__builtin_addcl: 1434 case Builtin::BI__builtin_addcll: 1435 case Builtin::BI__builtin_subcb: 1436 case Builtin::BI__builtin_subcs: 1437 case Builtin::BI__builtin_subc: 1438 case Builtin::BI__builtin_subcl: 1439 case Builtin::BI__builtin_subcll: { 1440 1441 // We translate all of these builtins from expressions of the form: 1442 // int x = ..., y = ..., carryin = ..., carryout, result; 1443 // result = __builtin_addc(x, y, carryin, &carryout); 1444 // 1445 // to LLVM IR of the form: 1446 // 1447 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1448 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1449 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1450 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1451 // i32 %carryin) 1452 // %result = extractvalue {i32, i1} %tmp2, 0 1453 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1454 // %tmp3 = or i1 %carry1, %carry2 1455 // %tmp4 = zext i1 %tmp3 to i32 1456 // store i32 %tmp4, i32* %carryout 1457 1458 // Scalarize our inputs. 1459 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1460 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1461 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1462 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1463 EmitPointerWithAlignment(E->getArg(3)); 1464 1465 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1466 llvm::Intrinsic::ID IntrinsicId; 1467 switch (BuiltinID) { 1468 default: llvm_unreachable("Unknown multiprecision builtin id."); 1469 case Builtin::BI__builtin_addcb: 1470 case Builtin::BI__builtin_addcs: 1471 case Builtin::BI__builtin_addc: 1472 case Builtin::BI__builtin_addcl: 1473 case Builtin::BI__builtin_addcll: 1474 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1475 break; 1476 case Builtin::BI__builtin_subcb: 1477 case Builtin::BI__builtin_subcs: 1478 case Builtin::BI__builtin_subc: 1479 case Builtin::BI__builtin_subcl: 1480 case Builtin::BI__builtin_subcll: 1481 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1482 break; 1483 } 1484 1485 // Construct our resulting LLVM IR expression. 1486 llvm::Value *Carry1; 1487 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1488 X, Y, Carry1); 1489 llvm::Value *Carry2; 1490 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1491 Sum1, Carryin, Carry2); 1492 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1493 X->getType()); 1494 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1495 CarryOutPtr.first); 1496 CarryOutStore->setAlignment(CarryOutPtr.second); 1497 return RValue::get(Sum2); 1498 } 1499 case Builtin::BI__builtin_uadd_overflow: 1500 case Builtin::BI__builtin_uaddl_overflow: 1501 case Builtin::BI__builtin_uaddll_overflow: 1502 case Builtin::BI__builtin_usub_overflow: 1503 case Builtin::BI__builtin_usubl_overflow: 1504 case Builtin::BI__builtin_usubll_overflow: 1505 case Builtin::BI__builtin_umul_overflow: 1506 case Builtin::BI__builtin_umull_overflow: 1507 case Builtin::BI__builtin_umulll_overflow: 1508 case Builtin::BI__builtin_sadd_overflow: 1509 case Builtin::BI__builtin_saddl_overflow: 1510 case Builtin::BI__builtin_saddll_overflow: 1511 case Builtin::BI__builtin_ssub_overflow: 1512 case Builtin::BI__builtin_ssubl_overflow: 1513 case Builtin::BI__builtin_ssubll_overflow: 1514 case Builtin::BI__builtin_smul_overflow: 1515 case Builtin::BI__builtin_smull_overflow: 1516 case Builtin::BI__builtin_smulll_overflow: { 1517 1518 // We translate all of these builtins directly to the relevant llvm IR node. 1519 1520 // Scalarize our inputs. 1521 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1522 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1523 std::pair<llvm::Value *, unsigned> SumOutPtr = 1524 EmitPointerWithAlignment(E->getArg(2)); 1525 1526 // Decide which of the overflow intrinsics we are lowering to: 1527 llvm::Intrinsic::ID IntrinsicId; 1528 switch (BuiltinID) { 1529 default: llvm_unreachable("Unknown security overflow builtin id."); 1530 case Builtin::BI__builtin_uadd_overflow: 1531 case Builtin::BI__builtin_uaddl_overflow: 1532 case Builtin::BI__builtin_uaddll_overflow: 1533 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1534 break; 1535 case Builtin::BI__builtin_usub_overflow: 1536 case Builtin::BI__builtin_usubl_overflow: 1537 case Builtin::BI__builtin_usubll_overflow: 1538 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1539 break; 1540 case Builtin::BI__builtin_umul_overflow: 1541 case Builtin::BI__builtin_umull_overflow: 1542 case Builtin::BI__builtin_umulll_overflow: 1543 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1544 break; 1545 case Builtin::BI__builtin_sadd_overflow: 1546 case Builtin::BI__builtin_saddl_overflow: 1547 case Builtin::BI__builtin_saddll_overflow: 1548 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1549 break; 1550 case Builtin::BI__builtin_ssub_overflow: 1551 case Builtin::BI__builtin_ssubl_overflow: 1552 case Builtin::BI__builtin_ssubll_overflow: 1553 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1554 break; 1555 case Builtin::BI__builtin_smul_overflow: 1556 case Builtin::BI__builtin_smull_overflow: 1557 case Builtin::BI__builtin_smulll_overflow: 1558 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1559 break; 1560 } 1561 1562 1563 llvm::Value *Carry; 1564 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1565 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1566 SumOutStore->setAlignment(SumOutPtr.second); 1567 1568 return RValue::get(Carry); 1569 } 1570 case Builtin::BI__builtin_addressof: 1571 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1572 case Builtin::BI__builtin_operator_new: 1573 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1574 E->getArg(0), false); 1575 case Builtin::BI__builtin_operator_delete: 1576 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1577 E->getArg(0), true); 1578 case Builtin::BI__noop: 1579 // __noop always evaluates to an integer literal zero. 1580 return RValue::get(ConstantInt::get(IntTy, 0)); 1581 case Builtin::BI__builtin_call_with_static_chain: { 1582 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1583 const Expr *Chain = E->getArg(1); 1584 return EmitCall(Call->getCallee()->getType(), 1585 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1586 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1587 } 1588 case Builtin::BI_InterlockedExchange: 1589 case Builtin::BI_InterlockedExchangePointer: 1590 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1591 case Builtin::BI_InterlockedCompareExchangePointer: { 1592 llvm::Type *RTy; 1593 llvm::IntegerType *IntType = 1594 IntegerType::get(getLLVMContext(), 1595 getContext().getTypeSize(E->getType())); 1596 llvm::Type *IntPtrType = IntType->getPointerTo(); 1597 1598 llvm::Value *Destination = 1599 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1600 1601 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1602 RTy = Exchange->getType(); 1603 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1604 1605 llvm::Value *Comparand = 1606 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1607 1608 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1609 SequentiallyConsistent, 1610 SequentiallyConsistent); 1611 Result->setVolatile(true); 1612 1613 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1614 0), 1615 RTy)); 1616 } 1617 case Builtin::BI_InterlockedCompareExchange: { 1618 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1619 EmitScalarExpr(E->getArg(0)), 1620 EmitScalarExpr(E->getArg(2)), 1621 EmitScalarExpr(E->getArg(1)), 1622 SequentiallyConsistent, 1623 SequentiallyConsistent); 1624 CXI->setVolatile(true); 1625 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1626 } 1627 case Builtin::BI_InterlockedIncrement: { 1628 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1629 AtomicRMWInst::Add, 1630 EmitScalarExpr(E->getArg(0)), 1631 ConstantInt::get(Int32Ty, 1), 1632 llvm::SequentiallyConsistent); 1633 RMWI->setVolatile(true); 1634 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1635 } 1636 case Builtin::BI_InterlockedDecrement: { 1637 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1638 AtomicRMWInst::Sub, 1639 EmitScalarExpr(E->getArg(0)), 1640 ConstantInt::get(Int32Ty, 1), 1641 llvm::SequentiallyConsistent); 1642 RMWI->setVolatile(true); 1643 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1644 } 1645 case Builtin::BI_InterlockedExchangeAdd: { 1646 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1647 AtomicRMWInst::Add, 1648 EmitScalarExpr(E->getArg(0)), 1649 EmitScalarExpr(E->getArg(1)), 1650 llvm::SequentiallyConsistent); 1651 RMWI->setVolatile(true); 1652 return RValue::get(RMWI); 1653 } 1654 case Builtin::BI__readfsdword: { 1655 Value *IntToPtr = 1656 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1657 llvm::PointerType::get(CGM.Int32Ty, 257)); 1658 LoadInst *Load = 1659 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1660 return RValue::get(Load); 1661 } 1662 1663 case Builtin::BI__exception_code: 1664 case Builtin::BI_exception_code: 1665 return RValue::get(EmitSEHExceptionCode()); 1666 case Builtin::BI__exception_info: 1667 case Builtin::BI_exception_info: 1668 return RValue::get(EmitSEHExceptionInfo()); 1669 case Builtin::BI__abnormal_termination: 1670 case Builtin::BI_abnormal_termination: 1671 return RValue::get(EmitSEHAbnormalTermination()); 1672 case Builtin::BI_setjmpex: { 1673 if (getTarget().getTriple().isOSMSVCRT()) { 1674 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1675 llvm::AttributeSet ReturnsTwiceAttr = 1676 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1677 llvm::Attribute::ReturnsTwice); 1678 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1679 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1680 "_setjmpex", ReturnsTwiceAttr); 1681 llvm::Value *Buf = 1682 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1683 llvm::Value *FrameAddr = 1684 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1685 ConstantInt::get(Int32Ty, 0)); 1686 llvm::Value *Args[] = {Buf, FrameAddr}; 1687 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1688 CS.setAttributes(ReturnsTwiceAttr); 1689 return RValue::get(CS.getInstruction()); 1690 } 1691 } 1692 case Builtin::BI_setjmp: { 1693 if (getTarget().getTriple().isOSMSVCRT()) { 1694 llvm::AttributeSet ReturnsTwiceAttr = 1695 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1696 llvm::Attribute::ReturnsTwice); 1697 llvm::Value *Buf = 1698 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1699 llvm::CallSite CS; 1700 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1701 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1702 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1703 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1704 "_setjmp3", ReturnsTwiceAttr); 1705 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1706 llvm::Value *Args[] = {Buf, Count}; 1707 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1708 } else { 1709 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1710 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1711 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1712 "_setjmp", ReturnsTwiceAttr); 1713 llvm::Value *FrameAddr = 1714 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1715 ConstantInt::get(Int32Ty, 0)); 1716 llvm::Value *Args[] = {Buf, FrameAddr}; 1717 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1718 } 1719 CS.setAttributes(ReturnsTwiceAttr); 1720 return RValue::get(CS.getInstruction()); 1721 } 1722 } 1723 } 1724 1725 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1726 // the call using the normal call path, but using the unmangled 1727 // version of the function name. 1728 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1729 return emitLibraryCall(*this, FD, E, 1730 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1731 1732 // If this is a predefined lib function (e.g. malloc), emit the call 1733 // using exactly the normal call path. 1734 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1735 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1736 1737 // See if we have a target specific intrinsic. 1738 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1739 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1740 if (const char *Prefix = 1741 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1742 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1743 // NOTE we dont need to perform a compatibility flag check here since the 1744 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1745 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1746 if (IntrinsicID == Intrinsic::not_intrinsic) 1747 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1748 } 1749 1750 if (IntrinsicID != Intrinsic::not_intrinsic) { 1751 SmallVector<Value*, 16> Args; 1752 1753 // Find out if any arguments are required to be integer constant 1754 // expressions. 1755 unsigned ICEArguments = 0; 1756 ASTContext::GetBuiltinTypeError Error; 1757 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1758 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1759 1760 Function *F = CGM.getIntrinsic(IntrinsicID); 1761 llvm::FunctionType *FTy = F->getFunctionType(); 1762 1763 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1764 Value *ArgValue; 1765 // If this is a normal argument, just emit it as a scalar. 1766 if ((ICEArguments & (1 << i)) == 0) { 1767 ArgValue = EmitScalarExpr(E->getArg(i)); 1768 } else { 1769 // If this is required to be a constant, constant fold it so that we 1770 // know that the generated intrinsic gets a ConstantInt. 1771 llvm::APSInt Result; 1772 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1773 assert(IsConst && "Constant arg isn't actually constant?"); 1774 (void)IsConst; 1775 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1776 } 1777 1778 // If the intrinsic arg type is different from the builtin arg type 1779 // we need to do a bit cast. 1780 llvm::Type *PTy = FTy->getParamType(i); 1781 if (PTy != ArgValue->getType()) { 1782 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1783 "Must be able to losslessly bit cast to param"); 1784 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1785 } 1786 1787 Args.push_back(ArgValue); 1788 } 1789 1790 Value *V = Builder.CreateCall(F, Args); 1791 QualType BuiltinRetType = E->getType(); 1792 1793 llvm::Type *RetTy = VoidTy; 1794 if (!BuiltinRetType->isVoidType()) 1795 RetTy = ConvertType(BuiltinRetType); 1796 1797 if (RetTy != V->getType()) { 1798 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1799 "Must be able to losslessly bit cast result type"); 1800 V = Builder.CreateBitCast(V, RetTy); 1801 } 1802 1803 return RValue::get(V); 1804 } 1805 1806 // See if we have a target specific builtin that needs to be lowered. 1807 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1808 return RValue::get(V); 1809 1810 ErrorUnsupported(E, "builtin function"); 1811 1812 // Unknown builtin, for now just dump it out and return undef. 1813 return GetUndefRValue(E->getType()); 1814 } 1815 1816 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1817 const CallExpr *E) { 1818 switch (getTarget().getTriple().getArch()) { 1819 case llvm::Triple::arm: 1820 case llvm::Triple::armeb: 1821 case llvm::Triple::thumb: 1822 case llvm::Triple::thumbeb: 1823 return EmitARMBuiltinExpr(BuiltinID, E); 1824 case llvm::Triple::aarch64: 1825 case llvm::Triple::aarch64_be: 1826 return EmitAArch64BuiltinExpr(BuiltinID, E); 1827 case llvm::Triple::x86: 1828 case llvm::Triple::x86_64: 1829 return EmitX86BuiltinExpr(BuiltinID, E); 1830 case llvm::Triple::ppc: 1831 case llvm::Triple::ppc64: 1832 case llvm::Triple::ppc64le: 1833 return EmitPPCBuiltinExpr(BuiltinID, E); 1834 case llvm::Triple::r600: 1835 case llvm::Triple::amdgcn: 1836 return EmitR600BuiltinExpr(BuiltinID, E); 1837 default: 1838 return nullptr; 1839 } 1840 } 1841 1842 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1843 NeonTypeFlags TypeFlags, 1844 bool V1Ty=false) { 1845 int IsQuad = TypeFlags.isQuad(); 1846 switch (TypeFlags.getEltType()) { 1847 case NeonTypeFlags::Int8: 1848 case NeonTypeFlags::Poly8: 1849 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1850 case NeonTypeFlags::Int16: 1851 case NeonTypeFlags::Poly16: 1852 case NeonTypeFlags::Float16: 1853 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1854 case NeonTypeFlags::Int32: 1855 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1856 case NeonTypeFlags::Int64: 1857 case NeonTypeFlags::Poly64: 1858 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1859 case NeonTypeFlags::Poly128: 1860 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1861 // There is a lot of i128 and f128 API missing. 1862 // so we use v16i8 to represent poly128 and get pattern matched. 1863 return llvm::VectorType::get(CGF->Int8Ty, 16); 1864 case NeonTypeFlags::Float32: 1865 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1866 case NeonTypeFlags::Float64: 1867 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1868 } 1869 llvm_unreachable("Unknown vector element type!"); 1870 } 1871 1872 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1873 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1874 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1875 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1876 } 1877 1878 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1879 const char *name, 1880 unsigned shift, bool rightshift) { 1881 unsigned j = 0; 1882 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1883 ai != ae; ++ai, ++j) 1884 if (shift > 0 && shift == j) 1885 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1886 else 1887 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1888 1889 return Builder.CreateCall(F, Ops, name); 1890 } 1891 1892 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1893 bool neg) { 1894 int SV = cast<ConstantInt>(V)->getSExtValue(); 1895 1896 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1897 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1898 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1899 } 1900 1901 // \brief Right-shift a vector by a constant. 1902 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1903 llvm::Type *Ty, bool usgn, 1904 const char *name) { 1905 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1906 1907 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1908 int EltSize = VTy->getScalarSizeInBits(); 1909 1910 Vec = Builder.CreateBitCast(Vec, Ty); 1911 1912 // lshr/ashr are undefined when the shift amount is equal to the vector 1913 // element size. 1914 if (ShiftAmt == EltSize) { 1915 if (usgn) { 1916 // Right-shifting an unsigned value by its size yields 0. 1917 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1918 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1919 } else { 1920 // Right-shifting a signed value by its size is equivalent 1921 // to a shift of size-1. 1922 --ShiftAmt; 1923 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1924 } 1925 } 1926 1927 Shift = EmitNeonShiftVector(Shift, Ty, false); 1928 if (usgn) 1929 return Builder.CreateLShr(Vec, Shift, name); 1930 else 1931 return Builder.CreateAShr(Vec, Shift, name); 1932 } 1933 1934 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1935 /// alignment of the type referenced by the pointer. Skip over implicit 1936 /// casts. 1937 std::pair<llvm::Value*, unsigned> 1938 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1939 assert(Addr->getType()->isPointerType()); 1940 Addr = Addr->IgnoreParens(); 1941 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1942 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1943 ICE->getSubExpr()->getType()->isPointerType()) { 1944 std::pair<llvm::Value*, unsigned> Ptr = 1945 EmitPointerWithAlignment(ICE->getSubExpr()); 1946 Ptr.first = Builder.CreateBitCast(Ptr.first, 1947 ConvertType(Addr->getType())); 1948 return Ptr; 1949 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1950 LValue LV = EmitLValue(ICE->getSubExpr()); 1951 unsigned Align = LV.getAlignment().getQuantity(); 1952 if (!Align) { 1953 // FIXME: Once LValues are fixed to always set alignment, 1954 // zap this code. 1955 QualType PtTy = ICE->getSubExpr()->getType(); 1956 if (!PtTy->isIncompleteType()) 1957 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1958 else 1959 Align = 1; 1960 } 1961 return std::make_pair(LV.getAddress(), Align); 1962 } 1963 } 1964 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1965 if (UO->getOpcode() == UO_AddrOf) { 1966 LValue LV = EmitLValue(UO->getSubExpr()); 1967 unsigned Align = LV.getAlignment().getQuantity(); 1968 if (!Align) { 1969 // FIXME: Once LValues are fixed to always set alignment, 1970 // zap this code. 1971 QualType PtTy = UO->getSubExpr()->getType(); 1972 if (!PtTy->isIncompleteType()) 1973 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1974 else 1975 Align = 1; 1976 } 1977 return std::make_pair(LV.getAddress(), Align); 1978 } 1979 } 1980 1981 unsigned Align = 1; 1982 QualType PtTy = Addr->getType()->getPointeeType(); 1983 if (!PtTy->isIncompleteType()) 1984 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1985 1986 return std::make_pair(EmitScalarExpr(Addr), Align); 1987 } 1988 1989 enum { 1990 AddRetType = (1 << 0), 1991 Add1ArgType = (1 << 1), 1992 Add2ArgTypes = (1 << 2), 1993 1994 VectorizeRetType = (1 << 3), 1995 VectorizeArgTypes = (1 << 4), 1996 1997 InventFloatType = (1 << 5), 1998 UnsignedAlts = (1 << 6), 1999 2000 Use64BitVectors = (1 << 7), 2001 Use128BitVectors = (1 << 8), 2002 2003 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2004 VectorRet = AddRetType | VectorizeRetType, 2005 VectorRetGetArgs01 = 2006 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2007 FpCmpzModifiers = 2008 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2009 }; 2010 2011 struct NeonIntrinsicInfo { 2012 unsigned BuiltinID; 2013 unsigned LLVMIntrinsic; 2014 unsigned AltLLVMIntrinsic; 2015 const char *NameHint; 2016 unsigned TypeModifier; 2017 2018 bool operator<(unsigned RHSBuiltinID) const { 2019 return BuiltinID < RHSBuiltinID; 2020 } 2021 }; 2022 2023 #define NEONMAP0(NameBase) \ 2024 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2025 2026 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2027 { NEON:: BI__builtin_neon_ ## NameBase, \ 2028 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2029 2030 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2031 { NEON:: BI__builtin_neon_ ## NameBase, \ 2032 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2033 #NameBase, TypeModifier } 2034 2035 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2036 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2037 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2038 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2039 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2040 NEONMAP0(vaddhn_v), 2041 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2042 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2043 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2044 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2045 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2046 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2047 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2048 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2049 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2050 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2051 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2052 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2053 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2054 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2055 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2056 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2057 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2058 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2059 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2060 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2061 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 2062 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2063 NEONMAP0(vcvt_f32_v), 2064 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2065 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2066 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2067 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2068 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2069 NEONMAP0(vcvt_s32_v), 2070 NEONMAP0(vcvt_s64_v), 2071 NEONMAP0(vcvt_u32_v), 2072 NEONMAP0(vcvt_u64_v), 2073 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2074 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2075 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2076 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2077 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2078 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2079 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2080 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2081 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2082 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2083 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2084 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2085 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2086 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2087 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2088 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2089 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2090 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2091 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2092 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2093 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2094 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2095 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2096 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2097 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2098 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2099 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2100 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2101 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2102 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2103 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2104 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2105 NEONMAP0(vcvtq_f32_v), 2106 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2107 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2108 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2109 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2110 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2111 NEONMAP0(vcvtq_s32_v), 2112 NEONMAP0(vcvtq_s64_v), 2113 NEONMAP0(vcvtq_u32_v), 2114 NEONMAP0(vcvtq_u64_v), 2115 NEONMAP0(vext_v), 2116 NEONMAP0(vextq_v), 2117 NEONMAP0(vfma_v), 2118 NEONMAP0(vfmaq_v), 2119 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2120 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2121 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2122 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2123 NEONMAP0(vld1_dup_v), 2124 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2125 NEONMAP0(vld1q_dup_v), 2126 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2127 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2128 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2129 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2130 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2131 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2132 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2133 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2134 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2135 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2136 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2137 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2138 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2139 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2140 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2141 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2142 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2143 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2144 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2145 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2146 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2147 NEONMAP0(vmovl_v), 2148 NEONMAP0(vmovn_v), 2149 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2150 NEONMAP0(vmull_v), 2151 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2152 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2153 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2154 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2155 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2156 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2157 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2158 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2159 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2160 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2161 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2162 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2163 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2164 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2165 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2166 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2167 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2168 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2169 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2170 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2171 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2172 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2173 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2174 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2175 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2176 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2177 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2178 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2179 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2180 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2181 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2182 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2183 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2184 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2185 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2186 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2187 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2188 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2189 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2190 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2191 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2192 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2193 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2194 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2195 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2196 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2197 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2198 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2199 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2200 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2201 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2202 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2203 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2204 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2205 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2206 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2207 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2208 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2209 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2210 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2211 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2212 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2213 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2214 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2215 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2216 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2217 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2218 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2219 NEONMAP0(vshl_n_v), 2220 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2221 NEONMAP0(vshll_n_v), 2222 NEONMAP0(vshlq_n_v), 2223 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2224 NEONMAP0(vshr_n_v), 2225 NEONMAP0(vshrn_n_v), 2226 NEONMAP0(vshrq_n_v), 2227 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2228 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2229 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2230 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2231 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2232 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2233 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2234 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2235 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2236 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2237 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2238 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2239 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2240 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2241 NEONMAP0(vsubhn_v), 2242 NEONMAP0(vtrn_v), 2243 NEONMAP0(vtrnq_v), 2244 NEONMAP0(vtst_v), 2245 NEONMAP0(vtstq_v), 2246 NEONMAP0(vuzp_v), 2247 NEONMAP0(vuzpq_v), 2248 NEONMAP0(vzip_v), 2249 NEONMAP0(vzipq_v) 2250 }; 2251 2252 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2253 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2254 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2255 NEONMAP0(vaddhn_v), 2256 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2257 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2258 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2259 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2260 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2261 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2262 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2263 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2264 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2265 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2266 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2267 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2268 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2269 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2270 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2271 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2272 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2273 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2274 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2275 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2276 NEONMAP0(vcvt_f32_v), 2277 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2278 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2279 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2280 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2281 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2282 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2283 NEONMAP0(vcvtq_f32_v), 2284 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2285 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2286 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2287 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2288 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2289 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2290 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2291 NEONMAP0(vext_v), 2292 NEONMAP0(vextq_v), 2293 NEONMAP0(vfma_v), 2294 NEONMAP0(vfmaq_v), 2295 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2296 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2297 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2298 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2299 NEONMAP0(vmovl_v), 2300 NEONMAP0(vmovn_v), 2301 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2302 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2303 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2304 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2305 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2306 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2307 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2308 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2309 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2310 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2311 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2312 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2313 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2314 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2315 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2316 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2317 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2318 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2319 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2320 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2321 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2322 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2323 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2324 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2325 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2326 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2327 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2328 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2329 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2330 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2331 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2332 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2333 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2334 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2335 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2336 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2337 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2338 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2339 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2340 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2341 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2342 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2343 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2344 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2345 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2346 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2347 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2348 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2349 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2350 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2351 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2352 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2353 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2354 NEONMAP0(vshl_n_v), 2355 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2356 NEONMAP0(vshll_n_v), 2357 NEONMAP0(vshlq_n_v), 2358 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2359 NEONMAP0(vshr_n_v), 2360 NEONMAP0(vshrn_n_v), 2361 NEONMAP0(vshrq_n_v), 2362 NEONMAP0(vsubhn_v), 2363 NEONMAP0(vtst_v), 2364 NEONMAP0(vtstq_v), 2365 }; 2366 2367 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2368 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2369 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2370 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2371 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2372 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2373 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2374 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2375 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2376 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2377 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2378 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2379 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2380 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2381 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2382 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2383 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2384 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2385 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2386 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2387 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2388 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2389 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2390 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2391 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2392 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2393 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2394 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2395 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2396 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2397 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2398 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2399 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2400 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2401 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2402 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2403 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2404 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2405 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2406 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2407 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2408 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2409 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2410 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2411 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2412 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2413 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2414 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2415 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2416 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2417 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2418 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2419 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2420 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2421 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2422 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2423 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2424 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2425 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2426 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2427 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2428 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2429 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2430 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2431 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2432 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2433 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2434 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2435 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2436 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2437 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2438 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2439 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2440 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2441 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2442 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2443 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2444 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2445 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2446 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2447 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2448 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2449 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2450 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2451 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2452 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2453 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2454 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2455 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2456 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2457 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2458 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2459 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2460 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2461 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2462 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2463 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2464 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2465 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2466 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2467 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2468 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2469 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2470 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2471 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2472 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2473 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2474 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2475 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2476 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2477 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2478 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2479 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2480 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2481 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2482 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2483 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2484 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2485 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2486 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2487 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2488 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2489 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2490 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2491 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2492 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2493 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2494 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2495 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2496 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2497 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2498 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2499 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2500 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2501 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2502 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2503 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2504 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2505 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2506 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2507 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2508 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2509 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2510 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2511 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2512 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2513 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2514 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2515 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2516 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2517 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2518 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2519 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2520 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2521 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2522 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2523 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2524 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2525 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2526 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2527 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2528 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2529 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2530 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2531 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2532 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2533 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2534 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2535 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2536 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2537 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2538 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2539 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2540 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2541 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2542 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2543 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2544 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2545 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2546 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2547 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2548 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2549 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2550 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2551 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2552 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2553 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2554 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2555 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2556 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2557 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2558 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2559 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2560 }; 2561 2562 #undef NEONMAP0 2563 #undef NEONMAP1 2564 #undef NEONMAP2 2565 2566 static bool NEONSIMDIntrinsicsProvenSorted = false; 2567 2568 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2569 static bool AArch64SISDIntrinsicsProvenSorted = false; 2570 2571 2572 static const NeonIntrinsicInfo * 2573 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2574 unsigned BuiltinID, bool &MapProvenSorted) { 2575 2576 #ifndef NDEBUG 2577 if (!MapProvenSorted) { 2578 // FIXME: use std::is_sorted once C++11 is allowed 2579 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2580 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2581 MapProvenSorted = true; 2582 } 2583 #endif 2584 2585 const NeonIntrinsicInfo *Builtin = 2586 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2587 2588 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2589 return Builtin; 2590 2591 return nullptr; 2592 } 2593 2594 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2595 unsigned Modifier, 2596 llvm::Type *ArgType, 2597 const CallExpr *E) { 2598 int VectorSize = 0; 2599 if (Modifier & Use64BitVectors) 2600 VectorSize = 64; 2601 else if (Modifier & Use128BitVectors) 2602 VectorSize = 128; 2603 2604 // Return type. 2605 SmallVector<llvm::Type *, 3> Tys; 2606 if (Modifier & AddRetType) { 2607 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 2608 if (Modifier & VectorizeRetType) 2609 Ty = llvm::VectorType::get( 2610 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2611 2612 Tys.push_back(Ty); 2613 } 2614 2615 // Arguments. 2616 if (Modifier & VectorizeArgTypes) { 2617 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2618 ArgType = llvm::VectorType::get(ArgType, Elts); 2619 } 2620 2621 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2622 Tys.push_back(ArgType); 2623 2624 if (Modifier & Add2ArgTypes) 2625 Tys.push_back(ArgType); 2626 2627 if (Modifier & InventFloatType) 2628 Tys.push_back(FloatTy); 2629 2630 return CGM.getIntrinsic(IntrinsicID, Tys); 2631 } 2632 2633 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2634 const NeonIntrinsicInfo &SISDInfo, 2635 SmallVectorImpl<Value *> &Ops, 2636 const CallExpr *E) { 2637 unsigned BuiltinID = SISDInfo.BuiltinID; 2638 unsigned int Int = SISDInfo.LLVMIntrinsic; 2639 unsigned Modifier = SISDInfo.TypeModifier; 2640 const char *s = SISDInfo.NameHint; 2641 2642 switch (BuiltinID) { 2643 case NEON::BI__builtin_neon_vcled_s64: 2644 case NEON::BI__builtin_neon_vcled_u64: 2645 case NEON::BI__builtin_neon_vcles_f32: 2646 case NEON::BI__builtin_neon_vcled_f64: 2647 case NEON::BI__builtin_neon_vcltd_s64: 2648 case NEON::BI__builtin_neon_vcltd_u64: 2649 case NEON::BI__builtin_neon_vclts_f32: 2650 case NEON::BI__builtin_neon_vcltd_f64: 2651 case NEON::BI__builtin_neon_vcales_f32: 2652 case NEON::BI__builtin_neon_vcaled_f64: 2653 case NEON::BI__builtin_neon_vcalts_f32: 2654 case NEON::BI__builtin_neon_vcaltd_f64: 2655 // Only one direction of comparisons actually exist, cmle is actually a cmge 2656 // with swapped operands. The table gives us the right intrinsic but we 2657 // still need to do the swap. 2658 std::swap(Ops[0], Ops[1]); 2659 break; 2660 } 2661 2662 assert(Int && "Generic code assumes a valid intrinsic"); 2663 2664 // Determine the type(s) of this overloaded AArch64 intrinsic. 2665 const Expr *Arg = E->getArg(0); 2666 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2667 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2668 2669 int j = 0; 2670 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2671 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2672 ai != ae; ++ai, ++j) { 2673 llvm::Type *ArgTy = ai->getType(); 2674 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2675 ArgTy->getPrimitiveSizeInBits()) 2676 continue; 2677 2678 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2679 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2680 // it before inserting. 2681 Ops[j] = 2682 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2683 Ops[j] = 2684 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2685 } 2686 2687 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2688 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2689 if (ResultType->getPrimitiveSizeInBits() < 2690 Result->getType()->getPrimitiveSizeInBits()) 2691 return CGF.Builder.CreateExtractElement(Result, C0); 2692 2693 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2694 } 2695 2696 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2697 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2698 const char *NameHint, unsigned Modifier, const CallExpr *E, 2699 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2700 // Get the last argument, which specifies the vector type. 2701 llvm::APSInt NeonTypeConst; 2702 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2703 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2704 return nullptr; 2705 2706 // Determine the type of this overloaded NEON intrinsic. 2707 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2708 bool Usgn = Type.isUnsigned(); 2709 bool Quad = Type.isQuad(); 2710 2711 llvm::VectorType *VTy = GetNeonType(this, Type); 2712 llvm::Type *Ty = VTy; 2713 if (!Ty) 2714 return nullptr; 2715 2716 unsigned Int = LLVMIntrinsic; 2717 if ((Modifier & UnsignedAlts) && !Usgn) 2718 Int = AltLLVMIntrinsic; 2719 2720 switch (BuiltinID) { 2721 default: break; 2722 case NEON::BI__builtin_neon_vabs_v: 2723 case NEON::BI__builtin_neon_vabsq_v: 2724 if (VTy->getElementType()->isFloatingPointTy()) 2725 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2726 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2727 case NEON::BI__builtin_neon_vaddhn_v: { 2728 llvm::VectorType *SrcTy = 2729 llvm::VectorType::getExtendedElementVectorType(VTy); 2730 2731 // %sum = add <4 x i32> %lhs, %rhs 2732 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2733 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2734 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2735 2736 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2737 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2738 SrcTy->getScalarSizeInBits() / 2); 2739 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2740 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2741 2742 // %res = trunc <4 x i32> %high to <4 x i16> 2743 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2744 } 2745 case NEON::BI__builtin_neon_vcale_v: 2746 case NEON::BI__builtin_neon_vcaleq_v: 2747 case NEON::BI__builtin_neon_vcalt_v: 2748 case NEON::BI__builtin_neon_vcaltq_v: 2749 std::swap(Ops[0], Ops[1]); 2750 case NEON::BI__builtin_neon_vcage_v: 2751 case NEON::BI__builtin_neon_vcageq_v: 2752 case NEON::BI__builtin_neon_vcagt_v: 2753 case NEON::BI__builtin_neon_vcagtq_v: { 2754 llvm::Type *VecFlt = llvm::VectorType::get( 2755 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2756 VTy->getNumElements()); 2757 llvm::Type *Tys[] = { VTy, VecFlt }; 2758 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2759 return EmitNeonCall(F, Ops, NameHint); 2760 } 2761 case NEON::BI__builtin_neon_vclz_v: 2762 case NEON::BI__builtin_neon_vclzq_v: 2763 // We generate target-independent intrinsic, which needs a second argument 2764 // for whether or not clz of zero is undefined; on ARM it isn't. 2765 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2766 break; 2767 case NEON::BI__builtin_neon_vcvt_f32_v: 2768 case NEON::BI__builtin_neon_vcvtq_f32_v: 2769 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2770 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2771 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2772 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2773 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2774 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2775 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2776 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2777 bool Double = 2778 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2779 llvm::Type *FloatTy = 2780 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2781 : NeonTypeFlags::Float32, 2782 false, Quad)); 2783 llvm::Type *Tys[2] = { FloatTy, Ty }; 2784 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2785 Function *F = CGM.getIntrinsic(Int, Tys); 2786 return EmitNeonCall(F, Ops, "vcvt_n"); 2787 } 2788 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2789 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2790 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2791 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2792 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2793 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2794 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2795 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2796 bool Double = 2797 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2798 llvm::Type *FloatTy = 2799 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2800 : NeonTypeFlags::Float32, 2801 false, Quad)); 2802 llvm::Type *Tys[2] = { Ty, FloatTy }; 2803 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2804 return EmitNeonCall(F, Ops, "vcvt_n"); 2805 } 2806 case NEON::BI__builtin_neon_vcvt_s32_v: 2807 case NEON::BI__builtin_neon_vcvt_u32_v: 2808 case NEON::BI__builtin_neon_vcvt_s64_v: 2809 case NEON::BI__builtin_neon_vcvt_u64_v: 2810 case NEON::BI__builtin_neon_vcvtq_s32_v: 2811 case NEON::BI__builtin_neon_vcvtq_u32_v: 2812 case NEON::BI__builtin_neon_vcvtq_s64_v: 2813 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2814 bool Double = 2815 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2816 llvm::Type *FloatTy = 2817 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2818 : NeonTypeFlags::Float32, 2819 false, Quad)); 2820 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2821 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2822 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2823 } 2824 case NEON::BI__builtin_neon_vcvta_s32_v: 2825 case NEON::BI__builtin_neon_vcvta_s64_v: 2826 case NEON::BI__builtin_neon_vcvta_u32_v: 2827 case NEON::BI__builtin_neon_vcvta_u64_v: 2828 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2829 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2830 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2831 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2832 case NEON::BI__builtin_neon_vcvtn_s32_v: 2833 case NEON::BI__builtin_neon_vcvtn_s64_v: 2834 case NEON::BI__builtin_neon_vcvtn_u32_v: 2835 case NEON::BI__builtin_neon_vcvtn_u64_v: 2836 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2837 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2838 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2839 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2840 case NEON::BI__builtin_neon_vcvtp_s32_v: 2841 case NEON::BI__builtin_neon_vcvtp_s64_v: 2842 case NEON::BI__builtin_neon_vcvtp_u32_v: 2843 case NEON::BI__builtin_neon_vcvtp_u64_v: 2844 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2845 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2846 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2847 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2848 case NEON::BI__builtin_neon_vcvtm_s32_v: 2849 case NEON::BI__builtin_neon_vcvtm_s64_v: 2850 case NEON::BI__builtin_neon_vcvtm_u32_v: 2851 case NEON::BI__builtin_neon_vcvtm_u64_v: 2852 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2853 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2854 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2855 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2856 bool Double = 2857 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2858 llvm::Type *InTy = 2859 GetNeonType(this, 2860 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2861 : NeonTypeFlags::Float32, false, Quad)); 2862 llvm::Type *Tys[2] = { Ty, InTy }; 2863 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2864 } 2865 case NEON::BI__builtin_neon_vext_v: 2866 case NEON::BI__builtin_neon_vextq_v: { 2867 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2868 SmallVector<Constant*, 16> Indices; 2869 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2870 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2871 2872 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2873 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2874 Value *SV = llvm::ConstantVector::get(Indices); 2875 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2876 } 2877 case NEON::BI__builtin_neon_vfma_v: 2878 case NEON::BI__builtin_neon_vfmaq_v: { 2879 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2880 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2881 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2882 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2883 2884 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2885 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2886 } 2887 case NEON::BI__builtin_neon_vld1_v: 2888 case NEON::BI__builtin_neon_vld1q_v: 2889 Ops.push_back(Align); 2890 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2891 case NEON::BI__builtin_neon_vld2_v: 2892 case NEON::BI__builtin_neon_vld2q_v: 2893 case NEON::BI__builtin_neon_vld3_v: 2894 case NEON::BI__builtin_neon_vld3q_v: 2895 case NEON::BI__builtin_neon_vld4_v: 2896 case NEON::BI__builtin_neon_vld4q_v: { 2897 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2898 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, NameHint); 2899 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2900 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2901 return Builder.CreateStore(Ops[1], Ops[0]); 2902 } 2903 case NEON::BI__builtin_neon_vld1_dup_v: 2904 case NEON::BI__builtin_neon_vld1q_dup_v: { 2905 Value *V = UndefValue::get(Ty); 2906 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2907 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2908 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2909 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2910 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2911 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2912 return EmitNeonSplat(Ops[0], CI); 2913 } 2914 case NEON::BI__builtin_neon_vld2_lane_v: 2915 case NEON::BI__builtin_neon_vld2q_lane_v: 2916 case NEON::BI__builtin_neon_vld3_lane_v: 2917 case NEON::BI__builtin_neon_vld3q_lane_v: 2918 case NEON::BI__builtin_neon_vld4_lane_v: 2919 case NEON::BI__builtin_neon_vld4q_lane_v: { 2920 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2921 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2922 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2923 Ops.push_back(Align); 2924 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2925 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2926 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2927 return Builder.CreateStore(Ops[1], Ops[0]); 2928 } 2929 case NEON::BI__builtin_neon_vmovl_v: { 2930 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2931 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2932 if (Usgn) 2933 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2934 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2935 } 2936 case NEON::BI__builtin_neon_vmovn_v: { 2937 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2938 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2939 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2940 } 2941 case NEON::BI__builtin_neon_vmull_v: 2942 // FIXME: the integer vmull operations could be emitted in terms of pure 2943 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2944 // hoisting the exts outside loops. Until global ISel comes along that can 2945 // see through such movement this leads to bad CodeGen. So we need an 2946 // intrinsic for now. 2947 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2948 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2949 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2950 case NEON::BI__builtin_neon_vpadal_v: 2951 case NEON::BI__builtin_neon_vpadalq_v: { 2952 // The source operand type has twice as many elements of half the size. 2953 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2954 llvm::Type *EltTy = 2955 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2956 llvm::Type *NarrowTy = 2957 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2958 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2959 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 2960 } 2961 case NEON::BI__builtin_neon_vpaddl_v: 2962 case NEON::BI__builtin_neon_vpaddlq_v: { 2963 // The source operand type has twice as many elements of half the size. 2964 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2965 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2966 llvm::Type *NarrowTy = 2967 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2968 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2969 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2970 } 2971 case NEON::BI__builtin_neon_vqdmlal_v: 2972 case NEON::BI__builtin_neon_vqdmlsl_v: { 2973 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2974 Value *Mul = EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), 2975 MulOps, "vqdmlal"); 2976 2977 SmallVector<Value *, 2> AccumOps; 2978 AccumOps.push_back(Ops[0]); 2979 AccumOps.push_back(Mul); 2980 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), 2981 AccumOps, NameHint); 2982 } 2983 case NEON::BI__builtin_neon_vqshl_n_v: 2984 case NEON::BI__builtin_neon_vqshlq_n_v: 2985 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2986 1, false); 2987 case NEON::BI__builtin_neon_vqshlu_n_v: 2988 case NEON::BI__builtin_neon_vqshluq_n_v: 2989 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 2990 1, false); 2991 case NEON::BI__builtin_neon_vrecpe_v: 2992 case NEON::BI__builtin_neon_vrecpeq_v: 2993 case NEON::BI__builtin_neon_vrsqrte_v: 2994 case NEON::BI__builtin_neon_vrsqrteq_v: 2995 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 2996 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 2997 2998 case NEON::BI__builtin_neon_vrshr_n_v: 2999 case NEON::BI__builtin_neon_vrshrq_n_v: 3000 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3001 1, true); 3002 case NEON::BI__builtin_neon_vshl_n_v: 3003 case NEON::BI__builtin_neon_vshlq_n_v: 3004 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3005 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3006 "vshl_n"); 3007 case NEON::BI__builtin_neon_vshll_n_v: { 3008 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3009 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3010 if (Usgn) 3011 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3012 else 3013 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3014 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3015 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3016 } 3017 case NEON::BI__builtin_neon_vshrn_n_v: { 3018 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3019 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3020 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3021 if (Usgn) 3022 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3023 else 3024 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3025 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3026 } 3027 case NEON::BI__builtin_neon_vshr_n_v: 3028 case NEON::BI__builtin_neon_vshrq_n_v: 3029 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3030 case NEON::BI__builtin_neon_vst1_v: 3031 case NEON::BI__builtin_neon_vst1q_v: 3032 case NEON::BI__builtin_neon_vst2_v: 3033 case NEON::BI__builtin_neon_vst2q_v: 3034 case NEON::BI__builtin_neon_vst3_v: 3035 case NEON::BI__builtin_neon_vst3q_v: 3036 case NEON::BI__builtin_neon_vst4_v: 3037 case NEON::BI__builtin_neon_vst4q_v: 3038 case NEON::BI__builtin_neon_vst2_lane_v: 3039 case NEON::BI__builtin_neon_vst2q_lane_v: 3040 case NEON::BI__builtin_neon_vst3_lane_v: 3041 case NEON::BI__builtin_neon_vst3q_lane_v: 3042 case NEON::BI__builtin_neon_vst4_lane_v: 3043 case NEON::BI__builtin_neon_vst4q_lane_v: 3044 Ops.push_back(Align); 3045 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3046 case NEON::BI__builtin_neon_vsubhn_v: { 3047 llvm::VectorType *SrcTy = 3048 llvm::VectorType::getExtendedElementVectorType(VTy); 3049 3050 // %sum = add <4 x i32> %lhs, %rhs 3051 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3052 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3053 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3054 3055 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3056 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 3057 SrcTy->getScalarSizeInBits() / 2); 3058 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 3059 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3060 3061 // %res = trunc <4 x i32> %high to <4 x i16> 3062 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3063 } 3064 case NEON::BI__builtin_neon_vtrn_v: 3065 case NEON::BI__builtin_neon_vtrnq_v: { 3066 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3067 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3068 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3069 Value *SV = nullptr; 3070 3071 for (unsigned vi = 0; vi != 2; ++vi) { 3072 SmallVector<Constant*, 16> Indices; 3073 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3074 Indices.push_back(Builder.getInt32(i+vi)); 3075 Indices.push_back(Builder.getInt32(i+e+vi)); 3076 } 3077 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3078 SV = llvm::ConstantVector::get(Indices); 3079 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3080 SV = Builder.CreateStore(SV, Addr); 3081 } 3082 return SV; 3083 } 3084 case NEON::BI__builtin_neon_vtst_v: 3085 case NEON::BI__builtin_neon_vtstq_v: { 3086 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3087 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3088 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3089 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3090 ConstantAggregateZero::get(Ty)); 3091 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3092 } 3093 case NEON::BI__builtin_neon_vuzp_v: 3094 case NEON::BI__builtin_neon_vuzpq_v: { 3095 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3096 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3097 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3098 Value *SV = nullptr; 3099 3100 for (unsigned vi = 0; vi != 2; ++vi) { 3101 SmallVector<Constant*, 16> Indices; 3102 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3103 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3104 3105 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3106 SV = llvm::ConstantVector::get(Indices); 3107 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3108 SV = Builder.CreateStore(SV, Addr); 3109 } 3110 return SV; 3111 } 3112 case NEON::BI__builtin_neon_vzip_v: 3113 case NEON::BI__builtin_neon_vzipq_v: { 3114 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3115 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3116 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3117 Value *SV = nullptr; 3118 3119 for (unsigned vi = 0; vi != 2; ++vi) { 3120 SmallVector<Constant*, 16> Indices; 3121 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3122 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3123 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3124 } 3125 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 3126 SV = llvm::ConstantVector::get(Indices); 3127 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3128 SV = Builder.CreateStore(SV, Addr); 3129 } 3130 return SV; 3131 } 3132 } 3133 3134 assert(Int && "Expected valid intrinsic number"); 3135 3136 // Determine the type(s) of this overloaded AArch64 intrinsic. 3137 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3138 3139 Value *Result = EmitNeonCall(F, Ops, NameHint); 3140 llvm::Type *ResultType = ConvertType(E->getType()); 3141 // AArch64 intrinsic one-element vector type cast to 3142 // scalar type expected by the builtin 3143 return Builder.CreateBitCast(Result, ResultType, NameHint); 3144 } 3145 3146 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3147 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3148 const CmpInst::Predicate Ip, const Twine &Name) { 3149 llvm::Type *OTy = Op->getType(); 3150 3151 // FIXME: this is utterly horrific. We should not be looking at previous 3152 // codegen context to find out what needs doing. Unfortunately TableGen 3153 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3154 // (etc). 3155 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3156 OTy = BI->getOperand(0)->getType(); 3157 3158 Op = Builder.CreateBitCast(Op, OTy); 3159 if (OTy->getScalarType()->isFloatingPointTy()) { 3160 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3161 } else { 3162 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3163 } 3164 return Builder.CreateSExt(Op, Ty, Name); 3165 } 3166 3167 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3168 Value *ExtOp, Value *IndexOp, 3169 llvm::Type *ResTy, unsigned IntID, 3170 const char *Name) { 3171 SmallVector<Value *, 2> TblOps; 3172 if (ExtOp) 3173 TblOps.push_back(ExtOp); 3174 3175 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3176 SmallVector<Constant*, 16> Indices; 3177 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3178 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3179 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3180 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3181 } 3182 Value *SV = llvm::ConstantVector::get(Indices); 3183 3184 int PairPos = 0, End = Ops.size() - 1; 3185 while (PairPos < End) { 3186 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3187 Ops[PairPos+1], SV, Name)); 3188 PairPos += 2; 3189 } 3190 3191 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3192 // of the 128-bit lookup table with zero. 3193 if (PairPos == End) { 3194 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3195 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3196 ZeroTbl, SV, Name)); 3197 } 3198 3199 Function *TblF; 3200 TblOps.push_back(IndexOp); 3201 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3202 3203 return CGF.EmitNeonCall(TblF, TblOps, Name); 3204 } 3205 3206 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3207 switch (BuiltinID) { 3208 default: 3209 return nullptr; 3210 case ARM::BI__builtin_arm_nop: 3211 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3212 llvm::ConstantInt::get(Int32Ty, 0)); 3213 case ARM::BI__builtin_arm_yield: 3214 case ARM::BI__yield: 3215 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3216 llvm::ConstantInt::get(Int32Ty, 1)); 3217 case ARM::BI__builtin_arm_wfe: 3218 case ARM::BI__wfe: 3219 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3220 llvm::ConstantInt::get(Int32Ty, 2)); 3221 case ARM::BI__builtin_arm_wfi: 3222 case ARM::BI__wfi: 3223 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3224 llvm::ConstantInt::get(Int32Ty, 3)); 3225 case ARM::BI__builtin_arm_sev: 3226 case ARM::BI__sev: 3227 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3228 llvm::ConstantInt::get(Int32Ty, 4)); 3229 case ARM::BI__builtin_arm_sevl: 3230 case ARM::BI__sevl: 3231 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3232 llvm::ConstantInt::get(Int32Ty, 5)); 3233 } 3234 } 3235 3236 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3237 const CallExpr *E) { 3238 if (auto Hint = GetValueForARMHint(BuiltinID)) 3239 return Hint; 3240 3241 if (BuiltinID == ARM::BI__emit) { 3242 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3243 llvm::FunctionType *FTy = 3244 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3245 3246 APSInt Value; 3247 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3248 llvm_unreachable("Sema will ensure that the parameter is constant"); 3249 3250 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3251 3252 llvm::InlineAsm *Emit = 3253 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3254 /*SideEffects=*/true) 3255 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3256 /*SideEffects=*/true); 3257 3258 return Builder.CreateCall(Emit); 3259 } 3260 3261 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3262 Value *Option = EmitScalarExpr(E->getArg(0)); 3263 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3264 } 3265 3266 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3267 Value *Address = EmitScalarExpr(E->getArg(0)); 3268 Value *RW = EmitScalarExpr(E->getArg(1)); 3269 Value *IsData = EmitScalarExpr(E->getArg(2)); 3270 3271 // Locality is not supported on ARM target 3272 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3273 3274 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3275 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 3276 } 3277 3278 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3279 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3280 EmitScalarExpr(E->getArg(0)), 3281 "rbit"); 3282 } 3283 3284 if (BuiltinID == ARM::BI__clear_cache) { 3285 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3286 const FunctionDecl *FD = E->getDirectCallee(); 3287 SmallVector<Value*, 2> Ops; 3288 for (unsigned i = 0; i < 2; i++) 3289 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3290 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3291 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3292 StringRef Name = FD->getName(); 3293 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3294 } 3295 3296 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3297 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3298 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3299 getContext().getTypeSize(E->getType()) == 64) || 3300 BuiltinID == ARM::BI__ldrexd) { 3301 Function *F; 3302 3303 switch (BuiltinID) { 3304 default: llvm_unreachable("unexpected builtin"); 3305 case ARM::BI__builtin_arm_ldaex: 3306 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3307 break; 3308 case ARM::BI__builtin_arm_ldrexd: 3309 case ARM::BI__builtin_arm_ldrex: 3310 case ARM::BI__ldrexd: 3311 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3312 break; 3313 } 3314 3315 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3316 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3317 "ldrexd"); 3318 3319 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3320 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3321 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3322 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3323 3324 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3325 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3326 Val = Builder.CreateOr(Val, Val1); 3327 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3328 } 3329 3330 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3331 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3332 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3333 3334 QualType Ty = E->getType(); 3335 llvm::Type *RealResTy = ConvertType(Ty); 3336 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3337 getContext().getTypeSize(Ty)); 3338 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3339 3340 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3341 ? Intrinsic::arm_ldaex 3342 : Intrinsic::arm_ldrex, 3343 LoadAddr->getType()); 3344 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3345 3346 if (RealResTy->isPointerTy()) 3347 return Builder.CreateIntToPtr(Val, RealResTy); 3348 else { 3349 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3350 return Builder.CreateBitCast(Val, RealResTy); 3351 } 3352 } 3353 3354 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3355 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3356 BuiltinID == ARM::BI__builtin_arm_strex) && 3357 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3358 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3359 ? Intrinsic::arm_stlexd 3360 : Intrinsic::arm_strexd); 3361 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3362 3363 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3364 Value *Val = EmitScalarExpr(E->getArg(0)); 3365 Builder.CreateStore(Val, Tmp); 3366 3367 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3368 Val = Builder.CreateLoad(LdPtr); 3369 3370 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3371 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3372 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3373 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 3374 } 3375 3376 if (BuiltinID == ARM::BI__builtin_arm_strex || 3377 BuiltinID == ARM::BI__builtin_arm_stlex) { 3378 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3379 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3380 3381 QualType Ty = E->getArg(0)->getType(); 3382 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3383 getContext().getTypeSize(Ty)); 3384 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3385 3386 if (StoreVal->getType()->isPointerTy()) 3387 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3388 else { 3389 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3390 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3391 } 3392 3393 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3394 ? Intrinsic::arm_stlex 3395 : Intrinsic::arm_strex, 3396 StoreAddr->getType()); 3397 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 3398 } 3399 3400 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3401 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3402 return Builder.CreateCall(F); 3403 } 3404 3405 // CRC32 3406 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3407 switch (BuiltinID) { 3408 case ARM::BI__builtin_arm_crc32b: 3409 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3410 case ARM::BI__builtin_arm_crc32cb: 3411 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3412 case ARM::BI__builtin_arm_crc32h: 3413 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3414 case ARM::BI__builtin_arm_crc32ch: 3415 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3416 case ARM::BI__builtin_arm_crc32w: 3417 case ARM::BI__builtin_arm_crc32d: 3418 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3419 case ARM::BI__builtin_arm_crc32cw: 3420 case ARM::BI__builtin_arm_crc32cd: 3421 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3422 } 3423 3424 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3425 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3426 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3427 3428 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3429 // intrinsics, hence we need different codegen for these cases. 3430 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3431 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3432 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3433 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3434 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3435 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3436 3437 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3438 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 3439 return Builder.CreateCall2(F, Res, Arg1b); 3440 } else { 3441 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3442 3443 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3444 return Builder.CreateCall2(F, Arg0, Arg1); 3445 } 3446 } 3447 3448 SmallVector<Value*, 4> Ops; 3449 llvm::Value *Align = nullptr; 3450 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3451 if (i == 0) { 3452 switch (BuiltinID) { 3453 case NEON::BI__builtin_neon_vld1_v: 3454 case NEON::BI__builtin_neon_vld1q_v: 3455 case NEON::BI__builtin_neon_vld1q_lane_v: 3456 case NEON::BI__builtin_neon_vld1_lane_v: 3457 case NEON::BI__builtin_neon_vld1_dup_v: 3458 case NEON::BI__builtin_neon_vld1q_dup_v: 3459 case NEON::BI__builtin_neon_vst1_v: 3460 case NEON::BI__builtin_neon_vst1q_v: 3461 case NEON::BI__builtin_neon_vst1q_lane_v: 3462 case NEON::BI__builtin_neon_vst1_lane_v: 3463 case NEON::BI__builtin_neon_vst2_v: 3464 case NEON::BI__builtin_neon_vst2q_v: 3465 case NEON::BI__builtin_neon_vst2_lane_v: 3466 case NEON::BI__builtin_neon_vst2q_lane_v: 3467 case NEON::BI__builtin_neon_vst3_v: 3468 case NEON::BI__builtin_neon_vst3q_v: 3469 case NEON::BI__builtin_neon_vst3_lane_v: 3470 case NEON::BI__builtin_neon_vst3q_lane_v: 3471 case NEON::BI__builtin_neon_vst4_v: 3472 case NEON::BI__builtin_neon_vst4q_v: 3473 case NEON::BI__builtin_neon_vst4_lane_v: 3474 case NEON::BI__builtin_neon_vst4q_lane_v: 3475 // Get the alignment for the argument in addition to the value; 3476 // we'll use it later. 3477 std::pair<llvm::Value*, unsigned> Src = 3478 EmitPointerWithAlignment(E->getArg(0)); 3479 Ops.push_back(Src.first); 3480 Align = Builder.getInt32(Src.second); 3481 continue; 3482 } 3483 } 3484 if (i == 1) { 3485 switch (BuiltinID) { 3486 case NEON::BI__builtin_neon_vld2_v: 3487 case NEON::BI__builtin_neon_vld2q_v: 3488 case NEON::BI__builtin_neon_vld3_v: 3489 case NEON::BI__builtin_neon_vld3q_v: 3490 case NEON::BI__builtin_neon_vld4_v: 3491 case NEON::BI__builtin_neon_vld4q_v: 3492 case NEON::BI__builtin_neon_vld2_lane_v: 3493 case NEON::BI__builtin_neon_vld2q_lane_v: 3494 case NEON::BI__builtin_neon_vld3_lane_v: 3495 case NEON::BI__builtin_neon_vld3q_lane_v: 3496 case NEON::BI__builtin_neon_vld4_lane_v: 3497 case NEON::BI__builtin_neon_vld4q_lane_v: 3498 case NEON::BI__builtin_neon_vld2_dup_v: 3499 case NEON::BI__builtin_neon_vld3_dup_v: 3500 case NEON::BI__builtin_neon_vld4_dup_v: 3501 // Get the alignment for the argument in addition to the value; 3502 // we'll use it later. 3503 std::pair<llvm::Value*, unsigned> Src = 3504 EmitPointerWithAlignment(E->getArg(1)); 3505 Ops.push_back(Src.first); 3506 Align = Builder.getInt32(Src.second); 3507 continue; 3508 } 3509 } 3510 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3511 } 3512 3513 switch (BuiltinID) { 3514 default: break; 3515 // vget_lane and vset_lane are not overloaded and do not have an extra 3516 // argument that specifies the vector type. 3517 case NEON::BI__builtin_neon_vget_lane_i8: 3518 case NEON::BI__builtin_neon_vget_lane_i16: 3519 case NEON::BI__builtin_neon_vget_lane_i32: 3520 case NEON::BI__builtin_neon_vget_lane_i64: 3521 case NEON::BI__builtin_neon_vget_lane_f32: 3522 case NEON::BI__builtin_neon_vgetq_lane_i8: 3523 case NEON::BI__builtin_neon_vgetq_lane_i16: 3524 case NEON::BI__builtin_neon_vgetq_lane_i32: 3525 case NEON::BI__builtin_neon_vgetq_lane_i64: 3526 case NEON::BI__builtin_neon_vgetq_lane_f32: 3527 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 3528 "vget_lane"); 3529 case NEON::BI__builtin_neon_vset_lane_i8: 3530 case NEON::BI__builtin_neon_vset_lane_i16: 3531 case NEON::BI__builtin_neon_vset_lane_i32: 3532 case NEON::BI__builtin_neon_vset_lane_i64: 3533 case NEON::BI__builtin_neon_vset_lane_f32: 3534 case NEON::BI__builtin_neon_vsetq_lane_i8: 3535 case NEON::BI__builtin_neon_vsetq_lane_i16: 3536 case NEON::BI__builtin_neon_vsetq_lane_i32: 3537 case NEON::BI__builtin_neon_vsetq_lane_i64: 3538 case NEON::BI__builtin_neon_vsetq_lane_f32: 3539 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3540 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3541 3542 // Non-polymorphic crypto instructions also not overloaded 3543 case NEON::BI__builtin_neon_vsha1h_u32: 3544 Ops.push_back(EmitScalarExpr(E->getArg(0))); 3545 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3546 "vsha1h"); 3547 case NEON::BI__builtin_neon_vsha1cq_u32: 3548 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3549 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3550 "vsha1h"); 3551 case NEON::BI__builtin_neon_vsha1pq_u32: 3552 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3553 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3554 "vsha1h"); 3555 case NEON::BI__builtin_neon_vsha1mq_u32: 3556 Ops.push_back(EmitScalarExpr(E->getArg(2))); 3557 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3558 "vsha1h"); 3559 } 3560 3561 // Get the last argument, which specifies the vector type. 3562 llvm::APSInt Result; 3563 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3564 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3565 return nullptr; 3566 3567 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3568 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3569 // Determine the overloaded type of this builtin. 3570 llvm::Type *Ty; 3571 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3572 Ty = FloatTy; 3573 else 3574 Ty = DoubleTy; 3575 3576 // Determine whether this is an unsigned conversion or not. 3577 bool usgn = Result.getZExtValue() == 1; 3578 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3579 3580 // Call the appropriate intrinsic. 3581 Function *F = CGM.getIntrinsic(Int, Ty); 3582 return Builder.CreateCall(F, Ops, "vcvtr"); 3583 } 3584 3585 // Determine the type of this overloaded NEON intrinsic. 3586 NeonTypeFlags Type(Result.getZExtValue()); 3587 bool usgn = Type.isUnsigned(); 3588 bool rightShift = false; 3589 3590 llvm::VectorType *VTy = GetNeonType(this, Type); 3591 llvm::Type *Ty = VTy; 3592 if (!Ty) 3593 return nullptr; 3594 3595 // Many NEON builtins have identical semantics and uses in ARM and 3596 // AArch64. Emit these in a single function. 3597 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3598 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3599 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3600 if (Builtin) 3601 return EmitCommonNeonBuiltinExpr( 3602 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3603 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3604 3605 unsigned Int; 3606 switch (BuiltinID) { 3607 default: return nullptr; 3608 case NEON::BI__builtin_neon_vld1q_lane_v: 3609 // Handle 64-bit integer elements as a special case. Use shuffles of 3610 // one-element vectors to avoid poor code for i64 in the backend. 3611 if (VTy->getElementType()->isIntegerTy(64)) { 3612 // Extract the other lane. 3613 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3614 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3615 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3616 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3617 // Load the value as a one-element vector. 3618 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3619 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3620 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 3621 // Combine them. 3622 SmallVector<Constant*, 2> Indices; 3623 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 3624 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 3625 SV = llvm::ConstantVector::get(Indices); 3626 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3627 } 3628 // fall through 3629 case NEON::BI__builtin_neon_vld1_lane_v: { 3630 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3631 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3632 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3633 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3634 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3635 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3636 } 3637 case NEON::BI__builtin_neon_vld2_dup_v: 3638 case NEON::BI__builtin_neon_vld3_dup_v: 3639 case NEON::BI__builtin_neon_vld4_dup_v: { 3640 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3641 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3642 switch (BuiltinID) { 3643 case NEON::BI__builtin_neon_vld2_dup_v: 3644 Int = Intrinsic::arm_neon_vld2; 3645 break; 3646 case NEON::BI__builtin_neon_vld3_dup_v: 3647 Int = Intrinsic::arm_neon_vld3; 3648 break; 3649 case NEON::BI__builtin_neon_vld4_dup_v: 3650 Int = Intrinsic::arm_neon_vld4; 3651 break; 3652 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3653 } 3654 Function *F = CGM.getIntrinsic(Int, Ty); 3655 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3656 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3657 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3658 return Builder.CreateStore(Ops[1], Ops[0]); 3659 } 3660 switch (BuiltinID) { 3661 case NEON::BI__builtin_neon_vld2_dup_v: 3662 Int = Intrinsic::arm_neon_vld2lane; 3663 break; 3664 case NEON::BI__builtin_neon_vld3_dup_v: 3665 Int = Intrinsic::arm_neon_vld3lane; 3666 break; 3667 case NEON::BI__builtin_neon_vld4_dup_v: 3668 Int = Intrinsic::arm_neon_vld4lane; 3669 break; 3670 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3671 } 3672 Function *F = CGM.getIntrinsic(Int, Ty); 3673 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3674 3675 SmallVector<Value*, 6> Args; 3676 Args.push_back(Ops[1]); 3677 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3678 3679 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3680 Args.push_back(CI); 3681 Args.push_back(Align); 3682 3683 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3684 // splat lane 0 to all elts in each vector of the result. 3685 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3686 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3687 Value *Elt = Builder.CreateBitCast(Val, Ty); 3688 Elt = EmitNeonSplat(Elt, CI); 3689 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3690 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3691 } 3692 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3693 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3694 return Builder.CreateStore(Ops[1], Ops[0]); 3695 } 3696 case NEON::BI__builtin_neon_vqrshrn_n_v: 3697 Int = 3698 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3699 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3700 1, true); 3701 case NEON::BI__builtin_neon_vqrshrun_n_v: 3702 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3703 Ops, "vqrshrun_n", 1, true); 3704 case NEON::BI__builtin_neon_vqshrn_n_v: 3705 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3706 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3707 1, true); 3708 case NEON::BI__builtin_neon_vqshrun_n_v: 3709 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3710 Ops, "vqshrun_n", 1, true); 3711 case NEON::BI__builtin_neon_vrecpe_v: 3712 case NEON::BI__builtin_neon_vrecpeq_v: 3713 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3714 Ops, "vrecpe"); 3715 case NEON::BI__builtin_neon_vrshrn_n_v: 3716 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3717 Ops, "vrshrn_n", 1, true); 3718 case NEON::BI__builtin_neon_vrsra_n_v: 3719 case NEON::BI__builtin_neon_vrsraq_n_v: 3720 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3721 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3722 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3723 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3724 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3725 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3726 case NEON::BI__builtin_neon_vsri_n_v: 3727 case NEON::BI__builtin_neon_vsriq_n_v: 3728 rightShift = true; 3729 case NEON::BI__builtin_neon_vsli_n_v: 3730 case NEON::BI__builtin_neon_vsliq_n_v: 3731 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3733 Ops, "vsli_n"); 3734 case NEON::BI__builtin_neon_vsra_n_v: 3735 case NEON::BI__builtin_neon_vsraq_n_v: 3736 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3737 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3738 return Builder.CreateAdd(Ops[0], Ops[1]); 3739 case NEON::BI__builtin_neon_vst1q_lane_v: 3740 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3741 // a one-element vector and avoid poor code for i64 in the backend. 3742 if (VTy->getElementType()->isIntegerTy(64)) { 3743 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3744 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3745 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3746 Ops[2] = Align; 3747 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3748 Ops[1]->getType()), Ops); 3749 } 3750 // fall through 3751 case NEON::BI__builtin_neon_vst1_lane_v: { 3752 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3753 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3754 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3755 StoreInst *St = Builder.CreateStore(Ops[1], 3756 Builder.CreateBitCast(Ops[0], Ty)); 3757 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3758 return St; 3759 } 3760 case NEON::BI__builtin_neon_vtbl1_v: 3761 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3762 Ops, "vtbl1"); 3763 case NEON::BI__builtin_neon_vtbl2_v: 3764 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3765 Ops, "vtbl2"); 3766 case NEON::BI__builtin_neon_vtbl3_v: 3767 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3768 Ops, "vtbl3"); 3769 case NEON::BI__builtin_neon_vtbl4_v: 3770 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3771 Ops, "vtbl4"); 3772 case NEON::BI__builtin_neon_vtbx1_v: 3773 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3774 Ops, "vtbx1"); 3775 case NEON::BI__builtin_neon_vtbx2_v: 3776 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3777 Ops, "vtbx2"); 3778 case NEON::BI__builtin_neon_vtbx3_v: 3779 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3780 Ops, "vtbx3"); 3781 case NEON::BI__builtin_neon_vtbx4_v: 3782 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3783 Ops, "vtbx4"); 3784 } 3785 } 3786 3787 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3788 const CallExpr *E, 3789 SmallVectorImpl<Value *> &Ops) { 3790 unsigned int Int = 0; 3791 const char *s = nullptr; 3792 3793 switch (BuiltinID) { 3794 default: 3795 return nullptr; 3796 case NEON::BI__builtin_neon_vtbl1_v: 3797 case NEON::BI__builtin_neon_vqtbl1_v: 3798 case NEON::BI__builtin_neon_vqtbl1q_v: 3799 case NEON::BI__builtin_neon_vtbl2_v: 3800 case NEON::BI__builtin_neon_vqtbl2_v: 3801 case NEON::BI__builtin_neon_vqtbl2q_v: 3802 case NEON::BI__builtin_neon_vtbl3_v: 3803 case NEON::BI__builtin_neon_vqtbl3_v: 3804 case NEON::BI__builtin_neon_vqtbl3q_v: 3805 case NEON::BI__builtin_neon_vtbl4_v: 3806 case NEON::BI__builtin_neon_vqtbl4_v: 3807 case NEON::BI__builtin_neon_vqtbl4q_v: 3808 break; 3809 case NEON::BI__builtin_neon_vtbx1_v: 3810 case NEON::BI__builtin_neon_vqtbx1_v: 3811 case NEON::BI__builtin_neon_vqtbx1q_v: 3812 case NEON::BI__builtin_neon_vtbx2_v: 3813 case NEON::BI__builtin_neon_vqtbx2_v: 3814 case NEON::BI__builtin_neon_vqtbx2q_v: 3815 case NEON::BI__builtin_neon_vtbx3_v: 3816 case NEON::BI__builtin_neon_vqtbx3_v: 3817 case NEON::BI__builtin_neon_vqtbx3q_v: 3818 case NEON::BI__builtin_neon_vtbx4_v: 3819 case NEON::BI__builtin_neon_vqtbx4_v: 3820 case NEON::BI__builtin_neon_vqtbx4q_v: 3821 break; 3822 } 3823 3824 assert(E->getNumArgs() >= 3); 3825 3826 // Get the last argument, which specifies the vector type. 3827 llvm::APSInt Result; 3828 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3829 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3830 return nullptr; 3831 3832 // Determine the type of this overloaded NEON intrinsic. 3833 NeonTypeFlags Type(Result.getZExtValue()); 3834 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3835 llvm::Type *Ty = VTy; 3836 if (!Ty) 3837 return nullptr; 3838 3839 unsigned nElts = VTy->getNumElements(); 3840 3841 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3842 3843 // AArch64 scalar builtins are not overloaded, they do not have an extra 3844 // argument that specifies the vector type, need to handle each case. 3845 SmallVector<Value *, 2> TblOps; 3846 switch (BuiltinID) { 3847 case NEON::BI__builtin_neon_vtbl1_v: { 3848 TblOps.push_back(Ops[0]); 3849 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[1], Ty, 3850 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3851 } 3852 case NEON::BI__builtin_neon_vtbl2_v: { 3853 TblOps.push_back(Ops[0]); 3854 TblOps.push_back(Ops[1]); 3855 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3856 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3857 } 3858 case NEON::BI__builtin_neon_vtbl3_v: { 3859 TblOps.push_back(Ops[0]); 3860 TblOps.push_back(Ops[1]); 3861 TblOps.push_back(Ops[2]); 3862 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[3], Ty, 3863 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3864 } 3865 case NEON::BI__builtin_neon_vtbl4_v: { 3866 TblOps.push_back(Ops[0]); 3867 TblOps.push_back(Ops[1]); 3868 TblOps.push_back(Ops[2]); 3869 TblOps.push_back(Ops[3]); 3870 return packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3871 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3872 } 3873 case NEON::BI__builtin_neon_vtbx1_v: { 3874 TblOps.push_back(Ops[1]); 3875 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[2], Ty, 3876 Intrinsic::aarch64_neon_tbl1, "vtbl1"); 3877 3878 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3879 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3880 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3881 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3882 3883 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3884 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3885 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3886 } 3887 case NEON::BI__builtin_neon_vtbx2_v: { 3888 TblOps.push_back(Ops[1]); 3889 TblOps.push_back(Ops[2]); 3890 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3891 Intrinsic::aarch64_neon_tbx1, "vtbx1"); 3892 } 3893 case NEON::BI__builtin_neon_vtbx3_v: { 3894 TblOps.push_back(Ops[1]); 3895 TblOps.push_back(Ops[2]); 3896 TblOps.push_back(Ops[3]); 3897 Value *TblRes = packTBLDVectorList(CGF, TblOps, nullptr, Ops[4], Ty, 3898 Intrinsic::aarch64_neon_tbl2, "vtbl2"); 3899 3900 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3901 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3902 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3903 TwentyFourV); 3904 CmpRes = Builder.CreateSExt(CmpRes, Ty); 3905 3906 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 3907 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 3908 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 3909 } 3910 case NEON::BI__builtin_neon_vtbx4_v: { 3911 TblOps.push_back(Ops[1]); 3912 TblOps.push_back(Ops[2]); 3913 TblOps.push_back(Ops[3]); 3914 TblOps.push_back(Ops[4]); 3915 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3916 Intrinsic::aarch64_neon_tbx2, "vtbx2"); 3917 } 3918 case NEON::BI__builtin_neon_vqtbl1_v: 3919 case NEON::BI__builtin_neon_vqtbl1q_v: 3920 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 3921 case NEON::BI__builtin_neon_vqtbl2_v: 3922 case NEON::BI__builtin_neon_vqtbl2q_v: { 3923 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 3924 case NEON::BI__builtin_neon_vqtbl3_v: 3925 case NEON::BI__builtin_neon_vqtbl3q_v: 3926 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 3927 case NEON::BI__builtin_neon_vqtbl4_v: 3928 case NEON::BI__builtin_neon_vqtbl4q_v: 3929 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 3930 case NEON::BI__builtin_neon_vqtbx1_v: 3931 case NEON::BI__builtin_neon_vqtbx1q_v: 3932 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 3933 case NEON::BI__builtin_neon_vqtbx2_v: 3934 case NEON::BI__builtin_neon_vqtbx2q_v: 3935 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 3936 case NEON::BI__builtin_neon_vqtbx3_v: 3937 case NEON::BI__builtin_neon_vqtbx3q_v: 3938 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 3939 case NEON::BI__builtin_neon_vqtbx4_v: 3940 case NEON::BI__builtin_neon_vqtbx4q_v: 3941 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 3942 } 3943 } 3944 3945 if (!Int) 3946 return nullptr; 3947 3948 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 3949 return CGF.EmitNeonCall(F, Ops, s); 3950 } 3951 3952 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 3953 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3954 Op = Builder.CreateBitCast(Op, Int16Ty); 3955 Value *V = UndefValue::get(VTy); 3956 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3957 Op = Builder.CreateInsertElement(V, Op, CI); 3958 return Op; 3959 } 3960 3961 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 3962 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3963 Op = Builder.CreateBitCast(Op, Int8Ty); 3964 Value *V = UndefValue::get(VTy); 3965 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3966 Op = Builder.CreateInsertElement(V, Op, CI); 3967 return Op; 3968 } 3969 3970 Value *CodeGenFunction:: 3971 emitVectorWrappedScalar8Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3972 const char *Name) { 3973 // i8 is not a legal types for AArch64, so we can't just use 3974 // a normal overloaded intrinsic call for these scalar types. Instead 3975 // we'll build 64-bit vectors w/ lane zero being our input values and 3976 // perform the operation on that. The back end can pattern match directly 3977 // to the scalar instruction. 3978 Ops[0] = vectorWrapScalar8(Ops[0]); 3979 Ops[1] = vectorWrapScalar8(Ops[1]); 3980 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 3981 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3982 Constant *CI = ConstantInt::get(SizeTy, 0); 3983 return Builder.CreateExtractElement(V, CI, "lane0"); 3984 } 3985 3986 Value *CodeGenFunction:: 3987 emitVectorWrappedScalar16Intrinsic(unsigned Int, SmallVectorImpl<Value*> &Ops, 3988 const char *Name) { 3989 // i16 is not a legal types for AArch64, so we can't just use 3990 // a normal overloaded intrinsic call for these scalar types. Instead 3991 // we'll build 64-bit vectors w/ lane zero being our input values and 3992 // perform the operation on that. The back end can pattern match directly 3993 // to the scalar instruction. 3994 Ops[0] = vectorWrapScalar16(Ops[0]); 3995 Ops[1] = vectorWrapScalar16(Ops[1]); 3996 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 3997 Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name); 3998 Constant *CI = ConstantInt::get(SizeTy, 0); 3999 return Builder.CreateExtractElement(V, CI, "lane0"); 4000 } 4001 4002 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4003 const CallExpr *E) { 4004 unsigned HintID = static_cast<unsigned>(-1); 4005 switch (BuiltinID) { 4006 default: break; 4007 case AArch64::BI__builtin_arm_nop: 4008 HintID = 0; 4009 break; 4010 case AArch64::BI__builtin_arm_yield: 4011 HintID = 1; 4012 break; 4013 case AArch64::BI__builtin_arm_wfe: 4014 HintID = 2; 4015 break; 4016 case AArch64::BI__builtin_arm_wfi: 4017 HintID = 3; 4018 break; 4019 case AArch64::BI__builtin_arm_sev: 4020 HintID = 4; 4021 break; 4022 case AArch64::BI__builtin_arm_sevl: 4023 HintID = 5; 4024 break; 4025 } 4026 4027 if (HintID != static_cast<unsigned>(-1)) { 4028 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4029 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4030 } 4031 4032 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4033 Value *Address = EmitScalarExpr(E->getArg(0)); 4034 Value *RW = EmitScalarExpr(E->getArg(1)); 4035 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4036 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4037 Value *IsData = EmitScalarExpr(E->getArg(4)); 4038 4039 Value *Locality = nullptr; 4040 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4041 // Temporal fetch, needs to convert cache level to locality. 4042 Locality = llvm::ConstantInt::get(Int32Ty, 4043 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4044 } else { 4045 // Streaming fetch. 4046 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4047 } 4048 4049 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4050 // PLDL3STRM or PLDL2STRM. 4051 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4052 return Builder.CreateCall4(F, Address, RW, Locality, IsData); 4053 } 4054 4055 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4056 assert((getContext().getTypeSize(E->getType()) == 32) && 4057 "rbit of unusual size!"); 4058 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4059 return Builder.CreateCall( 4060 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4061 } 4062 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4063 assert((getContext().getTypeSize(E->getType()) == 64) && 4064 "rbit of unusual size!"); 4065 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4066 return Builder.CreateCall( 4067 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4068 } 4069 4070 if (BuiltinID == AArch64::BI__clear_cache) { 4071 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4072 const FunctionDecl *FD = E->getDirectCallee(); 4073 SmallVector<Value*, 2> Ops; 4074 for (unsigned i = 0; i < 2; i++) 4075 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4076 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4077 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4078 StringRef Name = FD->getName(); 4079 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4080 } 4081 4082 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4083 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4084 getContext().getTypeSize(E->getType()) == 128) { 4085 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4086 ? Intrinsic::aarch64_ldaxp 4087 : Intrinsic::aarch64_ldxp); 4088 4089 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4090 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4091 "ldxp"); 4092 4093 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4094 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4095 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4096 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4097 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4098 4099 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4100 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4101 Val = Builder.CreateOr(Val, Val1); 4102 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4103 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4104 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4105 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4106 4107 QualType Ty = E->getType(); 4108 llvm::Type *RealResTy = ConvertType(Ty); 4109 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4110 getContext().getTypeSize(Ty)); 4111 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4112 4113 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4114 ? Intrinsic::aarch64_ldaxr 4115 : Intrinsic::aarch64_ldxr, 4116 LoadAddr->getType()); 4117 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4118 4119 if (RealResTy->isPointerTy()) 4120 return Builder.CreateIntToPtr(Val, RealResTy); 4121 4122 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4123 return Builder.CreateBitCast(Val, RealResTy); 4124 } 4125 4126 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4127 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4128 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4129 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4130 ? Intrinsic::aarch64_stlxp 4131 : Intrinsic::aarch64_stxp); 4132 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4133 4134 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4135 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4136 One); 4137 Value *Val = EmitScalarExpr(E->getArg(0)); 4138 Builder.CreateStore(Val, Tmp); 4139 4140 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4141 Val = Builder.CreateLoad(LdPtr); 4142 4143 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4144 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4145 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4146 Int8PtrTy); 4147 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "stxp"); 4148 } else if (BuiltinID == AArch64::BI__builtin_arm_strex || 4149 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4150 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4151 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4152 4153 QualType Ty = E->getArg(0)->getType(); 4154 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4155 getContext().getTypeSize(Ty)); 4156 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4157 4158 if (StoreVal->getType()->isPointerTy()) 4159 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4160 else { 4161 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4162 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4163 } 4164 4165 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4166 ? Intrinsic::aarch64_stlxr 4167 : Intrinsic::aarch64_stxr, 4168 StoreAddr->getType()); 4169 return Builder.CreateCall2(F, StoreVal, StoreAddr, "stxr"); 4170 } 4171 4172 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4173 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4174 return Builder.CreateCall(F); 4175 } 4176 4177 // CRC32 4178 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4179 switch (BuiltinID) { 4180 case AArch64::BI__builtin_arm_crc32b: 4181 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4182 case AArch64::BI__builtin_arm_crc32cb: 4183 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4184 case AArch64::BI__builtin_arm_crc32h: 4185 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4186 case AArch64::BI__builtin_arm_crc32ch: 4187 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4188 case AArch64::BI__builtin_arm_crc32w: 4189 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4190 case AArch64::BI__builtin_arm_crc32cw: 4191 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4192 case AArch64::BI__builtin_arm_crc32d: 4193 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4194 case AArch64::BI__builtin_arm_crc32cd: 4195 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4196 } 4197 4198 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4199 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4200 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4201 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4202 4203 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4204 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4205 4206 return Builder.CreateCall2(F, Arg0, Arg1); 4207 } 4208 4209 llvm::SmallVector<Value*, 4> Ops; 4210 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 4211 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4212 4213 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4214 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4215 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4216 4217 if (Builtin) { 4218 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4219 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4220 assert(Result && "SISD intrinsic should have been handled"); 4221 return Result; 4222 } 4223 4224 llvm::APSInt Result; 4225 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4226 NeonTypeFlags Type(0); 4227 if (Arg->isIntegerConstantExpr(Result, getContext())) 4228 // Determine the type of this overloaded NEON intrinsic. 4229 Type = NeonTypeFlags(Result.getZExtValue()); 4230 4231 bool usgn = Type.isUnsigned(); 4232 bool quad = Type.isQuad(); 4233 4234 // Handle non-overloaded intrinsics first. 4235 switch (BuiltinID) { 4236 default: break; 4237 case NEON::BI__builtin_neon_vldrq_p128: { 4238 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4239 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4240 return Builder.CreateLoad(Ptr); 4241 } 4242 case NEON::BI__builtin_neon_vstrq_p128: { 4243 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4244 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4245 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4246 } 4247 case NEON::BI__builtin_neon_vcvts_u32_f32: 4248 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4249 usgn = true; 4250 // FALL THROUGH 4251 case NEON::BI__builtin_neon_vcvts_s32_f32: 4252 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4253 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4254 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4255 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4256 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4257 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4258 if (usgn) 4259 return Builder.CreateFPToUI(Ops[0], InTy); 4260 return Builder.CreateFPToSI(Ops[0], InTy); 4261 } 4262 case NEON::BI__builtin_neon_vcvts_f32_u32: 4263 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4264 usgn = true; 4265 // FALL THROUGH 4266 case NEON::BI__builtin_neon_vcvts_f32_s32: 4267 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4268 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4269 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4270 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4271 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4272 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4273 if (usgn) 4274 return Builder.CreateUIToFP(Ops[0], FTy); 4275 return Builder.CreateSIToFP(Ops[0], FTy); 4276 } 4277 case NEON::BI__builtin_neon_vpaddd_s64: { 4278 llvm::Type *Ty = 4279 llvm::VectorType::get(llvm::Type::getInt64Ty(getLLVMContext()), 2); 4280 Value *Vec = EmitScalarExpr(E->getArg(0)); 4281 // The vector is v2f64, so make sure it's bitcast to that. 4282 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4283 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4284 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4285 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4286 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4287 // Pairwise addition of a v2f64 into a scalar f64. 4288 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4289 } 4290 case NEON::BI__builtin_neon_vpaddd_f64: { 4291 llvm::Type *Ty = 4292 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4293 Value *Vec = EmitScalarExpr(E->getArg(0)); 4294 // The vector is v2f64, so make sure it's bitcast to that. 4295 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4296 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4297 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4298 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4299 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4300 // Pairwise addition of a v2f64 into a scalar f64. 4301 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4302 } 4303 case NEON::BI__builtin_neon_vpadds_f32: { 4304 llvm::Type *Ty = 4305 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4306 Value *Vec = EmitScalarExpr(E->getArg(0)); 4307 // The vector is v2f32, so make sure it's bitcast to that. 4308 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4309 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4310 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4311 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4312 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4313 // Pairwise addition of a v2f32 into a scalar f32. 4314 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4315 } 4316 case NEON::BI__builtin_neon_vceqzd_s64: 4317 case NEON::BI__builtin_neon_vceqzd_f64: 4318 case NEON::BI__builtin_neon_vceqzs_f32: 4319 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4320 return EmitAArch64CompareBuiltinExpr( 4321 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OEQ, 4322 ICmpInst::ICMP_EQ, "vceqz"); 4323 case NEON::BI__builtin_neon_vcgezd_s64: 4324 case NEON::BI__builtin_neon_vcgezd_f64: 4325 case NEON::BI__builtin_neon_vcgezs_f32: 4326 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4327 return EmitAArch64CompareBuiltinExpr( 4328 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGE, 4329 ICmpInst::ICMP_SGE, "vcgez"); 4330 case NEON::BI__builtin_neon_vclezd_s64: 4331 case NEON::BI__builtin_neon_vclezd_f64: 4332 case NEON::BI__builtin_neon_vclezs_f32: 4333 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4334 return EmitAArch64CompareBuiltinExpr( 4335 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLE, 4336 ICmpInst::ICMP_SLE, "vclez"); 4337 case NEON::BI__builtin_neon_vcgtzd_s64: 4338 case NEON::BI__builtin_neon_vcgtzd_f64: 4339 case NEON::BI__builtin_neon_vcgtzs_f32: 4340 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4341 return EmitAArch64CompareBuiltinExpr( 4342 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OGT, 4343 ICmpInst::ICMP_SGT, "vcgtz"); 4344 case NEON::BI__builtin_neon_vcltzd_s64: 4345 case NEON::BI__builtin_neon_vcltzd_f64: 4346 case NEON::BI__builtin_neon_vcltzs_f32: 4347 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4348 return EmitAArch64CompareBuiltinExpr( 4349 Ops[0], ConvertType(E->getCallReturnType()), ICmpInst::FCMP_OLT, 4350 ICmpInst::ICMP_SLT, "vcltz"); 4351 4352 case NEON::BI__builtin_neon_vceqzd_u64: { 4353 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4354 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4355 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4356 Ops[0] = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, Ops[0], 4357 llvm::Constant::getNullValue(Ty)); 4358 return Builder.CreateSExt(Ops[0], Ty, "vceqzd"); 4359 } 4360 case NEON::BI__builtin_neon_vceqd_f64: 4361 case NEON::BI__builtin_neon_vcled_f64: 4362 case NEON::BI__builtin_neon_vcltd_f64: 4363 case NEON::BI__builtin_neon_vcged_f64: 4364 case NEON::BI__builtin_neon_vcgtd_f64: { 4365 llvm::CmpInst::Predicate P; 4366 switch (BuiltinID) { 4367 default: llvm_unreachable("missing builtin ID in switch!"); 4368 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4369 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4370 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4371 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4372 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4373 } 4374 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4375 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4376 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4377 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4378 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4379 } 4380 case NEON::BI__builtin_neon_vceqs_f32: 4381 case NEON::BI__builtin_neon_vcles_f32: 4382 case NEON::BI__builtin_neon_vclts_f32: 4383 case NEON::BI__builtin_neon_vcges_f32: 4384 case NEON::BI__builtin_neon_vcgts_f32: { 4385 llvm::CmpInst::Predicate P; 4386 switch (BuiltinID) { 4387 default: llvm_unreachable("missing builtin ID in switch!"); 4388 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4389 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4390 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4391 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4392 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4393 } 4394 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4395 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4396 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4397 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4398 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4399 } 4400 case NEON::BI__builtin_neon_vceqd_s64: 4401 case NEON::BI__builtin_neon_vceqd_u64: 4402 case NEON::BI__builtin_neon_vcgtd_s64: 4403 case NEON::BI__builtin_neon_vcgtd_u64: 4404 case NEON::BI__builtin_neon_vcltd_s64: 4405 case NEON::BI__builtin_neon_vcltd_u64: 4406 case NEON::BI__builtin_neon_vcged_u64: 4407 case NEON::BI__builtin_neon_vcged_s64: 4408 case NEON::BI__builtin_neon_vcled_u64: 4409 case NEON::BI__builtin_neon_vcled_s64: { 4410 llvm::CmpInst::Predicate P; 4411 switch (BuiltinID) { 4412 default: llvm_unreachable("missing builtin ID in switch!"); 4413 case NEON::BI__builtin_neon_vceqd_s64: 4414 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4415 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4416 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4417 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4418 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4419 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4420 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4421 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4422 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4423 } 4424 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4425 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4426 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4427 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4428 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4429 } 4430 case NEON::BI__builtin_neon_vtstd_s64: 4431 case NEON::BI__builtin_neon_vtstd_u64: { 4432 llvm::Type *Ty = llvm::Type::getInt64Ty(getLLVMContext()); 4433 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4434 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4435 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4436 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4437 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4438 llvm::Constant::getNullValue(Ty)); 4439 return Builder.CreateSExt(Ops[0], Ty, "vtstd"); 4440 } 4441 case NEON::BI__builtin_neon_vset_lane_i8: 4442 case NEON::BI__builtin_neon_vset_lane_i16: 4443 case NEON::BI__builtin_neon_vset_lane_i32: 4444 case NEON::BI__builtin_neon_vset_lane_i64: 4445 case NEON::BI__builtin_neon_vset_lane_f32: 4446 case NEON::BI__builtin_neon_vsetq_lane_i8: 4447 case NEON::BI__builtin_neon_vsetq_lane_i16: 4448 case NEON::BI__builtin_neon_vsetq_lane_i32: 4449 case NEON::BI__builtin_neon_vsetq_lane_i64: 4450 case NEON::BI__builtin_neon_vsetq_lane_f32: 4451 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4452 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4453 case NEON::BI__builtin_neon_vset_lane_f64: 4454 // The vector type needs a cast for the v1f64 variant. 4455 Ops[1] = Builder.CreateBitCast(Ops[1], 4456 llvm::VectorType::get(DoubleTy, 1)); 4457 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4458 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4459 case NEON::BI__builtin_neon_vsetq_lane_f64: 4460 // The vector type needs a cast for the v2f64 variant. 4461 Ops[1] = Builder.CreateBitCast(Ops[1], 4462 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4463 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4464 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4465 4466 case NEON::BI__builtin_neon_vget_lane_i8: 4467 case NEON::BI__builtin_neon_vdupb_lane_i8: 4468 Ops[0] = Builder.CreateBitCast(Ops[0], 4469 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8)); 4470 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4471 "vget_lane"); 4472 case NEON::BI__builtin_neon_vgetq_lane_i8: 4473 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4474 Ops[0] = Builder.CreateBitCast(Ops[0], 4475 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16)); 4476 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4477 "vgetq_lane"); 4478 case NEON::BI__builtin_neon_vget_lane_i16: 4479 case NEON::BI__builtin_neon_vduph_lane_i16: 4480 Ops[0] = Builder.CreateBitCast(Ops[0], 4481 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4)); 4482 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4483 "vget_lane"); 4484 case NEON::BI__builtin_neon_vgetq_lane_i16: 4485 case NEON::BI__builtin_neon_vduph_laneq_i16: 4486 Ops[0] = Builder.CreateBitCast(Ops[0], 4487 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8)); 4488 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4489 "vgetq_lane"); 4490 case NEON::BI__builtin_neon_vget_lane_i32: 4491 case NEON::BI__builtin_neon_vdups_lane_i32: 4492 Ops[0] = Builder.CreateBitCast( 4493 Ops[0], 4494 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 2)); 4495 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4496 "vget_lane"); 4497 case NEON::BI__builtin_neon_vdups_lane_f32: 4498 Ops[0] = Builder.CreateBitCast(Ops[0], 4499 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4500 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4501 "vdups_lane"); 4502 case NEON::BI__builtin_neon_vgetq_lane_i32: 4503 case NEON::BI__builtin_neon_vdups_laneq_i32: 4504 Ops[0] = Builder.CreateBitCast(Ops[0], 4505 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 32), 4)); 4506 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4507 "vgetq_lane"); 4508 case NEON::BI__builtin_neon_vget_lane_i64: 4509 case NEON::BI__builtin_neon_vdupd_lane_i64: 4510 Ops[0] = Builder.CreateBitCast(Ops[0], 4511 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 1)); 4512 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4513 "vget_lane"); 4514 case NEON::BI__builtin_neon_vdupd_lane_f64: 4515 Ops[0] = Builder.CreateBitCast(Ops[0], 4516 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4517 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4518 "vdupd_lane"); 4519 case NEON::BI__builtin_neon_vgetq_lane_i64: 4520 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4521 Ops[0] = Builder.CreateBitCast(Ops[0], 4522 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 64), 2)); 4523 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4524 "vgetq_lane"); 4525 case NEON::BI__builtin_neon_vget_lane_f32: 4526 Ops[0] = Builder.CreateBitCast(Ops[0], 4527 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4528 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4529 "vget_lane"); 4530 case NEON::BI__builtin_neon_vget_lane_f64: 4531 Ops[0] = Builder.CreateBitCast(Ops[0], 4532 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4533 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4534 "vget_lane"); 4535 case NEON::BI__builtin_neon_vgetq_lane_f32: 4536 case NEON::BI__builtin_neon_vdups_laneq_f32: 4537 Ops[0] = Builder.CreateBitCast(Ops[0], 4538 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4539 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4540 "vgetq_lane"); 4541 case NEON::BI__builtin_neon_vgetq_lane_f64: 4542 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4543 Ops[0] = Builder.CreateBitCast(Ops[0], 4544 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4545 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4546 "vgetq_lane"); 4547 case NEON::BI__builtin_neon_vaddd_s64: 4548 case NEON::BI__builtin_neon_vaddd_u64: 4549 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4550 case NEON::BI__builtin_neon_vsubd_s64: 4551 case NEON::BI__builtin_neon_vsubd_u64: 4552 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4553 case NEON::BI__builtin_neon_vqdmlalh_s16: 4554 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4555 SmallVector<Value *, 2> ProductOps; 4556 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4557 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4558 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4559 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4560 ProductOps, "vqdmlXl"); 4561 Constant *CI = ConstantInt::get(SizeTy, 0); 4562 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4563 4564 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4565 ? Intrinsic::aarch64_neon_sqadd 4566 : Intrinsic::aarch64_neon_sqsub; 4567 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4568 } 4569 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4570 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4571 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4572 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4573 Ops, "vqshlu_n"); 4574 } 4575 case NEON::BI__builtin_neon_vqshld_n_u64: 4576 case NEON::BI__builtin_neon_vqshld_n_s64: { 4577 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4578 ? Intrinsic::aarch64_neon_uqshl 4579 : Intrinsic::aarch64_neon_sqshl; 4580 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4581 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4582 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4583 } 4584 case NEON::BI__builtin_neon_vrshrd_n_u64: 4585 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4586 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4587 ? Intrinsic::aarch64_neon_urshl 4588 : Intrinsic::aarch64_neon_srshl; 4589 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4590 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4591 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4592 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4593 } 4594 case NEON::BI__builtin_neon_vrsrad_n_u64: 4595 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4596 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4597 ? Intrinsic::aarch64_neon_urshl 4598 : Intrinsic::aarch64_neon_srshl; 4599 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4600 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4601 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Int64Ty), Ops[1], 4602 Builder.CreateSExt(Ops[2], Int64Ty)); 4603 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4604 } 4605 case NEON::BI__builtin_neon_vshld_n_s64: 4606 case NEON::BI__builtin_neon_vshld_n_u64: { 4607 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4608 return Builder.CreateShl( 4609 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4610 } 4611 case NEON::BI__builtin_neon_vshrd_n_s64: { 4612 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4613 return Builder.CreateAShr( 4614 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4615 Amt->getZExtValue())), 4616 "shrd_n"); 4617 } 4618 case NEON::BI__builtin_neon_vshrd_n_u64: { 4619 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4620 uint64_t ShiftAmt = Amt->getZExtValue(); 4621 // Right-shifting an unsigned value by its size yields 0. 4622 if (ShiftAmt == 64) 4623 return ConstantInt::get(Int64Ty, 0); 4624 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4625 "shrd_n"); 4626 } 4627 case NEON::BI__builtin_neon_vsrad_n_s64: { 4628 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4629 Ops[1] = Builder.CreateAShr( 4630 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4631 Amt->getZExtValue())), 4632 "shrd_n"); 4633 return Builder.CreateAdd(Ops[0], Ops[1]); 4634 } 4635 case NEON::BI__builtin_neon_vsrad_n_u64: { 4636 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4637 uint64_t ShiftAmt = Amt->getZExtValue(); 4638 // Right-shifting an unsigned value by its size yields 0. 4639 // As Op + 0 = Op, return Ops[0] directly. 4640 if (ShiftAmt == 64) 4641 return Ops[0]; 4642 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4643 "shrd_n"); 4644 return Builder.CreateAdd(Ops[0], Ops[1]); 4645 } 4646 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4647 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4648 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4649 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4650 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4651 "lane"); 4652 SmallVector<Value *, 2> ProductOps; 4653 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4654 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4655 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4656 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4657 ProductOps, "vqdmlXl"); 4658 Constant *CI = ConstantInt::get(SizeTy, 0); 4659 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4660 Ops.pop_back(); 4661 4662 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4663 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4664 ? Intrinsic::aarch64_neon_sqadd 4665 : Intrinsic::aarch64_neon_sqsub; 4666 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4667 } 4668 case NEON::BI__builtin_neon_vqdmlals_s32: 4669 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4670 SmallVector<Value *, 2> ProductOps; 4671 ProductOps.push_back(Ops[1]); 4672 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4673 Ops[1] = 4674 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4675 ProductOps, "vqdmlXl"); 4676 4677 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4678 ? Intrinsic::aarch64_neon_sqadd 4679 : Intrinsic::aarch64_neon_sqsub; 4680 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4681 } 4682 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4683 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4684 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4685 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4686 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4687 "lane"); 4688 SmallVector<Value *, 2> ProductOps; 4689 ProductOps.push_back(Ops[1]); 4690 ProductOps.push_back(Ops[2]); 4691 Ops[1] = 4692 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4693 ProductOps, "vqdmlXl"); 4694 Ops.pop_back(); 4695 4696 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4697 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4698 ? Intrinsic::aarch64_neon_sqadd 4699 : Intrinsic::aarch64_neon_sqsub; 4700 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4701 } 4702 } 4703 4704 llvm::VectorType *VTy = GetNeonType(this, Type); 4705 llvm::Type *Ty = VTy; 4706 if (!Ty) 4707 return nullptr; 4708 4709 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4710 // defer to common code if it's been added to our special map. 4711 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4712 AArch64SIMDIntrinsicsProvenSorted); 4713 4714 if (Builtin) 4715 return EmitCommonNeonBuiltinExpr( 4716 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4717 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4718 4719 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4720 return V; 4721 4722 unsigned Int; 4723 switch (BuiltinID) { 4724 default: return nullptr; 4725 case NEON::BI__builtin_neon_vbsl_v: 4726 case NEON::BI__builtin_neon_vbslq_v: { 4727 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4728 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4729 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4730 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4731 4732 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4733 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4734 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4735 return Builder.CreateBitCast(Ops[0], Ty); 4736 } 4737 case NEON::BI__builtin_neon_vfma_lane_v: 4738 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4739 // The ARM builtins (and instructions) have the addend as the first 4740 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4741 Value *Addend = Ops[0]; 4742 Value *Multiplicand = Ops[1]; 4743 Value *LaneSource = Ops[2]; 4744 Ops[0] = Multiplicand; 4745 Ops[1] = LaneSource; 4746 Ops[2] = Addend; 4747 4748 // Now adjust things to handle the lane access. 4749 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4750 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4751 VTy; 4752 llvm::Constant *cst = cast<Constant>(Ops[3]); 4753 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4754 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4755 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4756 4757 Ops.pop_back(); 4758 Int = Intrinsic::fma; 4759 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4760 } 4761 case NEON::BI__builtin_neon_vfma_laneq_v: { 4762 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4763 // v1f64 fma should be mapped to Neon scalar f64 fma 4764 if (VTy && VTy->getElementType() == DoubleTy) { 4765 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4766 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4767 llvm::Type *VTy = GetNeonType(this, 4768 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4769 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4770 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4771 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4772 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4773 return Builder.CreateBitCast(Result, Ty); 4774 } 4775 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4776 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4777 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4778 4779 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4780 VTy->getNumElements() * 2); 4781 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4782 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4783 cast<ConstantInt>(Ops[3])); 4784 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4785 4786 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4787 } 4788 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4789 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4790 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4791 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4792 4793 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4794 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4795 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 4796 } 4797 case NEON::BI__builtin_neon_vfmas_lane_f32: 4798 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4799 case NEON::BI__builtin_neon_vfmad_lane_f64: 4800 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4801 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4802 llvm::Type *Ty = ConvertType(E->getCallReturnType()); 4803 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4804 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4805 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 4806 } 4807 case NEON::BI__builtin_neon_vfms_v: 4808 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4809 // FIXME: probably remove when we no longer support aarch64_simd.h 4810 // (arm_neon.h delegates to vfma). 4811 4812 // The ARM builtins (and instructions) have the addend as the first 4813 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4814 Value *Subtrahend = Ops[0]; 4815 Value *Multiplicand = Ops[2]; 4816 Ops[0] = Multiplicand; 4817 Ops[2] = Subtrahend; 4818 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 4819 Ops[1] = Builder.CreateFNeg(Ops[1]); 4820 Int = Intrinsic::fma; 4821 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 4822 } 4823 case NEON::BI__builtin_neon_vmull_v: 4824 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4825 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 4826 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 4827 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 4828 case NEON::BI__builtin_neon_vmax_v: 4829 case NEON::BI__builtin_neon_vmaxq_v: 4830 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4831 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 4832 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 4833 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 4834 case NEON::BI__builtin_neon_vmin_v: 4835 case NEON::BI__builtin_neon_vminq_v: 4836 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4837 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 4838 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 4839 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 4840 case NEON::BI__builtin_neon_vabd_v: 4841 case NEON::BI__builtin_neon_vabdq_v: 4842 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4843 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 4844 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 4845 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 4846 case NEON::BI__builtin_neon_vpadal_v: 4847 case NEON::BI__builtin_neon_vpadalq_v: { 4848 unsigned ArgElts = VTy->getNumElements(); 4849 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 4850 unsigned BitWidth = EltTy->getBitWidth(); 4851 llvm::Type *ArgTy = llvm::VectorType::get( 4852 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 4853 llvm::Type* Tys[2] = { VTy, ArgTy }; 4854 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 4855 SmallVector<llvm::Value*, 1> TmpOps; 4856 TmpOps.push_back(Ops[1]); 4857 Function *F = CGM.getIntrinsic(Int, Tys); 4858 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 4859 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 4860 return Builder.CreateAdd(tmp, addend); 4861 } 4862 case NEON::BI__builtin_neon_vpmin_v: 4863 case NEON::BI__builtin_neon_vpminq_v: 4864 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4865 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 4866 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 4867 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 4868 case NEON::BI__builtin_neon_vpmax_v: 4869 case NEON::BI__builtin_neon_vpmaxq_v: 4870 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 4871 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 4872 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 4873 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 4874 case NEON::BI__builtin_neon_vminnm_v: 4875 case NEON::BI__builtin_neon_vminnmq_v: 4876 Int = Intrinsic::aarch64_neon_fminnm; 4877 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 4878 case NEON::BI__builtin_neon_vmaxnm_v: 4879 case NEON::BI__builtin_neon_vmaxnmq_v: 4880 Int = Intrinsic::aarch64_neon_fmaxnm; 4881 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 4882 case NEON::BI__builtin_neon_vrecpss_f32: { 4883 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 4884 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4885 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 4886 Ops, "vrecps"); 4887 } 4888 case NEON::BI__builtin_neon_vrecpsd_f64: { 4889 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 4890 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4891 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 4892 Ops, "vrecps"); 4893 } 4894 case NEON::BI__builtin_neon_vqshrun_n_v: 4895 Int = Intrinsic::aarch64_neon_sqshrun; 4896 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 4897 case NEON::BI__builtin_neon_vqrshrun_n_v: 4898 Int = Intrinsic::aarch64_neon_sqrshrun; 4899 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 4900 case NEON::BI__builtin_neon_vqshrn_n_v: 4901 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 4902 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 4903 case NEON::BI__builtin_neon_vrshrn_n_v: 4904 Int = Intrinsic::aarch64_neon_rshrn; 4905 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 4906 case NEON::BI__builtin_neon_vqrshrn_n_v: 4907 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 4908 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 4909 case NEON::BI__builtin_neon_vrnda_v: 4910 case NEON::BI__builtin_neon_vrndaq_v: { 4911 Int = Intrinsic::round; 4912 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4913 } 4914 case NEON::BI__builtin_neon_vrndi_v: 4915 case NEON::BI__builtin_neon_vrndiq_v: { 4916 Int = Intrinsic::nearbyint; 4917 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4918 } 4919 case NEON::BI__builtin_neon_vrndm_v: 4920 case NEON::BI__builtin_neon_vrndmq_v: { 4921 Int = Intrinsic::floor; 4922 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4923 } 4924 case NEON::BI__builtin_neon_vrndn_v: 4925 case NEON::BI__builtin_neon_vrndnq_v: { 4926 Int = Intrinsic::aarch64_neon_frintn; 4927 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4928 } 4929 case NEON::BI__builtin_neon_vrndp_v: 4930 case NEON::BI__builtin_neon_vrndpq_v: { 4931 Int = Intrinsic::ceil; 4932 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4933 } 4934 case NEON::BI__builtin_neon_vrndx_v: 4935 case NEON::BI__builtin_neon_vrndxq_v: { 4936 Int = Intrinsic::rint; 4937 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4938 } 4939 case NEON::BI__builtin_neon_vrnd_v: 4940 case NEON::BI__builtin_neon_vrndq_v: { 4941 Int = Intrinsic::trunc; 4942 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 4943 } 4944 case NEON::BI__builtin_neon_vceqz_v: 4945 case NEON::BI__builtin_neon_vceqzq_v: 4946 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4947 ICmpInst::ICMP_EQ, "vceqz"); 4948 case NEON::BI__builtin_neon_vcgez_v: 4949 case NEON::BI__builtin_neon_vcgezq_v: 4950 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4951 ICmpInst::ICMP_SGE, "vcgez"); 4952 case NEON::BI__builtin_neon_vclez_v: 4953 case NEON::BI__builtin_neon_vclezq_v: 4954 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4955 ICmpInst::ICMP_SLE, "vclez"); 4956 case NEON::BI__builtin_neon_vcgtz_v: 4957 case NEON::BI__builtin_neon_vcgtzq_v: 4958 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4959 ICmpInst::ICMP_SGT, "vcgtz"); 4960 case NEON::BI__builtin_neon_vcltz_v: 4961 case NEON::BI__builtin_neon_vcltzq_v: 4962 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4963 ICmpInst::ICMP_SLT, "vcltz"); 4964 case NEON::BI__builtin_neon_vcvt_f64_v: 4965 case NEON::BI__builtin_neon_vcvtq_f64_v: 4966 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4967 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4968 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4969 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4970 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4971 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 4972 "unexpected vcvt_f64_f32 builtin"); 4973 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 4974 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4975 4976 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4977 } 4978 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4979 assert(Type.getEltType() == NeonTypeFlags::Float32 && 4980 "unexpected vcvt_f32_f64 builtin"); 4981 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 4982 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 4983 4984 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4985 } 4986 case NEON::BI__builtin_neon_vcvt_s32_v: 4987 case NEON::BI__builtin_neon_vcvt_u32_v: 4988 case NEON::BI__builtin_neon_vcvt_s64_v: 4989 case NEON::BI__builtin_neon_vcvt_u64_v: 4990 case NEON::BI__builtin_neon_vcvtq_s32_v: 4991 case NEON::BI__builtin_neon_vcvtq_u32_v: 4992 case NEON::BI__builtin_neon_vcvtq_s64_v: 4993 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4994 bool Double = 4995 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 4996 llvm::Type *InTy = 4997 GetNeonType(this, 4998 NeonTypeFlags(Double ? NeonTypeFlags::Float64 4999 : NeonTypeFlags::Float32, false, quad)); 5000 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5001 if (usgn) 5002 return Builder.CreateFPToUI(Ops[0], Ty); 5003 return Builder.CreateFPToSI(Ops[0], Ty); 5004 } 5005 case NEON::BI__builtin_neon_vcvta_s32_v: 5006 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5007 case NEON::BI__builtin_neon_vcvta_u32_v: 5008 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5009 case NEON::BI__builtin_neon_vcvta_s64_v: 5010 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5011 case NEON::BI__builtin_neon_vcvta_u64_v: 5012 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5013 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5014 bool Double = 5015 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5016 llvm::Type *InTy = 5017 GetNeonType(this, 5018 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5019 : NeonTypeFlags::Float32, false, quad)); 5020 llvm::Type *Tys[2] = { Ty, InTy }; 5021 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5022 } 5023 case NEON::BI__builtin_neon_vcvtm_s32_v: 5024 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5025 case NEON::BI__builtin_neon_vcvtm_u32_v: 5026 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5027 case NEON::BI__builtin_neon_vcvtm_s64_v: 5028 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5029 case NEON::BI__builtin_neon_vcvtm_u64_v: 5030 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5031 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5032 bool Double = 5033 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5034 llvm::Type *InTy = 5035 GetNeonType(this, 5036 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5037 : NeonTypeFlags::Float32, false, quad)); 5038 llvm::Type *Tys[2] = { Ty, InTy }; 5039 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5040 } 5041 case NEON::BI__builtin_neon_vcvtn_s32_v: 5042 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5043 case NEON::BI__builtin_neon_vcvtn_u32_v: 5044 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5045 case NEON::BI__builtin_neon_vcvtn_s64_v: 5046 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5047 case NEON::BI__builtin_neon_vcvtn_u64_v: 5048 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5049 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5050 bool Double = 5051 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5052 llvm::Type *InTy = 5053 GetNeonType(this, 5054 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5055 : NeonTypeFlags::Float32, false, quad)); 5056 llvm::Type *Tys[2] = { Ty, InTy }; 5057 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5058 } 5059 case NEON::BI__builtin_neon_vcvtp_s32_v: 5060 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5061 case NEON::BI__builtin_neon_vcvtp_u32_v: 5062 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5063 case NEON::BI__builtin_neon_vcvtp_s64_v: 5064 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5065 case NEON::BI__builtin_neon_vcvtp_u64_v: 5066 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5067 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5068 bool Double = 5069 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5070 llvm::Type *InTy = 5071 GetNeonType(this, 5072 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5073 : NeonTypeFlags::Float32, false, quad)); 5074 llvm::Type *Tys[2] = { Ty, InTy }; 5075 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5076 } 5077 case NEON::BI__builtin_neon_vmulx_v: 5078 case NEON::BI__builtin_neon_vmulxq_v: { 5079 Int = Intrinsic::aarch64_neon_fmulx; 5080 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5081 } 5082 case NEON::BI__builtin_neon_vmul_lane_v: 5083 case NEON::BI__builtin_neon_vmul_laneq_v: { 5084 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5085 bool Quad = false; 5086 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5087 Quad = true; 5088 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5089 llvm::Type *VTy = GetNeonType(this, 5090 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5091 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5092 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5093 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5094 return Builder.CreateBitCast(Result, Ty); 5095 } 5096 case NEON::BI__builtin_neon_vnegd_s64: 5097 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5098 case NEON::BI__builtin_neon_vpmaxnm_v: 5099 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5100 Int = Intrinsic::aarch64_neon_fmaxnmp; 5101 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5102 } 5103 case NEON::BI__builtin_neon_vpminnm_v: 5104 case NEON::BI__builtin_neon_vpminnmq_v: { 5105 Int = Intrinsic::aarch64_neon_fminnmp; 5106 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5107 } 5108 case NEON::BI__builtin_neon_vsqrt_v: 5109 case NEON::BI__builtin_neon_vsqrtq_v: { 5110 Int = Intrinsic::sqrt; 5111 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5112 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5113 } 5114 case NEON::BI__builtin_neon_vrbit_v: 5115 case NEON::BI__builtin_neon_vrbitq_v: { 5116 Int = Intrinsic::aarch64_neon_rbit; 5117 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5118 } 5119 case NEON::BI__builtin_neon_vaddv_u8: 5120 // FIXME: These are handled by the AArch64 scalar code. 5121 usgn = true; 5122 // FALLTHROUGH 5123 case NEON::BI__builtin_neon_vaddv_s8: { 5124 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5125 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5126 VTy = 5127 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5128 llvm::Type *Tys[2] = { Ty, VTy }; 5129 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5130 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5131 return Builder.CreateTrunc(Ops[0], 5132 llvm::IntegerType::get(getLLVMContext(), 8)); 5133 } 5134 case NEON::BI__builtin_neon_vaddv_u16: 5135 usgn = true; 5136 // FALLTHROUGH 5137 case NEON::BI__builtin_neon_vaddv_s16: { 5138 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5139 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5140 VTy = 5141 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5142 llvm::Type *Tys[2] = { Ty, VTy }; 5143 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5144 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5145 return Builder.CreateTrunc(Ops[0], 5146 llvm::IntegerType::get(getLLVMContext(), 16)); 5147 } 5148 case NEON::BI__builtin_neon_vaddvq_u8: 5149 usgn = true; 5150 // FALLTHROUGH 5151 case NEON::BI__builtin_neon_vaddvq_s8: { 5152 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5153 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5154 VTy = 5155 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5156 llvm::Type *Tys[2] = { Ty, VTy }; 5157 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5158 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5159 return Builder.CreateTrunc(Ops[0], 5160 llvm::IntegerType::get(getLLVMContext(), 8)); 5161 } 5162 case NEON::BI__builtin_neon_vaddvq_u16: 5163 usgn = true; 5164 // FALLTHROUGH 5165 case NEON::BI__builtin_neon_vaddvq_s16: { 5166 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5167 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5168 VTy = 5169 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5170 llvm::Type *Tys[2] = { Ty, VTy }; 5171 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5172 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5173 return Builder.CreateTrunc(Ops[0], 5174 llvm::IntegerType::get(getLLVMContext(), 16)); 5175 } 5176 case NEON::BI__builtin_neon_vmaxv_u8: { 5177 Int = Intrinsic::aarch64_neon_umaxv; 5178 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5179 VTy = 5180 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5181 llvm::Type *Tys[2] = { Ty, VTy }; 5182 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5183 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5184 return Builder.CreateTrunc(Ops[0], 5185 llvm::IntegerType::get(getLLVMContext(), 8)); 5186 } 5187 case NEON::BI__builtin_neon_vmaxv_u16: { 5188 Int = Intrinsic::aarch64_neon_umaxv; 5189 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5190 VTy = 5191 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5192 llvm::Type *Tys[2] = { Ty, VTy }; 5193 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5194 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5195 return Builder.CreateTrunc(Ops[0], 5196 llvm::IntegerType::get(getLLVMContext(), 16)); 5197 } 5198 case NEON::BI__builtin_neon_vmaxvq_u8: { 5199 Int = Intrinsic::aarch64_neon_umaxv; 5200 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5201 VTy = 5202 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5203 llvm::Type *Tys[2] = { Ty, VTy }; 5204 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5205 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5206 return Builder.CreateTrunc(Ops[0], 5207 llvm::IntegerType::get(getLLVMContext(), 8)); 5208 } 5209 case NEON::BI__builtin_neon_vmaxvq_u16: { 5210 Int = Intrinsic::aarch64_neon_umaxv; 5211 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5212 VTy = 5213 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5214 llvm::Type *Tys[2] = { Ty, VTy }; 5215 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5216 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5217 return Builder.CreateTrunc(Ops[0], 5218 llvm::IntegerType::get(getLLVMContext(), 16)); 5219 } 5220 case NEON::BI__builtin_neon_vmaxv_s8: { 5221 Int = Intrinsic::aarch64_neon_smaxv; 5222 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5223 VTy = 5224 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5225 llvm::Type *Tys[2] = { Ty, VTy }; 5226 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5227 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5228 return Builder.CreateTrunc(Ops[0], 5229 llvm::IntegerType::get(getLLVMContext(), 8)); 5230 } 5231 case NEON::BI__builtin_neon_vmaxv_s16: { 5232 Int = Intrinsic::aarch64_neon_smaxv; 5233 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5234 VTy = 5235 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5236 llvm::Type *Tys[2] = { Ty, VTy }; 5237 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5238 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5239 return Builder.CreateTrunc(Ops[0], 5240 llvm::IntegerType::get(getLLVMContext(), 16)); 5241 } 5242 case NEON::BI__builtin_neon_vmaxvq_s8: { 5243 Int = Intrinsic::aarch64_neon_smaxv; 5244 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5245 VTy = 5246 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5247 llvm::Type *Tys[2] = { Ty, VTy }; 5248 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5249 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5250 return Builder.CreateTrunc(Ops[0], 5251 llvm::IntegerType::get(getLLVMContext(), 8)); 5252 } 5253 case NEON::BI__builtin_neon_vmaxvq_s16: { 5254 Int = Intrinsic::aarch64_neon_smaxv; 5255 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5256 VTy = 5257 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5258 llvm::Type *Tys[2] = { Ty, VTy }; 5259 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5260 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5261 return Builder.CreateTrunc(Ops[0], 5262 llvm::IntegerType::get(getLLVMContext(), 16)); 5263 } 5264 case NEON::BI__builtin_neon_vminv_u8: { 5265 Int = Intrinsic::aarch64_neon_uminv; 5266 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5267 VTy = 5268 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5269 llvm::Type *Tys[2] = { Ty, VTy }; 5270 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5271 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5272 return Builder.CreateTrunc(Ops[0], 5273 llvm::IntegerType::get(getLLVMContext(), 8)); 5274 } 5275 case NEON::BI__builtin_neon_vminv_u16: { 5276 Int = Intrinsic::aarch64_neon_uminv; 5277 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5278 VTy = 5279 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5280 llvm::Type *Tys[2] = { Ty, VTy }; 5281 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5282 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5283 return Builder.CreateTrunc(Ops[0], 5284 llvm::IntegerType::get(getLLVMContext(), 16)); 5285 } 5286 case NEON::BI__builtin_neon_vminvq_u8: { 5287 Int = Intrinsic::aarch64_neon_uminv; 5288 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5289 VTy = 5290 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5291 llvm::Type *Tys[2] = { Ty, VTy }; 5292 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5293 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5294 return Builder.CreateTrunc(Ops[0], 5295 llvm::IntegerType::get(getLLVMContext(), 8)); 5296 } 5297 case NEON::BI__builtin_neon_vminvq_u16: { 5298 Int = Intrinsic::aarch64_neon_uminv; 5299 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5300 VTy = 5301 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5302 llvm::Type *Tys[2] = { Ty, VTy }; 5303 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5304 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5305 return Builder.CreateTrunc(Ops[0], 5306 llvm::IntegerType::get(getLLVMContext(), 16)); 5307 } 5308 case NEON::BI__builtin_neon_vminv_s8: { 5309 Int = Intrinsic::aarch64_neon_sminv; 5310 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5311 VTy = 5312 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5313 llvm::Type *Tys[2] = { Ty, VTy }; 5314 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5315 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5316 return Builder.CreateTrunc(Ops[0], 5317 llvm::IntegerType::get(getLLVMContext(), 8)); 5318 } 5319 case NEON::BI__builtin_neon_vminv_s16: { 5320 Int = Intrinsic::aarch64_neon_sminv; 5321 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5322 VTy = 5323 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5324 llvm::Type *Tys[2] = { Ty, VTy }; 5325 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5326 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5327 return Builder.CreateTrunc(Ops[0], 5328 llvm::IntegerType::get(getLLVMContext(), 16)); 5329 } 5330 case NEON::BI__builtin_neon_vminvq_s8: { 5331 Int = Intrinsic::aarch64_neon_sminv; 5332 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5333 VTy = 5334 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5335 llvm::Type *Tys[2] = { Ty, VTy }; 5336 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5337 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5338 return Builder.CreateTrunc(Ops[0], 5339 llvm::IntegerType::get(getLLVMContext(), 8)); 5340 } 5341 case NEON::BI__builtin_neon_vminvq_s16: { 5342 Int = Intrinsic::aarch64_neon_sminv; 5343 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5344 VTy = 5345 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5346 llvm::Type *Tys[2] = { Ty, VTy }; 5347 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5348 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5349 return Builder.CreateTrunc(Ops[0], 5350 llvm::IntegerType::get(getLLVMContext(), 16)); 5351 } 5352 case NEON::BI__builtin_neon_vmul_n_f64: { 5353 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5354 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5355 return Builder.CreateFMul(Ops[0], RHS); 5356 } 5357 case NEON::BI__builtin_neon_vaddlv_u8: { 5358 Int = Intrinsic::aarch64_neon_uaddlv; 5359 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5360 VTy = 5361 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5362 llvm::Type *Tys[2] = { Ty, VTy }; 5363 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5364 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5365 return Builder.CreateTrunc(Ops[0], 5366 llvm::IntegerType::get(getLLVMContext(), 16)); 5367 } 5368 case NEON::BI__builtin_neon_vaddlv_u16: { 5369 Int = Intrinsic::aarch64_neon_uaddlv; 5370 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5371 VTy = 5372 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5373 llvm::Type *Tys[2] = { Ty, VTy }; 5374 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5375 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5376 } 5377 case NEON::BI__builtin_neon_vaddlvq_u8: { 5378 Int = Intrinsic::aarch64_neon_uaddlv; 5379 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5380 VTy = 5381 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5382 llvm::Type *Tys[2] = { Ty, VTy }; 5383 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5384 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5385 return Builder.CreateTrunc(Ops[0], 5386 llvm::IntegerType::get(getLLVMContext(), 16)); 5387 } 5388 case NEON::BI__builtin_neon_vaddlvq_u16: { 5389 Int = Intrinsic::aarch64_neon_uaddlv; 5390 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5391 VTy = 5392 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5393 llvm::Type *Tys[2] = { Ty, VTy }; 5394 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5395 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5396 } 5397 case NEON::BI__builtin_neon_vaddlv_s8: { 5398 Int = Intrinsic::aarch64_neon_saddlv; 5399 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5400 VTy = 5401 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 8); 5402 llvm::Type *Tys[2] = { Ty, VTy }; 5403 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5404 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5405 return Builder.CreateTrunc(Ops[0], 5406 llvm::IntegerType::get(getLLVMContext(), 16)); 5407 } 5408 case NEON::BI__builtin_neon_vaddlv_s16: { 5409 Int = Intrinsic::aarch64_neon_saddlv; 5410 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5411 VTy = 5412 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 4); 5413 llvm::Type *Tys[2] = { Ty, VTy }; 5414 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5415 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5416 } 5417 case NEON::BI__builtin_neon_vaddlvq_s8: { 5418 Int = Intrinsic::aarch64_neon_saddlv; 5419 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5420 VTy = 5421 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 8), 16); 5422 llvm::Type *Tys[2] = { Ty, VTy }; 5423 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5424 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5425 return Builder.CreateTrunc(Ops[0], 5426 llvm::IntegerType::get(getLLVMContext(), 16)); 5427 } 5428 case NEON::BI__builtin_neon_vaddlvq_s16: { 5429 Int = Intrinsic::aarch64_neon_saddlv; 5430 Ty = llvm::IntegerType::get(getLLVMContext(), 32); 5431 VTy = 5432 llvm::VectorType::get(llvm::IntegerType::get(getLLVMContext(), 16), 8); 5433 llvm::Type *Tys[2] = { Ty, VTy }; 5434 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5435 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5436 } 5437 case NEON::BI__builtin_neon_vsri_n_v: 5438 case NEON::BI__builtin_neon_vsriq_n_v: { 5439 Int = Intrinsic::aarch64_neon_vsri; 5440 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5441 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5442 } 5443 case NEON::BI__builtin_neon_vsli_n_v: 5444 case NEON::BI__builtin_neon_vsliq_n_v: { 5445 Int = Intrinsic::aarch64_neon_vsli; 5446 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5447 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5448 } 5449 case NEON::BI__builtin_neon_vsra_n_v: 5450 case NEON::BI__builtin_neon_vsraq_n_v: 5451 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5452 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5453 return Builder.CreateAdd(Ops[0], Ops[1]); 5454 case NEON::BI__builtin_neon_vrsra_n_v: 5455 case NEON::BI__builtin_neon_vrsraq_n_v: { 5456 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5457 SmallVector<llvm::Value*,2> TmpOps; 5458 TmpOps.push_back(Ops[1]); 5459 TmpOps.push_back(Ops[2]); 5460 Function* F = CGM.getIntrinsic(Int, Ty); 5461 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5462 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5463 return Builder.CreateAdd(Ops[0], tmp); 5464 } 5465 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5466 // of an Align parameter here. 5467 case NEON::BI__builtin_neon_vld1_x2_v: 5468 case NEON::BI__builtin_neon_vld1q_x2_v: 5469 case NEON::BI__builtin_neon_vld1_x3_v: 5470 case NEON::BI__builtin_neon_vld1q_x3_v: 5471 case NEON::BI__builtin_neon_vld1_x4_v: 5472 case NEON::BI__builtin_neon_vld1q_x4_v: { 5473 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5474 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5475 llvm::Type *Tys[2] = { VTy, PTy }; 5476 unsigned Int; 5477 switch (BuiltinID) { 5478 case NEON::BI__builtin_neon_vld1_x2_v: 5479 case NEON::BI__builtin_neon_vld1q_x2_v: 5480 Int = Intrinsic::aarch64_neon_ld1x2; 5481 break; 5482 case NEON::BI__builtin_neon_vld1_x3_v: 5483 case NEON::BI__builtin_neon_vld1q_x3_v: 5484 Int = Intrinsic::aarch64_neon_ld1x3; 5485 break; 5486 case NEON::BI__builtin_neon_vld1_x4_v: 5487 case NEON::BI__builtin_neon_vld1q_x4_v: 5488 Int = Intrinsic::aarch64_neon_ld1x4; 5489 break; 5490 } 5491 Function *F = CGM.getIntrinsic(Int, Tys); 5492 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5493 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5494 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5495 return Builder.CreateStore(Ops[1], Ops[0]); 5496 } 5497 case NEON::BI__builtin_neon_vst1_x2_v: 5498 case NEON::BI__builtin_neon_vst1q_x2_v: 5499 case NEON::BI__builtin_neon_vst1_x3_v: 5500 case NEON::BI__builtin_neon_vst1q_x3_v: 5501 case NEON::BI__builtin_neon_vst1_x4_v: 5502 case NEON::BI__builtin_neon_vst1q_x4_v: { 5503 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5504 llvm::Type *Tys[2] = { VTy, PTy }; 5505 unsigned Int; 5506 switch (BuiltinID) { 5507 case NEON::BI__builtin_neon_vst1_x2_v: 5508 case NEON::BI__builtin_neon_vst1q_x2_v: 5509 Int = Intrinsic::aarch64_neon_st1x2; 5510 break; 5511 case NEON::BI__builtin_neon_vst1_x3_v: 5512 case NEON::BI__builtin_neon_vst1q_x3_v: 5513 Int = Intrinsic::aarch64_neon_st1x3; 5514 break; 5515 case NEON::BI__builtin_neon_vst1_x4_v: 5516 case NEON::BI__builtin_neon_vst1q_x4_v: 5517 Int = Intrinsic::aarch64_neon_st1x4; 5518 break; 5519 } 5520 SmallVector<Value *, 4> IntOps(Ops.begin()+1, Ops.end()); 5521 IntOps.push_back(Ops[0]); 5522 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), IntOps, ""); 5523 } 5524 case NEON::BI__builtin_neon_vld1_v: 5525 case NEON::BI__builtin_neon_vld1q_v: 5526 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5527 return Builder.CreateLoad(Ops[0]); 5528 case NEON::BI__builtin_neon_vst1_v: 5529 case NEON::BI__builtin_neon_vst1q_v: 5530 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5531 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5532 return Builder.CreateStore(Ops[1], Ops[0]); 5533 case NEON::BI__builtin_neon_vld1_lane_v: 5534 case NEON::BI__builtin_neon_vld1q_lane_v: 5535 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5536 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5537 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5538 Ops[0] = Builder.CreateLoad(Ops[0]); 5539 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5540 case NEON::BI__builtin_neon_vld1_dup_v: 5541 case NEON::BI__builtin_neon_vld1q_dup_v: { 5542 Value *V = UndefValue::get(Ty); 5543 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5544 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5545 Ops[0] = Builder.CreateLoad(Ops[0]); 5546 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5547 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5548 return EmitNeonSplat(Ops[0], CI); 5549 } 5550 case NEON::BI__builtin_neon_vst1_lane_v: 5551 case NEON::BI__builtin_neon_vst1q_lane_v: 5552 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5553 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5554 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5555 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5556 case NEON::BI__builtin_neon_vld2_v: 5557 case NEON::BI__builtin_neon_vld2q_v: { 5558 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5559 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5560 llvm::Type *Tys[2] = { VTy, PTy }; 5561 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5562 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5563 Ops[0] = Builder.CreateBitCast(Ops[0], 5564 llvm::PointerType::getUnqual(Ops[1]->getType())); 5565 return Builder.CreateStore(Ops[1], Ops[0]); 5566 } 5567 case NEON::BI__builtin_neon_vld3_v: 5568 case NEON::BI__builtin_neon_vld3q_v: { 5569 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5570 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5571 llvm::Type *Tys[2] = { VTy, PTy }; 5572 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5573 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5574 Ops[0] = Builder.CreateBitCast(Ops[0], 5575 llvm::PointerType::getUnqual(Ops[1]->getType())); 5576 return Builder.CreateStore(Ops[1], Ops[0]); 5577 } 5578 case NEON::BI__builtin_neon_vld4_v: 5579 case NEON::BI__builtin_neon_vld4q_v: { 5580 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5581 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5582 llvm::Type *Tys[2] = { VTy, PTy }; 5583 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5584 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5585 Ops[0] = Builder.CreateBitCast(Ops[0], 5586 llvm::PointerType::getUnqual(Ops[1]->getType())); 5587 return Builder.CreateStore(Ops[1], Ops[0]); 5588 } 5589 case NEON::BI__builtin_neon_vld2_dup_v: 5590 case NEON::BI__builtin_neon_vld2q_dup_v: { 5591 llvm::Type *PTy = 5592 llvm::PointerType::getUnqual(VTy->getElementType()); 5593 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5594 llvm::Type *Tys[2] = { VTy, PTy }; 5595 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5596 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5597 Ops[0] = Builder.CreateBitCast(Ops[0], 5598 llvm::PointerType::getUnqual(Ops[1]->getType())); 5599 return Builder.CreateStore(Ops[1], Ops[0]); 5600 } 5601 case NEON::BI__builtin_neon_vld3_dup_v: 5602 case NEON::BI__builtin_neon_vld3q_dup_v: { 5603 llvm::Type *PTy = 5604 llvm::PointerType::getUnqual(VTy->getElementType()); 5605 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5606 llvm::Type *Tys[2] = { VTy, PTy }; 5607 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5608 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5609 Ops[0] = Builder.CreateBitCast(Ops[0], 5610 llvm::PointerType::getUnqual(Ops[1]->getType())); 5611 return Builder.CreateStore(Ops[1], Ops[0]); 5612 } 5613 case NEON::BI__builtin_neon_vld4_dup_v: 5614 case NEON::BI__builtin_neon_vld4q_dup_v: { 5615 llvm::Type *PTy = 5616 llvm::PointerType::getUnqual(VTy->getElementType()); 5617 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5618 llvm::Type *Tys[2] = { VTy, PTy }; 5619 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5620 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5621 Ops[0] = Builder.CreateBitCast(Ops[0], 5622 llvm::PointerType::getUnqual(Ops[1]->getType())); 5623 return Builder.CreateStore(Ops[1], Ops[0]); 5624 } 5625 case NEON::BI__builtin_neon_vld2_lane_v: 5626 case NEON::BI__builtin_neon_vld2q_lane_v: { 5627 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5628 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5629 Ops.push_back(Ops[1]); 5630 Ops.erase(Ops.begin()+1); 5631 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5632 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5633 Ops[3] = Builder.CreateZExt(Ops[3], 5634 llvm::IntegerType::get(getLLVMContext(), 64)); 5635 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5636 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5637 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5638 return Builder.CreateStore(Ops[1], Ops[0]); 5639 } 5640 case NEON::BI__builtin_neon_vld3_lane_v: 5641 case NEON::BI__builtin_neon_vld3q_lane_v: { 5642 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5643 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5644 Ops.push_back(Ops[1]); 5645 Ops.erase(Ops.begin()+1); 5646 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5647 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5648 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5649 Ops[4] = Builder.CreateZExt(Ops[4], 5650 llvm::IntegerType::get(getLLVMContext(), 64)); 5651 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5652 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5653 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5654 return Builder.CreateStore(Ops[1], Ops[0]); 5655 } 5656 case NEON::BI__builtin_neon_vld4_lane_v: 5657 case NEON::BI__builtin_neon_vld4q_lane_v: { 5658 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5659 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5660 Ops.push_back(Ops[1]); 5661 Ops.erase(Ops.begin()+1); 5662 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5663 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5664 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5665 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5666 Ops[5] = Builder.CreateZExt(Ops[5], 5667 llvm::IntegerType::get(getLLVMContext(), 64)); 5668 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5669 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5670 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5671 return Builder.CreateStore(Ops[1], Ops[0]); 5672 } 5673 case NEON::BI__builtin_neon_vst2_v: 5674 case NEON::BI__builtin_neon_vst2q_v: { 5675 Ops.push_back(Ops[0]); 5676 Ops.erase(Ops.begin()); 5677 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5678 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5679 Ops, ""); 5680 } 5681 case NEON::BI__builtin_neon_vst2_lane_v: 5682 case NEON::BI__builtin_neon_vst2q_lane_v: { 5683 Ops.push_back(Ops[0]); 5684 Ops.erase(Ops.begin()); 5685 Ops[2] = Builder.CreateZExt(Ops[2], 5686 llvm::IntegerType::get(getLLVMContext(), 64)); 5687 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5688 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5689 Ops, ""); 5690 } 5691 case NEON::BI__builtin_neon_vst3_v: 5692 case NEON::BI__builtin_neon_vst3q_v: { 5693 Ops.push_back(Ops[0]); 5694 Ops.erase(Ops.begin()); 5695 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5696 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5697 Ops, ""); 5698 } 5699 case NEON::BI__builtin_neon_vst3_lane_v: 5700 case NEON::BI__builtin_neon_vst3q_lane_v: { 5701 Ops.push_back(Ops[0]); 5702 Ops.erase(Ops.begin()); 5703 Ops[3] = Builder.CreateZExt(Ops[3], 5704 llvm::IntegerType::get(getLLVMContext(), 64)); 5705 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5706 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5707 Ops, ""); 5708 } 5709 case NEON::BI__builtin_neon_vst4_v: 5710 case NEON::BI__builtin_neon_vst4q_v: { 5711 Ops.push_back(Ops[0]); 5712 Ops.erase(Ops.begin()); 5713 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5714 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5715 Ops, ""); 5716 } 5717 case NEON::BI__builtin_neon_vst4_lane_v: 5718 case NEON::BI__builtin_neon_vst4q_lane_v: { 5719 Ops.push_back(Ops[0]); 5720 Ops.erase(Ops.begin()); 5721 Ops[4] = Builder.CreateZExt(Ops[4], 5722 llvm::IntegerType::get(getLLVMContext(), 64)); 5723 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5724 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5725 Ops, ""); 5726 } 5727 case NEON::BI__builtin_neon_vtrn_v: 5728 case NEON::BI__builtin_neon_vtrnq_v: { 5729 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5730 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5731 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5732 Value *SV = nullptr; 5733 5734 for (unsigned vi = 0; vi != 2; ++vi) { 5735 SmallVector<Constant*, 16> Indices; 5736 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5737 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5738 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5739 } 5740 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5741 SV = llvm::ConstantVector::get(Indices); 5742 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5743 SV = Builder.CreateStore(SV, Addr); 5744 } 5745 return SV; 5746 } 5747 case NEON::BI__builtin_neon_vuzp_v: 5748 case NEON::BI__builtin_neon_vuzpq_v: { 5749 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5750 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5751 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5752 Value *SV = nullptr; 5753 5754 for (unsigned vi = 0; vi != 2; ++vi) { 5755 SmallVector<Constant*, 16> Indices; 5756 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5757 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5758 5759 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5760 SV = llvm::ConstantVector::get(Indices); 5761 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5762 SV = Builder.CreateStore(SV, Addr); 5763 } 5764 return SV; 5765 } 5766 case NEON::BI__builtin_neon_vzip_v: 5767 case NEON::BI__builtin_neon_vzipq_v: { 5768 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5769 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5770 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5771 Value *SV = nullptr; 5772 5773 for (unsigned vi = 0; vi != 2; ++vi) { 5774 SmallVector<Constant*, 16> Indices; 5775 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5776 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5777 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5778 } 5779 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 5780 SV = llvm::ConstantVector::get(Indices); 5781 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5782 SV = Builder.CreateStore(SV, Addr); 5783 } 5784 return SV; 5785 } 5786 case NEON::BI__builtin_neon_vqtbl1q_v: { 5787 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5788 Ops, "vtbl1"); 5789 } 5790 case NEON::BI__builtin_neon_vqtbl2q_v: { 5791 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5792 Ops, "vtbl2"); 5793 } 5794 case NEON::BI__builtin_neon_vqtbl3q_v: { 5795 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5796 Ops, "vtbl3"); 5797 } 5798 case NEON::BI__builtin_neon_vqtbl4q_v: { 5799 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5800 Ops, "vtbl4"); 5801 } 5802 case NEON::BI__builtin_neon_vqtbx1q_v: { 5803 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5804 Ops, "vtbx1"); 5805 } 5806 case NEON::BI__builtin_neon_vqtbx2q_v: { 5807 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5808 Ops, "vtbx2"); 5809 } 5810 case NEON::BI__builtin_neon_vqtbx3q_v: { 5811 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5812 Ops, "vtbx3"); 5813 } 5814 case NEON::BI__builtin_neon_vqtbx4q_v: { 5815 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5816 Ops, "vtbx4"); 5817 } 5818 case NEON::BI__builtin_neon_vsqadd_v: 5819 case NEON::BI__builtin_neon_vsqaddq_v: { 5820 Int = Intrinsic::aarch64_neon_usqadd; 5821 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5822 } 5823 case NEON::BI__builtin_neon_vuqadd_v: 5824 case NEON::BI__builtin_neon_vuqaddq_v: { 5825 Int = Intrinsic::aarch64_neon_suqadd; 5826 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5827 } 5828 } 5829 } 5830 5831 llvm::Value *CodeGenFunction:: 5832 BuildVector(ArrayRef<llvm::Value*> Ops) { 5833 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5834 "Not a power-of-two sized vector!"); 5835 bool AllConstants = true; 5836 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5837 AllConstants &= isa<Constant>(Ops[i]); 5838 5839 // If this is a constant vector, create a ConstantVector. 5840 if (AllConstants) { 5841 SmallVector<llvm::Constant*, 16> CstOps; 5842 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5843 CstOps.push_back(cast<Constant>(Ops[i])); 5844 return llvm::ConstantVector::get(CstOps); 5845 } 5846 5847 // Otherwise, insertelement the values to build the vector. 5848 Value *Result = 5849 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5850 5851 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5852 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5853 5854 return Result; 5855 } 5856 5857 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5858 const CallExpr *E) { 5859 SmallVector<Value*, 4> Ops; 5860 5861 // Find out if any arguments are required to be integer constant expressions. 5862 unsigned ICEArguments = 0; 5863 ASTContext::GetBuiltinTypeError Error; 5864 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5865 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5866 5867 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5868 // If this is a normal argument, just emit it as a scalar. 5869 if ((ICEArguments & (1 << i)) == 0) { 5870 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5871 continue; 5872 } 5873 5874 // If this is required to be a constant, constant fold it so that we know 5875 // that the generated intrinsic gets a ConstantInt. 5876 llvm::APSInt Result; 5877 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 5878 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 5879 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 5880 } 5881 5882 switch (BuiltinID) { 5883 default: return nullptr; 5884 case X86::BI_mm_prefetch: { 5885 Value *Address = EmitScalarExpr(E->getArg(0)); 5886 Value *RW = ConstantInt::get(Int32Ty, 0); 5887 Value *Locality = EmitScalarExpr(E->getArg(1)); 5888 Value *Data = ConstantInt::get(Int32Ty, 1); 5889 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 5890 return Builder.CreateCall4(F, Address, RW, Locality, Data); 5891 } 5892 case X86::BI__builtin_ia32_vec_init_v8qi: 5893 case X86::BI__builtin_ia32_vec_init_v4hi: 5894 case X86::BI__builtin_ia32_vec_init_v2si: 5895 return Builder.CreateBitCast(BuildVector(Ops), 5896 llvm::Type::getX86_MMXTy(getLLVMContext())); 5897 case X86::BI__builtin_ia32_vec_ext_v2si: 5898 return Builder.CreateExtractElement(Ops[0], 5899 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 5900 case X86::BI__builtin_ia32_ldmxcsr: { 5901 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 5902 Builder.CreateStore(Ops[0], Tmp); 5903 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 5904 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5905 } 5906 case X86::BI__builtin_ia32_stmxcsr: { 5907 Value *Tmp = CreateMemTemp(E->getType()); 5908 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 5909 Builder.CreateBitCast(Tmp, Int8PtrTy)); 5910 return Builder.CreateLoad(Tmp, "stmxcsr"); 5911 } 5912 case X86::BI__builtin_ia32_storehps: 5913 case X86::BI__builtin_ia32_storelps: { 5914 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 5915 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 5916 5917 // cast val v2i64 5918 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 5919 5920 // extract (0, 1) 5921 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 5922 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 5923 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 5924 5925 // cast pointer to i64 & store 5926 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 5927 return Builder.CreateStore(Ops[1], Ops[0]); 5928 } 5929 case X86::BI__builtin_ia32_palignr128: 5930 case X86::BI__builtin_ia32_palignr256: { 5931 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 5932 5933 unsigned NumElts = 5934 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 5935 assert(NumElts % 16 == 0); 5936 unsigned NumLanes = NumElts / 16; 5937 unsigned NumLaneElts = NumElts / NumLanes; 5938 5939 // If palignr is shifting the pair of vectors more than the size of two 5940 // lanes, emit zero. 5941 if (ShiftVal >= (2 * NumLaneElts)) 5942 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5943 5944 // If palignr is shifting the pair of input vectors more than one lane, 5945 // but less than two lanes, convert to shifting in zeroes. 5946 if (ShiftVal > NumLaneElts) { 5947 ShiftVal -= NumLaneElts; 5948 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 5949 } 5950 5951 SmallVector<llvm::Constant*, 32> Indices; 5952 // 256-bit palignr operates on 128-bit lanes so we need to handle that 5953 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 5954 for (unsigned i = 0; i != NumLaneElts; ++i) { 5955 unsigned Idx = ShiftVal + i; 5956 if (Idx >= NumLaneElts) 5957 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 5958 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 5959 } 5960 } 5961 5962 Value* SV = llvm::ConstantVector::get(Indices); 5963 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 5964 } 5965 case X86::BI__builtin_ia32_pslldqi256: { 5966 // Shift value is in bits so divide by 8. 5967 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 5968 5969 // If pslldq is shifting the vector more than 15 bytes, emit zero. 5970 if (shiftVal >= 16) 5971 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5972 5973 SmallVector<llvm::Constant*, 32> Indices; 5974 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 5975 for (unsigned l = 0; l != 32; l += 16) { 5976 for (unsigned i = 0; i != 16; ++i) { 5977 unsigned Idx = 32 + i - shiftVal; 5978 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 5979 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 5980 } 5981 } 5982 5983 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 5984 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 5985 Value *Zero = llvm::Constant::getNullValue(VecTy); 5986 5987 Value *SV = llvm::ConstantVector::get(Indices); 5988 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 5989 llvm::Type *ResultType = ConvertType(E->getType()); 5990 return Builder.CreateBitCast(SV, ResultType, "cast"); 5991 } 5992 case X86::BI__builtin_ia32_psrldqi256: { 5993 // Shift value is in bits so divide by 8. 5994 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 5995 5996 // If psrldq is shifting the vector more than 15 bytes, emit zero. 5997 if (shiftVal >= 16) 5998 return llvm::Constant::getNullValue(ConvertType(E->getType())); 5999 6000 SmallVector<llvm::Constant*, 32> Indices; 6001 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6002 for (unsigned l = 0; l != 32; l += 16) { 6003 for (unsigned i = 0; i != 16; ++i) { 6004 unsigned Idx = i + shiftVal; 6005 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6006 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx + l)); 6007 } 6008 } 6009 6010 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6011 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6012 Value *Zero = llvm::Constant::getNullValue(VecTy); 6013 6014 Value *SV = llvm::ConstantVector::get(Indices); 6015 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6016 llvm::Type *ResultType = ConvertType(E->getType()); 6017 return Builder.CreateBitCast(SV, ResultType, "cast"); 6018 } 6019 case X86::BI__builtin_ia32_movntps: 6020 case X86::BI__builtin_ia32_movntps256: 6021 case X86::BI__builtin_ia32_movntpd: 6022 case X86::BI__builtin_ia32_movntpd256: 6023 case X86::BI__builtin_ia32_movntdq: 6024 case X86::BI__builtin_ia32_movntdq256: 6025 case X86::BI__builtin_ia32_movnti: 6026 case X86::BI__builtin_ia32_movnti64: { 6027 llvm::MDNode *Node = llvm::MDNode::get( 6028 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6029 6030 // Convert the type of the pointer to a pointer to the stored type. 6031 Value *BC = Builder.CreateBitCast(Ops[0], 6032 llvm::PointerType::getUnqual(Ops[1]->getType()), 6033 "cast"); 6034 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6035 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6036 6037 // If the operand is an integer, we can't assume alignment. Otherwise, 6038 // assume natural alignment. 6039 QualType ArgTy = E->getArg(1)->getType(); 6040 unsigned Align; 6041 if (ArgTy->isIntegerType()) 6042 Align = 1; 6043 else 6044 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6045 SI->setAlignment(Align); 6046 return SI; 6047 } 6048 // 3DNow! 6049 case X86::BI__builtin_ia32_pswapdsf: 6050 case X86::BI__builtin_ia32_pswapdsi: { 6051 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6052 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6053 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6054 return Builder.CreateCall(F, Ops, "pswapd"); 6055 } 6056 case X86::BI__builtin_ia32_rdrand16_step: 6057 case X86::BI__builtin_ia32_rdrand32_step: 6058 case X86::BI__builtin_ia32_rdrand64_step: 6059 case X86::BI__builtin_ia32_rdseed16_step: 6060 case X86::BI__builtin_ia32_rdseed32_step: 6061 case X86::BI__builtin_ia32_rdseed64_step: { 6062 Intrinsic::ID ID; 6063 switch (BuiltinID) { 6064 default: llvm_unreachable("Unsupported intrinsic!"); 6065 case X86::BI__builtin_ia32_rdrand16_step: 6066 ID = Intrinsic::x86_rdrand_16; 6067 break; 6068 case X86::BI__builtin_ia32_rdrand32_step: 6069 ID = Intrinsic::x86_rdrand_32; 6070 break; 6071 case X86::BI__builtin_ia32_rdrand64_step: 6072 ID = Intrinsic::x86_rdrand_64; 6073 break; 6074 case X86::BI__builtin_ia32_rdseed16_step: 6075 ID = Intrinsic::x86_rdseed_16; 6076 break; 6077 case X86::BI__builtin_ia32_rdseed32_step: 6078 ID = Intrinsic::x86_rdseed_32; 6079 break; 6080 case X86::BI__builtin_ia32_rdseed64_step: 6081 ID = Intrinsic::x86_rdseed_64; 6082 break; 6083 } 6084 6085 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6086 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6087 return Builder.CreateExtractValue(Call, 1); 6088 } 6089 // AVX2 broadcast 6090 case X86::BI__builtin_ia32_vbroadcastsi256: { 6091 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 6092 Builder.CreateStore(Ops[0], VecTmp); 6093 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 6094 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 6095 } 6096 // SSE comparison intrisics 6097 case X86::BI__builtin_ia32_cmpeqps: 6098 case X86::BI__builtin_ia32_cmpltps: 6099 case X86::BI__builtin_ia32_cmpleps: 6100 case X86::BI__builtin_ia32_cmpunordps: 6101 case X86::BI__builtin_ia32_cmpneqps: 6102 case X86::BI__builtin_ia32_cmpnltps: 6103 case X86::BI__builtin_ia32_cmpnleps: 6104 case X86::BI__builtin_ia32_cmpordps: 6105 case X86::BI__builtin_ia32_cmpeqss: 6106 case X86::BI__builtin_ia32_cmpltss: 6107 case X86::BI__builtin_ia32_cmpless: 6108 case X86::BI__builtin_ia32_cmpunordss: 6109 case X86::BI__builtin_ia32_cmpneqss: 6110 case X86::BI__builtin_ia32_cmpnltss: 6111 case X86::BI__builtin_ia32_cmpnless: 6112 case X86::BI__builtin_ia32_cmpordss: 6113 case X86::BI__builtin_ia32_cmpeqpd: 6114 case X86::BI__builtin_ia32_cmpltpd: 6115 case X86::BI__builtin_ia32_cmplepd: 6116 case X86::BI__builtin_ia32_cmpunordpd: 6117 case X86::BI__builtin_ia32_cmpneqpd: 6118 case X86::BI__builtin_ia32_cmpnltpd: 6119 case X86::BI__builtin_ia32_cmpnlepd: 6120 case X86::BI__builtin_ia32_cmpordpd: 6121 case X86::BI__builtin_ia32_cmpeqsd: 6122 case X86::BI__builtin_ia32_cmpltsd: 6123 case X86::BI__builtin_ia32_cmplesd: 6124 case X86::BI__builtin_ia32_cmpunordsd: 6125 case X86::BI__builtin_ia32_cmpneqsd: 6126 case X86::BI__builtin_ia32_cmpnltsd: 6127 case X86::BI__builtin_ia32_cmpnlesd: 6128 case X86::BI__builtin_ia32_cmpordsd: 6129 // These exist so that the builtin that takes an immediate can be bounds 6130 // checked by clang to avoid passing bad immediates to the backend. Since 6131 // AVX has a larger immediate than SSE we would need separate builtins to 6132 // do the different bounds checking. Rather than create a clang specific 6133 // SSE only builtin, this implements eight separate builtins to match gcc 6134 // implementation. 6135 6136 // Choose the immediate. 6137 unsigned Imm; 6138 switch (BuiltinID) { 6139 default: llvm_unreachable("Unsupported intrinsic!"); 6140 case X86::BI__builtin_ia32_cmpeqps: 6141 case X86::BI__builtin_ia32_cmpeqss: 6142 case X86::BI__builtin_ia32_cmpeqpd: 6143 case X86::BI__builtin_ia32_cmpeqsd: 6144 Imm = 0; 6145 break; 6146 case X86::BI__builtin_ia32_cmpltps: 6147 case X86::BI__builtin_ia32_cmpltss: 6148 case X86::BI__builtin_ia32_cmpltpd: 6149 case X86::BI__builtin_ia32_cmpltsd: 6150 Imm = 1; 6151 break; 6152 case X86::BI__builtin_ia32_cmpleps: 6153 case X86::BI__builtin_ia32_cmpless: 6154 case X86::BI__builtin_ia32_cmplepd: 6155 case X86::BI__builtin_ia32_cmplesd: 6156 Imm = 2; 6157 break; 6158 case X86::BI__builtin_ia32_cmpunordps: 6159 case X86::BI__builtin_ia32_cmpunordss: 6160 case X86::BI__builtin_ia32_cmpunordpd: 6161 case X86::BI__builtin_ia32_cmpunordsd: 6162 Imm = 3; 6163 break; 6164 case X86::BI__builtin_ia32_cmpneqps: 6165 case X86::BI__builtin_ia32_cmpneqss: 6166 case X86::BI__builtin_ia32_cmpneqpd: 6167 case X86::BI__builtin_ia32_cmpneqsd: 6168 Imm = 4; 6169 break; 6170 case X86::BI__builtin_ia32_cmpnltps: 6171 case X86::BI__builtin_ia32_cmpnltss: 6172 case X86::BI__builtin_ia32_cmpnltpd: 6173 case X86::BI__builtin_ia32_cmpnltsd: 6174 Imm = 5; 6175 break; 6176 case X86::BI__builtin_ia32_cmpnleps: 6177 case X86::BI__builtin_ia32_cmpnless: 6178 case X86::BI__builtin_ia32_cmpnlepd: 6179 case X86::BI__builtin_ia32_cmpnlesd: 6180 Imm = 6; 6181 break; 6182 case X86::BI__builtin_ia32_cmpordps: 6183 case X86::BI__builtin_ia32_cmpordss: 6184 case X86::BI__builtin_ia32_cmpordpd: 6185 case X86::BI__builtin_ia32_cmpordsd: 6186 Imm = 7; 6187 break; 6188 } 6189 6190 // Choose the intrinsic ID. 6191 const char *name; 6192 Intrinsic::ID ID; 6193 switch (BuiltinID) { 6194 default: llvm_unreachable("Unsupported intrinsic!"); 6195 case X86::BI__builtin_ia32_cmpeqps: 6196 case X86::BI__builtin_ia32_cmpltps: 6197 case X86::BI__builtin_ia32_cmpleps: 6198 case X86::BI__builtin_ia32_cmpunordps: 6199 case X86::BI__builtin_ia32_cmpneqps: 6200 case X86::BI__builtin_ia32_cmpnltps: 6201 case X86::BI__builtin_ia32_cmpnleps: 6202 case X86::BI__builtin_ia32_cmpordps: 6203 name = "cmpps"; 6204 ID = Intrinsic::x86_sse_cmp_ps; 6205 break; 6206 case X86::BI__builtin_ia32_cmpeqss: 6207 case X86::BI__builtin_ia32_cmpltss: 6208 case X86::BI__builtin_ia32_cmpless: 6209 case X86::BI__builtin_ia32_cmpunordss: 6210 case X86::BI__builtin_ia32_cmpneqss: 6211 case X86::BI__builtin_ia32_cmpnltss: 6212 case X86::BI__builtin_ia32_cmpnless: 6213 case X86::BI__builtin_ia32_cmpordss: 6214 name = "cmpss"; 6215 ID = Intrinsic::x86_sse_cmp_ss; 6216 break; 6217 case X86::BI__builtin_ia32_cmpeqpd: 6218 case X86::BI__builtin_ia32_cmpltpd: 6219 case X86::BI__builtin_ia32_cmplepd: 6220 case X86::BI__builtin_ia32_cmpunordpd: 6221 case X86::BI__builtin_ia32_cmpneqpd: 6222 case X86::BI__builtin_ia32_cmpnltpd: 6223 case X86::BI__builtin_ia32_cmpnlepd: 6224 case X86::BI__builtin_ia32_cmpordpd: 6225 name = "cmppd"; 6226 ID = Intrinsic::x86_sse2_cmp_pd; 6227 break; 6228 case X86::BI__builtin_ia32_cmpeqsd: 6229 case X86::BI__builtin_ia32_cmpltsd: 6230 case X86::BI__builtin_ia32_cmplesd: 6231 case X86::BI__builtin_ia32_cmpunordsd: 6232 case X86::BI__builtin_ia32_cmpneqsd: 6233 case X86::BI__builtin_ia32_cmpnltsd: 6234 case X86::BI__builtin_ia32_cmpnlesd: 6235 case X86::BI__builtin_ia32_cmpordsd: 6236 name = "cmpsd"; 6237 ID = Intrinsic::x86_sse2_cmp_sd; 6238 break; 6239 } 6240 6241 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6242 llvm::Function *F = CGM.getIntrinsic(ID); 6243 return Builder.CreateCall(F, Ops, name); 6244 } 6245 } 6246 6247 6248 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6249 const CallExpr *E) { 6250 SmallVector<Value*, 4> Ops; 6251 6252 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6253 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6254 6255 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6256 6257 switch (BuiltinID) { 6258 default: return nullptr; 6259 6260 // vec_ld, vec_lvsl, vec_lvsr 6261 case PPC::BI__builtin_altivec_lvx: 6262 case PPC::BI__builtin_altivec_lvxl: 6263 case PPC::BI__builtin_altivec_lvebx: 6264 case PPC::BI__builtin_altivec_lvehx: 6265 case PPC::BI__builtin_altivec_lvewx: 6266 case PPC::BI__builtin_altivec_lvsl: 6267 case PPC::BI__builtin_altivec_lvsr: 6268 case PPC::BI__builtin_vsx_lxvd2x: 6269 case PPC::BI__builtin_vsx_lxvw4x: 6270 { 6271 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6272 6273 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6274 Ops.pop_back(); 6275 6276 switch (BuiltinID) { 6277 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6278 case PPC::BI__builtin_altivec_lvx: 6279 ID = Intrinsic::ppc_altivec_lvx; 6280 break; 6281 case PPC::BI__builtin_altivec_lvxl: 6282 ID = Intrinsic::ppc_altivec_lvxl; 6283 break; 6284 case PPC::BI__builtin_altivec_lvebx: 6285 ID = Intrinsic::ppc_altivec_lvebx; 6286 break; 6287 case PPC::BI__builtin_altivec_lvehx: 6288 ID = Intrinsic::ppc_altivec_lvehx; 6289 break; 6290 case PPC::BI__builtin_altivec_lvewx: 6291 ID = Intrinsic::ppc_altivec_lvewx; 6292 break; 6293 case PPC::BI__builtin_altivec_lvsl: 6294 ID = Intrinsic::ppc_altivec_lvsl; 6295 break; 6296 case PPC::BI__builtin_altivec_lvsr: 6297 ID = Intrinsic::ppc_altivec_lvsr; 6298 break; 6299 case PPC::BI__builtin_vsx_lxvd2x: 6300 ID = Intrinsic::ppc_vsx_lxvd2x; 6301 break; 6302 case PPC::BI__builtin_vsx_lxvw4x: 6303 ID = Intrinsic::ppc_vsx_lxvw4x; 6304 break; 6305 } 6306 llvm::Function *F = CGM.getIntrinsic(ID); 6307 return Builder.CreateCall(F, Ops, ""); 6308 } 6309 6310 // vec_st 6311 case PPC::BI__builtin_altivec_stvx: 6312 case PPC::BI__builtin_altivec_stvxl: 6313 case PPC::BI__builtin_altivec_stvebx: 6314 case PPC::BI__builtin_altivec_stvehx: 6315 case PPC::BI__builtin_altivec_stvewx: 6316 case PPC::BI__builtin_vsx_stxvd2x: 6317 case PPC::BI__builtin_vsx_stxvw4x: 6318 { 6319 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6320 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6321 Ops.pop_back(); 6322 6323 switch (BuiltinID) { 6324 default: llvm_unreachable("Unsupported st intrinsic!"); 6325 case PPC::BI__builtin_altivec_stvx: 6326 ID = Intrinsic::ppc_altivec_stvx; 6327 break; 6328 case PPC::BI__builtin_altivec_stvxl: 6329 ID = Intrinsic::ppc_altivec_stvxl; 6330 break; 6331 case PPC::BI__builtin_altivec_stvebx: 6332 ID = Intrinsic::ppc_altivec_stvebx; 6333 break; 6334 case PPC::BI__builtin_altivec_stvehx: 6335 ID = Intrinsic::ppc_altivec_stvehx; 6336 break; 6337 case PPC::BI__builtin_altivec_stvewx: 6338 ID = Intrinsic::ppc_altivec_stvewx; 6339 break; 6340 case PPC::BI__builtin_vsx_stxvd2x: 6341 ID = Intrinsic::ppc_vsx_stxvd2x; 6342 break; 6343 case PPC::BI__builtin_vsx_stxvw4x: 6344 ID = Intrinsic::ppc_vsx_stxvw4x; 6345 break; 6346 } 6347 llvm::Function *F = CGM.getIntrinsic(ID); 6348 return Builder.CreateCall(F, Ops, ""); 6349 } 6350 } 6351 } 6352 6353 // Emit an intrinsic that has 1 float or double. 6354 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6355 const CallExpr *E, 6356 unsigned IntrinsicID) { 6357 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6358 6359 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6360 return CGF.Builder.CreateCall(F, Src0); 6361 } 6362 6363 // Emit an intrinsic that has 3 float or double operands. 6364 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6365 const CallExpr *E, 6366 unsigned IntrinsicID) { 6367 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6368 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6369 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6370 6371 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6372 return CGF.Builder.CreateCall3(F, Src0, Src1, Src2); 6373 } 6374 6375 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6376 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6377 const CallExpr *E, 6378 unsigned IntrinsicID) { 6379 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6380 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6381 6382 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6383 return CGF.Builder.CreateCall2(F, Src0, Src1); 6384 } 6385 6386 Value *CodeGenFunction::EmitR600BuiltinExpr(unsigned BuiltinID, 6387 const CallExpr *E) { 6388 switch (BuiltinID) { 6389 case R600::BI__builtin_amdgpu_div_scale: 6390 case R600::BI__builtin_amdgpu_div_scalef: { 6391 // Translate from the intrinsics's struct return to the builtin's out 6392 // argument. 6393 6394 std::pair<llvm::Value *, unsigned> FlagOutPtr 6395 = EmitPointerWithAlignment(E->getArg(3)); 6396 6397 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6398 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6399 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6400 6401 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6402 X->getType()); 6403 6404 llvm::Value *Tmp = Builder.CreateCall3(Callee, X, Y, Z); 6405 6406 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6407 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6408 6409 llvm::Type *RealFlagType 6410 = FlagOutPtr.first->getType()->getPointerElementType(); 6411 6412 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6413 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6414 FlagStore->setAlignment(FlagOutPtr.second); 6415 return Result; 6416 } 6417 case R600::BI__builtin_amdgpu_div_fmas: 6418 case R600::BI__builtin_amdgpu_div_fmasf: { 6419 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6420 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6421 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6422 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6423 6424 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6425 Src0->getType()); 6426 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6427 return Builder.CreateCall4(F, Src0, Src1, Src2, Src3ToBool); 6428 } 6429 case R600::BI__builtin_amdgpu_div_fixup: 6430 case R600::BI__builtin_amdgpu_div_fixupf: 6431 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6432 case R600::BI__builtin_amdgpu_trig_preop: 6433 case R600::BI__builtin_amdgpu_trig_preopf: 6434 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6435 case R600::BI__builtin_amdgpu_rcp: 6436 case R600::BI__builtin_amdgpu_rcpf: 6437 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6438 case R600::BI__builtin_amdgpu_rsq: 6439 case R600::BI__builtin_amdgpu_rsqf: 6440 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6441 case R600::BI__builtin_amdgpu_rsq_clamped: 6442 case R600::BI__builtin_amdgpu_rsq_clampedf: 6443 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6444 case R600::BI__builtin_amdgpu_ldexp: 6445 case R600::BI__builtin_amdgpu_ldexpf: 6446 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6447 case R600::BI__builtin_amdgpu_class: 6448 case R600::BI__builtin_amdgpu_classf: 6449 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6450 default: 6451 return nullptr; 6452 } 6453 } 6454