CGBuiltin.cpp revision d4b32e46517358f34e8cfbea35010adfcc3786e0
1//===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This contains code to emit Builtin calls as LLVM code. 11// 12//===----------------------------------------------------------------------===// 13 14#include "CodeGenFunction.h" 15#include "CodeGenModule.h" 16#include "clang/Basic/TargetInfo.h" 17#include "clang/AST/ASTContext.h" 18#include "clang/AST/Builtins.h" 19#include "clang/AST/Expr.h" 20#include "clang/AST/TargetBuiltins.h" 21#include "llvm/Constants.h" 22#include "llvm/Function.h" 23#include "llvm/Intrinsics.h" 24using namespace clang; 25using namespace CodeGen; 26using namespace llvm; 27 28/// Utility to insert an atomic instruction based Instrinsic::ID and 29// the expression node 30static RValue EmitBinaryAtomic(CodeGenFunction& CFG, 31 Intrinsic::ID Id, const CallExpr *E) { 32 const llvm::Type *ResType = CFG.ConvertType(E->getType()); 33 Value *AtomF = CFG.CGM.getIntrinsic(Id, &ResType, 1); 34 return RValue::get(CFG.Builder.CreateCall2(AtomF, 35 CFG.EmitScalarExpr(E->getArg(0)), 36 CFG.EmitScalarExpr(E->getArg(1)))); 37} 38 39RValue CodeGenFunction::EmitBuiltinExpr(unsigned BuiltinID, const CallExpr *E) { 40 switch (BuiltinID) { 41 default: { 42 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 43 return EmitCallExpr(CGM.getBuiltinLibFunction(BuiltinID), 44 E->getCallee()->getType(), E->arg_begin(), 45 E->getNumArgs()); 46 47 // See if we have a target specific intrinsic. 48 Intrinsic::ID IntrinsicID; 49 const char *TargetPrefix = Target.getTargetPrefix(); 50 const char *BuiltinName = getContext().BuiltinInfo.GetName(BuiltinID); 51#define GET_LLVM_INTRINSIC_FOR_GCC_BUILTIN 52#include "llvm/Intrinsics.gen" 53#undef GET_LLVM_INTRINSIC_FOR_GCC_BUILTIN 54 55 if (IntrinsicID != Intrinsic::not_intrinsic) { 56 SmallVector<Value*, 16> Args; 57 58 Function *F = CGM.getIntrinsic(IntrinsicID); 59 const llvm::FunctionType *FTy = F->getFunctionType(); 60 61 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 62 Value *ArgValue = EmitScalarExpr(E->getArg(i)); 63 64 // If the intrinsic arg type is different from the builtin arg type 65 // we need to do a bit cast. 66 const llvm::Type *PTy = FTy->getParamType(i); 67 if (PTy != ArgValue->getType()) { 68 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 69 "Must be able to losslessly bit cast to param"); 70 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 71 } 72 73 Args.push_back(ArgValue); 74 } 75 76 Value *V = Builder.CreateCall(F, &Args[0], &Args[0] + Args.size()); 77 QualType BuiltinRetType = E->getType(); 78 79 const llvm::Type *RetTy = llvm::Type::VoidTy; 80 if (!BuiltinRetType->isVoidType()) RetTy = ConvertType(BuiltinRetType); 81 82 if (RetTy != V->getType()) { 83 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 84 "Must be able to losslessly bit cast result type"); 85 V = Builder.CreateBitCast(V, RetTy); 86 } 87 88 return RValue::get(V); 89 } 90 91 // See if we have a target specific builtin that needs to be lowered. 92 Value *V = 0; 93 94 if (strcmp(TargetPrefix, "x86") == 0) 95 V = EmitX86BuiltinExpr(BuiltinID, E); 96 else if (strcmp(TargetPrefix, "ppc") == 0) 97 V = EmitPPCBuiltinExpr(BuiltinID, E); 98 99 if (V) 100 return RValue::get(V); 101 102 WarnUnsupported(E, "builtin function"); 103 104 // Unknown builtin, for now just dump it out and return undef. 105 if (hasAggregateLLVMType(E->getType())) 106 return RValue::getAggregate(CreateTempAlloca(ConvertType(E->getType()))); 107 return RValue::get(UndefValue::get(ConvertType(E->getType()))); 108 } 109 case Builtin::BI__builtin___CFStringMakeConstantString: { 110 const Expr *Arg = E->getArg(0); 111 112 while (1) { 113 if (const ParenExpr *PE = dyn_cast<ParenExpr>(Arg)) 114 Arg = PE->getSubExpr(); 115 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Arg)) 116 Arg = CE->getSubExpr(); 117 else 118 break; 119 } 120 121 const StringLiteral *Literal = cast<StringLiteral>(Arg); 122 std::string S(Literal->getStrData(), Literal->getByteLength()); 123 124 return RValue::get(CGM.GetAddrOfConstantCFString(S)); 125 } 126 case Builtin::BI__builtin_va_start: 127 case Builtin::BI__builtin_va_end: { 128 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 129 const llvm::Type *DestType = 130 llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 131 if (ArgValue->getType() != DestType) 132 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 133 ArgValue->getNameStart()); 134 135 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_start) ? 136 Intrinsic::vastart : Intrinsic::vaend; 137 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 138 } 139 case Builtin::BI__builtin_va_copy: { 140 // FIXME: This does not yet handle architectures where va_list is a struct. 141 Value *DstPtr = EmitScalarExpr(E->getArg(0)); 142 Value *SrcValue = EmitScalarExpr(E->getArg(1)); 143 144 Value *SrcPtr = CreateTempAlloca(SrcValue->getType(), "dst_ptr"); 145 146 // FIXME: Volatile 147 Builder.CreateStore(SrcValue, SrcPtr, false); 148 149 const llvm::Type *Type = 150 llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 151 152 DstPtr = Builder.CreateBitCast(DstPtr, Type); 153 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 154 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 155 DstPtr, SrcPtr)); 156 } 157 case Builtin::BI__builtin_classify_type: { 158 APSInt Result(32); 159 if (!E->isBuiltinClassifyType(Result)) 160 assert(0 && "Expr not __builtin_classify_type!"); 161 return RValue::get(ConstantInt::get(Result)); 162 } 163 case Builtin::BI__builtin_constant_p: { 164 APSInt Result(32); 165 // FIXME: Analyze the parameter and check if it is a constant. 166 Result = 0; 167 return RValue::get(ConstantInt::get(Result)); 168 } 169 case Builtin::BI__builtin_abs: { 170 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 171 172 llvm::BinaryOperator *NegOp = 173 Builder.CreateNeg(ArgValue, (ArgValue->getName() + "neg").c_str()); 174 Value *CmpResult = 175 Builder.CreateICmpSGE(ArgValue, NegOp->getOperand(0), "abscond"); 176 Value *Result = 177 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 178 179 return RValue::get(Result); 180 } 181 case Builtin::BI__builtin_ctz: 182 case Builtin::BI__builtin_ctzl: 183 case Builtin::BI__builtin_ctzll: { 184 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 185 186 const llvm::Type *ArgType = ArgValue->getType(); 187 Value *F = CGM.getIntrinsic(Intrinsic::cttz, &ArgType, 1); 188 189 const llvm::Type *ResultType = ConvertType(E->getType()); 190 Value *Result = Builder.CreateCall(F, ArgValue, "tmp"); 191 if (Result->getType() != ResultType) 192 Result = Builder.CreateIntCast(Result, ResultType, "cast"); 193 return RValue::get(Result); 194 } 195 case Builtin::BI__builtin_expect: 196 return RValue::get(EmitScalarExpr(E->getArg(0))); 197 case Builtin::BI__builtin_bswap32: 198 case Builtin::BI__builtin_bswap64: { 199 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 200 const llvm::Type *ArgType = ArgValue->getType(); 201 Value *F = CGM.getIntrinsic(Intrinsic::bswap, &ArgType, 1); 202 return RValue::get(Builder.CreateCall(F, ArgValue, "tmp")); 203 } 204 case Builtin::BI__builtin_inff: { 205 APFloat f(APFloat::IEEEsingle, APFloat::fcInfinity, false); 206 return RValue::get(ConstantFP::get(f)); 207 } 208 case Builtin::BI__builtin_huge_val: 209 case Builtin::BI__builtin_inf: 210 // FIXME: mapping long double onto double. 211 case Builtin::BI__builtin_infl: { 212 APFloat f(APFloat::IEEEdouble, APFloat::fcInfinity, false); 213 return RValue::get(ConstantFP::get(f)); 214 } 215 case Builtin::BI__builtin_isgreater: 216 case Builtin::BI__builtin_isgreaterequal: 217 case Builtin::BI__builtin_isless: 218 case Builtin::BI__builtin_islessequal: 219 case Builtin::BI__builtin_islessgreater: 220 case Builtin::BI__builtin_isunordered: { 221 // Ordered comparisons: we know the arguments to these are matching scalar 222 // floating point values. 223 Value *LHS = EmitScalarExpr(E->getArg(0)); 224 Value *RHS = EmitScalarExpr(E->getArg(1)); 225 226 switch (BuiltinID) { 227 default: assert(0 && "Unknown ordered comparison"); 228 case Builtin::BI__builtin_isgreater: 229 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 230 break; 231 case Builtin::BI__builtin_isgreaterequal: 232 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 233 break; 234 case Builtin::BI__builtin_isless: 235 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 236 break; 237 case Builtin::BI__builtin_islessequal: 238 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 239 break; 240 case Builtin::BI__builtin_islessgreater: 241 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 242 break; 243 case Builtin::BI__builtin_isunordered: 244 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 245 break; 246 } 247 // ZExt bool to int type. 248 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()), 249 "tmp")); 250 } 251 case Builtin::BI__builtin_alloca: 252 return RValue::get(Builder.CreateAlloca(llvm::Type::Int8Ty, 253 EmitScalarExpr(E->getArg(0)), 254 "tmp")); 255 case Builtin::BI__builtin_memcpy: { 256 Value* MemCpyOps[4] = { 257 EmitScalarExpr(E->getArg(0)), 258 EmitScalarExpr(E->getArg(1)), 259 EmitScalarExpr(E->getArg(2)), 260 llvm::ConstantInt::get(llvm::Type::Int32Ty, 0) 261 }; 262 Builder.CreateCall(CGM.getMemCpyFn(), MemCpyOps, MemCpyOps+4); 263 return RValue::get(MemCpyOps[0]); 264 } 265 case Builtin::BI__sync_fetch_and_add: 266 return EmitBinaryAtomic(*this, Intrinsic::atomic_las, E); 267 case Builtin::BI__sync_fetch_and_sub: 268 return EmitBinaryAtomic(*this, Intrinsic::atomic_lss, E); 269 case Builtin::BI__sync_fetch_and_min: 270 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_min, E); 271 case Builtin::BI__sync_fetch_and_max: 272 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_max, E); 273 case Builtin::BI__sync_fetch_and_umin: 274 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_umin, E); 275 case Builtin::BI__sync_fetch_and_umax: 276 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_umax, E); 277 case Builtin::BI__sync_fetch_and_and: 278 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_and, E); 279 case Builtin::BI__sync_fetch_and_or: 280 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_or, E); 281 case Builtin::BI__sync_fetch_and_xor: 282 return EmitBinaryAtomic(*this, Intrinsic::atomic_load_xor, E); 283 case Builtin::BI__sync_val_compare_and_swap: { 284 Value *Args[3]; 285 Args[0]= EmitScalarExpr(E->getArg(0)); 286 Args[1] = EmitScalarExpr(E->getArg(1)); 287 Args[2] = EmitScalarExpr(E->getArg(2)); 288 const llvm::Type *ResType = ConvertType(E->getType()); 289 Value *AtomF = CGM.getIntrinsic(Intrinsic::atomic_lcs, &ResType, 1); 290 return RValue::get(Builder.CreateCall(AtomF, &Args[0], &Args[1]+2)); 291 } 292 case Builtin::BI__sync_lock_test_and_set: 293 return EmitBinaryAtomic(*this, Intrinsic::atomic_swap, E); 294 } 295 return RValue::get(0); 296} 297 298Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 299 const CallExpr *E) { 300 301 llvm::SmallVector<Value*, 4> Ops; 302 303 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 304 Ops.push_back(EmitScalarExpr(E->getArg(i))); 305 306 switch (BuiltinID) { 307 default: return 0; 308 case X86::BI__builtin_ia32_mulps: 309 return Builder.CreateMul(Ops[0], Ops[1], "mulps"); 310 case X86::BI__builtin_ia32_mulpd: 311 return Builder.CreateMul(Ops[0], Ops[1], "mulpd"); 312 case X86::BI__builtin_ia32_pand: 313 case X86::BI__builtin_ia32_pand128: 314 return Builder.CreateAnd(Ops[0], Ops[1], "pand"); 315 case X86::BI__builtin_ia32_por: 316 case X86::BI__builtin_ia32_por128: 317 return Builder.CreateOr(Ops[0], Ops[1], "por"); 318 case X86::BI__builtin_ia32_pxor: 319 case X86::BI__builtin_ia32_pxor128: 320 return Builder.CreateXor(Ops[0], Ops[1], "pxor"); 321 case X86::BI__builtin_ia32_pandn: 322 case X86::BI__builtin_ia32_pandn128: 323 Ops[0] = Builder.CreateNot(Ops[0], "tmp"); 324 return Builder.CreateAnd(Ops[0], Ops[1], "pandn"); 325 case X86::BI__builtin_ia32_paddb: 326 case X86::BI__builtin_ia32_paddb128: 327 case X86::BI__builtin_ia32_paddd: 328 case X86::BI__builtin_ia32_paddd128: 329 case X86::BI__builtin_ia32_paddq: 330 case X86::BI__builtin_ia32_paddq128: 331 case X86::BI__builtin_ia32_paddw: 332 case X86::BI__builtin_ia32_paddw128: 333 case X86::BI__builtin_ia32_addps: 334 case X86::BI__builtin_ia32_addpd: 335 return Builder.CreateAdd(Ops[0], Ops[1], "add"); 336 case X86::BI__builtin_ia32_psubb: 337 case X86::BI__builtin_ia32_psubb128: 338 case X86::BI__builtin_ia32_psubd: 339 case X86::BI__builtin_ia32_psubd128: 340 case X86::BI__builtin_ia32_psubq: 341 case X86::BI__builtin_ia32_psubq128: 342 case X86::BI__builtin_ia32_psubw: 343 case X86::BI__builtin_ia32_psubw128: 344 case X86::BI__builtin_ia32_subps: 345 case X86::BI__builtin_ia32_subpd: 346 return Builder.CreateSub(Ops[0], Ops[1], "sub"); 347 case X86::BI__builtin_ia32_divps: 348 return Builder.CreateFDiv(Ops[0], Ops[1], "divps"); 349 case X86::BI__builtin_ia32_divpd: 350 return Builder.CreateFDiv(Ops[0], Ops[1], "divpd"); 351 case X86::BI__builtin_ia32_pmullw: 352 case X86::BI__builtin_ia32_pmullw128: 353 return Builder.CreateMul(Ops[0], Ops[1], "pmul"); 354 case X86::BI__builtin_ia32_punpckhbw: 355 return EmitShuffleVector(Ops[0], Ops[1], 4, 12, 5, 13, 6, 14, 7, 15, 356 "punpckhbw"); 357 case X86::BI__builtin_ia32_punpckhbw128: 358 return EmitShuffleVector(Ops[0], Ops[1], 8, 24, 9, 25, 10, 26, 11, 27, 359 12, 28, 13, 29, 14, 30, 15, 31, 360 "punpckhbw"); 361 case X86::BI__builtin_ia32_punpckhwd: 362 return EmitShuffleVector(Ops[0], Ops[1], 2, 6, 3, 7, "punpckhwd"); 363 case X86::BI__builtin_ia32_punpckhwd128: 364 return EmitShuffleVector(Ops[0], Ops[1], 4, 12, 5, 13, 6, 14, 7, 15, 365 "punpckhwd"); 366 case X86::BI__builtin_ia32_punpckhdq: 367 return EmitShuffleVector(Ops[0], Ops[1], 1, 3, "punpckhdq"); 368 case X86::BI__builtin_ia32_punpckhdq128: 369 return EmitShuffleVector(Ops[0], Ops[1], 2, 6, 3, 7, "punpckhdq"); 370 case X86::BI__builtin_ia32_punpcklbw: 371 return EmitShuffleVector(Ops[0], Ops[1], 0, 8, 1, 9, 2, 10, 3, 11, 372 "punpcklbw"); 373 case X86::BI__builtin_ia32_punpcklwd: 374 return EmitShuffleVector(Ops[0], Ops[1], 0, 4, 1, 5, "punpcklwd"); 375 case X86::BI__builtin_ia32_punpckldq: 376 return EmitShuffleVector(Ops[0], Ops[1], 0, 2, "punpckldq"); 377 case X86::BI__builtin_ia32_punpckldq128: 378 return EmitShuffleVector(Ops[0], Ops[1], 0, 4, 1, 5, "punpckldq"); 379 case X86::BI__builtin_ia32_pslldi128: 380 case X86::BI__builtin_ia32_psllqi128: 381 case X86::BI__builtin_ia32_psllwi128: 382 case X86::BI__builtin_ia32_psradi128: 383 case X86::BI__builtin_ia32_psrawi128: 384 case X86::BI__builtin_ia32_psrldi128: 385 case X86::BI__builtin_ia32_psrlqi128: 386 case X86::BI__builtin_ia32_psrlwi128: { 387 Ops[1] = Builder.CreateZExt(Ops[1], llvm::Type::Int64Ty, "zext"); 388 const llvm::Type *Ty = llvm::VectorType::get(llvm::Type::Int64Ty, 2); 389 llvm::Value *Zero = llvm::ConstantInt::get(llvm::Type::Int32Ty, 0); 390 Ops[1] = Builder.CreateInsertElement(llvm::UndefValue::get(Ty), 391 Ops[1], Zero, "insert"); 392 Ops[1] = Builder.CreateBitCast(Ops[1], Ops[0]->getType(), "bitcast"); 393 const char *name = 0; 394 Intrinsic::ID ID = Intrinsic::not_intrinsic; 395 396 switch (BuiltinID) { 397 default: assert(0 && "Unsupported shift intrinsic!"); 398 case X86::BI__builtin_ia32_pslldi128: 399 name = "pslldi"; 400 ID = Intrinsic::x86_sse2_psll_d; 401 break; 402 case X86::BI__builtin_ia32_psllqi128: 403 name = "psllqi"; 404 ID = Intrinsic::x86_sse2_psll_q; 405 break; 406 case X86::BI__builtin_ia32_psllwi128: 407 name = "psllwi"; 408 ID = Intrinsic::x86_sse2_psll_w; 409 break; 410 case X86::BI__builtin_ia32_psradi128: 411 name = "psradi"; 412 ID = Intrinsic::x86_sse2_psra_d; 413 break; 414 case X86::BI__builtin_ia32_psrawi128: 415 name = "psrawi"; 416 ID = Intrinsic::x86_sse2_psra_w; 417 break; 418 case X86::BI__builtin_ia32_psrldi128: 419 name = "psrldi"; 420 ID = Intrinsic::x86_sse2_psrl_d; 421 break; 422 case X86::BI__builtin_ia32_psrlqi128: 423 name = "psrlqi"; 424 ID = Intrinsic::x86_sse2_psrl_q; 425 break; 426 case X86::BI__builtin_ia32_psrlwi128: 427 name = "psrlwi"; 428 ID = Intrinsic::x86_sse2_psrl_w; 429 break; 430 } 431 llvm::Function *F = CGM.getIntrinsic(ID); 432 return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name); 433 } 434 case X86::BI__builtin_ia32_pslldi: 435 case X86::BI__builtin_ia32_psllqi: 436 case X86::BI__builtin_ia32_psllwi: 437 case X86::BI__builtin_ia32_psradi: 438 case X86::BI__builtin_ia32_psrawi: 439 case X86::BI__builtin_ia32_psrldi: 440 case X86::BI__builtin_ia32_psrlqi: 441 case X86::BI__builtin_ia32_psrlwi: { 442 Ops[1] = Builder.CreateZExt(Ops[1], llvm::Type::Int64Ty, "zext"); 443 const llvm::Type *Ty = llvm::VectorType::get(llvm::Type::Int64Ty, 1); 444 Ops[1] = Builder.CreateBitCast(Ops[1], Ty, "bitcast"); 445 const char *name = 0; 446 Intrinsic::ID ID = Intrinsic::not_intrinsic; 447 448 switch (BuiltinID) { 449 default: assert(0 && "Unsupported shift intrinsic!"); 450 case X86::BI__builtin_ia32_pslldi: 451 name = "pslldi"; 452 ID = Intrinsic::x86_mmx_psll_d; 453 break; 454 case X86::BI__builtin_ia32_psllqi: 455 name = "psllqi"; 456 ID = Intrinsic::x86_mmx_psll_q; 457 break; 458 case X86::BI__builtin_ia32_psllwi: 459 name = "psllwi"; 460 ID = Intrinsic::x86_mmx_psll_w; 461 break; 462 case X86::BI__builtin_ia32_psradi: 463 name = "psradi"; 464 ID = Intrinsic::x86_mmx_psra_d; 465 break; 466 case X86::BI__builtin_ia32_psrawi: 467 name = "psrawi"; 468 ID = Intrinsic::x86_mmx_psra_w; 469 break; 470 case X86::BI__builtin_ia32_psrldi: 471 name = "psrldi"; 472 ID = Intrinsic::x86_mmx_psrl_d; 473 break; 474 case X86::BI__builtin_ia32_psrlqi: 475 name = "psrlqi"; 476 ID = Intrinsic::x86_mmx_psrl_q; 477 break; 478 case X86::BI__builtin_ia32_psrlwi: 479 name = "psrlwi"; 480 ID = Intrinsic::x86_mmx_psrl_w; 481 break; 482 case X86::BI__builtin_ia32_pslldi128: 483 name = "pslldi"; 484 ID = Intrinsic::x86_sse2_psll_d; 485 break; 486 case X86::BI__builtin_ia32_psllqi128: 487 name = "psllqi"; 488 ID = Intrinsic::x86_sse2_psll_q; 489 break; 490 case X86::BI__builtin_ia32_psllwi128: 491 name = "psllwi"; 492 ID = Intrinsic::x86_sse2_psll_w; 493 break; 494 case X86::BI__builtin_ia32_psradi128: 495 name = "psradi"; 496 ID = Intrinsic::x86_sse2_psra_d; 497 break; 498 case X86::BI__builtin_ia32_psrawi128: 499 name = "psrawi"; 500 ID = Intrinsic::x86_sse2_psra_w; 501 break; 502 case X86::BI__builtin_ia32_psrldi128: 503 name = "psrldi"; 504 ID = Intrinsic::x86_sse2_psrl_d; 505 break; 506 case X86::BI__builtin_ia32_psrlqi128: 507 name = "psrlqi"; 508 ID = Intrinsic::x86_sse2_psrl_q; 509 break; 510 case X86::BI__builtin_ia32_psrlwi128: 511 name = "psrlwi"; 512 ID = Intrinsic::x86_sse2_psrl_w; 513 break; 514 } 515 llvm::Function *F = CGM.getIntrinsic(ID); 516 return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name); 517 } 518 case X86::BI__builtin_ia32_pshuflw: { 519 unsigned i = cast<ConstantInt>(Ops[1])->getZExtValue(); 520 return EmitShuffleVector(Ops[0], Ops[0], 521 i & 0x3, (i & 0xc) >> 2, 522 (i & 0x30) >> 4, (i & 0xc0) >> 6, 4, 5, 6, 7, 523 "pshuflw"); 524 } 525 case X86::BI__builtin_ia32_pshufhw: { 526 unsigned i = cast<ConstantInt>(Ops[1])->getZExtValue(); 527 return EmitShuffleVector(Ops[0], Ops[0], 0, 1, 2, 3, 528 4 + (i & 0x3), 4 + ((i & 0xc) >> 2), 529 4 + ((i & 0x30) >> 4), 4 + ((i & 0xc0) >> 6), 530 "pshufhw"); 531 } 532 case X86::BI__builtin_ia32_pshufd: { 533 unsigned i = cast<ConstantInt>(Ops[1])->getZExtValue(); 534 return EmitShuffleVector(Ops[0], Ops[0], 535 i & 0x3, (i & 0xc) >> 2, 536 (i & 0x30) >> 4, (i & 0xc0) >> 6, 537 "pshufd"); 538 } 539 case X86::BI__builtin_ia32_vec_init_v4hi: 540 case X86::BI__builtin_ia32_vec_init_v8qi: 541 case X86::BI__builtin_ia32_vec_init_v2si: 542 return EmitVector(&Ops[0], Ops.size()); 543 case X86::BI__builtin_ia32_vec_ext_v2si: 544 case X86::BI__builtin_ia32_vec_ext_v2di: 545 case X86::BI__builtin_ia32_vec_ext_v4sf: 546 case X86::BI__builtin_ia32_vec_ext_v4si: 547 case X86::BI__builtin_ia32_vec_ext_v2df: 548 return Builder.CreateExtractElement(Ops[0], Ops[1], "result"); 549 case X86::BI__builtin_ia32_cmpordss: 550 case X86::BI__builtin_ia32_cmpordsd: 551 case X86::BI__builtin_ia32_cmpunordss: 552 case X86::BI__builtin_ia32_cmpunordsd: 553 case X86::BI__builtin_ia32_cmpeqss: 554 case X86::BI__builtin_ia32_cmpeqsd: 555 case X86::BI__builtin_ia32_cmpltss: 556 case X86::BI__builtin_ia32_cmpltsd: 557 case X86::BI__builtin_ia32_cmpless: 558 case X86::BI__builtin_ia32_cmplesd: 559 case X86::BI__builtin_ia32_cmpneqss: 560 case X86::BI__builtin_ia32_cmpneqsd: 561 case X86::BI__builtin_ia32_cmpnltss: 562 case X86::BI__builtin_ia32_cmpnltsd: 563 case X86::BI__builtin_ia32_cmpnless: 564 case X86::BI__builtin_ia32_cmpnlesd: { 565 unsigned i = 0; 566 const char *name = 0; 567 switch (BuiltinID) { 568 default: assert(0 && "Unknown compare builtin!"); 569 case X86::BI__builtin_ia32_cmpeqss: 570 case X86::BI__builtin_ia32_cmpeqsd: 571 i = 0; 572 name = "cmpeq"; 573 break; 574 case X86::BI__builtin_ia32_cmpltss: 575 case X86::BI__builtin_ia32_cmpltsd: 576 i = 1; 577 name = "cmplt"; 578 break; 579 case X86::BI__builtin_ia32_cmpless: 580 case X86::BI__builtin_ia32_cmplesd: 581 i = 2; 582 name = "cmple"; 583 break; 584 case X86::BI__builtin_ia32_cmpunordss: 585 case X86::BI__builtin_ia32_cmpunordsd: 586 i = 3; 587 name = "cmpunord"; 588 break; 589 case X86::BI__builtin_ia32_cmpneqss: 590 case X86::BI__builtin_ia32_cmpneqsd: 591 i = 4; 592 name = "cmpneq"; 593 break; 594 case X86::BI__builtin_ia32_cmpnltss: 595 case X86::BI__builtin_ia32_cmpnltsd: 596 i = 5; 597 name = "cmpntl"; 598 break; 599 case X86::BI__builtin_ia32_cmpnless: 600 case X86::BI__builtin_ia32_cmpnlesd: 601 i = 6; 602 name = "cmpnle"; 603 break; 604 case X86::BI__builtin_ia32_cmpordss: 605 case X86::BI__builtin_ia32_cmpordsd: 606 i = 7; 607 name = "cmpord"; 608 break; 609 } 610 611 llvm::Function *F; 612 if (cast<llvm::VectorType>(Ops[0]->getType())->getElementType() == 613 llvm::Type::FloatTy) 614 F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss); 615 else 616 F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd); 617 618 Ops.push_back(llvm::ConstantInt::get(llvm::Type::Int8Ty, i)); 619 return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name); 620 } 621 case X86::BI__builtin_ia32_ldmxcsr: { 622 llvm::Type *PtrTy = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 623 Value *One = llvm::ConstantInt::get(llvm::Type::Int32Ty, 1); 624 Value *Tmp = Builder.CreateAlloca(llvm::Type::Int32Ty, One, "tmp"); 625 Builder.CreateStore(Ops[0], Tmp); 626 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 627 Builder.CreateBitCast(Tmp, PtrTy)); 628 } 629 case X86::BI__builtin_ia32_stmxcsr: { 630 llvm::Type *PtrTy = llvm::PointerType::getUnqual(llvm::Type::Int8Ty); 631 Value *One = llvm::ConstantInt::get(llvm::Type::Int32Ty, 1); 632 Value *Tmp = Builder.CreateAlloca(llvm::Type::Int32Ty, One, "tmp"); 633 One = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 634 Builder.CreateBitCast(Tmp, PtrTy)); 635 return Builder.CreateLoad(Tmp, "stmxcsr"); 636 } 637 case X86::BI__builtin_ia32_cmpordps: 638 case X86::BI__builtin_ia32_cmpordpd: 639 case X86::BI__builtin_ia32_cmpunordps: 640 case X86::BI__builtin_ia32_cmpunordpd: 641 case X86::BI__builtin_ia32_cmpeqps: 642 case X86::BI__builtin_ia32_cmpeqpd: 643 case X86::BI__builtin_ia32_cmpltps: 644 case X86::BI__builtin_ia32_cmpltpd: 645 case X86::BI__builtin_ia32_cmpleps: 646 case X86::BI__builtin_ia32_cmplepd: 647 case X86::BI__builtin_ia32_cmpneqps: 648 case X86::BI__builtin_ia32_cmpneqpd: 649 case X86::BI__builtin_ia32_cmpngtps: 650 case X86::BI__builtin_ia32_cmpngtpd: 651 case X86::BI__builtin_ia32_cmpnltps: 652 case X86::BI__builtin_ia32_cmpnltpd: 653 case X86::BI__builtin_ia32_cmpgtps: 654 case X86::BI__builtin_ia32_cmpgtpd: 655 case X86::BI__builtin_ia32_cmpgeps: 656 case X86::BI__builtin_ia32_cmpgepd: 657 case X86::BI__builtin_ia32_cmpngeps: 658 case X86::BI__builtin_ia32_cmpngepd: 659 case X86::BI__builtin_ia32_cmpnleps: 660 case X86::BI__builtin_ia32_cmpnlepd: { 661 unsigned i = 0; 662 const char *name = 0; 663 bool ShouldSwap = false; 664 switch (BuiltinID) { 665 default: assert(0 && "Unknown compare builtin!"); 666 case X86::BI__builtin_ia32_cmpeqps: 667 case X86::BI__builtin_ia32_cmpeqpd: i = 0; name = "cmpeq"; break; 668 case X86::BI__builtin_ia32_cmpltps: 669 case X86::BI__builtin_ia32_cmpltpd: i = 1; name = "cmplt"; break; 670 case X86::BI__builtin_ia32_cmpleps: 671 case X86::BI__builtin_ia32_cmplepd: i = 2; name = "cmple"; break; 672 case X86::BI__builtin_ia32_cmpunordps: 673 case X86::BI__builtin_ia32_cmpunordpd: i = 3; name = "cmpunord"; break; 674 case X86::BI__builtin_ia32_cmpneqps: 675 case X86::BI__builtin_ia32_cmpneqpd: i = 4; name = "cmpneq"; break; 676 case X86::BI__builtin_ia32_cmpnltps: 677 case X86::BI__builtin_ia32_cmpnltpd: i = 5; name = "cmpntl"; break; 678 case X86::BI__builtin_ia32_cmpnleps: 679 case X86::BI__builtin_ia32_cmpnlepd: i = 6; name = "cmpnle"; break; 680 case X86::BI__builtin_ia32_cmpordps: 681 case X86::BI__builtin_ia32_cmpordpd: i = 7; name = "cmpord"; break; 682 case X86::BI__builtin_ia32_cmpgtps: 683 case X86::BI__builtin_ia32_cmpgtpd: 684 ShouldSwap = true; 685 i = 1; 686 name = "cmpgt"; 687 break; 688 case X86::BI__builtin_ia32_cmpgeps: 689 case X86::BI__builtin_ia32_cmpgepd: 690 i = 2; 691 name = "cmpge"; 692 ShouldSwap = true; 693 break; 694 case X86::BI__builtin_ia32_cmpngtps: 695 case X86::BI__builtin_ia32_cmpngtpd: 696 i = 5; 697 name = "cmpngt"; 698 ShouldSwap = true; 699 break; 700 case X86::BI__builtin_ia32_cmpngeps: 701 case X86::BI__builtin_ia32_cmpngepd: 702 i = 6; 703 name = "cmpnge"; 704 ShouldSwap = true; 705 break; 706 } 707 708 if (ShouldSwap) 709 std::swap(Ops[0], Ops[1]); 710 711 llvm::Function *F; 712 if (cast<llvm::VectorType>(Ops[0]->getType())->getElementType() == 713 llvm::Type::FloatTy) 714 F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps); 715 else 716 F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd); 717 718 Ops.push_back(llvm::ConstantInt::get(llvm::Type::Int8Ty, i)); 719 return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size(), name); 720 } 721 case X86::BI__builtin_ia32_movss: 722 return EmitShuffleVector(Ops[0], Ops[1], 4, 1, 2, 3, "movss"); 723 case X86::BI__builtin_ia32_shufps: { 724 unsigned i = cast<ConstantInt>(Ops[2])->getZExtValue(); 725 return EmitShuffleVector(Ops[0], Ops[1], 726 i & 0x3, (i & 0xc) >> 2, 727 ((i & 0x30) >> 4) + 4, 728 ((i & 0x60) >> 6) + 4, "shufps"); 729 } 730 case X86::BI__builtin_ia32_punpcklbw128: 731 return EmitShuffleVector(Ops[0], Ops[1], 0, 16, 1, 17, 2, 18, 3, 19, 732 4, 20, 5, 21, 6, 22, 7, 23, 733 "punpcklbw"); 734 case X86::BI__builtin_ia32_punpcklwd128: 735 return EmitShuffleVector(Ops[0], Ops[1], 0, 8, 1, 9, 2, 10, 3, 11, 736 "punpcklwd"); 737 case X86::BI__builtin_ia32_movlhps: 738 return EmitShuffleVector(Ops[0], Ops[1], 0, 1, 4, 5, "movlhps"); 739 case X86::BI__builtin_ia32_movhlps: 740 return EmitShuffleVector(Ops[0], Ops[1], 6, 7, 2, 3, "movhlps"); 741 case X86::BI__builtin_ia32_unpckhps: 742 return EmitShuffleVector(Ops[0], Ops[1], 2, 6, 3, 7, "unpckhps"); 743 case X86::BI__builtin_ia32_unpcklps: 744 return EmitShuffleVector(Ops[0], Ops[1], 0, 4, 1, 5, "unpcklps"); 745 case X86::BI__builtin_ia32_movqv4si: { 746 llvm::Type *Ty = llvm::VectorType::get(llvm::Type::Int64Ty, 2); 747 return Builder.CreateBitCast(Ops[0], Ty); 748 } 749 case X86::BI__builtin_ia32_loadlps: 750 case X86::BI__builtin_ia32_loadhps: { 751 // FIXME: This should probably be represented as 752 // shuffle (dst, (v4f32 (insert undef, (load i64), 0)), shuf mask hi/lo) 753 const llvm::Type *EltTy = llvm::Type::DoubleTy; 754 const llvm::Type *VecTy = llvm::VectorType::get(EltTy, 2); 755 const llvm::Type *OrigTy = Ops[0]->getType(); 756 unsigned Index = BuiltinID == X86::BI__builtin_ia32_loadlps ? 0 : 1; 757 llvm::Value *Idx = llvm::ConstantInt::get(llvm::Type::Int32Ty, Index); 758 Ops[1] = Builder.CreateBitCast(Ops[1], llvm::PointerType::getUnqual(EltTy)); 759 Ops[1] = Builder.CreateLoad(Ops[1], "tmp"); 760 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 761 Ops[0] = Builder.CreateInsertElement(Ops[0], Ops[1], Idx, "loadps"); 762 return Builder.CreateBitCast(Ops[0], OrigTy, "loadps"); 763 } 764 case X86::BI__builtin_ia32_storehps: 765 case X86::BI__builtin_ia32_storelps: { 766 const llvm::Type *EltTy = llvm::Type::Int64Ty; 767 llvm::Type *PtrTy = llvm::PointerType::getUnqual(EltTy); 768 llvm::Type *VecTy = llvm::VectorType::get(EltTy, 2); 769 770 // cast val v2i64 771 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 772 773 // extract (0, 1) 774 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 775 llvm::Value *Idx = llvm::ConstantInt::get(llvm::Type::Int32Ty, Index); 776 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 777 778 // cast pointer to i64 & store 779 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 780 return Builder.CreateStore(Ops[1], Ops[0]); 781 } 782 case X86::BI__builtin_ia32_loadlv4si: { 783 // load i64 784 const llvm::Type *EltTy = llvm::Type::Int64Ty; 785 llvm::Type *PtrTy = llvm::PointerType::getUnqual(EltTy); 786 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 787 Ops[0] = Builder.CreateLoad(Ops[0], "load"); 788 789 // scalar to vector: insert i64 into 2 x i64 undef 790 llvm::Type *VecTy = llvm::VectorType::get(EltTy, 2); 791 llvm::Value *Zero = llvm::ConstantInt::get(llvm::Type::Int32Ty, 0); 792 Ops[0] = Builder.CreateInsertElement(llvm::UndefValue::get(VecTy), 793 Ops[0], Zero, "s2v"); 794 795 // shuffle into zero vector. 796 std::vector<llvm::Constant *>Elts; 797 Elts.resize(2, llvm::ConstantInt::get(EltTy, 0)); 798 llvm::Value *ZV = ConstantVector::get(Elts); 799 Ops[0] = EmitShuffleVector(ZV, Ops[0], 2, 1, "loadl"); 800 801 // bitcast to result. 802 return Builder.CreateBitCast(Ops[0], 803 llvm::VectorType::get(llvm::Type::Int32Ty, 4)); 804 } 805 case X86::BI__builtin_ia32_vec_set_v4hi: 806 case X86::BI__builtin_ia32_vec_set_v8hi: 807 return Builder.CreateInsertElement(Ops[0], Ops[1], Ops[2], "pinsrw"); 808 case X86::BI__builtin_ia32_andps: 809 case X86::BI__builtin_ia32_andpd: 810 case X86::BI__builtin_ia32_andnps: 811 case X86::BI__builtin_ia32_andnpd: 812 case X86::BI__builtin_ia32_orps: 813 case X86::BI__builtin_ia32_orpd: 814 case X86::BI__builtin_ia32_xorpd: 815 case X86::BI__builtin_ia32_xorps: { 816 const llvm::Type *ITy = llvm::VectorType::get(llvm::Type::Int32Ty, 4); 817 const llvm::Type *FTy = Ops[0]->getType(); 818 Ops[0] = Builder.CreateBitCast(Ops[0], ITy, "bitcast"); 819 Ops[1] = Builder.CreateBitCast(Ops[1], ITy, "bitcast"); 820 switch (BuiltinID) { 821 case X86::BI__builtin_ia32_andps: 822 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1], "andps"); 823 break; 824 case X86::BI__builtin_ia32_andpd: 825 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1], "andpd"); 826 break; 827 case X86::BI__builtin_ia32_andnps: 828 Ops[0] = Builder.CreateNot(Ops[0], "not"); 829 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1], "andnps"); 830 break; 831 case X86::BI__builtin_ia32_andnpd: 832 Ops[0] = Builder.CreateNot(Ops[0], "not"); 833 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1], "andnpd"); 834 break; 835 case X86::BI__builtin_ia32_orps: 836 Ops[0] = Builder.CreateOr(Ops[0], Ops[1], "orps"); 837 break; 838 case X86::BI__builtin_ia32_orpd: 839 Ops[0] = Builder.CreateOr(Ops[0], Ops[1], "orpd"); 840 break; 841 case X86::BI__builtin_ia32_xorps: 842 Ops[0] = Builder.CreateXor(Ops[0], Ops[1], "xorps"); 843 break; 844 case X86::BI__builtin_ia32_xorpd: 845 Ops[0] = Builder.CreateXor(Ops[0], Ops[1], "xorpd"); 846 break; 847 } 848 return Builder.CreateBitCast(Ops[0], FTy, "bitcast"); 849 } 850 } 851} 852 853Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 854 const CallExpr *E) { 855 switch (BuiltinID) { 856 default: return 0; 857 } 858} 859