1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===// 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 file defines the interfaces that Mips uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 #include "MipsISelLowering.h" 15 #include "InstPrinter/MipsInstPrinter.h" 16 #include "MCTargetDesc/MipsBaseInfo.h" 17 #include "MipsCCState.h" 18 #include "MipsMachineFunction.h" 19 #include "MipsSubtarget.h" 20 #include "MipsTargetMachine.h" 21 #include "MipsTargetObjectFile.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/ADT/StringSwitch.h" 24 #include "llvm/CodeGen/CallingConvLower.h" 25 #include "llvm/CodeGen/MachineFrameInfo.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineJumpTableInfo.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/SelectionDAGISel.h" 31 #include "llvm/CodeGen/ValueTypes.h" 32 #include "llvm/IR/CallingConv.h" 33 #include "llvm/IR/DerivedTypes.h" 34 #include "llvm/IR/GlobalVariable.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/ErrorHandling.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include <cctype> 40 41 using namespace llvm; 42 43 #define DEBUG_TYPE "mips-lower" 44 45 STATISTIC(NumTailCalls, "Number of tail calls"); 46 47 static cl::opt<bool> 48 LargeGOT("mxgot", cl::Hidden, 49 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false)); 50 51 static cl::opt<bool> 52 NoZeroDivCheck("mno-check-zero-division", cl::Hidden, 53 cl::desc("MIPS: Don't trap on integer division by zero."), 54 cl::init(false)); 55 56 cl::opt<bool> 57 EnableMipsFastISel("mips-fast-isel", cl::Hidden, 58 cl::desc("Allow mips-fast-isel to be used"), 59 cl::init(false)); 60 61 static const MCPhysReg Mips64DPRegs[8] = { 62 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 63 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 64 }; 65 66 // If I is a shifted mask, set the size (Size) and the first bit of the 67 // mask (Pos), and return true. 68 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 69 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 70 if (!isShiftedMask_64(I)) 71 return false; 72 73 Size = CountPopulation_64(I); 74 Pos = countTrailingZeros(I); 75 return true; 76 } 77 78 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 79 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 80 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 81 } 82 83 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, 84 SelectionDAG &DAG, 85 unsigned Flag) const { 86 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag); 87 } 88 89 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, 90 SelectionDAG &DAG, 91 unsigned Flag) const { 92 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 93 } 94 95 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, 96 SelectionDAG &DAG, 97 unsigned Flag) const { 98 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 99 } 100 101 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, 102 SelectionDAG &DAG, 103 unsigned Flag) const { 104 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 105 } 106 107 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, 108 SelectionDAG &DAG, 109 unsigned Flag) const { 110 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 111 N->getOffset(), Flag); 112 } 113 114 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 115 switch (Opcode) { 116 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 117 case MipsISD::TailCall: return "MipsISD::TailCall"; 118 case MipsISD::Hi: return "MipsISD::Hi"; 119 case MipsISD::Lo: return "MipsISD::Lo"; 120 case MipsISD::GPRel: return "MipsISD::GPRel"; 121 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 122 case MipsISD::Ret: return "MipsISD::Ret"; 123 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 124 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 125 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 126 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 127 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 128 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP"; 129 case MipsISD::MFHI: return "MipsISD::MFHI"; 130 case MipsISD::MFLO: return "MipsISD::MFLO"; 131 case MipsISD::MTLOHI: return "MipsISD::MTLOHI"; 132 case MipsISD::Mult: return "MipsISD::Mult"; 133 case MipsISD::Multu: return "MipsISD::Multu"; 134 case MipsISD::MAdd: return "MipsISD::MAdd"; 135 case MipsISD::MAddu: return "MipsISD::MAddu"; 136 case MipsISD::MSub: return "MipsISD::MSub"; 137 case MipsISD::MSubu: return "MipsISD::MSubu"; 138 case MipsISD::DivRem: return "MipsISD::DivRem"; 139 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 140 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 141 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 142 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 143 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 144 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 145 case MipsISD::Sync: return "MipsISD::Sync"; 146 case MipsISD::Ext: return "MipsISD::Ext"; 147 case MipsISD::Ins: return "MipsISD::Ins"; 148 case MipsISD::LWL: return "MipsISD::LWL"; 149 case MipsISD::LWR: return "MipsISD::LWR"; 150 case MipsISD::SWL: return "MipsISD::SWL"; 151 case MipsISD::SWR: return "MipsISD::SWR"; 152 case MipsISD::LDL: return "MipsISD::LDL"; 153 case MipsISD::LDR: return "MipsISD::LDR"; 154 case MipsISD::SDL: return "MipsISD::SDL"; 155 case MipsISD::SDR: return "MipsISD::SDR"; 156 case MipsISD::EXTP: return "MipsISD::EXTP"; 157 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 158 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 159 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 160 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 161 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 162 case MipsISD::SHILO: return "MipsISD::SHILO"; 163 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 164 case MipsISD::MULT: return "MipsISD::MULT"; 165 case MipsISD::MULTU: return "MipsISD::MULTU"; 166 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 167 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 168 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 169 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 170 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP"; 171 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP"; 172 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP"; 173 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP"; 174 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP"; 175 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO"; 176 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO"; 177 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO"; 178 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO"; 179 case MipsISD::VCEQ: return "MipsISD::VCEQ"; 180 case MipsISD::VCLE_S: return "MipsISD::VCLE_S"; 181 case MipsISD::VCLE_U: return "MipsISD::VCLE_U"; 182 case MipsISD::VCLT_S: return "MipsISD::VCLT_S"; 183 case MipsISD::VCLT_U: return "MipsISD::VCLT_U"; 184 case MipsISD::VSMAX: return "MipsISD::VSMAX"; 185 case MipsISD::VSMIN: return "MipsISD::VSMIN"; 186 case MipsISD::VUMAX: return "MipsISD::VUMAX"; 187 case MipsISD::VUMIN: return "MipsISD::VUMIN"; 188 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT"; 189 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT"; 190 case MipsISD::VNOR: return "MipsISD::VNOR"; 191 case MipsISD::VSHF: return "MipsISD::VSHF"; 192 case MipsISD::SHF: return "MipsISD::SHF"; 193 case MipsISD::ILVEV: return "MipsISD::ILVEV"; 194 case MipsISD::ILVOD: return "MipsISD::ILVOD"; 195 case MipsISD::ILVL: return "MipsISD::ILVL"; 196 case MipsISD::ILVR: return "MipsISD::ILVR"; 197 case MipsISD::PCKEV: return "MipsISD::PCKEV"; 198 case MipsISD::PCKOD: return "MipsISD::PCKOD"; 199 case MipsISD::INSVE: return "MipsISD::INSVE"; 200 default: return nullptr; 201 } 202 } 203 204 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, 205 const MipsSubtarget &STI) 206 : TargetLowering(TM), Subtarget(STI) { 207 // Mips does not have i1 type, so use i32 for 208 // setcc operations results (slt, sgt, ...). 209 setBooleanContents(ZeroOrOneBooleanContent); 210 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 211 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA 212 // does. Integer booleans still use 0 and 1. 213 if (Subtarget.hasMips32r6()) 214 setBooleanContents(ZeroOrOneBooleanContent, 215 ZeroOrNegativeOneBooleanContent); 216 217 // Load extented operations for i1 types must be promoted 218 for (MVT VT : MVT::integer_valuetypes()) { 219 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote); 220 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); 221 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); 222 } 223 224 // MIPS doesn't have extending float->double load/store 225 for (MVT VT : MVT::fp_valuetypes()) 226 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); 227 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 228 229 // Used by legalize types to correctly generate the setcc result. 230 // Without this, every float setcc comes with a AND/OR with the result, 231 // we don't want this, since the fpcmp result goes to a flag register, 232 // which is used implicitly by brcond and select operations. 233 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 234 235 // Mips Custom Operations 236 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 237 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 238 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 239 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 240 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 241 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 242 setOperationAction(ISD::SELECT, MVT::f32, Custom); 243 setOperationAction(ISD::SELECT, MVT::f64, Custom); 244 setOperationAction(ISD::SELECT, MVT::i32, Custom); 245 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 246 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 247 setOperationAction(ISD::SETCC, MVT::f32, Custom); 248 setOperationAction(ISD::SETCC, MVT::f64, Custom); 249 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 250 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 251 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 252 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 253 254 if (Subtarget.isGP64bit()) { 255 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 256 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 257 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 258 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 259 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 260 setOperationAction(ISD::SELECT, MVT::i64, Custom); 261 setOperationAction(ISD::LOAD, MVT::i64, Custom); 262 setOperationAction(ISD::STORE, MVT::i64, Custom); 263 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 264 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); 265 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); 266 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); 267 } 268 269 if (!Subtarget.isGP64bit()) { 270 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 271 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 272 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 273 } 274 275 setOperationAction(ISD::ADD, MVT::i32, Custom); 276 if (Subtarget.isGP64bit()) 277 setOperationAction(ISD::ADD, MVT::i64, Custom); 278 279 setOperationAction(ISD::SDIV, MVT::i32, Expand); 280 setOperationAction(ISD::SREM, MVT::i32, Expand); 281 setOperationAction(ISD::UDIV, MVT::i32, Expand); 282 setOperationAction(ISD::UREM, MVT::i32, Expand); 283 setOperationAction(ISD::SDIV, MVT::i64, Expand); 284 setOperationAction(ISD::SREM, MVT::i64, Expand); 285 setOperationAction(ISD::UDIV, MVT::i64, Expand); 286 setOperationAction(ISD::UREM, MVT::i64, Expand); 287 288 // Operations not directly supported by Mips. 289 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 290 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 291 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 292 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 293 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand); 294 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand); 295 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 296 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 297 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 298 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 299 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 300 if (Subtarget.hasCnMips()) { 301 setOperationAction(ISD::CTPOP, MVT::i32, Legal); 302 setOperationAction(ISD::CTPOP, MVT::i64, Legal); 303 } else { 304 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 305 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 306 } 307 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 308 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 309 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 310 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); 311 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 312 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand); 313 setOperationAction(ISD::ROTL, MVT::i32, Expand); 314 setOperationAction(ISD::ROTL, MVT::i64, Expand); 315 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 316 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 317 318 if (!Subtarget.hasMips32r2()) 319 setOperationAction(ISD::ROTR, MVT::i32, Expand); 320 321 if (!Subtarget.hasMips64r2()) 322 setOperationAction(ISD::ROTR, MVT::i64, Expand); 323 324 setOperationAction(ISD::FSIN, MVT::f32, Expand); 325 setOperationAction(ISD::FSIN, MVT::f64, Expand); 326 setOperationAction(ISD::FCOS, MVT::f32, Expand); 327 setOperationAction(ISD::FCOS, MVT::f64, Expand); 328 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 329 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 330 setOperationAction(ISD::FPOWI, MVT::f32, Expand); 331 setOperationAction(ISD::FPOW, MVT::f32, Expand); 332 setOperationAction(ISD::FPOW, MVT::f64, Expand); 333 setOperationAction(ISD::FLOG, MVT::f32, Expand); 334 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 335 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 336 setOperationAction(ISD::FEXP, MVT::f32, Expand); 337 setOperationAction(ISD::FMA, MVT::f32, Expand); 338 setOperationAction(ISD::FMA, MVT::f64, Expand); 339 setOperationAction(ISD::FREM, MVT::f32, Expand); 340 setOperationAction(ISD::FREM, MVT::f64, Expand); 341 342 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 343 344 setOperationAction(ISD::VASTART, MVT::Other, Custom); 345 setOperationAction(ISD::VAARG, MVT::Other, Custom); 346 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 347 setOperationAction(ISD::VAEND, MVT::Other, Expand); 348 349 // Use the default for now 350 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 351 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 352 353 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand); 354 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 355 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand); 356 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 357 358 setInsertFencesForAtomic(true); 359 360 if (!Subtarget.hasMips32r2()) { 361 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 362 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 363 } 364 365 // MIPS16 lacks MIPS32's clz and clo instructions. 366 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) 367 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 368 if (!Subtarget.hasMips64()) 369 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 370 371 if (!Subtarget.hasMips32r2()) 372 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 373 if (!Subtarget.hasMips64r2()) 374 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 375 376 if (Subtarget.isGP64bit()) { 377 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom); 378 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom); 379 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom); 380 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 381 } 382 383 setOperationAction(ISD::TRAP, MVT::Other, Legal); 384 385 setTargetDAGCombine(ISD::SDIVREM); 386 setTargetDAGCombine(ISD::UDIVREM); 387 setTargetDAGCombine(ISD::SELECT); 388 setTargetDAGCombine(ISD::AND); 389 setTargetDAGCombine(ISD::OR); 390 setTargetDAGCombine(ISD::ADD); 391 392 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2); 393 394 // The arguments on the stack are defined in terms of 4-byte slots on O32 395 // and 8-byte slots on N32/N64. 396 setMinStackArgumentAlignment( 397 (Subtarget.isABI_N32() || Subtarget.isABI_N64()) ? 8 : 4); 398 399 setStackPointerRegisterToSaveRestore(Subtarget.isABI_N64() ? Mips::SP_64 400 : Mips::SP); 401 402 setExceptionPointerRegister(Subtarget.isABI_N64() ? Mips::A0_64 : Mips::A0); 403 setExceptionSelectorRegister(Subtarget.isABI_N64() ? Mips::A1_64 : Mips::A1); 404 405 MaxStoresPerMemcpy = 16; 406 407 isMicroMips = Subtarget.inMicroMipsMode(); 408 } 409 410 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM, 411 const MipsSubtarget &STI) { 412 if (STI.inMips16Mode()) 413 return llvm::createMips16TargetLowering(TM, STI); 414 415 return llvm::createMipsSETargetLowering(TM, STI); 416 } 417 418 // Create a fast isel object. 419 FastISel * 420 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 421 const TargetLibraryInfo *libInfo) const { 422 if (!EnableMipsFastISel) 423 return TargetLowering::createFastISel(funcInfo, libInfo); 424 return Mips::createFastISel(funcInfo, libInfo); 425 } 426 427 EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const { 428 if (!VT.isVector()) 429 return MVT::i32; 430 return VT.changeVectorElementTypeToInteger(); 431 } 432 433 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 434 TargetLowering::DAGCombinerInfo &DCI, 435 const MipsSubtarget &Subtarget) { 436 if (DCI.isBeforeLegalizeOps()) 437 return SDValue(); 438 439 EVT Ty = N->getValueType(0); 440 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; 441 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; 442 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 443 MipsISD::DivRemU16; 444 SDLoc DL(N); 445 446 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 447 N->getOperand(0), N->getOperand(1)); 448 SDValue InChain = DAG.getEntryNode(); 449 SDValue InGlue = DivRem; 450 451 // insert MFLO 452 if (N->hasAnyUseOfValue(0)) { 453 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 454 InGlue); 455 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 456 InChain = CopyFromLo.getValue(1); 457 InGlue = CopyFromLo.getValue(2); 458 } 459 460 // insert MFHI 461 if (N->hasAnyUseOfValue(1)) { 462 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 463 HI, Ty, InGlue); 464 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 465 } 466 467 return SDValue(); 468 } 469 470 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 471 switch (CC) { 472 default: llvm_unreachable("Unknown fp condition code!"); 473 case ISD::SETEQ: 474 case ISD::SETOEQ: return Mips::FCOND_OEQ; 475 case ISD::SETUNE: return Mips::FCOND_UNE; 476 case ISD::SETLT: 477 case ISD::SETOLT: return Mips::FCOND_OLT; 478 case ISD::SETGT: 479 case ISD::SETOGT: return Mips::FCOND_OGT; 480 case ISD::SETLE: 481 case ISD::SETOLE: return Mips::FCOND_OLE; 482 case ISD::SETGE: 483 case ISD::SETOGE: return Mips::FCOND_OGE; 484 case ISD::SETULT: return Mips::FCOND_ULT; 485 case ISD::SETULE: return Mips::FCOND_ULE; 486 case ISD::SETUGT: return Mips::FCOND_UGT; 487 case ISD::SETUGE: return Mips::FCOND_UGE; 488 case ISD::SETUO: return Mips::FCOND_UN; 489 case ISD::SETO: return Mips::FCOND_OR; 490 case ISD::SETNE: 491 case ISD::SETONE: return Mips::FCOND_ONE; 492 case ISD::SETUEQ: return Mips::FCOND_UEQ; 493 } 494 } 495 496 497 /// This function returns true if the floating point conditional branches and 498 /// conditional moves which use condition code CC should be inverted. 499 static bool invertFPCondCodeUser(Mips::CondCode CC) { 500 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 501 return false; 502 503 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 504 "Illegal Condition Code"); 505 506 return true; 507 } 508 509 // Creates and returns an FPCmp node from a setcc node. 510 // Returns Op if setcc is not a floating point comparison. 511 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 512 // must be a SETCC node 513 if (Op.getOpcode() != ISD::SETCC) 514 return Op; 515 516 SDValue LHS = Op.getOperand(0); 517 518 if (!LHS.getValueType().isFloatingPoint()) 519 return Op; 520 521 SDValue RHS = Op.getOperand(1); 522 SDLoc DL(Op); 523 524 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 525 // node if necessary. 526 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 527 528 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 529 DAG.getConstant(condCodeToFCC(CC), MVT::i32)); 530 } 531 532 // Creates and returns a CMovFPT/F node. 533 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 534 SDValue False, SDLoc DL) { 535 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 536 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 537 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 538 539 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 540 True.getValueType(), True, FCC0, False, Cond); 541 } 542 543 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 544 TargetLowering::DAGCombinerInfo &DCI, 545 const MipsSubtarget &Subtarget) { 546 if (DCI.isBeforeLegalizeOps()) 547 return SDValue(); 548 549 SDValue SetCC = N->getOperand(0); 550 551 if ((SetCC.getOpcode() != ISD::SETCC) || 552 !SetCC.getOperand(0).getValueType().isInteger()) 553 return SDValue(); 554 555 SDValue False = N->getOperand(2); 556 EVT FalseTy = False.getValueType(); 557 558 if (!FalseTy.isInteger()) 559 return SDValue(); 560 561 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False); 562 563 // If the RHS (False) is 0, we swap the order of the operands 564 // of ISD::SELECT (obviously also inverting the condition) so that we can 565 // take advantage of conditional moves using the $0 register. 566 // Example: 567 // return (a != 0) ? x : 0; 568 // load $reg, x 569 // movz $reg, $0, a 570 if (!FalseC) 571 return SDValue(); 572 573 const SDLoc DL(N); 574 575 if (!FalseC->getZExtValue()) { 576 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 577 SDValue True = N->getOperand(1); 578 579 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 580 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 581 582 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 583 } 584 585 // If both operands are integer constants there's a possibility that we 586 // can do some interesting optimizations. 587 SDValue True = N->getOperand(1); 588 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True); 589 590 if (!TrueC || !True.getValueType().isInteger()) 591 return SDValue(); 592 593 // We'll also ignore MVT::i64 operands as this optimizations proves 594 // to be ineffective because of the required sign extensions as the result 595 // of a SETCC operator is always MVT::i32 for non-vector types. 596 if (True.getValueType() == MVT::i64) 597 return SDValue(); 598 599 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); 600 601 // 1) (a < x) ? y : y-1 602 // slti $reg1, a, x 603 // addiu $reg2, $reg1, y-1 604 if (Diff == 1) 605 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False); 606 607 // 2) (a < x) ? y-1 : y 608 // slti $reg1, a, x 609 // xor $reg1, $reg1, 1 610 // addiu $reg2, $reg1, y-1 611 if (Diff == -1) { 612 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 613 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 614 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 615 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True); 616 } 617 618 // Couldn't optimize. 619 return SDValue(); 620 } 621 622 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, 623 TargetLowering::DAGCombinerInfo &DCI, 624 const MipsSubtarget &Subtarget) { 625 if (DCI.isBeforeLegalizeOps()) 626 return SDValue(); 627 628 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2); 629 630 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse); 631 if (!FalseC || FalseC->getZExtValue()) 632 return SDValue(); 633 634 // Since RHS (False) is 0, we swap the order of the True/False operands 635 // (obviously also inverting the condition) so that we can 636 // take advantage of conditional moves using the $0 register. 637 // Example: 638 // return (a != 0) ? x : 0; 639 // load $reg, x 640 // movz $reg, $0, a 641 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : 642 MipsISD::CMovFP_T; 643 644 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3); 645 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(), 646 ValueIfFalse, FCC, ValueIfTrue, Glue); 647 } 648 649 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 650 TargetLowering::DAGCombinerInfo &DCI, 651 const MipsSubtarget &Subtarget) { 652 // Pattern match EXT. 653 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 654 // => ext $dst, $src, size, pos 655 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 656 return SDValue(); 657 658 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1); 659 unsigned ShiftRightOpc = ShiftRight.getOpcode(); 660 661 // Op's first operand must be a shift right. 662 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL) 663 return SDValue(); 664 665 // The second operand of the shift must be an immediate. 666 ConstantSDNode *CN; 667 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1)))) 668 return SDValue(); 669 670 uint64_t Pos = CN->getZExtValue(); 671 uint64_t SMPos, SMSize; 672 673 // Op's second operand must be a shifted mask. 674 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 675 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 676 return SDValue(); 677 678 // Return if the shifted mask does not start at bit 0 or the sum of its size 679 // and Pos exceeds the word's size. 680 EVT ValTy = N->getValueType(0); 681 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 682 return SDValue(); 683 684 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy, 685 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32), 686 DAG.getConstant(SMSize, MVT::i32)); 687 } 688 689 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 690 TargetLowering::DAGCombinerInfo &DCI, 691 const MipsSubtarget &Subtarget) { 692 // Pattern match INS. 693 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 694 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 695 // => ins $dst, $src, size, pos, $src1 696 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) 697 return SDValue(); 698 699 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 700 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 701 ConstantSDNode *CN; 702 703 // See if Op's first operand matches (and $src1 , mask0). 704 if (And0.getOpcode() != ISD::AND) 705 return SDValue(); 706 707 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 708 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 709 return SDValue(); 710 711 // See if Op's second operand matches (and (shl $src, pos), mask1). 712 if (And1.getOpcode() != ISD::AND) 713 return SDValue(); 714 715 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 716 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 717 return SDValue(); 718 719 // The shift masks must have the same position and size. 720 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 721 return SDValue(); 722 723 SDValue Shl = And1.getOperand(0); 724 if (Shl.getOpcode() != ISD::SHL) 725 return SDValue(); 726 727 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 728 return SDValue(); 729 730 unsigned Shamt = CN->getZExtValue(); 731 732 // Return if the shift amount and the first bit position of mask are not the 733 // same. 734 EVT ValTy = N->getValueType(0); 735 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 736 return SDValue(); 737 738 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0), 739 DAG.getConstant(SMPos0, MVT::i32), 740 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0)); 741 } 742 743 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 744 TargetLowering::DAGCombinerInfo &DCI, 745 const MipsSubtarget &Subtarget) { 746 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 747 748 if (DCI.isBeforeLegalizeOps()) 749 return SDValue(); 750 751 SDValue Add = N->getOperand(1); 752 753 if (Add.getOpcode() != ISD::ADD) 754 return SDValue(); 755 756 SDValue Lo = Add.getOperand(1); 757 758 if ((Lo.getOpcode() != MipsISD::Lo) || 759 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 760 return SDValue(); 761 762 EVT ValTy = N->getValueType(0); 763 SDLoc DL(N); 764 765 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 766 Add.getOperand(0)); 767 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 768 } 769 770 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 771 const { 772 SelectionDAG &DAG = DCI.DAG; 773 unsigned Opc = N->getOpcode(); 774 775 switch (Opc) { 776 default: break; 777 case ISD::SDIVREM: 778 case ISD::UDIVREM: 779 return performDivRemCombine(N, DAG, DCI, Subtarget); 780 case ISD::SELECT: 781 return performSELECTCombine(N, DAG, DCI, Subtarget); 782 case MipsISD::CMovFP_F: 783 case MipsISD::CMovFP_T: 784 return performCMovFPCombine(N, DAG, DCI, Subtarget); 785 case ISD::AND: 786 return performANDCombine(N, DAG, DCI, Subtarget); 787 case ISD::OR: 788 return performORCombine(N, DAG, DCI, Subtarget); 789 case ISD::ADD: 790 return performADDCombine(N, DAG, DCI, Subtarget); 791 } 792 793 return SDValue(); 794 } 795 796 void 797 MipsTargetLowering::LowerOperationWrapper(SDNode *N, 798 SmallVectorImpl<SDValue> &Results, 799 SelectionDAG &DAG) const { 800 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 801 802 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 803 Results.push_back(Res.getValue(I)); 804 } 805 806 void 807 MipsTargetLowering::ReplaceNodeResults(SDNode *N, 808 SmallVectorImpl<SDValue> &Results, 809 SelectionDAG &DAG) const { 810 return LowerOperationWrapper(N, Results, DAG); 811 } 812 813 SDValue MipsTargetLowering:: 814 LowerOperation(SDValue Op, SelectionDAG &DAG) const 815 { 816 switch (Op.getOpcode()) 817 { 818 case ISD::BR_JT: return lowerBR_JT(Op, DAG); 819 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 820 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 821 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 822 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 823 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 824 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 825 case ISD::SELECT: return lowerSELECT(Op, DAG); 826 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG); 827 case ISD::SETCC: return lowerSETCC(Op, DAG); 828 case ISD::VASTART: return lowerVASTART(Op, DAG); 829 case ISD::VAARG: return lowerVAARG(Op, DAG); 830 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 831 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 832 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 833 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 834 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 835 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 836 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 837 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 838 case ISD::LOAD: return lowerLOAD(Op, DAG); 839 case ISD::STORE: return lowerSTORE(Op, DAG); 840 case ISD::ADD: return lowerADD(Op, DAG); 841 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 842 } 843 return SDValue(); 844 } 845 846 //===----------------------------------------------------------------------===// 847 // Lower helper functions 848 //===----------------------------------------------------------------------===// 849 850 // addLiveIn - This helper function adds the specified physical register to the 851 // MachineFunction as a live in value. It also creates a corresponding 852 // virtual register for it. 853 static unsigned 854 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 855 { 856 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 857 MF.getRegInfo().addLiveIn(PReg, VReg); 858 return VReg; 859 } 860 861 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr *MI, 862 MachineBasicBlock &MBB, 863 const TargetInstrInfo &TII, 864 bool Is64Bit) { 865 if (NoZeroDivCheck) 866 return &MBB; 867 868 // Insert instruction "teq $divisor_reg, $zero, 7". 869 MachineBasicBlock::iterator I(MI); 870 MachineInstrBuilder MIB; 871 MachineOperand &Divisor = MI->getOperand(2); 872 MIB = BuildMI(MBB, std::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ)) 873 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill())) 874 .addReg(Mips::ZERO).addImm(7); 875 876 // Use the 32-bit sub-register if this is a 64-bit division. 877 if (Is64Bit) 878 MIB->getOperand(0).setSubReg(Mips::sub_32); 879 880 // Clear Divisor's kill flag. 881 Divisor.setIsKill(false); 882 883 // We would normally delete the original instruction here but in this case 884 // we only needed to inject an additional instruction rather than replace it. 885 886 return &MBB; 887 } 888 889 MachineBasicBlock * 890 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 891 MachineBasicBlock *BB) const { 892 switch (MI->getOpcode()) { 893 default: 894 llvm_unreachable("Unexpected instr type to insert"); 895 case Mips::ATOMIC_LOAD_ADD_I8: 896 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 897 case Mips::ATOMIC_LOAD_ADD_I16: 898 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 899 case Mips::ATOMIC_LOAD_ADD_I32: 900 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 901 case Mips::ATOMIC_LOAD_ADD_I64: 902 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 903 904 case Mips::ATOMIC_LOAD_AND_I8: 905 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 906 case Mips::ATOMIC_LOAD_AND_I16: 907 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 908 case Mips::ATOMIC_LOAD_AND_I32: 909 return emitAtomicBinary(MI, BB, 4, Mips::AND); 910 case Mips::ATOMIC_LOAD_AND_I64: 911 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 912 913 case Mips::ATOMIC_LOAD_OR_I8: 914 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 915 case Mips::ATOMIC_LOAD_OR_I16: 916 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 917 case Mips::ATOMIC_LOAD_OR_I32: 918 return emitAtomicBinary(MI, BB, 4, Mips::OR); 919 case Mips::ATOMIC_LOAD_OR_I64: 920 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 921 922 case Mips::ATOMIC_LOAD_XOR_I8: 923 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 924 case Mips::ATOMIC_LOAD_XOR_I16: 925 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 926 case Mips::ATOMIC_LOAD_XOR_I32: 927 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 928 case Mips::ATOMIC_LOAD_XOR_I64: 929 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 930 931 case Mips::ATOMIC_LOAD_NAND_I8: 932 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 933 case Mips::ATOMIC_LOAD_NAND_I16: 934 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 935 case Mips::ATOMIC_LOAD_NAND_I32: 936 return emitAtomicBinary(MI, BB, 4, 0, true); 937 case Mips::ATOMIC_LOAD_NAND_I64: 938 return emitAtomicBinary(MI, BB, 8, 0, true); 939 940 case Mips::ATOMIC_LOAD_SUB_I8: 941 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 942 case Mips::ATOMIC_LOAD_SUB_I16: 943 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 944 case Mips::ATOMIC_LOAD_SUB_I32: 945 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 946 case Mips::ATOMIC_LOAD_SUB_I64: 947 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 948 949 case Mips::ATOMIC_SWAP_I8: 950 return emitAtomicBinaryPartword(MI, BB, 1, 0); 951 case Mips::ATOMIC_SWAP_I16: 952 return emitAtomicBinaryPartword(MI, BB, 2, 0); 953 case Mips::ATOMIC_SWAP_I32: 954 return emitAtomicBinary(MI, BB, 4, 0); 955 case Mips::ATOMIC_SWAP_I64: 956 return emitAtomicBinary(MI, BB, 8, 0); 957 958 case Mips::ATOMIC_CMP_SWAP_I8: 959 return emitAtomicCmpSwapPartword(MI, BB, 1); 960 case Mips::ATOMIC_CMP_SWAP_I16: 961 return emitAtomicCmpSwapPartword(MI, BB, 2); 962 case Mips::ATOMIC_CMP_SWAP_I32: 963 return emitAtomicCmpSwap(MI, BB, 4); 964 case Mips::ATOMIC_CMP_SWAP_I64: 965 return emitAtomicCmpSwap(MI, BB, 8); 966 case Mips::PseudoSDIV: 967 case Mips::PseudoUDIV: 968 case Mips::DIV: 969 case Mips::DIVU: 970 case Mips::MOD: 971 case Mips::MODU: 972 return insertDivByZeroTrap( 973 MI, *BB, *getTargetMachine().getSubtargetImpl()->getInstrInfo(), false); 974 case Mips::PseudoDSDIV: 975 case Mips::PseudoDUDIV: 976 case Mips::DDIV: 977 case Mips::DDIVU: 978 case Mips::DMOD: 979 case Mips::DMODU: 980 return insertDivByZeroTrap( 981 MI, *BB, *getTargetMachine().getSubtargetImpl()->getInstrInfo(), true); 982 case Mips::SEL_D: 983 return emitSEL_D(MI, BB); 984 985 case Mips::PseudoSELECT_I: 986 case Mips::PseudoSELECT_I64: 987 case Mips::PseudoSELECT_S: 988 case Mips::PseudoSELECT_D32: 989 case Mips::PseudoSELECT_D64: 990 return emitPseudoSELECT(MI, BB, false, Mips::BNE); 991 case Mips::PseudoSELECTFP_F_I: 992 case Mips::PseudoSELECTFP_F_I64: 993 case Mips::PseudoSELECTFP_F_S: 994 case Mips::PseudoSELECTFP_F_D32: 995 case Mips::PseudoSELECTFP_F_D64: 996 return emitPseudoSELECT(MI, BB, true, Mips::BC1F); 997 case Mips::PseudoSELECTFP_T_I: 998 case Mips::PseudoSELECTFP_T_I64: 999 case Mips::PseudoSELECTFP_T_S: 1000 case Mips::PseudoSELECTFP_T_D32: 1001 case Mips::PseudoSELECTFP_T_D64: 1002 return emitPseudoSELECT(MI, BB, true, Mips::BC1T); 1003 } 1004 } 1005 1006 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 1007 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 1008 MachineBasicBlock * 1009 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 1010 unsigned Size, unsigned BinOpcode, 1011 bool Nand) const { 1012 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 1013 1014 MachineFunction *MF = BB->getParent(); 1015 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1016 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1017 const TargetInstrInfo *TII = 1018 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1019 DebugLoc DL = MI->getDebugLoc(); 1020 unsigned LL, SC, AND, NOR, ZERO, BEQ; 1021 1022 if (Size == 4) { 1023 if (isMicroMips) { 1024 LL = Mips::LL_MM; 1025 SC = Mips::SC_MM; 1026 } else { 1027 LL = Subtarget.hasMips32r6() ? Mips::LL_R6 : Mips::LL; 1028 SC = Subtarget.hasMips32r6() ? Mips::SC_R6 : Mips::SC; 1029 } 1030 AND = Mips::AND; 1031 NOR = Mips::NOR; 1032 ZERO = Mips::ZERO; 1033 BEQ = Mips::BEQ; 1034 } else { 1035 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD; 1036 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD; 1037 AND = Mips::AND64; 1038 NOR = Mips::NOR64; 1039 ZERO = Mips::ZERO_64; 1040 BEQ = Mips::BEQ64; 1041 } 1042 1043 unsigned OldVal = MI->getOperand(0).getReg(); 1044 unsigned Ptr = MI->getOperand(1).getReg(); 1045 unsigned Incr = MI->getOperand(2).getReg(); 1046 1047 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1048 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1049 unsigned Success = RegInfo.createVirtualRegister(RC); 1050 1051 // insert new blocks after the current block 1052 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1053 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1054 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1055 MachineFunction::iterator It = BB; 1056 ++It; 1057 MF->insert(It, loopMBB); 1058 MF->insert(It, exitMBB); 1059 1060 // Transfer the remainder of BB and its successor edges to exitMBB. 1061 exitMBB->splice(exitMBB->begin(), BB, 1062 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1063 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1064 1065 // thisMBB: 1066 // ... 1067 // fallthrough --> loopMBB 1068 BB->addSuccessor(loopMBB); 1069 loopMBB->addSuccessor(loopMBB); 1070 loopMBB->addSuccessor(exitMBB); 1071 1072 // loopMBB: 1073 // ll oldval, 0(ptr) 1074 // <binop> storeval, oldval, incr 1075 // sc success, storeval, 0(ptr) 1076 // beq success, $0, loopMBB 1077 BB = loopMBB; 1078 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 1079 if (Nand) { 1080 // and andres, oldval, incr 1081 // nor storeval, $0, andres 1082 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 1083 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 1084 } else if (BinOpcode) { 1085 // <binop> storeval, oldval, incr 1086 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 1087 } else { 1088 StoreVal = Incr; 1089 } 1090 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 1091 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 1092 1093 MI->eraseFromParent(); // The instruction is gone now. 1094 1095 return exitMBB; 1096 } 1097 1098 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( 1099 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, 1100 unsigned SrcReg) const { 1101 const TargetInstrInfo *TII = 1102 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1103 DebugLoc DL = MI->getDebugLoc(); 1104 1105 if (Subtarget.hasMips32r2() && Size == 1) { 1106 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg); 1107 return BB; 1108 } 1109 1110 if (Subtarget.hasMips32r2() && Size == 2) { 1111 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg); 1112 return BB; 1113 } 1114 1115 MachineFunction *MF = BB->getParent(); 1116 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1117 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1118 unsigned ScrReg = RegInfo.createVirtualRegister(RC); 1119 1120 assert(Size < 32); 1121 int64_t ShiftImm = 32 - (Size * 8); 1122 1123 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm); 1124 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm); 1125 1126 return BB; 1127 } 1128 1129 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( 1130 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode, 1131 bool Nand) const { 1132 assert((Size == 1 || Size == 2) && 1133 "Unsupported size for EmitAtomicBinaryPartial."); 1134 1135 MachineFunction *MF = BB->getParent(); 1136 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1137 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1138 const TargetInstrInfo *TII = 1139 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1140 DebugLoc DL = MI->getDebugLoc(); 1141 1142 unsigned Dest = MI->getOperand(0).getReg(); 1143 unsigned Ptr = MI->getOperand(1).getReg(); 1144 unsigned Incr = MI->getOperand(2).getReg(); 1145 1146 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1147 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1148 unsigned Mask = RegInfo.createVirtualRegister(RC); 1149 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1150 unsigned NewVal = RegInfo.createVirtualRegister(RC); 1151 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1152 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 1153 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1154 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1155 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1156 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1157 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1158 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1159 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1160 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1161 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1162 unsigned Success = RegInfo.createVirtualRegister(RC); 1163 1164 // insert new blocks after the current block 1165 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1166 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1167 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1168 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1169 MachineFunction::iterator It = BB; 1170 ++It; 1171 MF->insert(It, loopMBB); 1172 MF->insert(It, sinkMBB); 1173 MF->insert(It, exitMBB); 1174 1175 // Transfer the remainder of BB and its successor edges to exitMBB. 1176 exitMBB->splice(exitMBB->begin(), BB, 1177 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1178 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1179 1180 BB->addSuccessor(loopMBB); 1181 loopMBB->addSuccessor(loopMBB); 1182 loopMBB->addSuccessor(sinkMBB); 1183 sinkMBB->addSuccessor(exitMBB); 1184 1185 // thisMBB: 1186 // addiu masklsb2,$0,-4 # 0xfffffffc 1187 // and alignedaddr,ptr,masklsb2 1188 // andi ptrlsb2,ptr,3 1189 // sll shiftamt,ptrlsb2,3 1190 // ori maskupper,$0,255 # 0xff 1191 // sll mask,maskupper,shiftamt 1192 // nor mask2,$0,mask 1193 // sll incr2,incr,shiftamt 1194 1195 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1196 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1197 .addReg(Mips::ZERO).addImm(-4); 1198 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1199 .addReg(Ptr).addReg(MaskLSB2); 1200 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1201 if (Subtarget.isLittle()) { 1202 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1203 } else { 1204 unsigned Off = RegInfo.createVirtualRegister(RC); 1205 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1206 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1207 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1208 } 1209 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1210 .addReg(Mips::ZERO).addImm(MaskImm); 1211 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1212 .addReg(MaskUpper).addReg(ShiftAmt); 1213 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1214 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1215 1216 // atomic.load.binop 1217 // loopMBB: 1218 // ll oldval,0(alignedaddr) 1219 // binop binopres,oldval,incr2 1220 // and newval,binopres,mask 1221 // and maskedoldval0,oldval,mask2 1222 // or storeval,maskedoldval0,newval 1223 // sc success,storeval,0(alignedaddr) 1224 // beq success,$0,loopMBB 1225 1226 // atomic.swap 1227 // loopMBB: 1228 // ll oldval,0(alignedaddr) 1229 // and newval,incr2,mask 1230 // and maskedoldval0,oldval,mask2 1231 // or storeval,maskedoldval0,newval 1232 // sc success,storeval,0(alignedaddr) 1233 // beq success,$0,loopMBB 1234 1235 BB = loopMBB; 1236 unsigned LL = isMicroMips ? Mips::LL_MM : Mips::LL; 1237 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1238 if (Nand) { 1239 // and andres, oldval, incr2 1240 // nor binopres, $0, andres 1241 // and newval, binopres, mask 1242 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1243 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1244 .addReg(Mips::ZERO).addReg(AndRes); 1245 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1246 } else if (BinOpcode) { 1247 // <binop> binopres, oldval, incr2 1248 // and newval, binopres, mask 1249 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1250 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1251 } else { // atomic.swap 1252 // and newval, incr2, mask 1253 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1254 } 1255 1256 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1257 .addReg(OldVal).addReg(Mask2); 1258 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1259 .addReg(MaskedOldVal0).addReg(NewVal); 1260 unsigned SC = isMicroMips ? Mips::SC_MM : Mips::SC; 1261 BuildMI(BB, DL, TII->get(SC), Success) 1262 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1263 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1264 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1265 1266 // sinkMBB: 1267 // and maskedoldval1,oldval,mask 1268 // srl srlres,maskedoldval1,shiftamt 1269 // sign_extend dest,srlres 1270 BB = sinkMBB; 1271 1272 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1273 .addReg(OldVal).addReg(Mask); 1274 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1275 .addReg(MaskedOldVal1).addReg(ShiftAmt); 1276 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1277 1278 MI->eraseFromParent(); // The instruction is gone now. 1279 1280 return exitMBB; 1281 } 1282 1283 MachineBasicBlock * MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI, 1284 MachineBasicBlock *BB, 1285 unsigned Size) const { 1286 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1287 1288 MachineFunction *MF = BB->getParent(); 1289 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1290 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1291 const TargetInstrInfo *TII = 1292 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1293 DebugLoc DL = MI->getDebugLoc(); 1294 unsigned LL, SC, ZERO, BNE, BEQ; 1295 1296 if (Size == 4) { 1297 LL = isMicroMips ? Mips::LL_MM : Mips::LL; 1298 SC = isMicroMips ? Mips::SC_MM : Mips::SC; 1299 ZERO = Mips::ZERO; 1300 BNE = Mips::BNE; 1301 BEQ = Mips::BEQ; 1302 } else { 1303 LL = Mips::LLD; 1304 SC = Mips::SCD; 1305 ZERO = Mips::ZERO_64; 1306 BNE = Mips::BNE64; 1307 BEQ = Mips::BEQ64; 1308 } 1309 1310 unsigned Dest = MI->getOperand(0).getReg(); 1311 unsigned Ptr = MI->getOperand(1).getReg(); 1312 unsigned OldVal = MI->getOperand(2).getReg(); 1313 unsigned NewVal = MI->getOperand(3).getReg(); 1314 1315 unsigned Success = RegInfo.createVirtualRegister(RC); 1316 1317 // insert new blocks after the current block 1318 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1319 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1320 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1321 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1322 MachineFunction::iterator It = BB; 1323 ++It; 1324 MF->insert(It, loop1MBB); 1325 MF->insert(It, loop2MBB); 1326 MF->insert(It, exitMBB); 1327 1328 // Transfer the remainder of BB and its successor edges to exitMBB. 1329 exitMBB->splice(exitMBB->begin(), BB, 1330 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1331 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1332 1333 // thisMBB: 1334 // ... 1335 // fallthrough --> loop1MBB 1336 BB->addSuccessor(loop1MBB); 1337 loop1MBB->addSuccessor(exitMBB); 1338 loop1MBB->addSuccessor(loop2MBB); 1339 loop2MBB->addSuccessor(loop1MBB); 1340 loop2MBB->addSuccessor(exitMBB); 1341 1342 // loop1MBB: 1343 // ll dest, 0(ptr) 1344 // bne dest, oldval, exitMBB 1345 BB = loop1MBB; 1346 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1347 BuildMI(BB, DL, TII->get(BNE)) 1348 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1349 1350 // loop2MBB: 1351 // sc success, newval, 0(ptr) 1352 // beq success, $0, loop1MBB 1353 BB = loop2MBB; 1354 BuildMI(BB, DL, TII->get(SC), Success) 1355 .addReg(NewVal).addReg(Ptr).addImm(0); 1356 BuildMI(BB, DL, TII->get(BEQ)) 1357 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1358 1359 MI->eraseFromParent(); // The instruction is gone now. 1360 1361 return exitMBB; 1362 } 1363 1364 MachineBasicBlock * 1365 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI, 1366 MachineBasicBlock *BB, 1367 unsigned Size) const { 1368 assert((Size == 1 || Size == 2) && 1369 "Unsupported size for EmitAtomicCmpSwapPartial."); 1370 1371 MachineFunction *MF = BB->getParent(); 1372 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1373 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1374 const TargetInstrInfo *TII = 1375 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1376 DebugLoc DL = MI->getDebugLoc(); 1377 1378 unsigned Dest = MI->getOperand(0).getReg(); 1379 unsigned Ptr = MI->getOperand(1).getReg(); 1380 unsigned CmpVal = MI->getOperand(2).getReg(); 1381 unsigned NewVal = MI->getOperand(3).getReg(); 1382 1383 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1384 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1385 unsigned Mask = RegInfo.createVirtualRegister(RC); 1386 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1387 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1388 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1389 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1390 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1391 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1392 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1393 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1394 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1395 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1396 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1397 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1398 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1399 unsigned Success = RegInfo.createVirtualRegister(RC); 1400 1401 // insert new blocks after the current block 1402 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1403 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1404 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1405 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1406 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1407 MachineFunction::iterator It = BB; 1408 ++It; 1409 MF->insert(It, loop1MBB); 1410 MF->insert(It, loop2MBB); 1411 MF->insert(It, sinkMBB); 1412 MF->insert(It, exitMBB); 1413 1414 // Transfer the remainder of BB and its successor edges to exitMBB. 1415 exitMBB->splice(exitMBB->begin(), BB, 1416 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 1417 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1418 1419 BB->addSuccessor(loop1MBB); 1420 loop1MBB->addSuccessor(sinkMBB); 1421 loop1MBB->addSuccessor(loop2MBB); 1422 loop2MBB->addSuccessor(loop1MBB); 1423 loop2MBB->addSuccessor(sinkMBB); 1424 sinkMBB->addSuccessor(exitMBB); 1425 1426 // FIXME: computation of newval2 can be moved to loop2MBB. 1427 // thisMBB: 1428 // addiu masklsb2,$0,-4 # 0xfffffffc 1429 // and alignedaddr,ptr,masklsb2 1430 // andi ptrlsb2,ptr,3 1431 // sll shiftamt,ptrlsb2,3 1432 // ori maskupper,$0,255 # 0xff 1433 // sll mask,maskupper,shiftamt 1434 // nor mask2,$0,mask 1435 // andi maskedcmpval,cmpval,255 1436 // sll shiftedcmpval,maskedcmpval,shiftamt 1437 // andi maskednewval,newval,255 1438 // sll shiftednewval,maskednewval,shiftamt 1439 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1440 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1441 .addReg(Mips::ZERO).addImm(-4); 1442 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1443 .addReg(Ptr).addReg(MaskLSB2); 1444 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1445 if (Subtarget.isLittle()) { 1446 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1447 } else { 1448 unsigned Off = RegInfo.createVirtualRegister(RC); 1449 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1450 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1451 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1452 } 1453 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1454 .addReg(Mips::ZERO).addImm(MaskImm); 1455 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1456 .addReg(MaskUpper).addReg(ShiftAmt); 1457 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1458 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1459 .addReg(CmpVal).addImm(MaskImm); 1460 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1461 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1462 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1463 .addReg(NewVal).addImm(MaskImm); 1464 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1465 .addReg(MaskedNewVal).addReg(ShiftAmt); 1466 1467 // loop1MBB: 1468 // ll oldval,0(alginedaddr) 1469 // and maskedoldval0,oldval,mask 1470 // bne maskedoldval0,shiftedcmpval,sinkMBB 1471 BB = loop1MBB; 1472 unsigned LL = isMicroMips ? Mips::LL_MM : Mips::LL; 1473 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1474 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1475 .addReg(OldVal).addReg(Mask); 1476 BuildMI(BB, DL, TII->get(Mips::BNE)) 1477 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1478 1479 // loop2MBB: 1480 // and maskedoldval1,oldval,mask2 1481 // or storeval,maskedoldval1,shiftednewval 1482 // sc success,storeval,0(alignedaddr) 1483 // beq success,$0,loop1MBB 1484 BB = loop2MBB; 1485 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1486 .addReg(OldVal).addReg(Mask2); 1487 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1488 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1489 unsigned SC = isMicroMips ? Mips::SC_MM : Mips::SC; 1490 BuildMI(BB, DL, TII->get(SC), Success) 1491 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1492 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1493 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1494 1495 // sinkMBB: 1496 // srl srlres,maskedoldval0,shiftamt 1497 // sign_extend dest,srlres 1498 BB = sinkMBB; 1499 1500 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1501 .addReg(MaskedOldVal0).addReg(ShiftAmt); 1502 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes); 1503 1504 MI->eraseFromParent(); // The instruction is gone now. 1505 1506 return exitMBB; 1507 } 1508 1509 MachineBasicBlock *MipsTargetLowering::emitSEL_D(MachineInstr *MI, 1510 MachineBasicBlock *BB) const { 1511 MachineFunction *MF = BB->getParent(); 1512 const TargetRegisterInfo *TRI = 1513 getTargetMachine().getSubtargetImpl()->getRegisterInfo(); 1514 const TargetInstrInfo *TII = 1515 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 1516 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1517 DebugLoc DL = MI->getDebugLoc(); 1518 MachineBasicBlock::iterator II(MI); 1519 1520 unsigned Fc = MI->getOperand(1).getReg(); 1521 const auto &FGR64RegClass = TRI->getRegClass(Mips::FGR64RegClassID); 1522 1523 unsigned Fc2 = RegInfo.createVirtualRegister(FGR64RegClass); 1524 1525 BuildMI(*BB, II, DL, TII->get(Mips::SUBREG_TO_REG), Fc2) 1526 .addImm(0) 1527 .addReg(Fc) 1528 .addImm(Mips::sub_lo); 1529 1530 // We don't erase the original instruction, we just replace the condition 1531 // register with the 64-bit super-register. 1532 MI->getOperand(1).setReg(Fc2); 1533 1534 return BB; 1535 } 1536 1537 //===----------------------------------------------------------------------===// 1538 // Misc Lower Operation implementation 1539 //===----------------------------------------------------------------------===// 1540 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 1541 SDValue Chain = Op.getOperand(0); 1542 SDValue Table = Op.getOperand(1); 1543 SDValue Index = Op.getOperand(2); 1544 SDLoc DL(Op); 1545 EVT PTy = getPointerTy(); 1546 unsigned EntrySize = 1547 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout()); 1548 1549 Index = DAG.getNode(ISD::MUL, DL, PTy, Index, 1550 DAG.getConstant(EntrySize, PTy)); 1551 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table); 1552 1553 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); 1554 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr, 1555 MachinePointerInfo::getJumpTable(), MemVT, false, false, 1556 false, 0); 1557 Chain = Addr.getValue(1); 1558 1559 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || 1560 Subtarget.isABI_N64()) { 1561 // For PIC, the sequence is: 1562 // BRIND(load(Jumptable + index) + RelocBase) 1563 // RelocBase can be JumpTable, GOT or some sort of global base. 1564 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr, 1565 getPICJumpTableRelocBase(Table, DAG)); 1566 } 1567 1568 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr); 1569 } 1570 1571 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { 1572 // The first operand is the chain, the second is the condition, the third is 1573 // the block to branch to if the condition is true. 1574 SDValue Chain = Op.getOperand(0); 1575 SDValue Dest = Op.getOperand(2); 1576 SDLoc DL(Op); 1577 1578 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1579 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1580 1581 // Return if flag is not set by a floating point comparison. 1582 if (CondRes.getOpcode() != MipsISD::FPCmp) 1583 return Op; 1584 1585 SDValue CCNode = CondRes.getOperand(2); 1586 Mips::CondCode CC = 1587 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1588 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1589 SDValue BrCode = DAG.getConstant(Opc, MVT::i32); 1590 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 1591 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1592 FCC0, Dest, CondRes); 1593 } 1594 1595 SDValue MipsTargetLowering:: 1596 lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1597 { 1598 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1599 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1600 1601 // Return if flag is not set by a floating point comparison. 1602 if (Cond.getOpcode() != MipsISD::FPCmp) 1603 return Op; 1604 1605 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1606 SDLoc(Op)); 1607 } 1608 1609 SDValue MipsTargetLowering:: 1610 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 1611 { 1612 SDLoc DL(Op); 1613 EVT Ty = Op.getOperand(0).getValueType(); 1614 SDValue Cond = DAG.getNode(ISD::SETCC, DL, 1615 getSetCCResultType(*DAG.getContext(), Ty), 1616 Op.getOperand(0), Op.getOperand(1), 1617 Op.getOperand(4)); 1618 1619 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2), 1620 Op.getOperand(3)); 1621 } 1622 1623 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 1624 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); 1625 SDValue Cond = createFPCmp(DAG, Op); 1626 1627 assert(Cond.getOpcode() == MipsISD::FPCmp && 1628 "Floating point operand expected."); 1629 1630 SDValue True = DAG.getConstant(1, MVT::i32); 1631 SDValue False = DAG.getConstant(0, MVT::i32); 1632 1633 return createCMovFP(DAG, Cond, True, False, SDLoc(Op)); 1634 } 1635 1636 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 1637 SelectionDAG &DAG) const { 1638 EVT Ty = Op.getValueType(); 1639 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); 1640 const GlobalValue *GV = N->getGlobal(); 1641 1642 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && 1643 !Subtarget.isABI_N64()) { 1644 const MipsTargetObjectFile &TLOF = 1645 (const MipsTargetObjectFile&)getObjFileLowering(); 1646 1647 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) 1648 // %gp_rel relocation 1649 return getAddrGPRel(N, SDLoc(N), Ty, DAG); 1650 1651 // %hi/%lo relocation 1652 return getAddrNonPIC(N, SDLoc(N), Ty, DAG); 1653 } 1654 1655 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))) 1656 return getAddrLocal(N, SDLoc(N), Ty, DAG, 1657 Subtarget.isABI_N32() || Subtarget.isABI_N64()); 1658 1659 if (LargeGOT) 1660 return getAddrGlobalLargeGOT(N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, 1661 MipsII::MO_GOT_LO16, DAG.getEntryNode(), 1662 MachinePointerInfo::getGOT()); 1663 1664 return getAddrGlobal(N, SDLoc(N), Ty, DAG, 1665 (Subtarget.isABI_N32() || Subtarget.isABI_N64()) 1666 ? MipsII::MO_GOT_DISP 1667 : MipsII::MO_GOT16, 1668 DAG.getEntryNode(), MachinePointerInfo::getGOT()); 1669 } 1670 1671 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 1672 SelectionDAG &DAG) const { 1673 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op); 1674 EVT Ty = Op.getValueType(); 1675 1676 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && 1677 !Subtarget.isABI_N64()) 1678 return getAddrNonPIC(N, SDLoc(N), Ty, DAG); 1679 1680 return getAddrLocal(N, SDLoc(N), Ty, DAG, 1681 Subtarget.isABI_N32() || Subtarget.isABI_N64()); 1682 } 1683 1684 SDValue MipsTargetLowering:: 1685 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 1686 { 1687 // If the relocation model is PIC, use the General Dynamic TLS Model or 1688 // Local Dynamic TLS model, otherwise use the Initial Exec or 1689 // Local Exec TLS Model. 1690 1691 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1692 SDLoc DL(GA); 1693 const GlobalValue *GV = GA->getGlobal(); 1694 EVT PtrVT = getPointerTy(); 1695 1696 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 1697 1698 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 1699 // General Dynamic and Local Dynamic TLS Model. 1700 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 1701 : MipsII::MO_TLSGD; 1702 1703 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 1704 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 1705 getGlobalReg(DAG, PtrVT), TGA); 1706 unsigned PtrSize = PtrVT.getSizeInBits(); 1707 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 1708 1709 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 1710 1711 ArgListTy Args; 1712 ArgListEntry Entry; 1713 Entry.Node = Argument; 1714 Entry.Ty = PtrTy; 1715 Args.push_back(Entry); 1716 1717 TargetLowering::CallLoweringInfo CLI(DAG); 1718 CLI.setDebugLoc(DL).setChain(DAG.getEntryNode()) 1719 .setCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args), 0); 1720 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 1721 1722 SDValue Ret = CallResult.first; 1723 1724 if (model != TLSModel::LocalDynamic) 1725 return Ret; 1726 1727 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1728 MipsII::MO_DTPREL_HI); 1729 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1730 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1731 MipsII::MO_DTPREL_LO); 1732 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1733 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 1734 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 1735 } 1736 1737 SDValue Offset; 1738 if (model == TLSModel::InitialExec) { 1739 // Initial Exec TLS Model 1740 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1741 MipsII::MO_GOTTPREL); 1742 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 1743 TGA); 1744 Offset = DAG.getLoad(PtrVT, DL, 1745 DAG.getEntryNode(), TGA, MachinePointerInfo(), 1746 false, false, false, 0); 1747 } else { 1748 // Local Exec TLS Model 1749 assert(model == TLSModel::LocalExec); 1750 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1751 MipsII::MO_TPREL_HI); 1752 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1753 MipsII::MO_TPREL_LO); 1754 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1755 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1756 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 1757 } 1758 1759 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 1760 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 1761 } 1762 1763 SDValue MipsTargetLowering:: 1764 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 1765 { 1766 JumpTableSDNode *N = cast<JumpTableSDNode>(Op); 1767 EVT Ty = Op.getValueType(); 1768 1769 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && 1770 !Subtarget.isABI_N64()) 1771 return getAddrNonPIC(N, SDLoc(N), Ty, DAG); 1772 1773 return getAddrLocal(N, SDLoc(N), Ty, DAG, 1774 Subtarget.isABI_N32() || Subtarget.isABI_N64()); 1775 } 1776 1777 SDValue MipsTargetLowering:: 1778 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 1779 { 1780 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op); 1781 EVT Ty = Op.getValueType(); 1782 1783 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && 1784 !Subtarget.isABI_N64()) { 1785 const MipsTargetObjectFile &TLOF = 1786 (const MipsTargetObjectFile&)getObjFileLowering(); 1787 1788 if (TLOF.IsConstantInSmallSection(N->getConstVal(), getTargetMachine())) 1789 // %gp_rel relocation 1790 return getAddrGPRel(N, SDLoc(N), Ty, DAG); 1791 1792 return getAddrNonPIC(N, SDLoc(N), Ty, DAG); 1793 } 1794 1795 return getAddrLocal(N, SDLoc(N), Ty, DAG, 1796 Subtarget.isABI_N32() || Subtarget.isABI_N64()); 1797 } 1798 1799 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 1800 MachineFunction &MF = DAG.getMachineFunction(); 1801 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 1802 1803 SDLoc DL(Op); 1804 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 1805 getPointerTy()); 1806 1807 // vastart just stores the address of the VarArgsFrameIndex slot into the 1808 // memory location argument. 1809 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1810 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 1811 MachinePointerInfo(SV), false, false, 0); 1812 } 1813 1814 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { 1815 SDNode *Node = Op.getNode(); 1816 EVT VT = Node->getValueType(0); 1817 SDValue Chain = Node->getOperand(0); 1818 SDValue VAListPtr = Node->getOperand(1); 1819 unsigned Align = Node->getConstantOperandVal(3); 1820 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); 1821 SDLoc DL(Node); 1822 unsigned ArgSlotSizeInBytes = 1823 (Subtarget.isABI_N32() || Subtarget.isABI_N64()) ? 8 : 4; 1824 1825 SDValue VAListLoad = DAG.getLoad(getPointerTy(), DL, Chain, VAListPtr, 1826 MachinePointerInfo(SV), false, false, false, 1827 0); 1828 SDValue VAList = VAListLoad; 1829 1830 // Re-align the pointer if necessary. 1831 // It should only ever be necessary for 64-bit types on O32 since the minimum 1832 // argument alignment is the same as the maximum type alignment for N32/N64. 1833 // 1834 // FIXME: We currently align too often. The code generator doesn't notice 1835 // when the pointer is still aligned from the last va_arg (or pair of 1836 // va_args for the i64 on O32 case). 1837 if (Align > getMinStackArgumentAlignment()) { 1838 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2"); 1839 1840 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 1841 DAG.getConstant(Align - 1, 1842 VAList.getValueType())); 1843 1844 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList, 1845 DAG.getConstant(-(int64_t)Align, 1846 VAList.getValueType())); 1847 } 1848 1849 // Increment the pointer, VAList, to the next vaarg. 1850 unsigned ArgSizeInBytes = getDataLayout()->getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())); 1851 SDValue Tmp3 = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList, 1852 DAG.getConstant(RoundUpToAlignment(ArgSizeInBytes, ArgSlotSizeInBytes), 1853 VAList.getValueType())); 1854 // Store the incremented VAList to the legalized pointer 1855 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr, 1856 MachinePointerInfo(SV), false, false, 0); 1857 1858 // In big-endian mode we must adjust the pointer when the load size is smaller 1859 // than the argument slot size. We must also reduce the known alignment to 1860 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get 1861 // the correct half of the slot, and reduce the alignment from 8 (slot 1862 // alignment) down to 4 (type alignment). 1863 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { 1864 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; 1865 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList, 1866 DAG.getIntPtrConstant(Adjustment)); 1867 } 1868 // Load the actual argument out of the pointer VAList 1869 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo(), false, false, 1870 false, 0); 1871 } 1872 1873 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, 1874 bool HasExtractInsert) { 1875 EVT TyX = Op.getOperand(0).getValueType(); 1876 EVT TyY = Op.getOperand(1).getValueType(); 1877 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1878 SDValue Const31 = DAG.getConstant(31, MVT::i32); 1879 SDLoc DL(Op); 1880 SDValue Res; 1881 1882 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1883 // to i32. 1884 SDValue X = (TyX == MVT::f32) ? 1885 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1886 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1887 Const1); 1888 SDValue Y = (TyY == MVT::f32) ? 1889 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 1890 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 1891 Const1); 1892 1893 if (HasExtractInsert) { 1894 // ext E, Y, 31, 1 ; extract bit31 of Y 1895 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 1896 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 1897 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 1898 } else { 1899 // sll SllX, X, 1 1900 // srl SrlX, SllX, 1 1901 // srl SrlY, Y, 31 1902 // sll SllY, SrlX, 31 1903 // or Or, SrlX, SllY 1904 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1905 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1906 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 1907 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 1908 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 1909 } 1910 1911 if (TyX == MVT::f32) 1912 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 1913 1914 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1915 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1916 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1917 } 1918 1919 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, 1920 bool HasExtractInsert) { 1921 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 1922 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 1923 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 1924 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1925 SDLoc DL(Op); 1926 1927 // Bitcast to integer nodes. 1928 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 1929 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 1930 1931 if (HasExtractInsert) { 1932 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 1933 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 1934 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 1935 DAG.getConstant(WidthY - 1, MVT::i32), Const1); 1936 1937 if (WidthX > WidthY) 1938 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 1939 else if (WidthY > WidthX) 1940 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 1941 1942 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 1943 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X); 1944 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 1945 } 1946 1947 // (d)sll SllX, X, 1 1948 // (d)srl SrlX, SllX, 1 1949 // (d)srl SrlY, Y, width(Y)-1 1950 // (d)sll SllY, SrlX, width(Y)-1 1951 // or Or, SrlX, SllY 1952 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 1953 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 1954 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 1955 DAG.getConstant(WidthY - 1, MVT::i32)); 1956 1957 if (WidthX > WidthY) 1958 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 1959 else if (WidthY > WidthX) 1960 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 1961 1962 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 1963 DAG.getConstant(WidthX - 1, MVT::i32)); 1964 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 1965 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 1966 } 1967 1968 SDValue 1969 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 1970 if (Subtarget.isGP64bit()) 1971 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert()); 1972 1973 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert()); 1974 } 1975 1976 SDValue MipsTargetLowering:: 1977 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 1978 // check the depth 1979 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1980 "Frame address can only be determined for current frame."); 1981 1982 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1983 MFI->setFrameAddressIsTaken(true); 1984 EVT VT = Op.getValueType(); 1985 SDLoc DL(Op); 1986 SDValue FrameAddr = 1987 DAG.getCopyFromReg(DAG.getEntryNode(), DL, 1988 Subtarget.isABI_N64() ? Mips::FP_64 : Mips::FP, VT); 1989 return FrameAddr; 1990 } 1991 1992 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 1993 SelectionDAG &DAG) const { 1994 if (verifyReturnAddressArgumentIsConstant(Op, DAG)) 1995 return SDValue(); 1996 1997 // check the depth 1998 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1999 "Return address can be determined only for current frame."); 2000 2001 MachineFunction &MF = DAG.getMachineFunction(); 2002 MachineFrameInfo *MFI = MF.getFrameInfo(); 2003 MVT VT = Op.getSimpleValueType(); 2004 unsigned RA = Subtarget.isABI_N64() ? Mips::RA_64 : Mips::RA; 2005 MFI->setReturnAddressIsTaken(true); 2006 2007 // Return RA, which contains the return address. Mark it an implicit live-in. 2008 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 2009 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 2010 } 2011 2012 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is 2013 // generated from __builtin_eh_return (offset, handler) 2014 // The effect of this is to adjust the stack pointer by "offset" 2015 // and then branch to "handler". 2016 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 2017 const { 2018 MachineFunction &MF = DAG.getMachineFunction(); 2019 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2020 2021 MipsFI->setCallsEhReturn(); 2022 SDValue Chain = Op.getOperand(0); 2023 SDValue Offset = Op.getOperand(1); 2024 SDValue Handler = Op.getOperand(2); 2025 SDLoc DL(Op); 2026 EVT Ty = Subtarget.isABI_N64() ? MVT::i64 : MVT::i32; 2027 2028 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 2029 // EH_RETURN nodes, so that instructions are emitted back-to-back. 2030 unsigned OffsetReg = Subtarget.isABI_N64() ? Mips::V1_64 : Mips::V1; 2031 unsigned AddrReg = Subtarget.isABI_N64() ? Mips::V0_64 : Mips::V0; 2032 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 2033 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 2034 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 2035 DAG.getRegister(OffsetReg, Ty), 2036 DAG.getRegister(AddrReg, getPointerTy()), 2037 Chain.getValue(1)); 2038 } 2039 2040 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 2041 SelectionDAG &DAG) const { 2042 // FIXME: Need pseudo-fence for 'singlethread' fences 2043 // FIXME: Set SType for weaker fences where supported/appropriate. 2044 unsigned SType = 0; 2045 SDLoc DL(Op); 2046 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 2047 DAG.getConstant(SType, MVT::i32)); 2048 } 2049 2050 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 2051 SelectionDAG &DAG) const { 2052 SDLoc DL(Op); 2053 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2054 2055 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2056 SDValue Shamt = Op.getOperand(2); 2057 // if shamt < (VT.bits): 2058 // lo = (shl lo, shamt) 2059 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 2060 // else: 2061 // lo = 0 2062 // hi = (shl lo, shamt[4:0]) 2063 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2064 DAG.getConstant(-1, MVT::i32)); 2065 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, 2066 DAG.getConstant(1, VT)); 2067 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not); 2068 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); 2069 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2070 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); 2071 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2072 DAG.getConstant(VT.getSizeInBits(), MVT::i32)); 2073 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2074 DAG.getConstant(0, VT), ShiftLeftLo); 2075 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or); 2076 2077 SDValue Ops[2] = {Lo, Hi}; 2078 return DAG.getMergeValues(Ops, DL); 2079 } 2080 2081 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 2082 bool IsSRA) const { 2083 SDLoc DL(Op); 2084 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 2085 SDValue Shamt = Op.getOperand(2); 2086 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; 2087 2088 // if shamt < (VT.bits): 2089 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 2090 // if isSRA: 2091 // hi = (sra hi, shamt) 2092 // else: 2093 // hi = (srl hi, shamt) 2094 // else: 2095 // if isSRA: 2096 // lo = (sra hi, shamt[4:0]) 2097 // hi = (sra hi, 31) 2098 // else: 2099 // lo = (srl hi, shamt[4:0]) 2100 // hi = 0 2101 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 2102 DAG.getConstant(-1, MVT::i32)); 2103 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi, 2104 DAG.getConstant(1, VT)); 2105 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not); 2106 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); 2107 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); 2108 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, 2109 DL, VT, Hi, Shamt); 2110 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 2111 DAG.getConstant(VT.getSizeInBits(), MVT::i32)); 2112 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi, 2113 DAG.getConstant(VT.getSizeInBits() - 1, VT)); 2114 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or); 2115 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, 2116 IsSRA ? Ext : DAG.getConstant(0, VT), ShiftRightHi); 2117 2118 SDValue Ops[2] = {Lo, Hi}; 2119 return DAG.getMergeValues(Ops, DL); 2120 } 2121 2122 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 2123 SDValue Chain, SDValue Src, unsigned Offset) { 2124 SDValue Ptr = LD->getBasePtr(); 2125 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 2126 EVT BasePtrVT = Ptr.getValueType(); 2127 SDLoc DL(LD); 2128 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 2129 2130 if (Offset) 2131 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2132 DAG.getConstant(Offset, BasePtrVT)); 2133 2134 SDValue Ops[] = { Chain, Ptr, Src }; 2135 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2136 LD->getMemOperand()); 2137 } 2138 2139 // Expand an unaligned 32 or 64-bit integer load node. 2140 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 2141 LoadSDNode *LD = cast<LoadSDNode>(Op); 2142 EVT MemVT = LD->getMemoryVT(); 2143 2144 if (Subtarget.systemSupportsUnalignedAccess()) 2145 return Op; 2146 2147 // Return if load is aligned or if MemVT is neither i32 nor i64. 2148 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 2149 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 2150 return SDValue(); 2151 2152 bool IsLittle = Subtarget.isLittle(); 2153 EVT VT = Op.getValueType(); 2154 ISD::LoadExtType ExtType = LD->getExtensionType(); 2155 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 2156 2157 assert((VT == MVT::i32) || (VT == MVT::i64)); 2158 2159 // Expand 2160 // (set dst, (i64 (load baseptr))) 2161 // to 2162 // (set tmp, (ldl (add baseptr, 7), undef)) 2163 // (set dst, (ldr baseptr, tmp)) 2164 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 2165 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 2166 IsLittle ? 7 : 0); 2167 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 2168 IsLittle ? 0 : 7); 2169 } 2170 2171 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 2172 IsLittle ? 3 : 0); 2173 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 2174 IsLittle ? 0 : 3); 2175 2176 // Expand 2177 // (set dst, (i32 (load baseptr))) or 2178 // (set dst, (i64 (sextload baseptr))) or 2179 // (set dst, (i64 (extload baseptr))) 2180 // to 2181 // (set tmp, (lwl (add baseptr, 3), undef)) 2182 // (set dst, (lwr baseptr, tmp)) 2183 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2184 (ExtType == ISD::EXTLOAD)) 2185 return LWR; 2186 2187 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2188 2189 // Expand 2190 // (set dst, (i64 (zextload baseptr))) 2191 // to 2192 // (set tmp0, (lwl (add baseptr, 3), undef)) 2193 // (set tmp1, (lwr baseptr, tmp0)) 2194 // (set tmp2, (shl tmp1, 32)) 2195 // (set dst, (srl tmp2, 32)) 2196 SDLoc DL(LD); 2197 SDValue Const32 = DAG.getConstant(32, MVT::i32); 2198 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2199 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2200 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2201 return DAG.getMergeValues(Ops, DL); 2202 } 2203 2204 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2205 SDValue Chain, unsigned Offset) { 2206 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2207 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2208 SDLoc DL(SD); 2209 SDVTList VTList = DAG.getVTList(MVT::Other); 2210 2211 if (Offset) 2212 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2213 DAG.getConstant(Offset, BasePtrVT)); 2214 2215 SDValue Ops[] = { Chain, Value, Ptr }; 2216 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT, 2217 SD->getMemOperand()); 2218 } 2219 2220 // Expand an unaligned 32 or 64-bit integer store node. 2221 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2222 bool IsLittle) { 2223 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2224 EVT VT = Value.getValueType(); 2225 2226 // Expand 2227 // (store val, baseptr) or 2228 // (truncstore val, baseptr) 2229 // to 2230 // (swl val, (add baseptr, 3)) 2231 // (swr val, baseptr) 2232 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2233 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2234 IsLittle ? 3 : 0); 2235 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2236 } 2237 2238 assert(VT == MVT::i64); 2239 2240 // Expand 2241 // (store val, baseptr) 2242 // to 2243 // (sdl val, (add baseptr, 7)) 2244 // (sdr val, baseptr) 2245 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2246 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2247 } 2248 2249 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2250 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) { 2251 SDValue Val = SD->getValue(); 2252 2253 if (Val.getOpcode() != ISD::FP_TO_SINT) 2254 return SDValue(); 2255 2256 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2257 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2258 Val.getOperand(0)); 2259 2260 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2261 SD->getPointerInfo(), SD->isVolatile(), 2262 SD->isNonTemporal(), SD->getAlignment()); 2263 } 2264 2265 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2266 StoreSDNode *SD = cast<StoreSDNode>(Op); 2267 EVT MemVT = SD->getMemoryVT(); 2268 2269 // Lower unaligned integer stores. 2270 if (!Subtarget.systemSupportsUnalignedAccess() && 2271 (SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2272 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2273 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle()); 2274 2275 return lowerFP_TO_SINT_STORE(SD, DAG); 2276 } 2277 2278 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const { 2279 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR 2280 || cast<ConstantSDNode> 2281 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0 2282 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET) 2283 return SDValue(); 2284 2285 // The pattern 2286 // (add (frameaddr 0), (frame_to_args_offset)) 2287 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to 2288 // (add FrameObject, 0) 2289 // where FrameObject is a fixed StackObject with offset 0 which points to 2290 // the old stack pointer. 2291 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2292 EVT ValTy = Op->getValueType(0); 2293 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2294 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy); 2295 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr, 2296 DAG.getConstant(0, ValTy)); 2297 } 2298 2299 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2300 SelectionDAG &DAG) const { 2301 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2302 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2303 Op.getOperand(0)); 2304 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2305 } 2306 2307 //===----------------------------------------------------------------------===// 2308 // Calling Convention Implementation 2309 //===----------------------------------------------------------------------===// 2310 2311 //===----------------------------------------------------------------------===// 2312 // TODO: Implement a generic logic using tblgen that can support this. 2313 // Mips O32 ABI rules: 2314 // --- 2315 // i32 - Passed in A0, A1, A2, A3 and stack 2316 // f32 - Only passed in f32 registers if no int reg has been used yet to hold 2317 // an argument. Otherwise, passed in A1, A2, A3 and stack. 2318 // f64 - Only passed in two aliased f32 registers if no int reg has been used 2319 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2320 // not used, it must be shadowed. If only A3 is available, shadow it and 2321 // go to stack. 2322 // 2323 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2324 //===----------------------------------------------------------------------===// 2325 2326 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2327 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2328 CCState &State, const MCPhysReg *F64Regs) { 2329 const MipsSubtarget &Subtarget = 2330 State.getMachineFunction().getTarget() 2331 .getSubtarget<const MipsSubtarget>(); 2332 2333 static const unsigned IntRegsSize = 4, FloatRegsSize = 2; 2334 2335 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; 2336 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; 2337 2338 // Do not process byval args here. 2339 if (ArgFlags.isByVal()) 2340 return true; 2341 2342 // Promote i8 and i16 2343 if (ArgFlags.isInReg() && !Subtarget.isLittle()) { 2344 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { 2345 LocVT = MVT::i32; 2346 if (ArgFlags.isSExt()) 2347 LocInfo = CCValAssign::SExtUpper; 2348 else if (ArgFlags.isZExt()) 2349 LocInfo = CCValAssign::ZExtUpper; 2350 else 2351 LocInfo = CCValAssign::AExtUpper; 2352 } 2353 } 2354 2355 // Promote i8 and i16 2356 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2357 LocVT = MVT::i32; 2358 if (ArgFlags.isSExt()) 2359 LocInfo = CCValAssign::SExt; 2360 else if (ArgFlags.isZExt()) 2361 LocInfo = CCValAssign::ZExt; 2362 else 2363 LocInfo = CCValAssign::AExt; 2364 } 2365 2366 unsigned Reg; 2367 2368 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2369 // is true: function is vararg, argument is 3rd or higher, there is previous 2370 // argument which is not f32 or f64. 2371 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 2372 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo; 2373 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2374 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2375 2376 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2377 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2378 // If this is the first part of an i64 arg, 2379 // the allocated register must be either A0 or A2. 2380 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2381 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2382 LocVT = MVT::i32; 2383 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2384 // Allocate int register and shadow next int register. If first 2385 // available register is Mips::A1 or Mips::A3, shadow it too. 2386 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2387 if (Reg == Mips::A1 || Reg == Mips::A3) 2388 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2389 State.AllocateReg(IntRegs, IntRegsSize); 2390 LocVT = MVT::i32; 2391 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2392 // we are guaranteed to find an available float register 2393 if (ValVT == MVT::f32) { 2394 Reg = State.AllocateReg(F32Regs, FloatRegsSize); 2395 // Shadow int register 2396 State.AllocateReg(IntRegs, IntRegsSize); 2397 } else { 2398 Reg = State.AllocateReg(F64Regs, FloatRegsSize); 2399 // Shadow int registers 2400 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize); 2401 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2402 State.AllocateReg(IntRegs, IntRegsSize); 2403 State.AllocateReg(IntRegs, IntRegsSize); 2404 } 2405 } else 2406 llvm_unreachable("Cannot handle this ValVT."); 2407 2408 if (!Reg) { 2409 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3, 2410 OrigAlign); 2411 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2412 } else 2413 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2414 2415 return false; 2416 } 2417 2418 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, 2419 MVT LocVT, CCValAssign::LocInfo LocInfo, 2420 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2421 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; 2422 2423 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2424 } 2425 2426 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, 2427 MVT LocVT, CCValAssign::LocInfo LocInfo, 2428 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2429 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; 2430 2431 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2432 } 2433 2434 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, 2435 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, 2436 CCState &State) LLVM_ATTRIBUTE_UNUSED; 2437 2438 #include "MipsGenCallingConv.inc" 2439 2440 //===----------------------------------------------------------------------===// 2441 // Call Calling Convention Implementation 2442 //===----------------------------------------------------------------------===// 2443 2444 // Return next O32 integer argument register. 2445 static unsigned getNextIntArgReg(unsigned Reg) { 2446 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 2447 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 2448 } 2449 2450 SDValue 2451 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 2452 SDValue Chain, SDValue Arg, SDLoc DL, 2453 bool IsTailCall, SelectionDAG &DAG) const { 2454 if (!IsTailCall) { 2455 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, 2456 DAG.getIntPtrConstant(Offset)); 2457 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false, 2458 false, 0); 2459 } 2460 2461 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2462 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 2463 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2464 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 2465 /*isVolatile=*/ true, false, 0); 2466 } 2467 2468 void MipsTargetLowering:: 2469 getOpndList(SmallVectorImpl<SDValue> &Ops, 2470 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 2471 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 2472 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, 2473 SDValue Chain) const { 2474 // Insert node "GP copy globalreg" before call to function. 2475 // 2476 // R_MIPS_CALL* operators (emitted when non-internal functions are called 2477 // in PIC mode) allow symbols to be resolved via lazy binding. 2478 // The lazy binding stub requires GP to point to the GOT. 2479 // Note that we don't need GP to point to the GOT for indirect calls 2480 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates 2481 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs 2482 // used for the function (that is, Mips linker doesn't generate lazy binding 2483 // stub for a function whose address is taken in the program). 2484 if (IsPICCall && !InternalLinkage && IsCallReloc) { 2485 unsigned GPReg = Subtarget.isABI_N64() ? Mips::GP_64 : Mips::GP; 2486 EVT Ty = Subtarget.isABI_N64() ? MVT::i64 : MVT::i32; 2487 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 2488 } 2489 2490 // Build a sequence of copy-to-reg nodes chained together with token 2491 // chain and flag operands which copy the outgoing args into registers. 2492 // The InFlag in necessary since all emitted instructions must be 2493 // stuck together. 2494 SDValue InFlag; 2495 2496 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 2497 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 2498 RegsToPass[i].second, InFlag); 2499 InFlag = Chain.getValue(1); 2500 } 2501 2502 // Add argument registers to the end of the list so that they are 2503 // known live into the call. 2504 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 2505 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 2506 RegsToPass[i].second.getValueType())); 2507 2508 // Add a register mask operand representing the call-preserved registers. 2509 const TargetRegisterInfo *TRI = 2510 getTargetMachine().getSubtargetImpl()->getRegisterInfo(); 2511 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv); 2512 assert(Mask && "Missing call preserved mask for calling convention"); 2513 if (Subtarget.inMips16HardFloat()) { 2514 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) { 2515 llvm::StringRef Sym = G->getGlobal()->getName(); 2516 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 2517 if (F && F->hasFnAttribute("__Mips16RetHelper")) { 2518 Mask = MipsRegisterInfo::getMips16RetHelperMask(); 2519 } 2520 } 2521 } 2522 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 2523 2524 if (InFlag.getNode()) 2525 Ops.push_back(InFlag); 2526 } 2527 2528 /// LowerCall - functions arguments are copied from virtual regs to 2529 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 2530 SDValue 2531 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 2532 SmallVectorImpl<SDValue> &InVals) const { 2533 SelectionDAG &DAG = CLI.DAG; 2534 SDLoc DL = CLI.DL; 2535 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 2536 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 2537 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 2538 SDValue Chain = CLI.Chain; 2539 SDValue Callee = CLI.Callee; 2540 bool &IsTailCall = CLI.IsTailCall; 2541 CallingConv::ID CallConv = CLI.CallConv; 2542 bool IsVarArg = CLI.IsVarArg; 2543 2544 MachineFunction &MF = DAG.getMachineFunction(); 2545 MachineFrameInfo *MFI = MF.getFrameInfo(); 2546 const TargetFrameLowering *TFL = MF.getSubtarget().getFrameLowering(); 2547 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 2548 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_; 2549 2550 // Analyze operands of the call, assigning locations to each operand. 2551 SmallVector<CCValAssign, 16> ArgLocs; 2552 MipsCCState CCInfo( 2553 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), 2554 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget)); 2555 2556 // Allocate the reserved argument area. It seems strange to do this from the 2557 // caller side but removing it breaks the frame size calculation. 2558 const MipsABIInfo &ABI = Subtarget.getABI(); 2559 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 2560 2561 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), Callee.getNode()); 2562 2563 // Get a count of how many bytes are to be pushed on the stack. 2564 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2565 2566 // Check if it's really possible to do a tail call. 2567 if (IsTailCall) 2568 IsTailCall = isEligibleForTailCallOptimization( 2569 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>()); 2570 2571 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall()) 2572 report_fatal_error("failed to perform tail call elimination on a call " 2573 "site marked musttail"); 2574 2575 if (IsTailCall) 2576 ++NumTailCalls; 2577 2578 // Chain is the output chain of the last Load/Store or CopyToReg node. 2579 // ByValChain is the output chain of the last Memcpy node created for copying 2580 // byval arguments to the stack. 2581 unsigned StackAlignment = TFL->getStackAlignment(); 2582 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment); 2583 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true); 2584 2585 if (!IsTailCall) 2586 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL); 2587 2588 SDValue StackPtr = DAG.getCopyFromReg( 2589 Chain, DL, Subtarget.isABI_N64() ? Mips::SP_64 : Mips::SP, 2590 getPointerTy()); 2591 2592 // With EABI is it possible to have 16 args on registers. 2593 std::deque< std::pair<unsigned, SDValue> > RegsToPass; 2594 SmallVector<SDValue, 8> MemOpChains; 2595 2596 CCInfo.rewindByValRegsInfo(); 2597 2598 // Walk the register/memloc assignments, inserting copies/loads. 2599 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2600 SDValue Arg = OutVals[i]; 2601 CCValAssign &VA = ArgLocs[i]; 2602 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 2603 ISD::ArgFlagsTy Flags = Outs[i].Flags; 2604 bool UseUpperBits = false; 2605 2606 // ByVal Arg. 2607 if (Flags.isByVal()) { 2608 unsigned FirstByValReg, LastByValReg; 2609 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 2610 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 2611 2612 assert(Flags.getByValSize() && 2613 "ByVal args of size 0 should have been ignored by front-end."); 2614 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 2615 assert(!IsTailCall && 2616 "Do not tail-call optimize if there is a byval argument."); 2617 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 2618 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(), 2619 VA); 2620 CCInfo.nextInRegsParam(); 2621 continue; 2622 } 2623 2624 // Promote the value if needed. 2625 switch (VA.getLocInfo()) { 2626 default: 2627 llvm_unreachable("Unknown loc info!"); 2628 case CCValAssign::Full: 2629 if (VA.isRegLoc()) { 2630 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 2631 (ValVT == MVT::f64 && LocVT == MVT::i64) || 2632 (ValVT == MVT::i64 && LocVT == MVT::f64)) 2633 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 2634 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 2635 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2636 Arg, DAG.getConstant(0, MVT::i32)); 2637 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2638 Arg, DAG.getConstant(1, MVT::i32)); 2639 if (!Subtarget.isLittle()) 2640 std::swap(Lo, Hi); 2641 unsigned LocRegLo = VA.getLocReg(); 2642 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 2643 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 2644 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 2645 continue; 2646 } 2647 } 2648 break; 2649 case CCValAssign::BCvt: 2650 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 2651 break; 2652 case CCValAssign::SExtUpper: 2653 UseUpperBits = true; 2654 // Fallthrough 2655 case CCValAssign::SExt: 2656 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 2657 break; 2658 case CCValAssign::ZExtUpper: 2659 UseUpperBits = true; 2660 // Fallthrough 2661 case CCValAssign::ZExt: 2662 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 2663 break; 2664 case CCValAssign::AExtUpper: 2665 UseUpperBits = true; 2666 // Fallthrough 2667 case CCValAssign::AExt: 2668 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 2669 break; 2670 } 2671 2672 if (UseUpperBits) { 2673 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 2674 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 2675 Arg = DAG.getNode( 2676 ISD::SHL, DL, VA.getLocVT(), Arg, 2677 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT())); 2678 } 2679 2680 // Arguments that can be passed on register must be kept at 2681 // RegsToPass vector 2682 if (VA.isRegLoc()) { 2683 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 2684 continue; 2685 } 2686 2687 // Register can't get to this point... 2688 assert(VA.isMemLoc()); 2689 2690 // emit ISD::STORE whichs stores the 2691 // parameter value to a stack Location 2692 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 2693 Chain, Arg, DL, IsTailCall, DAG)); 2694 } 2695 2696 // Transform all store nodes into one single node because all store 2697 // nodes are independent of each other. 2698 if (!MemOpChains.empty()) 2699 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); 2700 2701 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 2702 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 2703 // node so that legalize doesn't hack it. 2704 bool IsPICCall = 2705 (Subtarget.isABI_N64() || IsPIC); // true if calls are translated to 2706 // jalr $25 2707 bool GlobalOrExternal = false, InternalLinkage = false, IsCallReloc = false; 2708 SDValue CalleeLo; 2709 EVT Ty = Callee.getValueType(); 2710 2711 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2712 if (IsPICCall) { 2713 const GlobalValue *Val = G->getGlobal(); 2714 InternalLinkage = Val->hasInternalLinkage(); 2715 2716 if (InternalLinkage) 2717 Callee = getAddrLocal(G, DL, Ty, DAG, 2718 Subtarget.isABI_N32() || Subtarget.isABI_N64()); 2719 else if (LargeGOT) { 2720 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16, 2721 MipsII::MO_CALL_LO16, Chain, 2722 FuncInfo->callPtrInfo(Val)); 2723 IsCallReloc = true; 2724 } else { 2725 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 2726 FuncInfo->callPtrInfo(Val)); 2727 IsCallReloc = true; 2728 } 2729 } else 2730 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0, 2731 MipsII::MO_NO_FLAG); 2732 GlobalOrExternal = true; 2733 } 2734 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 2735 const char *Sym = S->getSymbol(); 2736 2737 if (!Subtarget.isABI_N64() && !IsPIC) // !N64 && static 2738 Callee = 2739 DAG.getTargetExternalSymbol(Sym, getPointerTy(), MipsII::MO_NO_FLAG); 2740 else if (LargeGOT) { 2741 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16, 2742 MipsII::MO_CALL_LO16, Chain, 2743 FuncInfo->callPtrInfo(Sym)); 2744 IsCallReloc = true; 2745 } else { // N64 || PIC 2746 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain, 2747 FuncInfo->callPtrInfo(Sym)); 2748 IsCallReloc = true; 2749 } 2750 2751 GlobalOrExternal = true; 2752 } 2753 2754 SmallVector<SDValue, 8> Ops(1, Chain); 2755 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 2756 2757 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage, 2758 IsCallReloc, CLI, Callee, Chain); 2759 2760 if (IsTailCall) 2761 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops); 2762 2763 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops); 2764 SDValue InFlag = Chain.getValue(1); 2765 2766 // Create the CALLSEQ_END node. 2767 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 2768 DAG.getIntPtrConstant(0, true), InFlag, DL); 2769 InFlag = Chain.getValue(1); 2770 2771 // Handle result values, copying them out of physregs into vregs that we 2772 // return. 2773 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG, 2774 InVals, CLI); 2775 } 2776 2777 /// LowerCallResult - Lower the result values of a call into the 2778 /// appropriate copies out of appropriate physical registers. 2779 SDValue MipsTargetLowering::LowerCallResult( 2780 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg, 2781 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG, 2782 SmallVectorImpl<SDValue> &InVals, 2783 TargetLowering::CallLoweringInfo &CLI) const { 2784 // Assign locations to each value returned by this call. 2785 SmallVector<CCValAssign, 16> RVLocs; 2786 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, 2787 *DAG.getContext()); 2788 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI); 2789 2790 // Copy all of the result registers out of their specified physreg. 2791 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2792 CCValAssign &VA = RVLocs[i]; 2793 assert(VA.isRegLoc() && "Can only return in registers!"); 2794 2795 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 2796 RVLocs[i].getLocVT(), InFlag); 2797 Chain = Val.getValue(1); 2798 InFlag = Val.getValue(2); 2799 2800 if (VA.isUpperBitsInLoc()) { 2801 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); 2802 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 2803 unsigned Shift = 2804 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 2805 Val = DAG.getNode( 2806 Shift, DL, VA.getLocVT(), Val, 2807 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT())); 2808 } 2809 2810 switch (VA.getLocInfo()) { 2811 default: 2812 llvm_unreachable("Unknown loc info!"); 2813 case CCValAssign::Full: 2814 break; 2815 case CCValAssign::BCvt: 2816 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); 2817 break; 2818 case CCValAssign::AExt: 2819 case CCValAssign::AExtUpper: 2820 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 2821 break; 2822 case CCValAssign::ZExt: 2823 case CCValAssign::ZExtUpper: 2824 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val, 2825 DAG.getValueType(VA.getValVT())); 2826 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 2827 break; 2828 case CCValAssign::SExt: 2829 case CCValAssign::SExtUpper: 2830 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val, 2831 DAG.getValueType(VA.getValVT())); 2832 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val); 2833 break; 2834 } 2835 2836 InVals.push_back(Val); 2837 } 2838 2839 return Chain; 2840 } 2841 2842 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, 2843 EVT ArgVT, SDLoc DL, SelectionDAG &DAG) { 2844 MVT LocVT = VA.getLocVT(); 2845 EVT ValVT = VA.getValVT(); 2846 2847 // Shift into the upper bits if necessary. 2848 switch (VA.getLocInfo()) { 2849 default: 2850 break; 2851 case CCValAssign::AExtUpper: 2852 case CCValAssign::SExtUpper: 2853 case CCValAssign::ZExtUpper: { 2854 unsigned ValSizeInBits = ArgVT.getSizeInBits(); 2855 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 2856 unsigned Opcode = 2857 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; 2858 Val = DAG.getNode( 2859 Opcode, DL, VA.getLocVT(), Val, 2860 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT())); 2861 break; 2862 } 2863 } 2864 2865 // If this is an value smaller than the argument slot size (32-bit for O32, 2866 // 64-bit for N32/N64), it has been promoted in some way to the argument slot 2867 // size. Extract the value and insert any appropriate assertions regarding 2868 // sign/zero extension. 2869 switch (VA.getLocInfo()) { 2870 default: 2871 llvm_unreachable("Unknown loc info!"); 2872 case CCValAssign::Full: 2873 break; 2874 case CCValAssign::AExtUpper: 2875 case CCValAssign::AExt: 2876 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 2877 break; 2878 case CCValAssign::SExtUpper: 2879 case CCValAssign::SExt: 2880 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT)); 2881 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 2882 break; 2883 case CCValAssign::ZExtUpper: 2884 case CCValAssign::ZExt: 2885 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT)); 2886 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val); 2887 break; 2888 case CCValAssign::BCvt: 2889 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val); 2890 break; 2891 } 2892 2893 return Val; 2894 } 2895 2896 //===----------------------------------------------------------------------===// 2897 // Formal Arguments Calling Convention Implementation 2898 //===----------------------------------------------------------------------===// 2899 /// LowerFormalArguments - transform physical registers into virtual registers 2900 /// and generate load operations for arguments places on the stack. 2901 SDValue 2902 MipsTargetLowering::LowerFormalArguments(SDValue Chain, 2903 CallingConv::ID CallConv, 2904 bool IsVarArg, 2905 const SmallVectorImpl<ISD::InputArg> &Ins, 2906 SDLoc DL, SelectionDAG &DAG, 2907 SmallVectorImpl<SDValue> &InVals) 2908 const { 2909 MachineFunction &MF = DAG.getMachineFunction(); 2910 MachineFrameInfo *MFI = MF.getFrameInfo(); 2911 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2912 2913 MipsFI->setVarArgsFrameIndex(0); 2914 2915 // Used with vargs to acumulate store chains. 2916 std::vector<SDValue> OutChains; 2917 2918 // Assign locations to all of the incoming arguments. 2919 SmallVector<CCValAssign, 16> ArgLocs; 2920 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, 2921 *DAG.getContext()); 2922 const MipsABIInfo &ABI = Subtarget.getABI(); 2923 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1); 2924 Function::const_arg_iterator FuncArg = 2925 DAG.getMachineFunction().getFunction()->arg_begin(); 2926 2927 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg); 2928 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 2929 CCInfo.getInRegsParamsCount() > 0); 2930 2931 unsigned CurArgIdx = 0; 2932 CCInfo.rewindByValRegsInfo(); 2933 2934 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2935 CCValAssign &VA = ArgLocs[i]; 2936 if (Ins[i].isOrigArg()) { 2937 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx); 2938 CurArgIdx = Ins[i].getOrigArgIndex(); 2939 } 2940 EVT ValVT = VA.getValVT(); 2941 ISD::ArgFlagsTy Flags = Ins[i].Flags; 2942 bool IsRegLoc = VA.isRegLoc(); 2943 2944 if (Flags.isByVal()) { 2945 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit"); 2946 unsigned FirstByValReg, LastByValReg; 2947 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); 2948 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg); 2949 2950 assert(Flags.getByValSize() && 2951 "ByVal args of size 0 should have been ignored by front-end."); 2952 assert(ByValIdx < CCInfo.getInRegsParamsCount()); 2953 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 2954 FirstByValReg, LastByValReg, VA, CCInfo); 2955 CCInfo.nextInRegsParam(); 2956 continue; 2957 } 2958 2959 // Arguments stored on registers 2960 if (IsRegLoc) { 2961 MVT RegVT = VA.getLocVT(); 2962 unsigned ArgReg = VA.getLocReg(); 2963 const TargetRegisterClass *RC = getRegClassFor(RegVT); 2964 2965 // Transform the arguments stored on 2966 // physical registers into virtual ones 2967 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 2968 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 2969 2970 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 2971 2972 // Handle floating point arguments passed in integer registers and 2973 // long double arguments passed in floating point registers. 2974 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 2975 (RegVT == MVT::i64 && ValVT == MVT::f64) || 2976 (RegVT == MVT::f64 && ValVT == MVT::i64)) 2977 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 2978 else if (Subtarget.isABI_O32() && RegVT == MVT::i32 && 2979 ValVT == MVT::f64) { 2980 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 2981 getNextIntArgReg(ArgReg), RC); 2982 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 2983 if (!Subtarget.isLittle()) 2984 std::swap(ArgValue, ArgValue2); 2985 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 2986 ArgValue, ArgValue2); 2987 } 2988 2989 InVals.push_back(ArgValue); 2990 } else { // VA.isRegLoc() 2991 MVT LocVT = VA.getLocVT(); 2992 2993 if (Subtarget.isABI_O32()) { 2994 // We ought to be able to use LocVT directly but O32 sets it to i32 2995 // when allocating floating point values to integer registers. 2996 // This shouldn't influence how we load the value into registers unless 2997 // we are targetting softfloat. 2998 if (VA.getValVT().isFloatingPoint() && !Subtarget.abiUsesSoftFloat()) 2999 LocVT = VA.getValVT(); 3000 } 3001 3002 // sanity check 3003 assert(VA.isMemLoc()); 3004 3005 // The stack pointer offset is relative to the caller stack frame. 3006 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8, 3007 VA.getLocMemOffset(), true); 3008 3009 // Create load nodes to retrieve arguments from the stack 3010 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 3011 SDValue ArgValue = DAG.getLoad(LocVT, DL, Chain, FIN, 3012 MachinePointerInfo::getFixedStack(FI), 3013 false, false, false, 0); 3014 OutChains.push_back(ArgValue.getValue(1)); 3015 3016 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG); 3017 3018 InVals.push_back(ArgValue); 3019 } 3020 } 3021 3022 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 3023 // The mips ABIs for returning structs by value requires that we copy 3024 // the sret argument into $v0 for the return. Save the argument into 3025 // a virtual register so that we can access it from the return points. 3026 if (Ins[i].Flags.isSRet()) { 3027 unsigned Reg = MipsFI->getSRetReturnReg(); 3028 if (!Reg) { 3029 Reg = MF.getRegInfo().createVirtualRegister( 3030 getRegClassFor(Subtarget.isABI_N64() ? MVT::i64 : MVT::i32)); 3031 MipsFI->setSRetReturnReg(Reg); 3032 } 3033 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]); 3034 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 3035 break; 3036 } 3037 } 3038 3039 if (IsVarArg) 3040 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo); 3041 3042 // All stores are grouped in one node to allow the matching between 3043 // the size of Ins and InVals. This only happens when on varg functions 3044 if (!OutChains.empty()) { 3045 OutChains.push_back(Chain); 3046 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); 3047 } 3048 3049 return Chain; 3050 } 3051 3052 //===----------------------------------------------------------------------===// 3053 // Return Value Calling Convention Implementation 3054 //===----------------------------------------------------------------------===// 3055 3056 bool 3057 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 3058 MachineFunction &MF, bool IsVarArg, 3059 const SmallVectorImpl<ISD::OutputArg> &Outs, 3060 LLVMContext &Context) const { 3061 SmallVector<CCValAssign, 16> RVLocs; 3062 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); 3063 return CCInfo.CheckReturn(Outs, RetCC_Mips); 3064 } 3065 3066 bool 3067 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const { 3068 if (Subtarget.hasMips3() && Subtarget.abiUsesSoftFloat()) { 3069 if (Type == MVT::i32) 3070 return true; 3071 } 3072 return IsSigned; 3073 } 3074 3075 SDValue 3076 MipsTargetLowering::LowerReturn(SDValue Chain, 3077 CallingConv::ID CallConv, bool IsVarArg, 3078 const SmallVectorImpl<ISD::OutputArg> &Outs, 3079 const SmallVectorImpl<SDValue> &OutVals, 3080 SDLoc DL, SelectionDAG &DAG) const { 3081 // CCValAssign - represent the assignment of 3082 // the return value to a location 3083 SmallVector<CCValAssign, 16> RVLocs; 3084 MachineFunction &MF = DAG.getMachineFunction(); 3085 3086 // CCState - Info about the registers and stack slot. 3087 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); 3088 3089 // Analyze return values. 3090 CCInfo.AnalyzeReturn(Outs, RetCC_Mips); 3091 3092 SDValue Flag; 3093 SmallVector<SDValue, 4> RetOps(1, Chain); 3094 3095 // Copy the result values into the output registers. 3096 for (unsigned i = 0; i != RVLocs.size(); ++i) { 3097 SDValue Val = OutVals[i]; 3098 CCValAssign &VA = RVLocs[i]; 3099 assert(VA.isRegLoc() && "Can only return in registers!"); 3100 bool UseUpperBits = false; 3101 3102 switch (VA.getLocInfo()) { 3103 default: 3104 llvm_unreachable("Unknown loc info!"); 3105 case CCValAssign::Full: 3106 break; 3107 case CCValAssign::BCvt: 3108 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val); 3109 break; 3110 case CCValAssign::AExtUpper: 3111 UseUpperBits = true; 3112 // Fallthrough 3113 case CCValAssign::AExt: 3114 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val); 3115 break; 3116 case CCValAssign::ZExtUpper: 3117 UseUpperBits = true; 3118 // Fallthrough 3119 case CCValAssign::ZExt: 3120 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val); 3121 break; 3122 case CCValAssign::SExtUpper: 3123 UseUpperBits = true; 3124 // Fallthrough 3125 case CCValAssign::SExt: 3126 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val); 3127 break; 3128 } 3129 3130 if (UseUpperBits) { 3131 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); 3132 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); 3133 Val = DAG.getNode( 3134 ISD::SHL, DL, VA.getLocVT(), Val, 3135 DAG.getConstant(LocSizeInBits - ValSizeInBits, VA.getLocVT())); 3136 } 3137 3138 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 3139 3140 // Guarantee that all emitted copies are stuck together with flags. 3141 Flag = Chain.getValue(1); 3142 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 3143 } 3144 3145 // The mips ABIs for returning structs by value requires that we copy 3146 // the sret argument into $v0 for the return. We saved the argument into 3147 // a virtual register in the entry block, so now we copy the value out 3148 // and into $v0. 3149 if (MF.getFunction()->hasStructRetAttr()) { 3150 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3151 unsigned Reg = MipsFI->getSRetReturnReg(); 3152 3153 if (!Reg) 3154 llvm_unreachable("sret virtual register not created in the entry block"); 3155 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy()); 3156 unsigned V0 = Subtarget.isABI_N64() ? Mips::V0_64 : Mips::V0; 3157 3158 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 3159 Flag = Chain.getValue(1); 3160 RetOps.push_back(DAG.getRegister(V0, getPointerTy())); 3161 } 3162 3163 RetOps[0] = Chain; // Update chain. 3164 3165 // Add the flag if we have it. 3166 if (Flag.getNode()) 3167 RetOps.push_back(Flag); 3168 3169 // Return on Mips is always a "jr $ra" 3170 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps); 3171 } 3172 3173 //===----------------------------------------------------------------------===// 3174 // Mips Inline Assembly Support 3175 //===----------------------------------------------------------------------===// 3176 3177 /// getConstraintType - Given a constraint letter, return the type of 3178 /// constraint it is for this target. 3179 MipsTargetLowering::ConstraintType MipsTargetLowering:: 3180 getConstraintType(const std::string &Constraint) const 3181 { 3182 // Mips specific constraints 3183 // GCC config/mips/constraints.md 3184 // 3185 // 'd' : An address register. Equivalent to r 3186 // unless generating MIPS16 code. 3187 // 'y' : Equivalent to r; retained for 3188 // backwards compatibility. 3189 // 'c' : A register suitable for use in an indirect 3190 // jump. This will always be $25 for -mabicalls. 3191 // 'l' : The lo register. 1 word storage. 3192 // 'x' : The hilo register pair. Double word storage. 3193 if (Constraint.size() == 1) { 3194 switch (Constraint[0]) { 3195 default : break; 3196 case 'd': 3197 case 'y': 3198 case 'f': 3199 case 'c': 3200 case 'l': 3201 case 'x': 3202 return C_RegisterClass; 3203 case 'R': 3204 return C_Memory; 3205 } 3206 } 3207 return TargetLowering::getConstraintType(Constraint); 3208 } 3209 3210 /// Examine constraint type and operand type and determine a weight value. 3211 /// This object must already have been set up with the operand type 3212 /// and the current alternative constraint selected. 3213 TargetLowering::ConstraintWeight 3214 MipsTargetLowering::getSingleConstraintMatchWeight( 3215 AsmOperandInfo &info, const char *constraint) const { 3216 ConstraintWeight weight = CW_Invalid; 3217 Value *CallOperandVal = info.CallOperandVal; 3218 // If we don't have a value, we can't do a match, 3219 // but allow it at the lowest weight. 3220 if (!CallOperandVal) 3221 return CW_Default; 3222 Type *type = CallOperandVal->getType(); 3223 // Look at the constraint type. 3224 switch (*constraint) { 3225 default: 3226 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 3227 break; 3228 case 'd': 3229 case 'y': 3230 if (type->isIntegerTy()) 3231 weight = CW_Register; 3232 break; 3233 case 'f': // FPU or MSA register 3234 if (Subtarget.hasMSA() && type->isVectorTy() && 3235 cast<VectorType>(type)->getBitWidth() == 128) 3236 weight = CW_Register; 3237 else if (type->isFloatTy()) 3238 weight = CW_Register; 3239 break; 3240 case 'c': // $25 for indirect jumps 3241 case 'l': // lo register 3242 case 'x': // hilo register pair 3243 if (type->isIntegerTy()) 3244 weight = CW_SpecificReg; 3245 break; 3246 case 'I': // signed 16 bit immediate 3247 case 'J': // integer zero 3248 case 'K': // unsigned 16 bit immediate 3249 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3250 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3251 case 'O': // signed 15 bit immediate (+- 16383) 3252 case 'P': // immediate in the range of 65535 to 1 (inclusive) 3253 if (isa<ConstantInt>(CallOperandVal)) 3254 weight = CW_Constant; 3255 break; 3256 case 'R': 3257 weight = CW_Memory; 3258 break; 3259 } 3260 return weight; 3261 } 3262 3263 /// This is a helper function to parse a physical register string and split it 3264 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 3265 /// that is returned indicates whether parsing was successful. The second flag 3266 /// is true if the numeric part exists. 3267 static std::pair<bool, bool> 3268 parsePhysicalReg(StringRef C, std::string &Prefix, 3269 unsigned long long &Reg) { 3270 if (C.front() != '{' || C.back() != '}') 3271 return std::make_pair(false, false); 3272 3273 // Search for the first numeric character. 3274 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 3275 I = std::find_if(B, E, std::ptr_fun(isdigit)); 3276 3277 Prefix.assign(B, I - B); 3278 3279 // The second flag is set to false if no numeric characters were found. 3280 if (I == E) 3281 return std::make_pair(true, false); 3282 3283 // Parse the numeric characters. 3284 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 3285 true); 3286 } 3287 3288 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 3289 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { 3290 const TargetRegisterInfo *TRI = 3291 getTargetMachine().getSubtargetImpl()->getRegisterInfo(); 3292 const TargetRegisterClass *RC; 3293 std::string Prefix; 3294 unsigned long long Reg; 3295 3296 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 3297 3298 if (!R.first) 3299 return std::make_pair(0U, nullptr); 3300 3301 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 3302 // No numeric characters follow "hi" or "lo". 3303 if (R.second) 3304 return std::make_pair(0U, nullptr); 3305 3306 RC = TRI->getRegClass(Prefix == "hi" ? 3307 Mips::HI32RegClassID : Mips::LO32RegClassID); 3308 return std::make_pair(*(RC->begin()), RC); 3309 } else if (Prefix.compare(0, 4, "$msa") == 0) { 3310 // Parse $msa(ir|csr|access|save|modify|request|map|unmap) 3311 3312 // No numeric characters follow the name. 3313 if (R.second) 3314 return std::make_pair(0U, nullptr); 3315 3316 Reg = StringSwitch<unsigned long long>(Prefix) 3317 .Case("$msair", Mips::MSAIR) 3318 .Case("$msacsr", Mips::MSACSR) 3319 .Case("$msaaccess", Mips::MSAAccess) 3320 .Case("$msasave", Mips::MSASave) 3321 .Case("$msamodify", Mips::MSAModify) 3322 .Case("$msarequest", Mips::MSARequest) 3323 .Case("$msamap", Mips::MSAMap) 3324 .Case("$msaunmap", Mips::MSAUnmap) 3325 .Default(0); 3326 3327 if (!Reg) 3328 return std::make_pair(0U, nullptr); 3329 3330 RC = TRI->getRegClass(Mips::MSACtrlRegClassID); 3331 return std::make_pair(Reg, RC); 3332 } 3333 3334 if (!R.second) 3335 return std::make_pair(0U, nullptr); 3336 3337 if (Prefix == "$f") { // Parse $f0-$f31. 3338 // If the size of FP registers is 64-bit or Reg is an even number, select 3339 // the 64-bit register class. Otherwise, select the 32-bit register class. 3340 if (VT == MVT::Other) 3341 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 3342 3343 RC = getRegClassFor(VT); 3344 3345 if (RC == &Mips::AFGR64RegClass) { 3346 assert(Reg % 2 == 0); 3347 Reg >>= 1; 3348 } 3349 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7. 3350 RC = TRI->getRegClass(Mips::FCCRegClassID); 3351 else if (Prefix == "$w") { // Parse $w0-$w31. 3352 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT); 3353 } else { // Parse $0-$31. 3354 assert(Prefix == "$"); 3355 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 3356 } 3357 3358 assert(Reg < RC->getNumRegs()); 3359 return std::make_pair(*(RC->begin() + Reg), RC); 3360 } 3361 3362 /// Given a register class constraint, like 'r', if this corresponds directly 3363 /// to an LLVM register class, return a register of 0 and the register class 3364 /// pointer. 3365 std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering:: 3366 getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const 3367 { 3368 if (Constraint.size() == 1) { 3369 switch (Constraint[0]) { 3370 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 3371 case 'y': // Same as 'r'. Exists for compatibility. 3372 case 'r': 3373 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 3374 if (Subtarget.inMips16Mode()) 3375 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 3376 return std::make_pair(0U, &Mips::GPR32RegClass); 3377 } 3378 if (VT == MVT::i64 && !Subtarget.isGP64bit()) 3379 return std::make_pair(0U, &Mips::GPR32RegClass); 3380 if (VT == MVT::i64 && Subtarget.isGP64bit()) 3381 return std::make_pair(0U, &Mips::GPR64RegClass); 3382 // This will generate an error message 3383 return std::make_pair(0U, nullptr); 3384 case 'f': // FPU or MSA register 3385 if (VT == MVT::v16i8) 3386 return std::make_pair(0U, &Mips::MSA128BRegClass); 3387 else if (VT == MVT::v8i16 || VT == MVT::v8f16) 3388 return std::make_pair(0U, &Mips::MSA128HRegClass); 3389 else if (VT == MVT::v4i32 || VT == MVT::v4f32) 3390 return std::make_pair(0U, &Mips::MSA128WRegClass); 3391 else if (VT == MVT::v2i64 || VT == MVT::v2f64) 3392 return std::make_pair(0U, &Mips::MSA128DRegClass); 3393 else if (VT == MVT::f32) 3394 return std::make_pair(0U, &Mips::FGR32RegClass); 3395 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { 3396 if (Subtarget.isFP64bit()) 3397 return std::make_pair(0U, &Mips::FGR64RegClass); 3398 return std::make_pair(0U, &Mips::AFGR64RegClass); 3399 } 3400 break; 3401 case 'c': // register suitable for indirect jump 3402 if (VT == MVT::i32) 3403 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 3404 assert(VT == MVT::i64 && "Unexpected type."); 3405 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 3406 case 'l': // register suitable for indirect jump 3407 if (VT == MVT::i32) 3408 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass); 3409 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass); 3410 case 'x': // register suitable for indirect jump 3411 // Fixme: Not triggering the use of both hi and low 3412 // This will generate an error message 3413 return std::make_pair(0U, nullptr); 3414 } 3415 } 3416 3417 std::pair<unsigned, const TargetRegisterClass *> R; 3418 R = parseRegForInlineAsmConstraint(Constraint, VT); 3419 3420 if (R.second) 3421 return R; 3422 3423 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3424 } 3425 3426 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3427 /// vector. If it is invalid, don't add anything to Ops. 3428 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3429 std::string &Constraint, 3430 std::vector<SDValue>&Ops, 3431 SelectionDAG &DAG) const { 3432 SDValue Result; 3433 3434 // Only support length 1 constraints for now. 3435 if (Constraint.length() > 1) return; 3436 3437 char ConstraintLetter = Constraint[0]; 3438 switch (ConstraintLetter) { 3439 default: break; // This will fall through to the generic implementation 3440 case 'I': // Signed 16 bit constant 3441 // If this fails, the parent routine will give an error 3442 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3443 EVT Type = Op.getValueType(); 3444 int64_t Val = C->getSExtValue(); 3445 if (isInt<16>(Val)) { 3446 Result = DAG.getTargetConstant(Val, Type); 3447 break; 3448 } 3449 } 3450 return; 3451 case 'J': // integer zero 3452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3453 EVT Type = Op.getValueType(); 3454 int64_t Val = C->getZExtValue(); 3455 if (Val == 0) { 3456 Result = DAG.getTargetConstant(0, Type); 3457 break; 3458 } 3459 } 3460 return; 3461 case 'K': // unsigned 16 bit immediate 3462 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3463 EVT Type = Op.getValueType(); 3464 uint64_t Val = (uint64_t)C->getZExtValue(); 3465 if (isUInt<16>(Val)) { 3466 Result = DAG.getTargetConstant(Val, Type); 3467 break; 3468 } 3469 } 3470 return; 3471 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3472 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3473 EVT Type = Op.getValueType(); 3474 int64_t Val = C->getSExtValue(); 3475 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3476 Result = DAG.getTargetConstant(Val, Type); 3477 break; 3478 } 3479 } 3480 return; 3481 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3482 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3483 EVT Type = Op.getValueType(); 3484 int64_t Val = C->getSExtValue(); 3485 if ((Val >= -65535) && (Val <= -1)) { 3486 Result = DAG.getTargetConstant(Val, Type); 3487 break; 3488 } 3489 } 3490 return; 3491 case 'O': // signed 15 bit immediate 3492 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3493 EVT Type = Op.getValueType(); 3494 int64_t Val = C->getSExtValue(); 3495 if ((isInt<15>(Val))) { 3496 Result = DAG.getTargetConstant(Val, Type); 3497 break; 3498 } 3499 } 3500 return; 3501 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3502 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3503 EVT Type = Op.getValueType(); 3504 int64_t Val = C->getSExtValue(); 3505 if ((Val <= 65535) && (Val >= 1)) { 3506 Result = DAG.getTargetConstant(Val, Type); 3507 break; 3508 } 3509 } 3510 return; 3511 } 3512 3513 if (Result.getNode()) { 3514 Ops.push_back(Result); 3515 return; 3516 } 3517 3518 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 3519 } 3520 3521 bool MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, 3522 Type *Ty) const { 3523 // No global is ever allowed as a base. 3524 if (AM.BaseGV) 3525 return false; 3526 3527 switch (AM.Scale) { 3528 case 0: // "r+i" or just "i", depending on HasBaseReg. 3529 break; 3530 case 1: 3531 if (!AM.HasBaseReg) // allow "r+i". 3532 break; 3533 return false; // disallow "r+r" or "r+r+i". 3534 default: 3535 return false; 3536 } 3537 3538 return true; 3539 } 3540 3541 bool 3542 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 3543 // The Mips target isn't yet aware of offsets. 3544 return false; 3545 } 3546 3547 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 3548 unsigned SrcAlign, 3549 bool IsMemset, bool ZeroMemset, 3550 bool MemcpyStrSrc, 3551 MachineFunction &MF) const { 3552 if (Subtarget.hasMips64()) 3553 return MVT::i64; 3554 3555 return MVT::i32; 3556 } 3557 3558 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 3559 if (VT != MVT::f32 && VT != MVT::f64) 3560 return false; 3561 if (Imm.isNegZero()) 3562 return false; 3563 return Imm.isZero(); 3564 } 3565 3566 unsigned MipsTargetLowering::getJumpTableEncoding() const { 3567 if (Subtarget.isABI_N64()) 3568 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 3569 3570 return TargetLowering::getJumpTableEncoding(); 3571 } 3572 3573 void MipsTargetLowering::copyByValRegs( 3574 SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains, SelectionDAG &DAG, 3575 const ISD::ArgFlagsTy &Flags, SmallVectorImpl<SDValue> &InVals, 3576 const Argument *FuncArg, unsigned FirstReg, unsigned LastReg, 3577 const CCValAssign &VA, MipsCCState &State) const { 3578 MachineFunction &MF = DAG.getMachineFunction(); 3579 MachineFrameInfo *MFI = MF.getFrameInfo(); 3580 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); 3581 unsigned NumRegs = LastReg - FirstReg; 3582 unsigned RegAreaSize = NumRegs * GPRSizeInBytes; 3583 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 3584 int FrameObjOffset; 3585 const MipsABIInfo &ABI = Subtarget.getABI(); 3586 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); 3587 3588 if (RegAreaSize) 3589 FrameObjOffset = 3590 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 3591 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); 3592 else 3593 FrameObjOffset = VA.getLocMemOffset(); 3594 3595 // Create frame object. 3596 EVT PtrTy = getPointerTy(); 3597 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true); 3598 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 3599 InVals.push_back(FIN); 3600 3601 if (!NumRegs) 3602 return; 3603 3604 // Copy arg registers. 3605 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8); 3606 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3607 3608 for (unsigned I = 0; I < NumRegs; ++I) { 3609 unsigned ArgReg = ByValArgRegs[FirstReg + I]; 3610 unsigned VReg = addLiveIn(MF, ArgReg, RC); 3611 unsigned Offset = I * GPRSizeInBytes; 3612 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 3613 DAG.getConstant(Offset, PtrTy)); 3614 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 3615 StorePtr, MachinePointerInfo(FuncArg, Offset), 3616 false, false, 0); 3617 OutChains.push_back(Store); 3618 } 3619 } 3620 3621 // Copy byVal arg to registers and stack. 3622 void MipsTargetLowering::passByValArg( 3623 SDValue Chain, SDLoc DL, 3624 std::deque<std::pair<unsigned, SDValue>> &RegsToPass, 3625 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 3626 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, 3627 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, 3628 const CCValAssign &VA) const { 3629 unsigned ByValSizeInBytes = Flags.getByValSize(); 3630 unsigned OffsetInBytes = 0; // From beginning of struct 3631 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 3632 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes); 3633 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 3634 unsigned NumRegs = LastReg - FirstReg; 3635 3636 if (NumRegs) { 3637 const ArrayRef<MCPhysReg> ArgRegs = Subtarget.getABI().GetByValArgRegs(); 3638 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); 3639 unsigned I = 0; 3640 3641 // Copy words to registers. 3642 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { 3643 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3644 DAG.getConstant(OffsetInBytes, PtrTy)); 3645 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 3646 MachinePointerInfo(), false, false, false, 3647 Alignment); 3648 MemOpChains.push_back(LoadVal.getValue(1)); 3649 unsigned ArgReg = ArgRegs[FirstReg + I]; 3650 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 3651 } 3652 3653 // Return if the struct has been fully copied. 3654 if (ByValSizeInBytes == OffsetInBytes) 3655 return; 3656 3657 // Copy the remainder of the byval argument with sub-word loads and shifts. 3658 if (LeftoverBytes) { 3659 SDValue Val; 3660 3661 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; 3662 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { 3663 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; 3664 3665 if (RemainingSizeInBytes < LoadSizeInBytes) 3666 continue; 3667 3668 // Load subword. 3669 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3670 DAG.getConstant(OffsetInBytes, PtrTy)); 3671 SDValue LoadVal = DAG.getExtLoad( 3672 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(), 3673 MVT::getIntegerVT(LoadSizeInBytes * 8), false, false, false, 3674 Alignment); 3675 MemOpChains.push_back(LoadVal.getValue(1)); 3676 3677 // Shift the loaded value. 3678 unsigned Shamt; 3679 3680 if (isLittle) 3681 Shamt = TotalBytesLoaded * 8; 3682 else 3683 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; 3684 3685 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 3686 DAG.getConstant(Shamt, MVT::i32)); 3687 3688 if (Val.getNode()) 3689 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 3690 else 3691 Val = Shift; 3692 3693 OffsetInBytes += LoadSizeInBytes; 3694 TotalBytesLoaded += LoadSizeInBytes; 3695 Alignment = std::min(Alignment, LoadSizeInBytes); 3696 } 3697 3698 unsigned ArgReg = ArgRegs[FirstReg + I]; 3699 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 3700 return; 3701 } 3702 } 3703 3704 // Copy remainder of byval arg to it with memcpy. 3705 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; 3706 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3707 DAG.getConstant(OffsetInBytes, PtrTy)); 3708 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 3709 DAG.getIntPtrConstant(VA.getLocMemOffset())); 3710 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, DAG.getConstant(MemCpySize, PtrTy), 3711 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false, 3712 MachinePointerInfo(), MachinePointerInfo()); 3713 MemOpChains.push_back(Chain); 3714 } 3715 3716 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 3717 SDValue Chain, SDLoc DL, 3718 SelectionDAG &DAG, 3719 CCState &State) const { 3720 const ArrayRef<MCPhysReg> ArgRegs = Subtarget.getABI().GetVarArgRegs(); 3721 unsigned Idx = State.getFirstUnallocated(ArgRegs.data(), ArgRegs.size()); 3722 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 3723 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8); 3724 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3725 MachineFunction &MF = DAG.getMachineFunction(); 3726 MachineFrameInfo *MFI = MF.getFrameInfo(); 3727 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3728 3729 // Offset of the first variable argument from stack pointer. 3730 int VaArgOffset; 3731 3732 if (ArgRegs.size() == Idx) 3733 VaArgOffset = 3734 RoundUpToAlignment(State.getNextStackOffset(), RegSizeInBytes); 3735 else { 3736 const MipsABIInfo &ABI = Subtarget.getABI(); 3737 VaArgOffset = 3738 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) - 3739 (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); 3740 } 3741 3742 // Record the frame index of the first variable argument 3743 // which is a value necessary to VASTART. 3744 int FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 3745 MipsFI->setVarArgsFrameIndex(FI); 3746 3747 // Copy the integer registers that have not been used for argument passing 3748 // to the argument register save area. For O32, the save area is allocated 3749 // in the caller's stack frame, while for N32/64, it is allocated in the 3750 // callee's stack frame. 3751 for (unsigned I = Idx; I < ArgRegs.size(); 3752 ++I, VaArgOffset += RegSizeInBytes) { 3753 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 3754 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 3755 FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true); 3756 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy()); 3757 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, 3758 MachinePointerInfo(), false, false, 0); 3759 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue( 3760 (Value *)nullptr); 3761 OutChains.push_back(Store); 3762 } 3763 } 3764 3765 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, 3766 unsigned Align) const { 3767 MachineFunction &MF = State->getMachineFunction(); 3768 const TargetFrameLowering *TFL = MF.getSubtarget().getFrameLowering(); 3769 3770 assert(Size && "Byval argument's size shouldn't be 0."); 3771 3772 Align = std::min(Align, TFL->getStackAlignment()); 3773 3774 unsigned FirstReg = 0; 3775 unsigned NumRegs = 0; 3776 3777 if (State->getCallingConv() != CallingConv::Fast) { 3778 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); 3779 const ArrayRef<MCPhysReg> IntArgRegs = Subtarget.getABI().GetByValArgRegs(); 3780 // FIXME: The O32 case actually describes no shadow registers. 3781 const MCPhysReg *ShadowRegs = 3782 Subtarget.isABI_O32() ? IntArgRegs.data() : Mips64DPRegs; 3783 3784 // We used to check the size as well but we can't do that anymore since 3785 // CCState::HandleByVal() rounds up the size after calling this function. 3786 assert(!(Align % RegSizeInBytes) && 3787 "Byval argument's alignment should be a multiple of" 3788 "RegSizeInBytes."); 3789 3790 FirstReg = State->getFirstUnallocated(IntArgRegs.data(), IntArgRegs.size()); 3791 3792 // If Align > RegSizeInBytes, the first arg register must be even. 3793 // FIXME: This condition happens to do the right thing but it's not the 3794 // right way to test it. We want to check that the stack frame offset 3795 // of the register is aligned. 3796 if ((Align > RegSizeInBytes) && (FirstReg % 2)) { 3797 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]); 3798 ++FirstReg; 3799 } 3800 3801 // Mark the registers allocated. 3802 Size = RoundUpToAlignment(Size, RegSizeInBytes); 3803 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); 3804 Size -= RegSizeInBytes, ++I, ++NumRegs) 3805 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]); 3806 } 3807 3808 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs); 3809 } 3810 3811 MachineBasicBlock * 3812 MipsTargetLowering::emitPseudoSELECT(MachineInstr *MI, MachineBasicBlock *BB, 3813 bool isFPCmp, unsigned Opc) const { 3814 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && 3815 "Subtarget already supports SELECT nodes with the use of" 3816 "conditional-move instructions."); 3817 3818 const TargetInstrInfo *TII = 3819 getTargetMachine().getSubtargetImpl()->getInstrInfo(); 3820 DebugLoc DL = MI->getDebugLoc(); 3821 3822 // To "insert" a SELECT instruction, we actually have to insert the 3823 // diamond control-flow pattern. The incoming instruction knows the 3824 // destination vreg to set, the condition code register to branch on, the 3825 // true/false values to select between, and a branch opcode to use. 3826 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 3827 MachineFunction::iterator It = BB; 3828 ++It; 3829 3830 // thisMBB: 3831 // ... 3832 // TrueVal = ... 3833 // setcc r1, r2, r3 3834 // bNE r1, r0, copy1MBB 3835 // fallthrough --> copy0MBB 3836 MachineBasicBlock *thisMBB = BB; 3837 MachineFunction *F = BB->getParent(); 3838 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 3839 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 3840 F->insert(It, copy0MBB); 3841 F->insert(It, sinkMBB); 3842 3843 // Transfer the remainder of BB and its successor edges to sinkMBB. 3844 sinkMBB->splice(sinkMBB->begin(), BB, 3845 std::next(MachineBasicBlock::iterator(MI)), BB->end()); 3846 sinkMBB->transferSuccessorsAndUpdatePHIs(BB); 3847 3848 // Next, add the true and fallthrough blocks as its successors. 3849 BB->addSuccessor(copy0MBB); 3850 BB->addSuccessor(sinkMBB); 3851 3852 if (isFPCmp) { 3853 // bc1[tf] cc, sinkMBB 3854 BuildMI(BB, DL, TII->get(Opc)) 3855 .addReg(MI->getOperand(1).getReg()) 3856 .addMBB(sinkMBB); 3857 } else { 3858 // bne rs, $0, sinkMBB 3859 BuildMI(BB, DL, TII->get(Opc)) 3860 .addReg(MI->getOperand(1).getReg()) 3861 .addReg(Mips::ZERO) 3862 .addMBB(sinkMBB); 3863 } 3864 3865 // copy0MBB: 3866 // %FalseValue = ... 3867 // # fallthrough to sinkMBB 3868 BB = copy0MBB; 3869 3870 // Update machine-CFG edges 3871 BB->addSuccessor(sinkMBB); 3872 3873 // sinkMBB: 3874 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] 3875 // ... 3876 BB = sinkMBB; 3877 3878 BuildMI(*BB, BB->begin(), DL, 3879 TII->get(Mips::PHI), MI->getOperand(0).getReg()) 3880 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB) 3881 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB); 3882 3883 MI->eraseFromParent(); // The pseudo instruction is gone now. 3884 3885 return BB; 3886 } 3887 3888 // FIXME? Maybe this could be a TableGen attribute on some registers and 3889 // this table could be generated automatically from RegInfo. 3890 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, 3891 EVT VT) const { 3892 // Named registers is expected to be fairly rare. For now, just support $28 3893 // since the linux kernel uses it. 3894 if (Subtarget.isGP64bit()) { 3895 unsigned Reg = StringSwitch<unsigned>(RegName) 3896 .Case("$28", Mips::GP_64) 3897 .Default(0); 3898 if (Reg) 3899 return Reg; 3900 } else { 3901 unsigned Reg = StringSwitch<unsigned>(RegName) 3902 .Case("$28", Mips::GP) 3903 .Default(0); 3904 if (Reg) 3905 return Reg; 3906 } 3907 report_fatal_error("Invalid register name global variable"); 3908 } 3909