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 "TargetInfo.h" 15 #include "CodeGenFunction.h" 16 #include "CodeGenModule.h" 17 #include "CGObjCRuntime.h" 18 #include "clang/Basic/TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "llvm/Intrinsics.h" 23 #include "llvm/Target/TargetData.h" 24 25 using namespace clang; 26 using namespace CodeGen; 27 using namespace llvm; 28 29 /// getBuiltinLibFunction - Given a builtin id for a function like 30 /// "__builtin_fabsf", return a Function* for "fabsf". 31 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 32 unsigned BuiltinID) { 33 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 34 35 // Get the name, skip over the __builtin_ prefix (if necessary). 36 StringRef Name; 37 GlobalDecl D(FD); 38 39 // If the builtin has been declared explicitly with an assembler label, 40 // use the mangled name. This differs from the plain label on platforms 41 // that prefix labels. 42 if (FD->hasAttr<AsmLabelAttr>()) 43 Name = getMangledName(D); 44 else 45 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 46 47 llvm::FunctionType *Ty = 48 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 49 50 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 51 } 52 53 /// Emit the conversions required to turn the given value into an 54 /// integer of the given size. 55 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 56 QualType T, llvm::IntegerType *IntType) { 57 V = CGF.EmitToMemory(V, T); 58 59 if (V->getType()->isPointerTy()) 60 return CGF.Builder.CreatePtrToInt(V, IntType); 61 62 assert(V->getType() == IntType); 63 return V; 64 } 65 66 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 67 QualType T, llvm::Type *ResultType) { 68 V = CGF.EmitFromMemory(V, T); 69 70 if (ResultType->isPointerTy()) 71 return CGF.Builder.CreateIntToPtr(V, ResultType); 72 73 assert(V->getType() == ResultType); 74 return V; 75 } 76 77 /// Utility to insert an atomic instruction based on Instrinsic::ID 78 /// and the expression node. 79 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 80 llvm::AtomicRMWInst::BinOp Kind, 81 const CallExpr *E) { 82 QualType T = E->getType(); 83 assert(E->getArg(0)->getType()->isPointerType()); 84 assert(CGF.getContext().hasSameUnqualifiedType(T, 85 E->getArg(0)->getType()->getPointeeType())); 86 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 87 88 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 89 unsigned AddrSpace = 90 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 91 92 llvm::IntegerType *IntType = 93 llvm::IntegerType::get(CGF.getLLVMContext(), 94 CGF.getContext().getTypeSize(T)); 95 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 96 97 llvm::Value *Args[2]; 98 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 99 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 100 llvm::Type *ValueType = Args[1]->getType(); 101 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 102 103 llvm::Value *Result = 104 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 105 llvm::SequentiallyConsistent); 106 Result = EmitFromInt(CGF, Result, T, ValueType); 107 return RValue::get(Result); 108 } 109 110 /// Utility to insert an atomic instruction based Instrinsic::ID and 111 /// the expression node, where the return value is the result of the 112 /// operation. 113 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 114 llvm::AtomicRMWInst::BinOp Kind, 115 const CallExpr *E, 116 Instruction::BinaryOps Op) { 117 QualType T = E->getType(); 118 assert(E->getArg(0)->getType()->isPointerType()); 119 assert(CGF.getContext().hasSameUnqualifiedType(T, 120 E->getArg(0)->getType()->getPointeeType())); 121 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 122 123 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 124 unsigned AddrSpace = 125 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 126 127 llvm::IntegerType *IntType = 128 llvm::IntegerType::get(CGF.getLLVMContext(), 129 CGF.getContext().getTypeSize(T)); 130 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 131 132 llvm::Value *Args[2]; 133 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 134 llvm::Type *ValueType = Args[1]->getType(); 135 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 136 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 137 138 llvm::Value *Result = 139 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 140 llvm::SequentiallyConsistent); 141 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 142 Result = EmitFromInt(CGF, Result, T, ValueType); 143 return RValue::get(Result); 144 } 145 146 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy, 147 /// which must be a scalar floating point type. 148 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 149 const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>(); 150 assert(ValTyP && "isn't scalar fp type!"); 151 152 StringRef FnName; 153 switch (ValTyP->getKind()) { 154 default: llvm_unreachable("Isn't a scalar fp type!"); 155 case BuiltinType::Float: FnName = "fabsf"; break; 156 case BuiltinType::Double: FnName = "fabs"; break; 157 case BuiltinType::LongDouble: FnName = "fabsl"; break; 158 } 159 160 // The prototype is something that takes and returns whatever V's type is. 161 llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(), 162 false); 163 llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName); 164 165 return CGF.Builder.CreateCall(Fn, V, "abs"); 166 } 167 168 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 169 const CallExpr *E, llvm::Value *calleeValue) { 170 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, 171 ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn); 172 } 173 174 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 175 unsigned BuiltinID, const CallExpr *E) { 176 // See if we can constant fold this builtin. If so, don't emit it at all. 177 Expr::EvalResult Result; 178 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 179 !Result.hasSideEffects()) { 180 if (Result.Val.isInt()) 181 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 182 Result.Val.getInt())); 183 if (Result.Val.isFloat()) 184 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 185 Result.Val.getFloat())); 186 } 187 188 switch (BuiltinID) { 189 default: break; // Handle intrinsics and libm functions below. 190 case Builtin::BI__builtin___CFStringMakeConstantString: 191 case Builtin::BI__builtin___NSStringMakeConstantString: 192 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 193 case Builtin::BI__builtin_stdarg_start: 194 case Builtin::BI__builtin_va_start: 195 case Builtin::BI__builtin_va_end: { 196 Value *ArgValue = EmitVAListRef(E->getArg(0)); 197 llvm::Type *DestType = Int8PtrTy; 198 if (ArgValue->getType() != DestType) 199 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 200 ArgValue->getName().data()); 201 202 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 203 Intrinsic::vaend : Intrinsic::vastart; 204 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 205 } 206 case Builtin::BI__builtin_va_copy: { 207 Value *DstPtr = EmitVAListRef(E->getArg(0)); 208 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 209 210 llvm::Type *Type = Int8PtrTy; 211 212 DstPtr = Builder.CreateBitCast(DstPtr, Type); 213 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 214 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 215 DstPtr, SrcPtr)); 216 } 217 case Builtin::BI__builtin_abs: 218 case Builtin::BI__builtin_labs: 219 case Builtin::BI__builtin_llabs: { 220 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 221 222 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 223 Value *CmpResult = 224 Builder.CreateICmpSGE(ArgValue, 225 llvm::Constant::getNullValue(ArgValue->getType()), 226 "abscond"); 227 Value *Result = 228 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 229 230 return RValue::get(Result); 231 } 232 case Builtin::BI__builtin_ctzs: 233 case Builtin::BI__builtin_ctz: 234 case Builtin::BI__builtin_ctzl: 235 case Builtin::BI__builtin_ctzll: { 236 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 237 238 llvm::Type *ArgType = ArgValue->getType(); 239 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 240 241 llvm::Type *ResultType = ConvertType(E->getType()); 242 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 243 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 244 if (Result->getType() != ResultType) 245 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 246 "cast"); 247 return RValue::get(Result); 248 } 249 case Builtin::BI__builtin_clzs: 250 case Builtin::BI__builtin_clz: 251 case Builtin::BI__builtin_clzl: 252 case Builtin::BI__builtin_clzll: { 253 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 254 255 llvm::Type *ArgType = ArgValue->getType(); 256 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 257 258 llvm::Type *ResultType = ConvertType(E->getType()); 259 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 260 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 261 if (Result->getType() != ResultType) 262 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 263 "cast"); 264 return RValue::get(Result); 265 } 266 case Builtin::BI__builtin_ffs: 267 case Builtin::BI__builtin_ffsl: 268 case Builtin::BI__builtin_ffsll: { 269 // ffs(x) -> x ? cttz(x) + 1 : 0 270 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 271 272 llvm::Type *ArgType = ArgValue->getType(); 273 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 274 275 llvm::Type *ResultType = ConvertType(E->getType()); 276 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 277 Builder.getTrue()), 278 llvm::ConstantInt::get(ArgType, 1)); 279 Value *Zero = llvm::Constant::getNullValue(ArgType); 280 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 281 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 282 if (Result->getType() != ResultType) 283 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 284 "cast"); 285 return RValue::get(Result); 286 } 287 case Builtin::BI__builtin_parity: 288 case Builtin::BI__builtin_parityl: 289 case Builtin::BI__builtin_parityll: { 290 // parity(x) -> ctpop(x) & 1 291 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 292 293 llvm::Type *ArgType = ArgValue->getType(); 294 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 295 296 llvm::Type *ResultType = ConvertType(E->getType()); 297 Value *Tmp = Builder.CreateCall(F, ArgValue); 298 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 299 if (Result->getType() != ResultType) 300 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 301 "cast"); 302 return RValue::get(Result); 303 } 304 case Builtin::BI__builtin_popcount: 305 case Builtin::BI__builtin_popcountl: 306 case Builtin::BI__builtin_popcountll: { 307 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 308 309 llvm::Type *ArgType = ArgValue->getType(); 310 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 311 312 llvm::Type *ResultType = ConvertType(E->getType()); 313 Value *Result = Builder.CreateCall(F, ArgValue); 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_expect: { 320 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 321 llvm::Type *ArgType = ArgValue->getType(); 322 323 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 324 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 325 326 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 327 "expval"); 328 return RValue::get(Result); 329 } 330 case Builtin::BI__builtin_bswap32: 331 case Builtin::BI__builtin_bswap64: { 332 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 333 llvm::Type *ArgType = ArgValue->getType(); 334 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 335 return RValue::get(Builder.CreateCall(F, ArgValue)); 336 } 337 case Builtin::BI__builtin_object_size: { 338 // We rely on constant folding to deal with expressions with side effects. 339 assert(!E->getArg(0)->HasSideEffects(getContext()) && 340 "should have been constant folded"); 341 342 // We pass this builtin onto the optimizer so that it can 343 // figure out the object size in more complex cases. 344 llvm::Type *ResType = ConvertType(E->getType()); 345 346 // LLVM only supports 0 and 2, make sure that we pass along that 347 // as a boolean. 348 Value *Ty = EmitScalarExpr(E->getArg(1)); 349 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 350 assert(CI); 351 uint64_t val = CI->getZExtValue(); 352 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 353 354 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType); 355 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 356 } 357 case Builtin::BI__builtin_prefetch: { 358 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 359 // FIXME: Technically these constants should of type 'int', yes? 360 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 361 llvm::ConstantInt::get(Int32Ty, 0); 362 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 363 llvm::ConstantInt::get(Int32Ty, 3); 364 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 365 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 366 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 367 } 368 case Builtin::BI__builtin_trap: { 369 Value *F = CGM.getIntrinsic(Intrinsic::trap); 370 return RValue::get(Builder.CreateCall(F)); 371 } 372 case Builtin::BI__builtin_unreachable: { 373 if (CatchUndefined) 374 EmitBranch(getTrapBB()); 375 else 376 Builder.CreateUnreachable(); 377 378 // We do need to preserve an insertion point. 379 EmitBlock(createBasicBlock("unreachable.cont")); 380 381 return RValue::get(0); 382 } 383 384 case Builtin::BI__builtin_powi: 385 case Builtin::BI__builtin_powif: 386 case Builtin::BI__builtin_powil: { 387 Value *Base = EmitScalarExpr(E->getArg(0)); 388 Value *Exponent = EmitScalarExpr(E->getArg(1)); 389 llvm::Type *ArgType = Base->getType(); 390 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 391 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 392 } 393 394 case Builtin::BI__builtin_isgreater: 395 case Builtin::BI__builtin_isgreaterequal: 396 case Builtin::BI__builtin_isless: 397 case Builtin::BI__builtin_islessequal: 398 case Builtin::BI__builtin_islessgreater: 399 case Builtin::BI__builtin_isunordered: { 400 // Ordered comparisons: we know the arguments to these are matching scalar 401 // floating point values. 402 Value *LHS = EmitScalarExpr(E->getArg(0)); 403 Value *RHS = EmitScalarExpr(E->getArg(1)); 404 405 switch (BuiltinID) { 406 default: llvm_unreachable("Unknown ordered comparison"); 407 case Builtin::BI__builtin_isgreater: 408 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 409 break; 410 case Builtin::BI__builtin_isgreaterequal: 411 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 412 break; 413 case Builtin::BI__builtin_isless: 414 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 415 break; 416 case Builtin::BI__builtin_islessequal: 417 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 418 break; 419 case Builtin::BI__builtin_islessgreater: 420 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 421 break; 422 case Builtin::BI__builtin_isunordered: 423 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 424 break; 425 } 426 // ZExt bool to int type. 427 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 428 } 429 case Builtin::BI__builtin_isnan: { 430 Value *V = EmitScalarExpr(E->getArg(0)); 431 V = Builder.CreateFCmpUNO(V, V, "cmp"); 432 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 433 } 434 435 case Builtin::BI__builtin_isinf: { 436 // isinf(x) --> fabs(x) == infinity 437 Value *V = EmitScalarExpr(E->getArg(0)); 438 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 439 440 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 441 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 442 } 443 444 // TODO: BI__builtin_isinf_sign 445 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 446 447 case Builtin::BI__builtin_isnormal: { 448 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 449 Value *V = EmitScalarExpr(E->getArg(0)); 450 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 451 452 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 453 Value *IsLessThanInf = 454 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 455 APFloat Smallest = APFloat::getSmallestNormalized( 456 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 457 Value *IsNormal = 458 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 459 "isnormal"); 460 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 461 V = Builder.CreateAnd(V, IsNormal, "and"); 462 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 463 } 464 465 case Builtin::BI__builtin_isfinite: { 466 // isfinite(x) --> x == x && fabs(x) != infinity; 467 Value *V = EmitScalarExpr(E->getArg(0)); 468 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 469 470 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 471 Value *IsNotInf = 472 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 473 474 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 475 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 476 } 477 478 case Builtin::BI__builtin_fpclassify: { 479 Value *V = EmitScalarExpr(E->getArg(5)); 480 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 481 482 // Create Result 483 BasicBlock *Begin = Builder.GetInsertBlock(); 484 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 485 Builder.SetInsertPoint(End); 486 PHINode *Result = 487 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 488 "fpclassify_result"); 489 490 // if (V==0) return FP_ZERO 491 Builder.SetInsertPoint(Begin); 492 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 493 "iszero"); 494 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 495 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 496 Builder.CreateCondBr(IsZero, End, NotZero); 497 Result->addIncoming(ZeroLiteral, Begin); 498 499 // if (V != V) return FP_NAN 500 Builder.SetInsertPoint(NotZero); 501 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 502 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 503 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 504 Builder.CreateCondBr(IsNan, End, NotNan); 505 Result->addIncoming(NanLiteral, NotZero); 506 507 // if (fabs(V) == infinity) return FP_INFINITY 508 Builder.SetInsertPoint(NotNan); 509 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 510 Value *IsInf = 511 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 512 "isinf"); 513 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 514 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 515 Builder.CreateCondBr(IsInf, End, NotInf); 516 Result->addIncoming(InfLiteral, NotNan); 517 518 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 519 Builder.SetInsertPoint(NotInf); 520 APFloat Smallest = APFloat::getSmallestNormalized( 521 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 522 Value *IsNormal = 523 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 524 "isnormal"); 525 Value *NormalResult = 526 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 527 EmitScalarExpr(E->getArg(3))); 528 Builder.CreateBr(End); 529 Result->addIncoming(NormalResult, NotInf); 530 531 // return Result 532 Builder.SetInsertPoint(End); 533 return RValue::get(Result); 534 } 535 536 case Builtin::BIalloca: 537 case Builtin::BI__builtin_alloca: { 538 Value *Size = EmitScalarExpr(E->getArg(0)); 539 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 540 } 541 case Builtin::BIbzero: 542 case Builtin::BI__builtin_bzero: { 543 Value *Address = EmitScalarExpr(E->getArg(0)); 544 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 545 unsigned Align = GetPointeeAlignment(E->getArg(0)); 546 Builder.CreateMemSet(Address, Builder.getInt8(0), SizeVal, Align, false); 547 return RValue::get(Address); 548 } 549 case Builtin::BImemcpy: 550 case Builtin::BI__builtin_memcpy: { 551 Value *Address = EmitScalarExpr(E->getArg(0)); 552 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 553 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 554 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 555 GetPointeeAlignment(E->getArg(1))); 556 Builder.CreateMemCpy(Address, SrcAddr, SizeVal, Align, false); 557 return RValue::get(Address); 558 } 559 560 case Builtin::BI__builtin___memcpy_chk: { 561 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 562 llvm::APSInt Size, DstSize; 563 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 564 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 565 break; 566 if (Size.ugt(DstSize)) 567 break; 568 Value *Dest = EmitScalarExpr(E->getArg(0)); 569 Value *Src = EmitScalarExpr(E->getArg(1)); 570 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 571 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 572 GetPointeeAlignment(E->getArg(1))); 573 Builder.CreateMemCpy(Dest, Src, SizeVal, Align, false); 574 return RValue::get(Dest); 575 } 576 577 case Builtin::BI__builtin_objc_memmove_collectable: { 578 Value *Address = EmitScalarExpr(E->getArg(0)); 579 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 580 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 581 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 582 Address, SrcAddr, SizeVal); 583 return RValue::get(Address); 584 } 585 586 case Builtin::BI__builtin___memmove_chk: { 587 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 588 llvm::APSInt Size, DstSize; 589 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 590 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 591 break; 592 if (Size.ugt(DstSize)) 593 break; 594 Value *Dest = EmitScalarExpr(E->getArg(0)); 595 Value *Src = EmitScalarExpr(E->getArg(1)); 596 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 597 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 598 GetPointeeAlignment(E->getArg(1))); 599 Builder.CreateMemMove(Dest, Src, SizeVal, Align, false); 600 return RValue::get(Dest); 601 } 602 603 case Builtin::BImemmove: 604 case Builtin::BI__builtin_memmove: { 605 Value *Address = EmitScalarExpr(E->getArg(0)); 606 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 607 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 608 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 609 GetPointeeAlignment(E->getArg(1))); 610 Builder.CreateMemMove(Address, SrcAddr, SizeVal, Align, false); 611 return RValue::get(Address); 612 } 613 case Builtin::BImemset: 614 case Builtin::BI__builtin_memset: { 615 Value *Address = EmitScalarExpr(E->getArg(0)); 616 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 617 Builder.getInt8Ty()); 618 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 619 unsigned Align = GetPointeeAlignment(E->getArg(0)); 620 Builder.CreateMemSet(Address, ByteVal, SizeVal, Align, false); 621 return RValue::get(Address); 622 } 623 case Builtin::BI__builtin___memset_chk: { 624 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 625 llvm::APSInt Size, DstSize; 626 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 627 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 628 break; 629 if (Size.ugt(DstSize)) 630 break; 631 Value *Address = EmitScalarExpr(E->getArg(0)); 632 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 633 Builder.getInt8Ty()); 634 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 635 unsigned Align = GetPointeeAlignment(E->getArg(0)); 636 Builder.CreateMemSet(Address, ByteVal, SizeVal, Align, false); 637 638 return RValue::get(Address); 639 } 640 case Builtin::BI__builtin_dwarf_cfa: { 641 // The offset in bytes from the first argument to the CFA. 642 // 643 // Why on earth is this in the frontend? Is there any reason at 644 // all that the backend can't reasonably determine this while 645 // lowering llvm.eh.dwarf.cfa()? 646 // 647 // TODO: If there's a satisfactory reason, add a target hook for 648 // this instead of hard-coding 0, which is correct for most targets. 649 int32_t Offset = 0; 650 651 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 652 return RValue::get(Builder.CreateCall(F, 653 llvm::ConstantInt::get(Int32Ty, Offset))); 654 } 655 case Builtin::BI__builtin_return_address: { 656 Value *Depth = EmitScalarExpr(E->getArg(0)); 657 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 658 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 659 return RValue::get(Builder.CreateCall(F, Depth)); 660 } 661 case Builtin::BI__builtin_frame_address: { 662 Value *Depth = EmitScalarExpr(E->getArg(0)); 663 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 664 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 665 return RValue::get(Builder.CreateCall(F, Depth)); 666 } 667 case Builtin::BI__builtin_extract_return_addr: { 668 Value *Address = EmitScalarExpr(E->getArg(0)); 669 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 670 return RValue::get(Result); 671 } 672 case Builtin::BI__builtin_frob_return_addr: { 673 Value *Address = EmitScalarExpr(E->getArg(0)); 674 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 675 return RValue::get(Result); 676 } 677 case Builtin::BI__builtin_dwarf_sp_column: { 678 llvm::IntegerType *Ty 679 = cast<llvm::IntegerType>(ConvertType(E->getType())); 680 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 681 if (Column == -1) { 682 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 683 return RValue::get(llvm::UndefValue::get(Ty)); 684 } 685 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 686 } 687 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 688 Value *Address = EmitScalarExpr(E->getArg(0)); 689 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 690 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 691 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 692 } 693 case Builtin::BI__builtin_eh_return: { 694 Value *Int = EmitScalarExpr(E->getArg(0)); 695 Value *Ptr = EmitScalarExpr(E->getArg(1)); 696 697 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 698 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 699 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 700 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 701 ? Intrinsic::eh_return_i32 702 : Intrinsic::eh_return_i64); 703 Builder.CreateCall2(F, Int, Ptr); 704 Builder.CreateUnreachable(); 705 706 // We do need to preserve an insertion point. 707 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 708 709 return RValue::get(0); 710 } 711 case Builtin::BI__builtin_unwind_init: { 712 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 713 return RValue::get(Builder.CreateCall(F)); 714 } 715 case Builtin::BI__builtin_extend_pointer: { 716 // Extends a pointer to the size of an _Unwind_Word, which is 717 // uint64_t on all platforms. Generally this gets poked into a 718 // register and eventually used as an address, so if the 719 // addressing registers are wider than pointers and the platform 720 // doesn't implicitly ignore high-order bits when doing 721 // addressing, we need to make sure we zext / sext based on 722 // the platform's expectations. 723 // 724 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 725 726 // Cast the pointer to intptr_t. 727 Value *Ptr = EmitScalarExpr(E->getArg(0)); 728 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 729 730 // If that's 64 bits, we're done. 731 if (IntPtrTy->getBitWidth() == 64) 732 return RValue::get(Result); 733 734 // Otherwise, ask the codegen data what to do. 735 if (getTargetHooks().extendPointerWithSExt()) 736 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 737 else 738 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 739 } 740 case Builtin::BI__builtin_setjmp: { 741 // Buffer is a void**. 742 Value *Buf = EmitScalarExpr(E->getArg(0)); 743 744 // Store the frame pointer to the setjmp buffer. 745 Value *FrameAddr = 746 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 747 ConstantInt::get(Int32Ty, 0)); 748 Builder.CreateStore(FrameAddr, Buf); 749 750 // Store the stack pointer to the setjmp buffer. 751 Value *StackAddr = 752 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 753 Value *StackSaveSlot = 754 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 755 Builder.CreateStore(StackAddr, StackSaveSlot); 756 757 // Call LLVM's EH setjmp, which is lightweight. 758 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 759 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 760 return RValue::get(Builder.CreateCall(F, Buf)); 761 } 762 case Builtin::BI__builtin_longjmp: { 763 Value *Buf = EmitScalarExpr(E->getArg(0)); 764 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 765 766 // Call LLVM's EH longjmp, which is lightweight. 767 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 768 769 // longjmp doesn't return; mark this as unreachable. 770 Builder.CreateUnreachable(); 771 772 // We do need to preserve an insertion point. 773 EmitBlock(createBasicBlock("longjmp.cont")); 774 775 return RValue::get(0); 776 } 777 case Builtin::BI__sync_fetch_and_add: 778 case Builtin::BI__sync_fetch_and_sub: 779 case Builtin::BI__sync_fetch_and_or: 780 case Builtin::BI__sync_fetch_and_and: 781 case Builtin::BI__sync_fetch_and_xor: 782 case Builtin::BI__sync_add_and_fetch: 783 case Builtin::BI__sync_sub_and_fetch: 784 case Builtin::BI__sync_and_and_fetch: 785 case Builtin::BI__sync_or_and_fetch: 786 case Builtin::BI__sync_xor_and_fetch: 787 case Builtin::BI__sync_val_compare_and_swap: 788 case Builtin::BI__sync_bool_compare_and_swap: 789 case Builtin::BI__sync_lock_test_and_set: 790 case Builtin::BI__sync_lock_release: 791 case Builtin::BI__sync_swap: 792 llvm_unreachable("Shouldn't make it through sema"); 793 case Builtin::BI__sync_fetch_and_add_1: 794 case Builtin::BI__sync_fetch_and_add_2: 795 case Builtin::BI__sync_fetch_and_add_4: 796 case Builtin::BI__sync_fetch_and_add_8: 797 case Builtin::BI__sync_fetch_and_add_16: 798 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 799 case Builtin::BI__sync_fetch_and_sub_1: 800 case Builtin::BI__sync_fetch_and_sub_2: 801 case Builtin::BI__sync_fetch_and_sub_4: 802 case Builtin::BI__sync_fetch_and_sub_8: 803 case Builtin::BI__sync_fetch_and_sub_16: 804 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 805 case Builtin::BI__sync_fetch_and_or_1: 806 case Builtin::BI__sync_fetch_and_or_2: 807 case Builtin::BI__sync_fetch_and_or_4: 808 case Builtin::BI__sync_fetch_and_or_8: 809 case Builtin::BI__sync_fetch_and_or_16: 810 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 811 case Builtin::BI__sync_fetch_and_and_1: 812 case Builtin::BI__sync_fetch_and_and_2: 813 case Builtin::BI__sync_fetch_and_and_4: 814 case Builtin::BI__sync_fetch_and_and_8: 815 case Builtin::BI__sync_fetch_and_and_16: 816 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 817 case Builtin::BI__sync_fetch_and_xor_1: 818 case Builtin::BI__sync_fetch_and_xor_2: 819 case Builtin::BI__sync_fetch_and_xor_4: 820 case Builtin::BI__sync_fetch_and_xor_8: 821 case Builtin::BI__sync_fetch_and_xor_16: 822 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 823 824 // Clang extensions: not overloaded yet. 825 case Builtin::BI__sync_fetch_and_min: 826 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 827 case Builtin::BI__sync_fetch_and_max: 828 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 829 case Builtin::BI__sync_fetch_and_umin: 830 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 831 case Builtin::BI__sync_fetch_and_umax: 832 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 833 834 case Builtin::BI__sync_add_and_fetch_1: 835 case Builtin::BI__sync_add_and_fetch_2: 836 case Builtin::BI__sync_add_and_fetch_4: 837 case Builtin::BI__sync_add_and_fetch_8: 838 case Builtin::BI__sync_add_and_fetch_16: 839 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 840 llvm::Instruction::Add); 841 case Builtin::BI__sync_sub_and_fetch_1: 842 case Builtin::BI__sync_sub_and_fetch_2: 843 case Builtin::BI__sync_sub_and_fetch_4: 844 case Builtin::BI__sync_sub_and_fetch_8: 845 case Builtin::BI__sync_sub_and_fetch_16: 846 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 847 llvm::Instruction::Sub); 848 case Builtin::BI__sync_and_and_fetch_1: 849 case Builtin::BI__sync_and_and_fetch_2: 850 case Builtin::BI__sync_and_and_fetch_4: 851 case Builtin::BI__sync_and_and_fetch_8: 852 case Builtin::BI__sync_and_and_fetch_16: 853 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 854 llvm::Instruction::And); 855 case Builtin::BI__sync_or_and_fetch_1: 856 case Builtin::BI__sync_or_and_fetch_2: 857 case Builtin::BI__sync_or_and_fetch_4: 858 case Builtin::BI__sync_or_and_fetch_8: 859 case Builtin::BI__sync_or_and_fetch_16: 860 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 861 llvm::Instruction::Or); 862 case Builtin::BI__sync_xor_and_fetch_1: 863 case Builtin::BI__sync_xor_and_fetch_2: 864 case Builtin::BI__sync_xor_and_fetch_4: 865 case Builtin::BI__sync_xor_and_fetch_8: 866 case Builtin::BI__sync_xor_and_fetch_16: 867 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 868 llvm::Instruction::Xor); 869 870 case Builtin::BI__sync_val_compare_and_swap_1: 871 case Builtin::BI__sync_val_compare_and_swap_2: 872 case Builtin::BI__sync_val_compare_and_swap_4: 873 case Builtin::BI__sync_val_compare_and_swap_8: 874 case Builtin::BI__sync_val_compare_and_swap_16: { 875 QualType T = E->getType(); 876 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 877 unsigned AddrSpace = 878 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 879 880 llvm::IntegerType *IntType = 881 llvm::IntegerType::get(getLLVMContext(), 882 getContext().getTypeSize(T)); 883 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 884 885 Value *Args[3]; 886 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 887 Args[1] = EmitScalarExpr(E->getArg(1)); 888 llvm::Type *ValueType = Args[1]->getType(); 889 Args[1] = EmitToInt(*this, Args[1], T, IntType); 890 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 891 892 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 893 llvm::SequentiallyConsistent); 894 Result = EmitFromInt(*this, Result, T, ValueType); 895 return RValue::get(Result); 896 } 897 898 case Builtin::BI__sync_bool_compare_and_swap_1: 899 case Builtin::BI__sync_bool_compare_and_swap_2: 900 case Builtin::BI__sync_bool_compare_and_swap_4: 901 case Builtin::BI__sync_bool_compare_and_swap_8: 902 case Builtin::BI__sync_bool_compare_and_swap_16: { 903 QualType T = E->getArg(1)->getType(); 904 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 905 unsigned AddrSpace = 906 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 907 908 llvm::IntegerType *IntType = 909 llvm::IntegerType::get(getLLVMContext(), 910 getContext().getTypeSize(T)); 911 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 912 913 Value *Args[3]; 914 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 915 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 916 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 917 918 Value *OldVal = Args[1]; 919 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 920 llvm::SequentiallyConsistent); 921 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 922 // zext bool to int. 923 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 924 return RValue::get(Result); 925 } 926 927 case Builtin::BI__sync_swap_1: 928 case Builtin::BI__sync_swap_2: 929 case Builtin::BI__sync_swap_4: 930 case Builtin::BI__sync_swap_8: 931 case Builtin::BI__sync_swap_16: 932 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 933 934 case Builtin::BI__sync_lock_test_and_set_1: 935 case Builtin::BI__sync_lock_test_and_set_2: 936 case Builtin::BI__sync_lock_test_and_set_4: 937 case Builtin::BI__sync_lock_test_and_set_8: 938 case Builtin::BI__sync_lock_test_and_set_16: 939 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 940 941 case Builtin::BI__sync_lock_release_1: 942 case Builtin::BI__sync_lock_release_2: 943 case Builtin::BI__sync_lock_release_4: 944 case Builtin::BI__sync_lock_release_8: 945 case Builtin::BI__sync_lock_release_16: { 946 Value *Ptr = EmitScalarExpr(E->getArg(0)); 947 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 948 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 949 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 950 StoreSize.getQuantity() * 8); 951 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 952 llvm::StoreInst *Store = 953 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 954 Store->setAlignment(StoreSize.getQuantity()); 955 Store->setAtomic(llvm::Release); 956 return RValue::get(0); 957 } 958 959 case Builtin::BI__sync_synchronize: { 960 // We assume this is supposed to correspond to a C++0x-style 961 // sequentially-consistent fence (i.e. this is only usable for 962 // synchonization, not device I/O or anything like that). This intrinsic 963 // is really badly designed in the sense that in theory, there isn't 964 // any way to safely use it... but in practice, it mostly works 965 // to use it with non-atomic loads and stores to get acquire/release 966 // semantics. 967 Builder.CreateFence(llvm::SequentiallyConsistent); 968 return RValue::get(0); 969 } 970 971 case Builtin::BI__c11_atomic_is_lock_free: 972 case Builtin::BI__atomic_is_lock_free: { 973 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 974 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 975 // _Atomic(T) is always properly-aligned. 976 const char *LibCallName = "__atomic_is_lock_free"; 977 CallArgList Args; 978 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 979 getContext().getSizeType()); 980 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 981 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 982 getContext().VoidPtrTy); 983 else 984 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 985 getContext().VoidPtrTy); 986 const CGFunctionInfo &FuncInfo = 987 CGM.getTypes().arrangeFunctionCall(E->getType(), Args, 988 FunctionType::ExtInfo(), 989 RequiredArgs::All); 990 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 991 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 992 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 993 } 994 995 case Builtin::BI__atomic_test_and_set: { 996 // Look at the argument type to determine whether this is a volatile 997 // operation. The parameter type is always volatile. 998 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 999 bool Volatile = 1000 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1001 1002 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1003 unsigned AddrSpace = 1004 cast<llvm::PointerType>(Ptr->getType())->getAddressSpace(); 1005 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1006 Value *NewVal = Builder.getInt8(1); 1007 Value *Order = EmitScalarExpr(E->getArg(1)); 1008 if (isa<llvm::ConstantInt>(Order)) { 1009 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1010 AtomicRMWInst *Result = 0; 1011 switch (ord) { 1012 case 0: // memory_order_relaxed 1013 default: // invalid order 1014 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1015 Ptr, NewVal, 1016 llvm::Monotonic); 1017 break; 1018 case 1: // memory_order_consume 1019 case 2: // memory_order_acquire 1020 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1021 Ptr, NewVal, 1022 llvm::Acquire); 1023 break; 1024 case 3: // memory_order_release 1025 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1026 Ptr, NewVal, 1027 llvm::Release); 1028 break; 1029 case 4: // memory_order_acq_rel 1030 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1031 Ptr, NewVal, 1032 llvm::AcquireRelease); 1033 break; 1034 case 5: // memory_order_seq_cst 1035 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1036 Ptr, NewVal, 1037 llvm::SequentiallyConsistent); 1038 break; 1039 } 1040 Result->setVolatile(Volatile); 1041 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1042 } 1043 1044 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1045 1046 llvm::BasicBlock *BBs[5] = { 1047 createBasicBlock("monotonic", CurFn), 1048 createBasicBlock("acquire", CurFn), 1049 createBasicBlock("release", CurFn), 1050 createBasicBlock("acqrel", CurFn), 1051 createBasicBlock("seqcst", CurFn) 1052 }; 1053 llvm::AtomicOrdering Orders[5] = { 1054 llvm::Monotonic, llvm::Acquire, llvm::Release, 1055 llvm::AcquireRelease, llvm::SequentiallyConsistent 1056 }; 1057 1058 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1059 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1060 1061 Builder.SetInsertPoint(ContBB); 1062 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1063 1064 for (unsigned i = 0; i < 5; ++i) { 1065 Builder.SetInsertPoint(BBs[i]); 1066 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1067 Ptr, NewVal, Orders[i]); 1068 RMW->setVolatile(Volatile); 1069 Result->addIncoming(RMW, BBs[i]); 1070 Builder.CreateBr(ContBB); 1071 } 1072 1073 SI->addCase(Builder.getInt32(0), BBs[0]); 1074 SI->addCase(Builder.getInt32(1), BBs[1]); 1075 SI->addCase(Builder.getInt32(2), BBs[1]); 1076 SI->addCase(Builder.getInt32(3), BBs[2]); 1077 SI->addCase(Builder.getInt32(4), BBs[3]); 1078 SI->addCase(Builder.getInt32(5), BBs[4]); 1079 1080 Builder.SetInsertPoint(ContBB); 1081 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1082 } 1083 1084 case Builtin::BI__atomic_clear: { 1085 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1086 bool Volatile = 1087 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1088 1089 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1090 unsigned AddrSpace = 1091 cast<llvm::PointerType>(Ptr->getType())->getAddressSpace(); 1092 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1093 Value *NewVal = Builder.getInt8(0); 1094 Value *Order = EmitScalarExpr(E->getArg(1)); 1095 if (isa<llvm::ConstantInt>(Order)) { 1096 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1097 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1098 Store->setAlignment(1); 1099 switch (ord) { 1100 case 0: // memory_order_relaxed 1101 default: // invalid order 1102 Store->setOrdering(llvm::Monotonic); 1103 break; 1104 case 3: // memory_order_release 1105 Store->setOrdering(llvm::Release); 1106 break; 1107 case 5: // memory_order_seq_cst 1108 Store->setOrdering(llvm::SequentiallyConsistent); 1109 break; 1110 } 1111 return RValue::get(0); 1112 } 1113 1114 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1115 1116 llvm::BasicBlock *BBs[3] = { 1117 createBasicBlock("monotonic", CurFn), 1118 createBasicBlock("release", CurFn), 1119 createBasicBlock("seqcst", CurFn) 1120 }; 1121 llvm::AtomicOrdering Orders[3] = { 1122 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1123 }; 1124 1125 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1126 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1127 1128 for (unsigned i = 0; i < 3; ++i) { 1129 Builder.SetInsertPoint(BBs[i]); 1130 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1131 Store->setAlignment(1); 1132 Store->setOrdering(Orders[i]); 1133 Builder.CreateBr(ContBB); 1134 } 1135 1136 SI->addCase(Builder.getInt32(0), BBs[0]); 1137 SI->addCase(Builder.getInt32(3), BBs[1]); 1138 SI->addCase(Builder.getInt32(5), BBs[2]); 1139 1140 Builder.SetInsertPoint(ContBB); 1141 return RValue::get(0); 1142 } 1143 1144 case Builtin::BI__atomic_thread_fence: 1145 case Builtin::BI__atomic_signal_fence: 1146 case Builtin::BI__c11_atomic_thread_fence: 1147 case Builtin::BI__c11_atomic_signal_fence: { 1148 llvm::SynchronizationScope Scope; 1149 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1150 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1151 Scope = llvm::SingleThread; 1152 else 1153 Scope = llvm::CrossThread; 1154 Value *Order = EmitScalarExpr(E->getArg(0)); 1155 if (isa<llvm::ConstantInt>(Order)) { 1156 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1157 switch (ord) { 1158 case 0: // memory_order_relaxed 1159 default: // invalid order 1160 break; 1161 case 1: // memory_order_consume 1162 case 2: // memory_order_acquire 1163 Builder.CreateFence(llvm::Acquire, Scope); 1164 break; 1165 case 3: // memory_order_release 1166 Builder.CreateFence(llvm::Release, Scope); 1167 break; 1168 case 4: // memory_order_acq_rel 1169 Builder.CreateFence(llvm::AcquireRelease, Scope); 1170 break; 1171 case 5: // memory_order_seq_cst 1172 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1173 break; 1174 } 1175 return RValue::get(0); 1176 } 1177 1178 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1179 AcquireBB = createBasicBlock("acquire", CurFn); 1180 ReleaseBB = createBasicBlock("release", CurFn); 1181 AcqRelBB = createBasicBlock("acqrel", CurFn); 1182 SeqCstBB = createBasicBlock("seqcst", CurFn); 1183 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1184 1185 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1186 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1187 1188 Builder.SetInsertPoint(AcquireBB); 1189 Builder.CreateFence(llvm::Acquire, Scope); 1190 Builder.CreateBr(ContBB); 1191 SI->addCase(Builder.getInt32(1), AcquireBB); 1192 SI->addCase(Builder.getInt32(2), AcquireBB); 1193 1194 Builder.SetInsertPoint(ReleaseBB); 1195 Builder.CreateFence(llvm::Release, Scope); 1196 Builder.CreateBr(ContBB); 1197 SI->addCase(Builder.getInt32(3), ReleaseBB); 1198 1199 Builder.SetInsertPoint(AcqRelBB); 1200 Builder.CreateFence(llvm::AcquireRelease, Scope); 1201 Builder.CreateBr(ContBB); 1202 SI->addCase(Builder.getInt32(4), AcqRelBB); 1203 1204 Builder.SetInsertPoint(SeqCstBB); 1205 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1206 Builder.CreateBr(ContBB); 1207 SI->addCase(Builder.getInt32(5), SeqCstBB); 1208 1209 Builder.SetInsertPoint(ContBB); 1210 return RValue::get(0); 1211 } 1212 1213 // Library functions with special handling. 1214 case Builtin::BIsqrt: 1215 case Builtin::BIsqrtf: 1216 case Builtin::BIsqrtl: { 1217 // TODO: there is currently no set of optimizer flags 1218 // sufficient for us to rewrite sqrt to @llvm.sqrt. 1219 // -fmath-errno=0 is not good enough; we need finiteness. 1220 // We could probably precondition the call with an ult 1221 // against 0, but is that worth the complexity? 1222 break; 1223 } 1224 1225 case Builtin::BIpow: 1226 case Builtin::BIpowf: 1227 case Builtin::BIpowl: { 1228 // Rewrite sqrt to intrinsic if allowed. 1229 if (!FD->hasAttr<ConstAttr>()) 1230 break; 1231 Value *Base = EmitScalarExpr(E->getArg(0)); 1232 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1233 llvm::Type *ArgType = Base->getType(); 1234 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1235 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1236 } 1237 1238 case Builtin::BIfma: 1239 case Builtin::BIfmaf: 1240 case Builtin::BIfmal: 1241 case Builtin::BI__builtin_fma: 1242 case Builtin::BI__builtin_fmaf: 1243 case Builtin::BI__builtin_fmal: { 1244 // Rewrite fma to intrinsic. 1245 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1246 llvm::Type *ArgType = FirstArg->getType(); 1247 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1248 return RValue::get(Builder.CreateCall3(F, FirstArg, 1249 EmitScalarExpr(E->getArg(1)), 1250 EmitScalarExpr(E->getArg(2)))); 1251 } 1252 1253 case Builtin::BI__builtin_signbit: 1254 case Builtin::BI__builtin_signbitf: 1255 case Builtin::BI__builtin_signbitl: { 1256 LLVMContext &C = CGM.getLLVMContext(); 1257 1258 Value *Arg = EmitScalarExpr(E->getArg(0)); 1259 llvm::Type *ArgTy = Arg->getType(); 1260 if (ArgTy->isPPC_FP128Ty()) 1261 break; // FIXME: I'm not sure what the right implementation is here. 1262 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1263 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1264 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1265 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1266 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1267 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1268 } 1269 case Builtin::BI__builtin_annotation: { 1270 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1271 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1272 AnnVal->getType()); 1273 1274 // Get the annotation string, go through casts. Sema requires this to be a 1275 // non-wide string literal, potentially casted, so the cast<> is safe. 1276 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1277 llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1278 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1279 } 1280 } 1281 1282 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1283 // the call using the normal call path, but using the unmangled 1284 // version of the function name. 1285 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1286 return emitLibraryCall(*this, FD, E, 1287 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1288 1289 // If this is a predefined lib function (e.g. malloc), emit the call 1290 // using exactly the normal call path. 1291 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1292 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1293 1294 // See if we have a target specific intrinsic. 1295 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1296 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1297 if (const char *Prefix = 1298 llvm::Triple::getArchTypePrefix(Target.getTriple().getArch())) 1299 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1300 1301 if (IntrinsicID != Intrinsic::not_intrinsic) { 1302 SmallVector<Value*, 16> Args; 1303 1304 // Find out if any arguments are required to be integer constant 1305 // expressions. 1306 unsigned ICEArguments = 0; 1307 ASTContext::GetBuiltinTypeError Error; 1308 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1309 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1310 1311 Function *F = CGM.getIntrinsic(IntrinsicID); 1312 llvm::FunctionType *FTy = F->getFunctionType(); 1313 1314 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1315 Value *ArgValue; 1316 // If this is a normal argument, just emit it as a scalar. 1317 if ((ICEArguments & (1 << i)) == 0) { 1318 ArgValue = EmitScalarExpr(E->getArg(i)); 1319 } else { 1320 // If this is required to be a constant, constant fold it so that we 1321 // know that the generated intrinsic gets a ConstantInt. 1322 llvm::APSInt Result; 1323 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1324 assert(IsConst && "Constant arg isn't actually constant?"); 1325 (void)IsConst; 1326 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1327 } 1328 1329 // If the intrinsic arg type is different from the builtin arg type 1330 // we need to do a bit cast. 1331 llvm::Type *PTy = FTy->getParamType(i); 1332 if (PTy != ArgValue->getType()) { 1333 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1334 "Must be able to losslessly bit cast to param"); 1335 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1336 } 1337 1338 Args.push_back(ArgValue); 1339 } 1340 1341 Value *V = Builder.CreateCall(F, Args); 1342 QualType BuiltinRetType = E->getType(); 1343 1344 llvm::Type *RetTy = VoidTy; 1345 if (!BuiltinRetType->isVoidType()) 1346 RetTy = ConvertType(BuiltinRetType); 1347 1348 if (RetTy != V->getType()) { 1349 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1350 "Must be able to losslessly bit cast result type"); 1351 V = Builder.CreateBitCast(V, RetTy); 1352 } 1353 1354 return RValue::get(V); 1355 } 1356 1357 // See if we have a target specific builtin that needs to be lowered. 1358 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1359 return RValue::get(V); 1360 1361 ErrorUnsupported(E, "builtin function"); 1362 1363 // Unknown builtin, for now just dump it out and return undef. 1364 if (hasAggregateLLVMType(E->getType())) 1365 return RValue::getAggregate(CreateMemTemp(E->getType())); 1366 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1367 } 1368 1369 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1370 const CallExpr *E) { 1371 switch (Target.getTriple().getArch()) { 1372 case llvm::Triple::arm: 1373 case llvm::Triple::thumb: 1374 return EmitARMBuiltinExpr(BuiltinID, E); 1375 case llvm::Triple::x86: 1376 case llvm::Triple::x86_64: 1377 return EmitX86BuiltinExpr(BuiltinID, E); 1378 case llvm::Triple::ppc: 1379 case llvm::Triple::ppc64: 1380 return EmitPPCBuiltinExpr(BuiltinID, E); 1381 default: 1382 return 0; 1383 } 1384 } 1385 1386 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1387 NeonTypeFlags TypeFlags) { 1388 int IsQuad = TypeFlags.isQuad(); 1389 switch (TypeFlags.getEltType()) { 1390 case NeonTypeFlags::Int8: 1391 case NeonTypeFlags::Poly8: 1392 return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad); 1393 case NeonTypeFlags::Int16: 1394 case NeonTypeFlags::Poly16: 1395 case NeonTypeFlags::Float16: 1396 return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad); 1397 case NeonTypeFlags::Int32: 1398 return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad); 1399 case NeonTypeFlags::Int64: 1400 return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad); 1401 case NeonTypeFlags::Float32: 1402 return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad); 1403 } 1404 llvm_unreachable("Invalid NeonTypeFlags element type!"); 1405 } 1406 1407 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1408 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1409 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1410 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1411 } 1412 1413 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1414 const char *name, 1415 unsigned shift, bool rightshift) { 1416 unsigned j = 0; 1417 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1418 ai != ae; ++ai, ++j) 1419 if (shift > 0 && shift == j) 1420 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1421 else 1422 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1423 1424 return Builder.CreateCall(F, Ops, name); 1425 } 1426 1427 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1428 bool neg) { 1429 int SV = cast<ConstantInt>(V)->getSExtValue(); 1430 1431 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1432 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1433 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1434 } 1435 1436 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1437 /// alignment of the type referenced by the pointer. Skip over implicit 1438 /// casts. 1439 unsigned CodeGenFunction::GetPointeeAlignment(const Expr *Addr) { 1440 unsigned Align = 1; 1441 // Check if the type is a pointer. The implicit cast operand might not be. 1442 while (Addr->getType()->isPointerType()) { 1443 QualType PtTy = Addr->getType()->getPointeeType(); 1444 1445 // Can't get alignment of incomplete types. 1446 if (!PtTy->isIncompleteType()) { 1447 unsigned NewA = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1448 if (NewA > Align) 1449 Align = NewA; 1450 } 1451 1452 // If the address is an implicit cast, repeat with the cast operand. 1453 if (const ImplicitCastExpr *CastAddr = dyn_cast<ImplicitCastExpr>(Addr)) { 1454 Addr = CastAddr->getSubExpr(); 1455 continue; 1456 } 1457 break; 1458 } 1459 return Align; 1460 } 1461 1462 /// GetPointeeAlignmentValue - Given an expression with a pointer type, find 1463 /// the alignment of the type referenced by the pointer. Skip over implicit 1464 /// casts. Return the alignment as an llvm::Value. 1465 Value *CodeGenFunction::GetPointeeAlignmentValue(const Expr *Addr) { 1466 return llvm::ConstantInt::get(Int32Ty, GetPointeeAlignment(Addr)); 1467 } 1468 1469 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1470 const CallExpr *E) { 1471 if (BuiltinID == ARM::BI__clear_cache) { 1472 const FunctionDecl *FD = E->getDirectCallee(); 1473 // Oddly people write this call without args on occasion and gcc accepts 1474 // it - it's also marked as varargs in the description file. 1475 SmallVector<Value*, 2> Ops; 1476 for (unsigned i = 0; i < E->getNumArgs(); i++) 1477 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1478 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1479 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1480 StringRef Name = FD->getName(); 1481 return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1482 } 1483 1484 if (BuiltinID == ARM::BI__builtin_arm_ldrexd) { 1485 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1486 1487 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1488 Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd"); 1489 1490 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1491 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1492 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1493 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1494 1495 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1496 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1497 return Builder.CreateOr(Val, Val1); 1498 } 1499 1500 if (BuiltinID == ARM::BI__builtin_arm_strexd) { 1501 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 1502 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 1503 1504 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 1505 Value *Tmp = Builder.CreateAlloca(Int64Ty, One); 1506 Value *Val = EmitScalarExpr(E->getArg(0)); 1507 Builder.CreateStore(Val, Tmp); 1508 1509 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 1510 Val = Builder.CreateLoad(LdPtr); 1511 1512 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 1513 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 1514 Value *StPtr = EmitScalarExpr(E->getArg(1)); 1515 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 1516 } 1517 1518 SmallVector<Value*, 4> Ops; 1519 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 1520 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1521 1522 // vget_lane and vset_lane are not overloaded and do not have an extra 1523 // argument that specifies the vector type. 1524 switch (BuiltinID) { 1525 default: break; 1526 case ARM::BI__builtin_neon_vget_lane_i8: 1527 case ARM::BI__builtin_neon_vget_lane_i16: 1528 case ARM::BI__builtin_neon_vget_lane_i32: 1529 case ARM::BI__builtin_neon_vget_lane_i64: 1530 case ARM::BI__builtin_neon_vget_lane_f32: 1531 case ARM::BI__builtin_neon_vgetq_lane_i8: 1532 case ARM::BI__builtin_neon_vgetq_lane_i16: 1533 case ARM::BI__builtin_neon_vgetq_lane_i32: 1534 case ARM::BI__builtin_neon_vgetq_lane_i64: 1535 case ARM::BI__builtin_neon_vgetq_lane_f32: 1536 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 1537 "vget_lane"); 1538 case ARM::BI__builtin_neon_vset_lane_i8: 1539 case ARM::BI__builtin_neon_vset_lane_i16: 1540 case ARM::BI__builtin_neon_vset_lane_i32: 1541 case ARM::BI__builtin_neon_vset_lane_i64: 1542 case ARM::BI__builtin_neon_vset_lane_f32: 1543 case ARM::BI__builtin_neon_vsetq_lane_i8: 1544 case ARM::BI__builtin_neon_vsetq_lane_i16: 1545 case ARM::BI__builtin_neon_vsetq_lane_i32: 1546 case ARM::BI__builtin_neon_vsetq_lane_i64: 1547 case ARM::BI__builtin_neon_vsetq_lane_f32: 1548 Ops.push_back(EmitScalarExpr(E->getArg(2))); 1549 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 1550 } 1551 1552 // Get the last argument, which specifies the vector type. 1553 llvm::APSInt Result; 1554 const Expr *Arg = E->getArg(E->getNumArgs()-1); 1555 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1556 return 0; 1557 1558 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 1559 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 1560 // Determine the overloaded type of this builtin. 1561 llvm::Type *Ty; 1562 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 1563 Ty = FloatTy; 1564 else 1565 Ty = DoubleTy; 1566 1567 // Determine whether this is an unsigned conversion or not. 1568 bool usgn = Result.getZExtValue() == 1; 1569 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 1570 1571 // Call the appropriate intrinsic. 1572 Function *F = CGM.getIntrinsic(Int, Ty); 1573 return Builder.CreateCall(F, Ops, "vcvtr"); 1574 } 1575 1576 // Determine the type of this overloaded NEON intrinsic. 1577 NeonTypeFlags Type(Result.getZExtValue()); 1578 bool usgn = Type.isUnsigned(); 1579 bool quad = Type.isQuad(); 1580 bool rightShift = false; 1581 1582 llvm::VectorType *VTy = GetNeonType(this, Type); 1583 llvm::Type *Ty = VTy; 1584 if (!Ty) 1585 return 0; 1586 1587 unsigned Int; 1588 switch (BuiltinID) { 1589 default: return 0; 1590 case ARM::BI__builtin_neon_vabd_v: 1591 case ARM::BI__builtin_neon_vabdq_v: 1592 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1593 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1594 case ARM::BI__builtin_neon_vabs_v: 1595 case ARM::BI__builtin_neon_vabsq_v: 1596 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 1597 Ops, "vabs"); 1598 case ARM::BI__builtin_neon_vaddhn_v: 1599 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 1600 Ops, "vaddhn"); 1601 case ARM::BI__builtin_neon_vcale_v: 1602 std::swap(Ops[0], Ops[1]); 1603 case ARM::BI__builtin_neon_vcage_v: { 1604 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 1605 return EmitNeonCall(F, Ops, "vcage"); 1606 } 1607 case ARM::BI__builtin_neon_vcaleq_v: 1608 std::swap(Ops[0], Ops[1]); 1609 case ARM::BI__builtin_neon_vcageq_v: { 1610 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1611 return EmitNeonCall(F, Ops, "vcage"); 1612 } 1613 case ARM::BI__builtin_neon_vcalt_v: 1614 std::swap(Ops[0], Ops[1]); 1615 case ARM::BI__builtin_neon_vcagt_v: { 1616 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 1617 return EmitNeonCall(F, Ops, "vcagt"); 1618 } 1619 case ARM::BI__builtin_neon_vcaltq_v: 1620 std::swap(Ops[0], Ops[1]); 1621 case ARM::BI__builtin_neon_vcagtq_v: { 1622 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1623 return EmitNeonCall(F, Ops, "vcagt"); 1624 } 1625 case ARM::BI__builtin_neon_vcls_v: 1626 case ARM::BI__builtin_neon_vclsq_v: { 1627 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1628 return EmitNeonCall(F, Ops, "vcls"); 1629 } 1630 case ARM::BI__builtin_neon_vclz_v: 1631 case ARM::BI__builtin_neon_vclzq_v: { 1632 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vclz, Ty); 1633 return EmitNeonCall(F, Ops, "vclz"); 1634 } 1635 case ARM::BI__builtin_neon_vcnt_v: 1636 case ARM::BI__builtin_neon_vcntq_v: { 1637 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcnt, Ty); 1638 return EmitNeonCall(F, Ops, "vcnt"); 1639 } 1640 case ARM::BI__builtin_neon_vcvt_f16_v: { 1641 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1642 "unexpected vcvt_f16_v builtin"); 1643 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1644 return EmitNeonCall(F, Ops, "vcvt"); 1645 } 1646 case ARM::BI__builtin_neon_vcvt_f32_f16: { 1647 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1648 "unexpected vcvt_f32_f16 builtin"); 1649 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1650 return EmitNeonCall(F, Ops, "vcvt"); 1651 } 1652 case ARM::BI__builtin_neon_vcvt_f32_v: 1653 case ARM::BI__builtin_neon_vcvtq_f32_v: 1654 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1655 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1656 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1657 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1658 case ARM::BI__builtin_neon_vcvt_s32_v: 1659 case ARM::BI__builtin_neon_vcvt_u32_v: 1660 case ARM::BI__builtin_neon_vcvtq_s32_v: 1661 case ARM::BI__builtin_neon_vcvtq_u32_v: { 1662 llvm::Type *FloatTy = 1663 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1664 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 1665 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1666 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1667 } 1668 case ARM::BI__builtin_neon_vcvt_n_f32_v: 1669 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 1670 llvm::Type *FloatTy = 1671 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1672 llvm::Type *Tys[2] = { FloatTy, Ty }; 1673 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 1674 : Intrinsic::arm_neon_vcvtfxs2fp; 1675 Function *F = CGM.getIntrinsic(Int, Tys); 1676 return EmitNeonCall(F, Ops, "vcvt_n"); 1677 } 1678 case ARM::BI__builtin_neon_vcvt_n_s32_v: 1679 case ARM::BI__builtin_neon_vcvt_n_u32_v: 1680 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 1681 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 1682 llvm::Type *FloatTy = 1683 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1684 llvm::Type *Tys[2] = { Ty, FloatTy }; 1685 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu 1686 : Intrinsic::arm_neon_vcvtfp2fxs; 1687 Function *F = CGM.getIntrinsic(Int, Tys); 1688 return EmitNeonCall(F, Ops, "vcvt_n"); 1689 } 1690 case ARM::BI__builtin_neon_vext_v: 1691 case ARM::BI__builtin_neon_vextq_v: { 1692 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1693 SmallVector<Constant*, 16> Indices; 1694 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1695 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1696 1697 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1698 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1699 Value *SV = llvm::ConstantVector::get(Indices); 1700 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1701 } 1702 case ARM::BI__builtin_neon_vhadd_v: 1703 case ARM::BI__builtin_neon_vhaddq_v: 1704 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1705 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1706 case ARM::BI__builtin_neon_vhsub_v: 1707 case ARM::BI__builtin_neon_vhsubq_v: 1708 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1709 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1710 case ARM::BI__builtin_neon_vld1_v: 1711 case ARM::BI__builtin_neon_vld1q_v: 1712 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 1713 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1714 Ops, "vld1"); 1715 case ARM::BI__builtin_neon_vld1_lane_v: 1716 case ARM::BI__builtin_neon_vld1q_lane_v: { 1717 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1718 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1719 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1720 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1721 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 1722 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1723 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 1724 } 1725 case ARM::BI__builtin_neon_vld1_dup_v: 1726 case ARM::BI__builtin_neon_vld1q_dup_v: { 1727 Value *V = UndefValue::get(Ty); 1728 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1729 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1730 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1731 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 1732 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1733 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1734 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 1735 return EmitNeonSplat(Ops[0], CI); 1736 } 1737 case ARM::BI__builtin_neon_vld2_v: 1738 case ARM::BI__builtin_neon_vld2q_v: { 1739 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1740 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1741 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1742 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1743 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1744 return Builder.CreateStore(Ops[1], Ops[0]); 1745 } 1746 case ARM::BI__builtin_neon_vld3_v: 1747 case ARM::BI__builtin_neon_vld3q_v: { 1748 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1749 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1750 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1751 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1752 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1753 return Builder.CreateStore(Ops[1], Ops[0]); 1754 } 1755 case ARM::BI__builtin_neon_vld4_v: 1756 case ARM::BI__builtin_neon_vld4q_v: { 1757 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1758 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1759 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1760 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1761 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1762 return Builder.CreateStore(Ops[1], Ops[0]); 1763 } 1764 case ARM::BI__builtin_neon_vld2_lane_v: 1765 case ARM::BI__builtin_neon_vld2q_lane_v: { 1766 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1767 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1768 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1769 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1770 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 1771 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1772 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1773 return Builder.CreateStore(Ops[1], Ops[0]); 1774 } 1775 case ARM::BI__builtin_neon_vld3_lane_v: 1776 case ARM::BI__builtin_neon_vld3q_lane_v: { 1777 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 1778 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1779 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1780 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1781 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1782 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1783 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1784 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1785 return Builder.CreateStore(Ops[1], Ops[0]); 1786 } 1787 case ARM::BI__builtin_neon_vld4_lane_v: 1788 case ARM::BI__builtin_neon_vld4q_lane_v: { 1789 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 1790 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1791 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1792 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1793 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 1794 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1795 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1796 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1797 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1798 return Builder.CreateStore(Ops[1], Ops[0]); 1799 } 1800 case ARM::BI__builtin_neon_vld2_dup_v: 1801 case ARM::BI__builtin_neon_vld3_dup_v: 1802 case ARM::BI__builtin_neon_vld4_dup_v: { 1803 // Handle 64-bit elements as a special-case. There is no "dup" needed. 1804 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 1805 switch (BuiltinID) { 1806 case ARM::BI__builtin_neon_vld2_dup_v: 1807 Int = Intrinsic::arm_neon_vld2; 1808 break; 1809 case ARM::BI__builtin_neon_vld3_dup_v: 1810 Int = Intrinsic::arm_neon_vld3; 1811 break; 1812 case ARM::BI__builtin_neon_vld4_dup_v: 1813 Int = Intrinsic::arm_neon_vld4; 1814 break; 1815 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1816 } 1817 Function *F = CGM.getIntrinsic(Int, Ty); 1818 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1819 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 1820 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1821 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1822 return Builder.CreateStore(Ops[1], Ops[0]); 1823 } 1824 switch (BuiltinID) { 1825 case ARM::BI__builtin_neon_vld2_dup_v: 1826 Int = Intrinsic::arm_neon_vld2lane; 1827 break; 1828 case ARM::BI__builtin_neon_vld3_dup_v: 1829 Int = Intrinsic::arm_neon_vld3lane; 1830 break; 1831 case ARM::BI__builtin_neon_vld4_dup_v: 1832 Int = Intrinsic::arm_neon_vld4lane; 1833 break; 1834 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1835 } 1836 Function *F = CGM.getIntrinsic(Int, Ty); 1837 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 1838 1839 SmallVector<Value*, 6> Args; 1840 Args.push_back(Ops[1]); 1841 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 1842 1843 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1844 Args.push_back(CI); 1845 Args.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1846 1847 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 1848 // splat lane 0 to all elts in each vector of the result. 1849 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1850 Value *Val = Builder.CreateExtractValue(Ops[1], i); 1851 Value *Elt = Builder.CreateBitCast(Val, Ty); 1852 Elt = EmitNeonSplat(Elt, CI); 1853 Elt = Builder.CreateBitCast(Elt, Val->getType()); 1854 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 1855 } 1856 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1857 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1858 return Builder.CreateStore(Ops[1], Ops[0]); 1859 } 1860 case ARM::BI__builtin_neon_vmax_v: 1861 case ARM::BI__builtin_neon_vmaxq_v: 1862 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 1863 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 1864 case ARM::BI__builtin_neon_vmin_v: 1865 case ARM::BI__builtin_neon_vminq_v: 1866 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 1867 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 1868 case ARM::BI__builtin_neon_vmovl_v: { 1869 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 1870 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 1871 if (usgn) 1872 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 1873 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 1874 } 1875 case ARM::BI__builtin_neon_vmovn_v: { 1876 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 1877 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 1878 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 1879 } 1880 case ARM::BI__builtin_neon_vmul_v: 1881 case ARM::BI__builtin_neon_vmulq_v: 1882 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 1883 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 1884 Ops, "vmul"); 1885 case ARM::BI__builtin_neon_vmull_v: 1886 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 1887 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 1888 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 1889 case ARM::BI__builtin_neon_vpadal_v: 1890 case ARM::BI__builtin_neon_vpadalq_v: { 1891 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 1892 // The source operand type has twice as many elements of half the size. 1893 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1894 llvm::Type *EltTy = 1895 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1896 llvm::Type *NarrowTy = 1897 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1898 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1899 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 1900 } 1901 case ARM::BI__builtin_neon_vpadd_v: 1902 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 1903 Ops, "vpadd"); 1904 case ARM::BI__builtin_neon_vpaddl_v: 1905 case ARM::BI__builtin_neon_vpaddlq_v: { 1906 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 1907 // The source operand type has twice as many elements of half the size. 1908 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1909 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1910 llvm::Type *NarrowTy = 1911 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1912 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1913 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 1914 } 1915 case ARM::BI__builtin_neon_vpmax_v: 1916 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 1917 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 1918 case ARM::BI__builtin_neon_vpmin_v: 1919 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 1920 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 1921 case ARM::BI__builtin_neon_vqabs_v: 1922 case ARM::BI__builtin_neon_vqabsq_v: 1923 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 1924 Ops, "vqabs"); 1925 case ARM::BI__builtin_neon_vqadd_v: 1926 case ARM::BI__builtin_neon_vqaddq_v: 1927 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 1928 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 1929 case ARM::BI__builtin_neon_vqdmlal_v: 1930 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 1931 Ops, "vqdmlal"); 1932 case ARM::BI__builtin_neon_vqdmlsl_v: 1933 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 1934 Ops, "vqdmlsl"); 1935 case ARM::BI__builtin_neon_vqdmulh_v: 1936 case ARM::BI__builtin_neon_vqdmulhq_v: 1937 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 1938 Ops, "vqdmulh"); 1939 case ARM::BI__builtin_neon_vqdmull_v: 1940 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 1941 Ops, "vqdmull"); 1942 case ARM::BI__builtin_neon_vqmovn_v: 1943 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 1944 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 1945 case ARM::BI__builtin_neon_vqmovun_v: 1946 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 1947 Ops, "vqdmull"); 1948 case ARM::BI__builtin_neon_vqneg_v: 1949 case ARM::BI__builtin_neon_vqnegq_v: 1950 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 1951 Ops, "vqneg"); 1952 case ARM::BI__builtin_neon_vqrdmulh_v: 1953 case ARM::BI__builtin_neon_vqrdmulhq_v: 1954 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 1955 Ops, "vqrdmulh"); 1956 case ARM::BI__builtin_neon_vqrshl_v: 1957 case ARM::BI__builtin_neon_vqrshlq_v: 1958 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 1959 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 1960 case ARM::BI__builtin_neon_vqrshrn_n_v: 1961 Int = usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 1962 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 1963 1, true); 1964 case ARM::BI__builtin_neon_vqrshrun_n_v: 1965 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 1966 Ops, "vqrshrun_n", 1, true); 1967 case ARM::BI__builtin_neon_vqshl_v: 1968 case ARM::BI__builtin_neon_vqshlq_v: 1969 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1970 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 1971 case ARM::BI__builtin_neon_vqshl_n_v: 1972 case ARM::BI__builtin_neon_vqshlq_n_v: 1973 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1974 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 1975 1, false); 1976 case ARM::BI__builtin_neon_vqshlu_n_v: 1977 case ARM::BI__builtin_neon_vqshluq_n_v: 1978 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 1979 Ops, "vqshlu", 1, false); 1980 case ARM::BI__builtin_neon_vqshrn_n_v: 1981 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 1982 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 1983 1, true); 1984 case ARM::BI__builtin_neon_vqshrun_n_v: 1985 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 1986 Ops, "vqshrun_n", 1, true); 1987 case ARM::BI__builtin_neon_vqsub_v: 1988 case ARM::BI__builtin_neon_vqsubq_v: 1989 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 1990 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 1991 case ARM::BI__builtin_neon_vraddhn_v: 1992 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 1993 Ops, "vraddhn"); 1994 case ARM::BI__builtin_neon_vrecpe_v: 1995 case ARM::BI__builtin_neon_vrecpeq_v: 1996 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 1997 Ops, "vrecpe"); 1998 case ARM::BI__builtin_neon_vrecps_v: 1999 case ARM::BI__builtin_neon_vrecpsq_v: 2000 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 2001 Ops, "vrecps"); 2002 case ARM::BI__builtin_neon_vrhadd_v: 2003 case ARM::BI__builtin_neon_vrhaddq_v: 2004 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 2005 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 2006 case ARM::BI__builtin_neon_vrshl_v: 2007 case ARM::BI__builtin_neon_vrshlq_v: 2008 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2009 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 2010 case ARM::BI__builtin_neon_vrshrn_n_v: 2011 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 2012 Ops, "vrshrn_n", 1, true); 2013 case ARM::BI__builtin_neon_vrshr_n_v: 2014 case ARM::BI__builtin_neon_vrshrq_n_v: 2015 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2016 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 2017 case ARM::BI__builtin_neon_vrsqrte_v: 2018 case ARM::BI__builtin_neon_vrsqrteq_v: 2019 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 2020 Ops, "vrsqrte"); 2021 case ARM::BI__builtin_neon_vrsqrts_v: 2022 case ARM::BI__builtin_neon_vrsqrtsq_v: 2023 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 2024 Ops, "vrsqrts"); 2025 case ARM::BI__builtin_neon_vrsra_n_v: 2026 case ARM::BI__builtin_neon_vrsraq_n_v: 2027 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2028 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2029 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 2030 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2031 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 2032 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 2033 case ARM::BI__builtin_neon_vrsubhn_v: 2034 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 2035 Ops, "vrsubhn"); 2036 case ARM::BI__builtin_neon_vshl_v: 2037 case ARM::BI__builtin_neon_vshlq_v: 2038 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 2039 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 2040 case ARM::BI__builtin_neon_vshll_n_v: 2041 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 2042 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 2043 case ARM::BI__builtin_neon_vshl_n_v: 2044 case ARM::BI__builtin_neon_vshlq_n_v: 2045 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2046 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], "vshl_n"); 2047 case ARM::BI__builtin_neon_vshrn_n_v: 2048 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 2049 Ops, "vshrn_n", 1, true); 2050 case ARM::BI__builtin_neon_vshr_n_v: 2051 case ARM::BI__builtin_neon_vshrq_n_v: 2052 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2053 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2054 if (usgn) 2055 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 2056 else 2057 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 2058 case ARM::BI__builtin_neon_vsri_n_v: 2059 case ARM::BI__builtin_neon_vsriq_n_v: 2060 rightShift = true; 2061 case ARM::BI__builtin_neon_vsli_n_v: 2062 case ARM::BI__builtin_neon_vsliq_n_v: 2063 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 2064 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 2065 Ops, "vsli_n"); 2066 case ARM::BI__builtin_neon_vsra_n_v: 2067 case ARM::BI__builtin_neon_vsraq_n_v: 2068 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2069 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2070 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 2071 if (usgn) 2072 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 2073 else 2074 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 2075 return Builder.CreateAdd(Ops[0], Ops[1]); 2076 case ARM::BI__builtin_neon_vst1_v: 2077 case ARM::BI__builtin_neon_vst1q_v: 2078 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2079 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 2080 Ops, ""); 2081 case ARM::BI__builtin_neon_vst1_lane_v: 2082 case ARM::BI__builtin_neon_vst1q_lane_v: { 2083 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2084 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 2085 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2086 StoreInst *St = Builder.CreateStore(Ops[1], 2087 Builder.CreateBitCast(Ops[0], Ty)); 2088 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 2089 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2090 return St; 2091 } 2092 case ARM::BI__builtin_neon_vst2_v: 2093 case ARM::BI__builtin_neon_vst2q_v: 2094 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2095 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 2096 Ops, ""); 2097 case ARM::BI__builtin_neon_vst2_lane_v: 2098 case ARM::BI__builtin_neon_vst2q_lane_v: 2099 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2100 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 2101 Ops, ""); 2102 case ARM::BI__builtin_neon_vst3_v: 2103 case ARM::BI__builtin_neon_vst3q_v: 2104 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2105 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 2106 Ops, ""); 2107 case ARM::BI__builtin_neon_vst3_lane_v: 2108 case ARM::BI__builtin_neon_vst3q_lane_v: 2109 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2110 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 2111 Ops, ""); 2112 case ARM::BI__builtin_neon_vst4_v: 2113 case ARM::BI__builtin_neon_vst4q_v: 2114 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2115 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 2116 Ops, ""); 2117 case ARM::BI__builtin_neon_vst4_lane_v: 2118 case ARM::BI__builtin_neon_vst4q_lane_v: 2119 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2120 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 2121 Ops, ""); 2122 case ARM::BI__builtin_neon_vsubhn_v: 2123 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 2124 Ops, "vsubhn"); 2125 case ARM::BI__builtin_neon_vtbl1_v: 2126 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 2127 Ops, "vtbl1"); 2128 case ARM::BI__builtin_neon_vtbl2_v: 2129 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 2130 Ops, "vtbl2"); 2131 case ARM::BI__builtin_neon_vtbl3_v: 2132 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 2133 Ops, "vtbl3"); 2134 case ARM::BI__builtin_neon_vtbl4_v: 2135 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 2136 Ops, "vtbl4"); 2137 case ARM::BI__builtin_neon_vtbx1_v: 2138 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 2139 Ops, "vtbx1"); 2140 case ARM::BI__builtin_neon_vtbx2_v: 2141 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 2142 Ops, "vtbx2"); 2143 case ARM::BI__builtin_neon_vtbx3_v: 2144 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 2145 Ops, "vtbx3"); 2146 case ARM::BI__builtin_neon_vtbx4_v: 2147 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 2148 Ops, "vtbx4"); 2149 case ARM::BI__builtin_neon_vtst_v: 2150 case ARM::BI__builtin_neon_vtstq_v: { 2151 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2152 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2153 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2154 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2155 ConstantAggregateZero::get(Ty)); 2156 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2157 } 2158 case ARM::BI__builtin_neon_vtrn_v: 2159 case ARM::BI__builtin_neon_vtrnq_v: { 2160 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2161 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2162 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2163 Value *SV = 0; 2164 2165 for (unsigned vi = 0; vi != 2; ++vi) { 2166 SmallVector<Constant*, 16> Indices; 2167 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2168 Indices.push_back(Builder.getInt32(i+vi)); 2169 Indices.push_back(Builder.getInt32(i+e+vi)); 2170 } 2171 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2172 SV = llvm::ConstantVector::get(Indices); 2173 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2174 SV = Builder.CreateStore(SV, Addr); 2175 } 2176 return SV; 2177 } 2178 case ARM::BI__builtin_neon_vuzp_v: 2179 case ARM::BI__builtin_neon_vuzpq_v: { 2180 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2181 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2182 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2183 Value *SV = 0; 2184 2185 for (unsigned vi = 0; vi != 2; ++vi) { 2186 SmallVector<Constant*, 16> Indices; 2187 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2188 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2189 2190 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2191 SV = llvm::ConstantVector::get(Indices); 2192 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2193 SV = Builder.CreateStore(SV, Addr); 2194 } 2195 return SV; 2196 } 2197 case ARM::BI__builtin_neon_vzip_v: 2198 case ARM::BI__builtin_neon_vzipq_v: { 2199 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2200 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2201 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2202 Value *SV = 0; 2203 2204 for (unsigned vi = 0; vi != 2; ++vi) { 2205 SmallVector<Constant*, 16> Indices; 2206 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2207 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2208 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2209 } 2210 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2211 SV = llvm::ConstantVector::get(Indices); 2212 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2213 SV = Builder.CreateStore(SV, Addr); 2214 } 2215 return SV; 2216 } 2217 } 2218 } 2219 2220 llvm::Value *CodeGenFunction:: 2221 BuildVector(ArrayRef<llvm::Value*> Ops) { 2222 assert((Ops.size() & (Ops.size() - 1)) == 0 && 2223 "Not a power-of-two sized vector!"); 2224 bool AllConstants = true; 2225 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 2226 AllConstants &= isa<Constant>(Ops[i]); 2227 2228 // If this is a constant vector, create a ConstantVector. 2229 if (AllConstants) { 2230 SmallVector<llvm::Constant*, 16> CstOps; 2231 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2232 CstOps.push_back(cast<Constant>(Ops[i])); 2233 return llvm::ConstantVector::get(CstOps); 2234 } 2235 2236 // Otherwise, insertelement the values to build the vector. 2237 Value *Result = 2238 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 2239 2240 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2241 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 2242 2243 return Result; 2244 } 2245 2246 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 2247 const CallExpr *E) { 2248 SmallVector<Value*, 4> Ops; 2249 2250 // Find out if any arguments are required to be integer constant expressions. 2251 unsigned ICEArguments = 0; 2252 ASTContext::GetBuiltinTypeError Error; 2253 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2254 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2255 2256 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2257 // If this is a normal argument, just emit it as a scalar. 2258 if ((ICEArguments & (1 << i)) == 0) { 2259 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2260 continue; 2261 } 2262 2263 // If this is required to be a constant, constant fold it so that we know 2264 // that the generated intrinsic gets a ConstantInt. 2265 llvm::APSInt Result; 2266 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 2267 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 2268 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 2269 } 2270 2271 switch (BuiltinID) { 2272 default: return 0; 2273 case X86::BI__builtin_ia32_vec_init_v8qi: 2274 case X86::BI__builtin_ia32_vec_init_v4hi: 2275 case X86::BI__builtin_ia32_vec_init_v2si: 2276 return Builder.CreateBitCast(BuildVector(Ops), 2277 llvm::Type::getX86_MMXTy(getLLVMContext())); 2278 case X86::BI__builtin_ia32_vec_ext_v2si: 2279 return Builder.CreateExtractElement(Ops[0], 2280 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2281 case X86::BI__builtin_ia32_ldmxcsr: { 2282 llvm::Type *PtrTy = Int8PtrTy; 2283 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2284 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2285 Builder.CreateStore(Ops[0], Tmp); 2286 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2287 Builder.CreateBitCast(Tmp, PtrTy)); 2288 } 2289 case X86::BI__builtin_ia32_stmxcsr: { 2290 llvm::Type *PtrTy = Int8PtrTy; 2291 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2292 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2293 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2294 Builder.CreateBitCast(Tmp, PtrTy)); 2295 return Builder.CreateLoad(Tmp, "stmxcsr"); 2296 } 2297 case X86::BI__builtin_ia32_storehps: 2298 case X86::BI__builtin_ia32_storelps: { 2299 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2300 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2301 2302 // cast val v2i64 2303 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2304 2305 // extract (0, 1) 2306 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2307 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2308 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2309 2310 // cast pointer to i64 & store 2311 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2312 return Builder.CreateStore(Ops[1], Ops[0]); 2313 } 2314 case X86::BI__builtin_ia32_palignr: { 2315 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2316 2317 // If palignr is shifting the pair of input vectors less than 9 bytes, 2318 // emit a shuffle instruction. 2319 if (shiftVal <= 8) { 2320 SmallVector<llvm::Constant*, 8> Indices; 2321 for (unsigned i = 0; i != 8; ++i) 2322 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2323 2324 Value* SV = llvm::ConstantVector::get(Indices); 2325 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2326 } 2327 2328 // If palignr is shifting the pair of input vectors more than 8 but less 2329 // than 16 bytes, emit a logical right shift of the destination. 2330 if (shiftVal < 16) { 2331 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2332 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2333 2334 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2335 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2336 2337 // create i32 constant 2338 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2339 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2340 } 2341 2342 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2343 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2344 } 2345 case X86::BI__builtin_ia32_palignr128: { 2346 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2347 2348 // If palignr is shifting the pair of input vectors less than 17 bytes, 2349 // emit a shuffle instruction. 2350 if (shiftVal <= 16) { 2351 SmallVector<llvm::Constant*, 16> Indices; 2352 for (unsigned i = 0; i != 16; ++i) 2353 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2354 2355 Value* SV = llvm::ConstantVector::get(Indices); 2356 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2357 } 2358 2359 // If palignr is shifting the pair of input vectors more than 16 but less 2360 // than 32 bytes, emit a logical right shift of the destination. 2361 if (shiftVal < 32) { 2362 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2363 2364 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2365 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2366 2367 // create i32 constant 2368 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 2369 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2370 } 2371 2372 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2373 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2374 } 2375 case X86::BI__builtin_ia32_palignr256: { 2376 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2377 2378 // If palignr is shifting the pair of input vectors less than 17 bytes, 2379 // emit a shuffle instruction. 2380 if (shiftVal <= 16) { 2381 SmallVector<llvm::Constant*, 32> Indices; 2382 // 256-bit palignr operates on 128-bit lanes so we need to handle that 2383 for (unsigned l = 0; l != 2; ++l) { 2384 unsigned LaneStart = l * 16; 2385 unsigned LaneEnd = (l+1) * 16; 2386 for (unsigned i = 0; i != 16; ++i) { 2387 unsigned Idx = shiftVal + i + LaneStart; 2388 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 2389 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 2390 } 2391 } 2392 2393 Value* SV = llvm::ConstantVector::get(Indices); 2394 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2395 } 2396 2397 // If palignr is shifting the pair of input vectors more than 16 but less 2398 // than 32 bytes, emit a logical right shift of the destination. 2399 if (shiftVal < 32) { 2400 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 2401 2402 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2403 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2404 2405 // create i32 constant 2406 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 2407 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2408 } 2409 2410 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2411 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2412 } 2413 case X86::BI__builtin_ia32_movntps: 2414 case X86::BI__builtin_ia32_movntps256: 2415 case X86::BI__builtin_ia32_movntpd: 2416 case X86::BI__builtin_ia32_movntpd256: 2417 case X86::BI__builtin_ia32_movntdq: 2418 case X86::BI__builtin_ia32_movntdq256: 2419 case X86::BI__builtin_ia32_movnti: { 2420 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 2421 Builder.getInt32(1)); 2422 2423 // Convert the type of the pointer to a pointer to the stored type. 2424 Value *BC = Builder.CreateBitCast(Ops[0], 2425 llvm::PointerType::getUnqual(Ops[1]->getType()), 2426 "cast"); 2427 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 2428 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 2429 SI->setAlignment(16); 2430 return SI; 2431 } 2432 // 3DNow! 2433 case X86::BI__builtin_ia32_pswapdsf: 2434 case X86::BI__builtin_ia32_pswapdsi: { 2435 const char *name = 0; 2436 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2437 switch(BuiltinID) { 2438 default: llvm_unreachable("Unsupported intrinsic!"); 2439 case X86::BI__builtin_ia32_pswapdsf: 2440 case X86::BI__builtin_ia32_pswapdsi: 2441 name = "pswapd"; 2442 ID = Intrinsic::x86_3dnowa_pswapd; 2443 break; 2444 } 2445 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 2446 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 2447 llvm::Function *F = CGM.getIntrinsic(ID); 2448 return Builder.CreateCall(F, Ops, name); 2449 } 2450 } 2451 } 2452 2453 2454 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 2455 const CallExpr *E) { 2456 SmallVector<Value*, 4> Ops; 2457 2458 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 2459 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2460 2461 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2462 2463 switch (BuiltinID) { 2464 default: return 0; 2465 2466 // vec_ld, vec_lvsl, vec_lvsr 2467 case PPC::BI__builtin_altivec_lvx: 2468 case PPC::BI__builtin_altivec_lvxl: 2469 case PPC::BI__builtin_altivec_lvebx: 2470 case PPC::BI__builtin_altivec_lvehx: 2471 case PPC::BI__builtin_altivec_lvewx: 2472 case PPC::BI__builtin_altivec_lvsl: 2473 case PPC::BI__builtin_altivec_lvsr: 2474 { 2475 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 2476 2477 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 2478 Ops.pop_back(); 2479 2480 switch (BuiltinID) { 2481 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 2482 case PPC::BI__builtin_altivec_lvx: 2483 ID = Intrinsic::ppc_altivec_lvx; 2484 break; 2485 case PPC::BI__builtin_altivec_lvxl: 2486 ID = Intrinsic::ppc_altivec_lvxl; 2487 break; 2488 case PPC::BI__builtin_altivec_lvebx: 2489 ID = Intrinsic::ppc_altivec_lvebx; 2490 break; 2491 case PPC::BI__builtin_altivec_lvehx: 2492 ID = Intrinsic::ppc_altivec_lvehx; 2493 break; 2494 case PPC::BI__builtin_altivec_lvewx: 2495 ID = Intrinsic::ppc_altivec_lvewx; 2496 break; 2497 case PPC::BI__builtin_altivec_lvsl: 2498 ID = Intrinsic::ppc_altivec_lvsl; 2499 break; 2500 case PPC::BI__builtin_altivec_lvsr: 2501 ID = Intrinsic::ppc_altivec_lvsr; 2502 break; 2503 } 2504 llvm::Function *F = CGM.getIntrinsic(ID); 2505 return Builder.CreateCall(F, Ops, ""); 2506 } 2507 2508 // vec_st 2509 case PPC::BI__builtin_altivec_stvx: 2510 case PPC::BI__builtin_altivec_stvxl: 2511 case PPC::BI__builtin_altivec_stvebx: 2512 case PPC::BI__builtin_altivec_stvehx: 2513 case PPC::BI__builtin_altivec_stvewx: 2514 { 2515 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 2516 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 2517 Ops.pop_back(); 2518 2519 switch (BuiltinID) { 2520 default: llvm_unreachable("Unsupported st intrinsic!"); 2521 case PPC::BI__builtin_altivec_stvx: 2522 ID = Intrinsic::ppc_altivec_stvx; 2523 break; 2524 case PPC::BI__builtin_altivec_stvxl: 2525 ID = Intrinsic::ppc_altivec_stvxl; 2526 break; 2527 case PPC::BI__builtin_altivec_stvebx: 2528 ID = Intrinsic::ppc_altivec_stvebx; 2529 break; 2530 case PPC::BI__builtin_altivec_stvehx: 2531 ID = Intrinsic::ppc_altivec_stvehx; 2532 break; 2533 case PPC::BI__builtin_altivec_stvewx: 2534 ID = Intrinsic::ppc_altivec_stvewx; 2535 break; 2536 } 2537 llvm::Function *F = CGM.getIntrinsic(ID); 2538 return Builder.CreateCall(F, Ops, ""); 2539 } 2540 } 2541 } 2542