MipsISelLowering.cpp revision 2c2c33a167c82db9abd9b6173c1cdfdaa40c2071
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#define DEBUG_TYPE "mips-lower" 15#include "MipsISelLowering.h" 16#include "InstPrinter/MipsInstPrinter.h" 17#include "MCTargetDesc/MipsBaseInfo.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/CodeGen/CallingConvLower.h" 24#include "llvm/CodeGen/MachineFrameInfo.h" 25#include "llvm/CodeGen/MachineFunction.h" 26#include "llvm/CodeGen/MachineInstrBuilder.h" 27#include "llvm/CodeGen/MachineRegisterInfo.h" 28#include "llvm/CodeGen/SelectionDAGISel.h" 29#include "llvm/CodeGen/ValueTypes.h" 30#include "llvm/IR/CallingConv.h" 31#include "llvm/IR/DerivedTypes.h" 32#include "llvm/IR/GlobalVariable.h" 33#include "llvm/IR/Intrinsics.h" 34#include "llvm/Support/CommandLine.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ErrorHandling.h" 37#include "llvm/Support/raw_ostream.h" 38 39using namespace llvm; 40 41STATISTIC(NumTailCalls, "Number of tail calls"); 42 43static cl::opt<bool> 44LargeGOT("mxgot", cl::Hidden, 45 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false)); 46 47static const uint16_t O32IntRegs[4] = { 48 Mips::A0, Mips::A1, Mips::A2, Mips::A3 49}; 50 51static const uint16_t Mips64IntRegs[8] = { 52 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64, 53 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64 54}; 55 56static const uint16_t Mips64DPRegs[8] = { 57 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 58 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 59}; 60 61// If I is a shifted mask, set the size (Size) and the first bit of the 62// mask (Pos), and return true. 63// For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 64static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 65 if (!isShiftedMask_64(I)) 66 return false; 67 68 Size = CountPopulation_64(I); 69 Pos = CountTrailingZeros_64(I); 70 return true; 71} 72 73SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 74 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 75 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 76} 77 78static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) { 79 EVT Ty = Op.getValueType(); 80 81 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op)) 82 return DAG.getTargetGlobalAddress(N->getGlobal(), Op.getDebugLoc(), Ty, 0, 83 Flag); 84 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op)) 85 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 86 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op)) 87 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 88 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op)) 89 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 90 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op)) 91 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 92 N->getOffset(), Flag); 93 94 llvm_unreachable("Unexpected node type."); 95 return SDValue(); 96} 97 98static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) { 99 DebugLoc DL = Op.getDebugLoc(); 100 EVT Ty = Op.getValueType(); 101 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI); 102 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO); 103 return DAG.getNode(ISD::ADD, DL, Ty, 104 DAG.getNode(MipsISD::Hi, DL, Ty, Hi), 105 DAG.getNode(MipsISD::Lo, DL, Ty, Lo)); 106} 107 108SDValue MipsTargetLowering::getAddrLocal(SDValue Op, SelectionDAG &DAG, 109 bool HasMips64) const { 110 DebugLoc DL = Op.getDebugLoc(); 111 EVT Ty = Op.getValueType(); 112 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT; 113 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty), 114 getTargetNode(Op, DAG, GOTFlag)); 115 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT, 116 MachinePointerInfo::getGOT(), false, false, false, 117 0); 118 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO; 119 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag)); 120 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo); 121} 122 123SDValue MipsTargetLowering::getAddrGlobal(SDValue Op, SelectionDAG &DAG, 124 unsigned Flag) const { 125 DebugLoc DL = Op.getDebugLoc(); 126 EVT Ty = Op.getValueType(); 127 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty), 128 getTargetNode(Op, DAG, Flag)); 129 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt, 130 MachinePointerInfo::getGOT(), false, false, false, 0); 131} 132 133SDValue MipsTargetLowering::getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG, 134 unsigned HiFlag, 135 unsigned LoFlag) const { 136 DebugLoc DL = Op.getDebugLoc(); 137 EVT Ty = Op.getValueType(); 138 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag)); 139 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty)); 140 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi, 141 getTargetNode(Op, DAG, LoFlag)); 142 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper, 143 MachinePointerInfo::getGOT(), false, false, false, 0); 144} 145 146const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 147 switch (Opcode) { 148 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 149 case MipsISD::TailCall: return "MipsISD::TailCall"; 150 case MipsISD::Hi: return "MipsISD::Hi"; 151 case MipsISD::Lo: return "MipsISD::Lo"; 152 case MipsISD::GPRel: return "MipsISD::GPRel"; 153 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 154 case MipsISD::Ret: return "MipsISD::Ret"; 155 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 156 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 157 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 158 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 159 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 160 case MipsISD::FPRound: return "MipsISD::FPRound"; 161 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI"; 162 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI"; 163 case MipsISD::Mult: return "MipsISD::Mult"; 164 case MipsISD::Multu: return "MipsISD::Multu"; 165 case MipsISD::MAdd: return "MipsISD::MAdd"; 166 case MipsISD::MAddu: return "MipsISD::MAddu"; 167 case MipsISD::MSub: return "MipsISD::MSub"; 168 case MipsISD::MSubu: return "MipsISD::MSubu"; 169 case MipsISD::DivRem: return "MipsISD::DivRem"; 170 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 171 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 172 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 173 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 174 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 175 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 176 case MipsISD::Sync: return "MipsISD::Sync"; 177 case MipsISD::Ext: return "MipsISD::Ext"; 178 case MipsISD::Ins: return "MipsISD::Ins"; 179 case MipsISD::LWL: return "MipsISD::LWL"; 180 case MipsISD::LWR: return "MipsISD::LWR"; 181 case MipsISD::SWL: return "MipsISD::SWL"; 182 case MipsISD::SWR: return "MipsISD::SWR"; 183 case MipsISD::LDL: return "MipsISD::LDL"; 184 case MipsISD::LDR: return "MipsISD::LDR"; 185 case MipsISD::SDL: return "MipsISD::SDL"; 186 case MipsISD::SDR: return "MipsISD::SDR"; 187 case MipsISD::EXTP: return "MipsISD::EXTP"; 188 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 189 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 190 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 191 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 192 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 193 case MipsISD::SHILO: return "MipsISD::SHILO"; 194 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 195 case MipsISD::MULT: return "MipsISD::MULT"; 196 case MipsISD::MULTU: return "MipsISD::MULTU"; 197 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 198 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 199 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 200 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 201 default: return NULL; 202 } 203} 204 205MipsTargetLowering:: 206MipsTargetLowering(MipsTargetMachine &TM) 207 : TargetLowering(TM, new MipsTargetObjectFile()), 208 Subtarget(&TM.getSubtarget<MipsSubtarget>()), 209 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()), 210 IsO32(Subtarget->isABI_O32()) { 211 // Mips does not have i1 type, so use i32 for 212 // setcc operations results (slt, sgt, ...). 213 setBooleanContents(ZeroOrOneBooleanContent); 214 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct? 215 216 // Load extented operations for i1 types must be promoted 217 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 218 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 219 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 220 221 // MIPS doesn't have extending float->double load/store 222 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 223 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 224 225 // Used by legalize types to correctly generate the setcc result. 226 // Without this, every float setcc comes with a AND/OR with the result, 227 // we don't want this, since the fpcmp result goes to a flag register, 228 // which is used implicitly by brcond and select operations. 229 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 230 231 // Mips Custom Operations 232 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 233 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 234 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 235 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 236 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 237 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 238 setOperationAction(ISD::SELECT, MVT::f32, Custom); 239 setOperationAction(ISD::SELECT, MVT::f64, Custom); 240 setOperationAction(ISD::SELECT, MVT::i32, Custom); 241 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 242 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 243 setOperationAction(ISD::SETCC, MVT::f32, Custom); 244 setOperationAction(ISD::SETCC, MVT::f64, Custom); 245 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 246 setOperationAction(ISD::VASTART, MVT::Other, Custom); 247 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 248 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 249 250 if (!TM.Options.NoNaNsFPMath) { 251 setOperationAction(ISD::FABS, MVT::f32, Custom); 252 setOperationAction(ISD::FABS, MVT::f64, Custom); 253 } 254 255 if (HasMips64) { 256 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 257 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 258 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 259 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 260 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 261 setOperationAction(ISD::SELECT, MVT::i64, Custom); 262 setOperationAction(ISD::LOAD, MVT::i64, Custom); 263 setOperationAction(ISD::STORE, MVT::i64, Custom); 264 } 265 266 if (!HasMips64) { 267 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 268 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 269 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 270 } 271 272 setOperationAction(ISD::ADD, MVT::i32, Custom); 273 if (HasMips64) 274 setOperationAction(ISD::ADD, MVT::i64, Custom); 275 276 setOperationAction(ISD::SDIV, MVT::i32, Expand); 277 setOperationAction(ISD::SREM, MVT::i32, Expand); 278 setOperationAction(ISD::UDIV, MVT::i32, Expand); 279 setOperationAction(ISD::UREM, MVT::i32, Expand); 280 setOperationAction(ISD::SDIV, MVT::i64, Expand); 281 setOperationAction(ISD::SREM, MVT::i64, Expand); 282 setOperationAction(ISD::UDIV, MVT::i64, Expand); 283 setOperationAction(ISD::UREM, MVT::i64, Expand); 284 285 // Operations not directly supported by Mips. 286 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 287 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 288 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 289 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 290 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); 291 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 292 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 293 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 294 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 295 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 296 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 297 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 298 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 299 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 300 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 301 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); 302 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 303 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand); 304 setOperationAction(ISD::ROTL, MVT::i32, Expand); 305 setOperationAction(ISD::ROTL, MVT::i64, Expand); 306 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 307 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 308 309 if (!Subtarget->hasMips32r2()) 310 setOperationAction(ISD::ROTR, MVT::i32, Expand); 311 312 if (!Subtarget->hasMips64r2()) 313 setOperationAction(ISD::ROTR, MVT::i64, Expand); 314 315 setOperationAction(ISD::FSIN, MVT::f32, Expand); 316 setOperationAction(ISD::FSIN, MVT::f64, Expand); 317 setOperationAction(ISD::FCOS, MVT::f32, Expand); 318 setOperationAction(ISD::FCOS, MVT::f64, Expand); 319 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 320 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 321 setOperationAction(ISD::FPOWI, MVT::f32, Expand); 322 setOperationAction(ISD::FPOW, MVT::f32, Expand); 323 setOperationAction(ISD::FPOW, MVT::f64, Expand); 324 setOperationAction(ISD::FLOG, MVT::f32, Expand); 325 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 326 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 327 setOperationAction(ISD::FEXP, MVT::f32, Expand); 328 setOperationAction(ISD::FMA, MVT::f32, Expand); 329 setOperationAction(ISD::FMA, MVT::f64, Expand); 330 setOperationAction(ISD::FREM, MVT::f32, Expand); 331 setOperationAction(ISD::FREM, MVT::f64, Expand); 332 333 if (!TM.Options.NoNaNsFPMath) { 334 setOperationAction(ISD::FNEG, MVT::f32, Expand); 335 setOperationAction(ISD::FNEG, MVT::f64, Expand); 336 } 337 338 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); 339 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand); 340 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); 341 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand); 342 343 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 344 345 setOperationAction(ISD::VAARG, MVT::Other, Expand); 346 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 347 setOperationAction(ISD::VAEND, MVT::Other, Expand); 348 349 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom); 350 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); 351 352 // Use the default for now 353 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 354 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 355 356 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand); 357 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 358 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand); 359 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 360 361 setInsertFencesForAtomic(true); 362 363 if (!Subtarget->hasSEInReg()) { 364 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 365 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 366 } 367 368 if (!Subtarget->hasBitCount()) { 369 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 370 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 371 } 372 373 if (!Subtarget->hasSwap()) { 374 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 375 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 376 } 377 378 if (HasMips64) { 379 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom); 380 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom); 381 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom); 382 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 383 } 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(HasMips64 ? 3 : 2); 393 394 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP); 395 396 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0); 397 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1); 398 399 MaxStoresPerMemcpy = 16; 400} 401 402const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) { 403 if (TM.getSubtargetImpl()->inMips16Mode()) 404 return llvm::createMips16TargetLowering(TM); 405 406 return llvm::createMipsSETargetLowering(TM); 407} 408 409EVT MipsTargetLowering::getSetCCResultType(EVT VT) const { 410 if (!VT.isVector()) 411 return MVT::i32; 412 return VT.changeVectorElementTypeToInteger(); 413} 414 415static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 416 TargetLowering::DAGCombinerInfo &DCI, 417 const MipsSubtarget *Subtarget) { 418 if (DCI.isBeforeLegalizeOps()) 419 return SDValue(); 420 421 EVT Ty = N->getValueType(0); 422 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64; 423 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64; 424 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 425 MipsISD::DivRemU16; 426 DebugLoc DL = N->getDebugLoc(); 427 428 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 429 N->getOperand(0), N->getOperand(1)); 430 SDValue InChain = DAG.getEntryNode(); 431 SDValue InGlue = DivRem; 432 433 // insert MFLO 434 if (N->hasAnyUseOfValue(0)) { 435 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 436 InGlue); 437 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 438 InChain = CopyFromLo.getValue(1); 439 InGlue = CopyFromLo.getValue(2); 440 } 441 442 // insert MFHI 443 if (N->hasAnyUseOfValue(1)) { 444 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 445 HI, Ty, InGlue); 446 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 447 } 448 449 return SDValue(); 450} 451 452static Mips::CondCode FPCondCCodeToFCC(ISD::CondCode CC) { 453 switch (CC) { 454 default: llvm_unreachable("Unknown fp condition code!"); 455 case ISD::SETEQ: 456 case ISD::SETOEQ: return Mips::FCOND_OEQ; 457 case ISD::SETUNE: return Mips::FCOND_UNE; 458 case ISD::SETLT: 459 case ISD::SETOLT: return Mips::FCOND_OLT; 460 case ISD::SETGT: 461 case ISD::SETOGT: return Mips::FCOND_OGT; 462 case ISD::SETLE: 463 case ISD::SETOLE: return Mips::FCOND_OLE; 464 case ISD::SETGE: 465 case ISD::SETOGE: return Mips::FCOND_OGE; 466 case ISD::SETULT: return Mips::FCOND_ULT; 467 case ISD::SETULE: return Mips::FCOND_ULE; 468 case ISD::SETUGT: return Mips::FCOND_UGT; 469 case ISD::SETUGE: return Mips::FCOND_UGE; 470 case ISD::SETUO: return Mips::FCOND_UN; 471 case ISD::SETO: return Mips::FCOND_OR; 472 case ISD::SETNE: 473 case ISD::SETONE: return Mips::FCOND_ONE; 474 case ISD::SETUEQ: return Mips::FCOND_UEQ; 475 } 476} 477 478 479/// This function returns true if the floating point conditional branches and 480/// conditional moves which use condition code CC should be inverted. 481static bool invertFPCondCodeUser(Mips::CondCode CC) { 482 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 483 return false; 484 485 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 486 "Illegal Condition Code"); 487 488 return true; 489} 490 491// Creates and returns an FPCmp node from a setcc node. 492// Returns Op if setcc is not a floating point comparison. 493static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 494 // must be a SETCC node 495 if (Op.getOpcode() != ISD::SETCC) 496 return Op; 497 498 SDValue LHS = Op.getOperand(0); 499 500 if (!LHS.getValueType().isFloatingPoint()) 501 return Op; 502 503 SDValue RHS = Op.getOperand(1); 504 DebugLoc DL = Op.getDebugLoc(); 505 506 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 507 // node if necessary. 508 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 509 510 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 511 DAG.getConstant(FPCondCCodeToFCC(CC), MVT::i32)); 512} 513 514// Creates and returns a CMovFPT/F node. 515static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 516 SDValue False, DebugLoc DL) { 517 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 518 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 519 520 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 521 True.getValueType(), True, False, Cond); 522} 523 524static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 525 TargetLowering::DAGCombinerInfo &DCI, 526 const MipsSubtarget *Subtarget) { 527 if (DCI.isBeforeLegalizeOps()) 528 return SDValue(); 529 530 SDValue SetCC = N->getOperand(0); 531 532 if ((SetCC.getOpcode() != ISD::SETCC) || 533 !SetCC.getOperand(0).getValueType().isInteger()) 534 return SDValue(); 535 536 SDValue False = N->getOperand(2); 537 EVT FalseTy = False.getValueType(); 538 539 if (!FalseTy.isInteger()) 540 return SDValue(); 541 542 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False); 543 544 if (!CN || CN->getZExtValue()) 545 return SDValue(); 546 547 const DebugLoc DL = N->getDebugLoc(); 548 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 549 SDValue True = N->getOperand(1); 550 551 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 552 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 553 554 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 555} 556 557static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 558 TargetLowering::DAGCombinerInfo &DCI, 559 const MipsSubtarget *Subtarget) { 560 // Pattern match EXT. 561 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 562 // => ext $dst, $src, size, pos 563 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 564 return SDValue(); 565 566 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1); 567 unsigned ShiftRightOpc = ShiftRight.getOpcode(); 568 569 // Op's first operand must be a shift right. 570 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL) 571 return SDValue(); 572 573 // The second operand of the shift must be an immediate. 574 ConstantSDNode *CN; 575 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1)))) 576 return SDValue(); 577 578 uint64_t Pos = CN->getZExtValue(); 579 uint64_t SMPos, SMSize; 580 581 // Op's second operand must be a shifted mask. 582 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 583 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 584 return SDValue(); 585 586 // Return if the shifted mask does not start at bit 0 or the sum of its size 587 // and Pos exceeds the word's size. 588 EVT ValTy = N->getValueType(0); 589 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 590 return SDValue(); 591 592 return DAG.getNode(MipsISD::Ext, N->getDebugLoc(), ValTy, 593 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32), 594 DAG.getConstant(SMSize, MVT::i32)); 595} 596 597static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 598 TargetLowering::DAGCombinerInfo &DCI, 599 const MipsSubtarget *Subtarget) { 600 // Pattern match INS. 601 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 602 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 603 // => ins $dst, $src, size, pos, $src1 604 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 605 return SDValue(); 606 607 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 608 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 609 ConstantSDNode *CN; 610 611 // See if Op's first operand matches (and $src1 , mask0). 612 if (And0.getOpcode() != ISD::AND) 613 return SDValue(); 614 615 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 616 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 617 return SDValue(); 618 619 // See if Op's second operand matches (and (shl $src, pos), mask1). 620 if (And1.getOpcode() != ISD::AND) 621 return SDValue(); 622 623 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 624 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 625 return SDValue(); 626 627 // The shift masks must have the same position and size. 628 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 629 return SDValue(); 630 631 SDValue Shl = And1.getOperand(0); 632 if (Shl.getOpcode() != ISD::SHL) 633 return SDValue(); 634 635 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 636 return SDValue(); 637 638 unsigned Shamt = CN->getZExtValue(); 639 640 // Return if the shift amount and the first bit position of mask are not the 641 // same. 642 EVT ValTy = N->getValueType(0); 643 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 644 return SDValue(); 645 646 return DAG.getNode(MipsISD::Ins, N->getDebugLoc(), ValTy, Shl.getOperand(0), 647 DAG.getConstant(SMPos0, MVT::i32), 648 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0)); 649} 650 651static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 652 TargetLowering::DAGCombinerInfo &DCI, 653 const MipsSubtarget *Subtarget) { 654 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 655 656 if (DCI.isBeforeLegalizeOps()) 657 return SDValue(); 658 659 SDValue Add = N->getOperand(1); 660 661 if (Add.getOpcode() != ISD::ADD) 662 return SDValue(); 663 664 SDValue Lo = Add.getOperand(1); 665 666 if ((Lo.getOpcode() != MipsISD::Lo) || 667 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 668 return SDValue(); 669 670 EVT ValTy = N->getValueType(0); 671 DebugLoc DL = N->getDebugLoc(); 672 673 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 674 Add.getOperand(0)); 675 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 676} 677 678SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 679 const { 680 SelectionDAG &DAG = DCI.DAG; 681 unsigned Opc = N->getOpcode(); 682 683 switch (Opc) { 684 default: break; 685 case ISD::SDIVREM: 686 case ISD::UDIVREM: 687 return performDivRemCombine(N, DAG, DCI, Subtarget); 688 case ISD::SELECT: 689 return performSELECTCombine(N, DAG, DCI, Subtarget); 690 case ISD::AND: 691 return performANDCombine(N, DAG, DCI, Subtarget); 692 case ISD::OR: 693 return performORCombine(N, DAG, DCI, Subtarget); 694 case ISD::ADD: 695 return performADDCombine(N, DAG, DCI, Subtarget); 696 } 697 698 return SDValue(); 699} 700 701void 702MipsTargetLowering::LowerOperationWrapper(SDNode *N, 703 SmallVectorImpl<SDValue> &Results, 704 SelectionDAG &DAG) const { 705 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 706 707 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 708 Results.push_back(Res.getValue(I)); 709} 710 711void 712MipsTargetLowering::ReplaceNodeResults(SDNode *N, 713 SmallVectorImpl<SDValue> &Results, 714 SelectionDAG &DAG) const { 715 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 716 717 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 718 Results.push_back(Res.getValue(I)); 719} 720 721SDValue MipsTargetLowering:: 722LowerOperation(SDValue Op, SelectionDAG &DAG) const 723{ 724 switch (Op.getOpcode()) 725 { 726 case ISD::BR_JT: return lowerBR_JT(Op, DAG); 727 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 728 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 729 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 730 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 731 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 732 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 733 case ISD::SELECT: return lowerSELECT(Op, DAG); 734 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG); 735 case ISD::SETCC: return lowerSETCC(Op, DAG); 736 case ISD::VASTART: return lowerVASTART(Op, DAG); 737 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 738 case ISD::FABS: return lowerFABS(Op, DAG); 739 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 740 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 741 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 742 case ISD::MEMBARRIER: return lowerMEMBARRIER(Op, DAG); 743 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 744 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 745 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 746 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 747 case ISD::LOAD: return lowerLOAD(Op, DAG); 748 case ISD::STORE: return lowerSTORE(Op, DAG); 749 case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG); 750 case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG); 751 case ISD::ADD: return lowerADD(Op, DAG); 752 } 753 return SDValue(); 754} 755 756//===----------------------------------------------------------------------===// 757// Lower helper functions 758//===----------------------------------------------------------------------===// 759 760// addLiveIn - This helper function adds the specified physical register to the 761// MachineFunction as a live in value. It also creates a corresponding 762// virtual register for it. 763static unsigned 764addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 765{ 766 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 767 MF.getRegInfo().addLiveIn(PReg, VReg); 768 return VReg; 769} 770 771MachineBasicBlock * 772MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 773 MachineBasicBlock *BB) const { 774 switch (MI->getOpcode()) { 775 default: 776 llvm_unreachable("Unexpected instr type to insert"); 777 case Mips::ATOMIC_LOAD_ADD_I8: 778 case Mips::ATOMIC_LOAD_ADD_I8_P8: 779 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 780 case Mips::ATOMIC_LOAD_ADD_I16: 781 case Mips::ATOMIC_LOAD_ADD_I16_P8: 782 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 783 case Mips::ATOMIC_LOAD_ADD_I32: 784 case Mips::ATOMIC_LOAD_ADD_I32_P8: 785 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 786 case Mips::ATOMIC_LOAD_ADD_I64: 787 case Mips::ATOMIC_LOAD_ADD_I64_P8: 788 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 789 790 case Mips::ATOMIC_LOAD_AND_I8: 791 case Mips::ATOMIC_LOAD_AND_I8_P8: 792 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 793 case Mips::ATOMIC_LOAD_AND_I16: 794 case Mips::ATOMIC_LOAD_AND_I16_P8: 795 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 796 case Mips::ATOMIC_LOAD_AND_I32: 797 case Mips::ATOMIC_LOAD_AND_I32_P8: 798 return emitAtomicBinary(MI, BB, 4, Mips::AND); 799 case Mips::ATOMIC_LOAD_AND_I64: 800 case Mips::ATOMIC_LOAD_AND_I64_P8: 801 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 802 803 case Mips::ATOMIC_LOAD_OR_I8: 804 case Mips::ATOMIC_LOAD_OR_I8_P8: 805 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 806 case Mips::ATOMIC_LOAD_OR_I16: 807 case Mips::ATOMIC_LOAD_OR_I16_P8: 808 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 809 case Mips::ATOMIC_LOAD_OR_I32: 810 case Mips::ATOMIC_LOAD_OR_I32_P8: 811 return emitAtomicBinary(MI, BB, 4, Mips::OR); 812 case Mips::ATOMIC_LOAD_OR_I64: 813 case Mips::ATOMIC_LOAD_OR_I64_P8: 814 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 815 816 case Mips::ATOMIC_LOAD_XOR_I8: 817 case Mips::ATOMIC_LOAD_XOR_I8_P8: 818 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 819 case Mips::ATOMIC_LOAD_XOR_I16: 820 case Mips::ATOMIC_LOAD_XOR_I16_P8: 821 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 822 case Mips::ATOMIC_LOAD_XOR_I32: 823 case Mips::ATOMIC_LOAD_XOR_I32_P8: 824 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 825 case Mips::ATOMIC_LOAD_XOR_I64: 826 case Mips::ATOMIC_LOAD_XOR_I64_P8: 827 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 828 829 case Mips::ATOMIC_LOAD_NAND_I8: 830 case Mips::ATOMIC_LOAD_NAND_I8_P8: 831 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 832 case Mips::ATOMIC_LOAD_NAND_I16: 833 case Mips::ATOMIC_LOAD_NAND_I16_P8: 834 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 835 case Mips::ATOMIC_LOAD_NAND_I32: 836 case Mips::ATOMIC_LOAD_NAND_I32_P8: 837 return emitAtomicBinary(MI, BB, 4, 0, true); 838 case Mips::ATOMIC_LOAD_NAND_I64: 839 case Mips::ATOMIC_LOAD_NAND_I64_P8: 840 return emitAtomicBinary(MI, BB, 8, 0, true); 841 842 case Mips::ATOMIC_LOAD_SUB_I8: 843 case Mips::ATOMIC_LOAD_SUB_I8_P8: 844 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 845 case Mips::ATOMIC_LOAD_SUB_I16: 846 case Mips::ATOMIC_LOAD_SUB_I16_P8: 847 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 848 case Mips::ATOMIC_LOAD_SUB_I32: 849 case Mips::ATOMIC_LOAD_SUB_I32_P8: 850 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 851 case Mips::ATOMIC_LOAD_SUB_I64: 852 case Mips::ATOMIC_LOAD_SUB_I64_P8: 853 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 854 855 case Mips::ATOMIC_SWAP_I8: 856 case Mips::ATOMIC_SWAP_I8_P8: 857 return emitAtomicBinaryPartword(MI, BB, 1, 0); 858 case Mips::ATOMIC_SWAP_I16: 859 case Mips::ATOMIC_SWAP_I16_P8: 860 return emitAtomicBinaryPartword(MI, BB, 2, 0); 861 case Mips::ATOMIC_SWAP_I32: 862 case Mips::ATOMIC_SWAP_I32_P8: 863 return emitAtomicBinary(MI, BB, 4, 0); 864 case Mips::ATOMIC_SWAP_I64: 865 case Mips::ATOMIC_SWAP_I64_P8: 866 return emitAtomicBinary(MI, BB, 8, 0); 867 868 case Mips::ATOMIC_CMP_SWAP_I8: 869 case Mips::ATOMIC_CMP_SWAP_I8_P8: 870 return emitAtomicCmpSwapPartword(MI, BB, 1); 871 case Mips::ATOMIC_CMP_SWAP_I16: 872 case Mips::ATOMIC_CMP_SWAP_I16_P8: 873 return emitAtomicCmpSwapPartword(MI, BB, 2); 874 case Mips::ATOMIC_CMP_SWAP_I32: 875 case Mips::ATOMIC_CMP_SWAP_I32_P8: 876 return emitAtomicCmpSwap(MI, BB, 4); 877 case Mips::ATOMIC_CMP_SWAP_I64: 878 case Mips::ATOMIC_CMP_SWAP_I64_P8: 879 return emitAtomicCmpSwap(MI, BB, 8); 880 } 881} 882 883// This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 884// Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 885MachineBasicBlock * 886MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 887 unsigned Size, unsigned BinOpcode, 888 bool Nand) const { 889 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 890 891 MachineFunction *MF = BB->getParent(); 892 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 893 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 894 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 895 DebugLoc DL = MI->getDebugLoc(); 896 unsigned LL, SC, AND, NOR, ZERO, BEQ; 897 898 if (Size == 4) { 899 LL = IsN64 ? Mips::LL_P8 : Mips::LL; 900 SC = IsN64 ? Mips::SC_P8 : Mips::SC; 901 AND = Mips::AND; 902 NOR = Mips::NOR; 903 ZERO = Mips::ZERO; 904 BEQ = Mips::BEQ; 905 } 906 else { 907 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD; 908 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD; 909 AND = Mips::AND64; 910 NOR = Mips::NOR64; 911 ZERO = Mips::ZERO_64; 912 BEQ = Mips::BEQ64; 913 } 914 915 unsigned OldVal = MI->getOperand(0).getReg(); 916 unsigned Ptr = MI->getOperand(1).getReg(); 917 unsigned Incr = MI->getOperand(2).getReg(); 918 919 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 920 unsigned AndRes = RegInfo.createVirtualRegister(RC); 921 unsigned Success = RegInfo.createVirtualRegister(RC); 922 923 // insert new blocks after the current block 924 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 925 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 926 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 927 MachineFunction::iterator It = BB; 928 ++It; 929 MF->insert(It, loopMBB); 930 MF->insert(It, exitMBB); 931 932 // Transfer the remainder of BB and its successor edges to exitMBB. 933 exitMBB->splice(exitMBB->begin(), BB, 934 llvm::next(MachineBasicBlock::iterator(MI)), 935 BB->end()); 936 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 937 938 // thisMBB: 939 // ... 940 // fallthrough --> loopMBB 941 BB->addSuccessor(loopMBB); 942 loopMBB->addSuccessor(loopMBB); 943 loopMBB->addSuccessor(exitMBB); 944 945 // loopMBB: 946 // ll oldval, 0(ptr) 947 // <binop> storeval, oldval, incr 948 // sc success, storeval, 0(ptr) 949 // beq success, $0, loopMBB 950 BB = loopMBB; 951 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 952 if (Nand) { 953 // and andres, oldval, incr 954 // nor storeval, $0, andres 955 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 956 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 957 } else if (BinOpcode) { 958 // <binop> storeval, oldval, incr 959 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 960 } else { 961 StoreVal = Incr; 962 } 963 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 964 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 965 966 MI->eraseFromParent(); // The instruction is gone now. 967 968 return exitMBB; 969} 970 971MachineBasicBlock * 972MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI, 973 MachineBasicBlock *BB, 974 unsigned Size, unsigned BinOpcode, 975 bool Nand) const { 976 assert((Size == 1 || Size == 2) && 977 "Unsupported size for EmitAtomicBinaryPartial."); 978 979 MachineFunction *MF = BB->getParent(); 980 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 981 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 982 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 983 DebugLoc DL = MI->getDebugLoc(); 984 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL; 985 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC; 986 987 unsigned Dest = MI->getOperand(0).getReg(); 988 unsigned Ptr = MI->getOperand(1).getReg(); 989 unsigned Incr = MI->getOperand(2).getReg(); 990 991 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 992 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 993 unsigned Mask = RegInfo.createVirtualRegister(RC); 994 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 995 unsigned NewVal = RegInfo.createVirtualRegister(RC); 996 unsigned OldVal = RegInfo.createVirtualRegister(RC); 997 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 998 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 999 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1000 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1001 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1002 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1003 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1004 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1005 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1006 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1007 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1008 unsigned Success = RegInfo.createVirtualRegister(RC); 1009 1010 // insert new blocks after the current block 1011 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1012 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1013 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1014 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1015 MachineFunction::iterator It = BB; 1016 ++It; 1017 MF->insert(It, loopMBB); 1018 MF->insert(It, sinkMBB); 1019 MF->insert(It, exitMBB); 1020 1021 // Transfer the remainder of BB and its successor edges to exitMBB. 1022 exitMBB->splice(exitMBB->begin(), BB, 1023 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1024 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1025 1026 BB->addSuccessor(loopMBB); 1027 loopMBB->addSuccessor(loopMBB); 1028 loopMBB->addSuccessor(sinkMBB); 1029 sinkMBB->addSuccessor(exitMBB); 1030 1031 // thisMBB: 1032 // addiu masklsb2,$0,-4 # 0xfffffffc 1033 // and alignedaddr,ptr,masklsb2 1034 // andi ptrlsb2,ptr,3 1035 // sll shiftamt,ptrlsb2,3 1036 // ori maskupper,$0,255 # 0xff 1037 // sll mask,maskupper,shiftamt 1038 // nor mask2,$0,mask 1039 // sll incr2,incr,shiftamt 1040 1041 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1042 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1043 .addReg(Mips::ZERO).addImm(-4); 1044 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1045 .addReg(Ptr).addReg(MaskLSB2); 1046 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1047 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1048 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1049 .addReg(Mips::ZERO).addImm(MaskImm); 1050 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1051 .addReg(ShiftAmt).addReg(MaskUpper); 1052 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1053 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(ShiftAmt).addReg(Incr); 1054 1055 // atomic.load.binop 1056 // loopMBB: 1057 // ll oldval,0(alignedaddr) 1058 // binop binopres,oldval,incr2 1059 // and newval,binopres,mask 1060 // and maskedoldval0,oldval,mask2 1061 // or storeval,maskedoldval0,newval 1062 // sc success,storeval,0(alignedaddr) 1063 // beq success,$0,loopMBB 1064 1065 // atomic.swap 1066 // loopMBB: 1067 // ll oldval,0(alignedaddr) 1068 // and newval,incr2,mask 1069 // and maskedoldval0,oldval,mask2 1070 // or storeval,maskedoldval0,newval 1071 // sc success,storeval,0(alignedaddr) 1072 // beq success,$0,loopMBB 1073 1074 BB = loopMBB; 1075 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1076 if (Nand) { 1077 // and andres, oldval, incr2 1078 // nor binopres, $0, andres 1079 // and newval, binopres, mask 1080 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1081 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1082 .addReg(Mips::ZERO).addReg(AndRes); 1083 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1084 } else if (BinOpcode) { 1085 // <binop> binopres, oldval, incr2 1086 // and newval, binopres, mask 1087 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1088 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1089 } else {// atomic.swap 1090 // and newval, incr2, mask 1091 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1092 } 1093 1094 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1095 .addReg(OldVal).addReg(Mask2); 1096 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1097 .addReg(MaskedOldVal0).addReg(NewVal); 1098 BuildMI(BB, DL, TII->get(SC), Success) 1099 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1100 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1101 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1102 1103 // sinkMBB: 1104 // and maskedoldval1,oldval,mask 1105 // srl srlres,maskedoldval1,shiftamt 1106 // sll sllres,srlres,24 1107 // sra dest,sllres,24 1108 BB = sinkMBB; 1109 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1110 1111 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1112 .addReg(OldVal).addReg(Mask); 1113 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1114 .addReg(ShiftAmt).addReg(MaskedOldVal1); 1115 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1116 .addReg(SrlRes).addImm(ShiftImm); 1117 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1118 .addReg(SllRes).addImm(ShiftImm); 1119 1120 MI->eraseFromParent(); // The instruction is gone now. 1121 1122 return exitMBB; 1123} 1124 1125MachineBasicBlock * 1126MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI, 1127 MachineBasicBlock *BB, 1128 unsigned Size) const { 1129 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1130 1131 MachineFunction *MF = BB->getParent(); 1132 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1133 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1134 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1135 DebugLoc DL = MI->getDebugLoc(); 1136 unsigned LL, SC, ZERO, BNE, BEQ; 1137 1138 if (Size == 4) { 1139 LL = IsN64 ? Mips::LL_P8 : Mips::LL; 1140 SC = IsN64 ? Mips::SC_P8 : Mips::SC; 1141 ZERO = Mips::ZERO; 1142 BNE = Mips::BNE; 1143 BEQ = Mips::BEQ; 1144 } 1145 else { 1146 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD; 1147 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD; 1148 ZERO = Mips::ZERO_64; 1149 BNE = Mips::BNE64; 1150 BEQ = Mips::BEQ64; 1151 } 1152 1153 unsigned Dest = MI->getOperand(0).getReg(); 1154 unsigned Ptr = MI->getOperand(1).getReg(); 1155 unsigned OldVal = MI->getOperand(2).getReg(); 1156 unsigned NewVal = MI->getOperand(3).getReg(); 1157 1158 unsigned Success = RegInfo.createVirtualRegister(RC); 1159 1160 // insert new blocks after the current block 1161 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1162 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1163 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1164 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1165 MachineFunction::iterator It = BB; 1166 ++It; 1167 MF->insert(It, loop1MBB); 1168 MF->insert(It, loop2MBB); 1169 MF->insert(It, exitMBB); 1170 1171 // Transfer the remainder of BB and its successor edges to exitMBB. 1172 exitMBB->splice(exitMBB->begin(), BB, 1173 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1174 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1175 1176 // thisMBB: 1177 // ... 1178 // fallthrough --> loop1MBB 1179 BB->addSuccessor(loop1MBB); 1180 loop1MBB->addSuccessor(exitMBB); 1181 loop1MBB->addSuccessor(loop2MBB); 1182 loop2MBB->addSuccessor(loop1MBB); 1183 loop2MBB->addSuccessor(exitMBB); 1184 1185 // loop1MBB: 1186 // ll dest, 0(ptr) 1187 // bne dest, oldval, exitMBB 1188 BB = loop1MBB; 1189 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1190 BuildMI(BB, DL, TII->get(BNE)) 1191 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1192 1193 // loop2MBB: 1194 // sc success, newval, 0(ptr) 1195 // beq success, $0, loop1MBB 1196 BB = loop2MBB; 1197 BuildMI(BB, DL, TII->get(SC), Success) 1198 .addReg(NewVal).addReg(Ptr).addImm(0); 1199 BuildMI(BB, DL, TII->get(BEQ)) 1200 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1201 1202 MI->eraseFromParent(); // The instruction is gone now. 1203 1204 return exitMBB; 1205} 1206 1207MachineBasicBlock * 1208MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI, 1209 MachineBasicBlock *BB, 1210 unsigned Size) const { 1211 assert((Size == 1 || Size == 2) && 1212 "Unsupported size for EmitAtomicCmpSwapPartial."); 1213 1214 MachineFunction *MF = BB->getParent(); 1215 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1216 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1217 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1218 DebugLoc DL = MI->getDebugLoc(); 1219 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL; 1220 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC; 1221 1222 unsigned Dest = MI->getOperand(0).getReg(); 1223 unsigned Ptr = MI->getOperand(1).getReg(); 1224 unsigned CmpVal = MI->getOperand(2).getReg(); 1225 unsigned NewVal = MI->getOperand(3).getReg(); 1226 1227 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1228 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1229 unsigned Mask = RegInfo.createVirtualRegister(RC); 1230 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1231 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1232 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1233 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1234 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1235 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1236 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1237 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1238 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1239 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1240 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1241 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1242 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1243 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1244 unsigned Success = RegInfo.createVirtualRegister(RC); 1245 1246 // insert new blocks after the current block 1247 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1248 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1249 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1250 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1251 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1252 MachineFunction::iterator It = BB; 1253 ++It; 1254 MF->insert(It, loop1MBB); 1255 MF->insert(It, loop2MBB); 1256 MF->insert(It, sinkMBB); 1257 MF->insert(It, exitMBB); 1258 1259 // Transfer the remainder of BB and its successor edges to exitMBB. 1260 exitMBB->splice(exitMBB->begin(), BB, 1261 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1262 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1263 1264 BB->addSuccessor(loop1MBB); 1265 loop1MBB->addSuccessor(sinkMBB); 1266 loop1MBB->addSuccessor(loop2MBB); 1267 loop2MBB->addSuccessor(loop1MBB); 1268 loop2MBB->addSuccessor(sinkMBB); 1269 sinkMBB->addSuccessor(exitMBB); 1270 1271 // FIXME: computation of newval2 can be moved to loop2MBB. 1272 // thisMBB: 1273 // addiu masklsb2,$0,-4 # 0xfffffffc 1274 // and alignedaddr,ptr,masklsb2 1275 // andi ptrlsb2,ptr,3 1276 // sll shiftamt,ptrlsb2,3 1277 // ori maskupper,$0,255 # 0xff 1278 // sll mask,maskupper,shiftamt 1279 // nor mask2,$0,mask 1280 // andi maskedcmpval,cmpval,255 1281 // sll shiftedcmpval,maskedcmpval,shiftamt 1282 // andi maskednewval,newval,255 1283 // sll shiftednewval,maskednewval,shiftamt 1284 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1285 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1286 .addReg(Mips::ZERO).addImm(-4); 1287 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1288 .addReg(Ptr).addReg(MaskLSB2); 1289 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1290 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1291 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1292 .addReg(Mips::ZERO).addImm(MaskImm); 1293 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1294 .addReg(ShiftAmt).addReg(MaskUpper); 1295 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1296 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1297 .addReg(CmpVal).addImm(MaskImm); 1298 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1299 .addReg(ShiftAmt).addReg(MaskedCmpVal); 1300 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1301 .addReg(NewVal).addImm(MaskImm); 1302 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1303 .addReg(ShiftAmt).addReg(MaskedNewVal); 1304 1305 // loop1MBB: 1306 // ll oldval,0(alginedaddr) 1307 // and maskedoldval0,oldval,mask 1308 // bne maskedoldval0,shiftedcmpval,sinkMBB 1309 BB = loop1MBB; 1310 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1311 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1312 .addReg(OldVal).addReg(Mask); 1313 BuildMI(BB, DL, TII->get(Mips::BNE)) 1314 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1315 1316 // loop2MBB: 1317 // and maskedoldval1,oldval,mask2 1318 // or storeval,maskedoldval1,shiftednewval 1319 // sc success,storeval,0(alignedaddr) 1320 // beq success,$0,loop1MBB 1321 BB = loop2MBB; 1322 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1323 .addReg(OldVal).addReg(Mask2); 1324 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1325 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1326 BuildMI(BB, DL, TII->get(SC), Success) 1327 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1328 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1329 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1330 1331 // sinkMBB: 1332 // srl srlres,maskedoldval0,shiftamt 1333 // sll sllres,srlres,24 1334 // sra dest,sllres,24 1335 BB = sinkMBB; 1336 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1337 1338 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1339 .addReg(ShiftAmt).addReg(MaskedOldVal0); 1340 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1341 .addReg(SrlRes).addImm(ShiftImm); 1342 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1343 .addReg(SllRes).addImm(ShiftImm); 1344 1345 MI->eraseFromParent(); // The instruction is gone now. 1346 1347 return exitMBB; 1348} 1349 1350//===----------------------------------------------------------------------===// 1351// Misc Lower Operation implementation 1352//===----------------------------------------------------------------------===// 1353SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 1354 SDValue Chain = Op.getOperand(0); 1355 SDValue Table = Op.getOperand(1); 1356 SDValue Index = Op.getOperand(2); 1357 DebugLoc DL = Op.getDebugLoc(); 1358 EVT PTy = getPointerTy(); 1359 unsigned EntrySize = 1360 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout()); 1361 1362 Index = DAG.getNode(ISD::MUL, DL, PTy, Index, 1363 DAG.getConstant(EntrySize, PTy)); 1364 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table); 1365 1366 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); 1367 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr, 1368 MachinePointerInfo::getJumpTable(), MemVT, false, false, 1369 0); 1370 Chain = Addr.getValue(1); 1371 1372 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) { 1373 // For PIC, the sequence is: 1374 // BRIND(load(Jumptable + index) + RelocBase) 1375 // RelocBase can be JumpTable, GOT or some sort of global base. 1376 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr, 1377 getPICJumpTableRelocBase(Table, DAG)); 1378 } 1379 1380 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr); 1381} 1382 1383SDValue MipsTargetLowering:: 1384lowerBRCOND(SDValue Op, SelectionDAG &DAG) const 1385{ 1386 // The first operand is the chain, the second is the condition, the third is 1387 // the block to branch to if the condition is true. 1388 SDValue Chain = Op.getOperand(0); 1389 SDValue Dest = Op.getOperand(2); 1390 DebugLoc DL = Op.getDebugLoc(); 1391 1392 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1393 1394 // Return if flag is not set by a floating point comparison. 1395 if (CondRes.getOpcode() != MipsISD::FPCmp) 1396 return Op; 1397 1398 SDValue CCNode = CondRes.getOperand(2); 1399 Mips::CondCode CC = 1400 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1401 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1402 SDValue BrCode = DAG.getConstant(Opc, MVT::i32); 1403 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1404 Dest, CondRes); 1405} 1406 1407SDValue MipsTargetLowering:: 1408lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1409{ 1410 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1411 1412 // Return if flag is not set by a floating point comparison. 1413 if (Cond.getOpcode() != MipsISD::FPCmp) 1414 return Op; 1415 1416 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1417 Op.getDebugLoc()); 1418} 1419 1420SDValue MipsTargetLowering:: 1421lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 1422{ 1423 DebugLoc DL = Op.getDebugLoc(); 1424 EVT Ty = Op.getOperand(0).getValueType(); 1425 SDValue Cond = DAG.getNode(ISD::SETCC, DL, getSetCCResultType(Ty), 1426 Op.getOperand(0), Op.getOperand(1), 1427 Op.getOperand(4)); 1428 1429 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2), 1430 Op.getOperand(3)); 1431} 1432 1433SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 1434 SDValue Cond = createFPCmp(DAG, Op); 1435 1436 assert(Cond.getOpcode() == MipsISD::FPCmp && 1437 "Floating point operand expected."); 1438 1439 SDValue True = DAG.getConstant(1, MVT::i32); 1440 SDValue False = DAG.getConstant(0, MVT::i32); 1441 1442 return createCMovFP(DAG, Cond, True, False, Op.getDebugLoc()); 1443} 1444 1445SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 1446 SelectionDAG &DAG) const { 1447 // FIXME there isn't actually debug info here 1448 DebugLoc DL = Op.getDebugLoc(); 1449 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1450 1451 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) { 1452 const MipsTargetObjectFile &TLOF = 1453 (const MipsTargetObjectFile&)getObjFileLowering(); 1454 1455 // %gp_rel relocation 1456 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) { 1457 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0, 1458 MipsII::MO_GPREL); 1459 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL, 1460 DAG.getVTList(MVT::i32), &GA, 1); 1461 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32); 1462 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode); 1463 } 1464 1465 // %hi/%lo relocation 1466 return getAddrNonPIC(Op, DAG); 1467 } 1468 1469 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))) 1470 return getAddrLocal(Op, DAG, HasMips64); 1471 1472 if (LargeGOT) 1473 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16, 1474 MipsII::MO_GOT_LO16); 1475 1476 return getAddrGlobal(Op, DAG, 1477 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16); 1478} 1479 1480SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 1481 SelectionDAG &DAG) const { 1482 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1483 return getAddrNonPIC(Op, DAG); 1484 1485 return getAddrLocal(Op, DAG, HasMips64); 1486} 1487 1488SDValue MipsTargetLowering:: 1489lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 1490{ 1491 // If the relocation model is PIC, use the General Dynamic TLS Model or 1492 // Local Dynamic TLS model, otherwise use the Initial Exec or 1493 // Local Exec TLS Model. 1494 1495 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1496 DebugLoc DL = GA->getDebugLoc(); 1497 const GlobalValue *GV = GA->getGlobal(); 1498 EVT PtrVT = getPointerTy(); 1499 1500 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 1501 1502 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 1503 // General Dynamic and Local Dynamic TLS Model. 1504 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 1505 : MipsII::MO_TLSGD; 1506 1507 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 1508 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 1509 getGlobalReg(DAG, PtrVT), TGA); 1510 unsigned PtrSize = PtrVT.getSizeInBits(); 1511 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 1512 1513 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 1514 1515 ArgListTy Args; 1516 ArgListEntry Entry; 1517 Entry.Node = Argument; 1518 Entry.Ty = PtrTy; 1519 Args.push_back(Entry); 1520 1521 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy, 1522 false, false, false, false, 0, CallingConv::C, 1523 /*IsTailCall=*/false, /*doesNotRet=*/false, 1524 /*isReturnValueUsed=*/true, 1525 TlsGetAddr, Args, DAG, DL); 1526 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 1527 1528 SDValue Ret = CallResult.first; 1529 1530 if (model != TLSModel::LocalDynamic) 1531 return Ret; 1532 1533 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1534 MipsII::MO_DTPREL_HI); 1535 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1536 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1537 MipsII::MO_DTPREL_LO); 1538 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1539 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 1540 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 1541 } 1542 1543 SDValue Offset; 1544 if (model == TLSModel::InitialExec) { 1545 // Initial Exec TLS Model 1546 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1547 MipsII::MO_GOTTPREL); 1548 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 1549 TGA); 1550 Offset = DAG.getLoad(PtrVT, DL, 1551 DAG.getEntryNode(), TGA, MachinePointerInfo(), 1552 false, false, false, 0); 1553 } else { 1554 // Local Exec TLS Model 1555 assert(model == TLSModel::LocalExec); 1556 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1557 MipsII::MO_TPREL_HI); 1558 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1559 MipsII::MO_TPREL_LO); 1560 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1561 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1562 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 1563 } 1564 1565 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 1566 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 1567} 1568 1569SDValue MipsTargetLowering:: 1570lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 1571{ 1572 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1573 return getAddrNonPIC(Op, DAG); 1574 1575 return getAddrLocal(Op, DAG, HasMips64); 1576} 1577 1578SDValue MipsTargetLowering:: 1579lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 1580{ 1581 // gp_rel relocation 1582 // FIXME: we should reference the constant pool using small data sections, 1583 // but the asm printer currently doesn't support this feature without 1584 // hacking it. This feature should come soon so we can uncomment the 1585 // stuff below. 1586 //if (IsInSmallSection(C->getType())) { 1587 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP); 1588 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32); 1589 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode); 1590 1591 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1592 return getAddrNonPIC(Op, DAG); 1593 1594 return getAddrLocal(Op, DAG, HasMips64); 1595} 1596 1597SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 1598 MachineFunction &MF = DAG.getMachineFunction(); 1599 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 1600 1601 DebugLoc DL = Op.getDebugLoc(); 1602 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 1603 getPointerTy()); 1604 1605 // vastart just stores the address of the VarArgsFrameIndex slot into the 1606 // memory location argument. 1607 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1608 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 1609 MachinePointerInfo(SV), false, false, 0); 1610} 1611 1612static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1613 EVT TyX = Op.getOperand(0).getValueType(); 1614 EVT TyY = Op.getOperand(1).getValueType(); 1615 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1616 SDValue Const31 = DAG.getConstant(31, MVT::i32); 1617 DebugLoc DL = Op.getDebugLoc(); 1618 SDValue Res; 1619 1620 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1621 // to i32. 1622 SDValue X = (TyX == MVT::f32) ? 1623 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1624 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1625 Const1); 1626 SDValue Y = (TyY == MVT::f32) ? 1627 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 1628 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 1629 Const1); 1630 1631 if (HasR2) { 1632 // ext E, Y, 31, 1 ; extract bit31 of Y 1633 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 1634 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 1635 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 1636 } else { 1637 // sll SllX, X, 1 1638 // srl SrlX, SllX, 1 1639 // srl SrlY, Y, 31 1640 // sll SllY, SrlX, 31 1641 // or Or, SrlX, SllY 1642 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1643 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1644 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 1645 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 1646 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 1647 } 1648 1649 if (TyX == MVT::f32) 1650 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 1651 1652 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1653 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1654 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1655} 1656 1657static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1658 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 1659 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 1660 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 1661 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1662 DebugLoc DL = Op.getDebugLoc(); 1663 1664 // Bitcast to integer nodes. 1665 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 1666 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 1667 1668 if (HasR2) { 1669 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 1670 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 1671 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 1672 DAG.getConstant(WidthY - 1, MVT::i32), Const1); 1673 1674 if (WidthX > WidthY) 1675 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 1676 else if (WidthY > WidthX) 1677 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 1678 1679 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 1680 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X); 1681 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 1682 } 1683 1684 // (d)sll SllX, X, 1 1685 // (d)srl SrlX, SllX, 1 1686 // (d)srl SrlY, Y, width(Y)-1 1687 // (d)sll SllY, SrlX, width(Y)-1 1688 // or Or, SrlX, SllY 1689 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 1690 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 1691 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 1692 DAG.getConstant(WidthY - 1, MVT::i32)); 1693 1694 if (WidthX > WidthY) 1695 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 1696 else if (WidthY > WidthX) 1697 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 1698 1699 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 1700 DAG.getConstant(WidthX - 1, MVT::i32)); 1701 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 1702 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 1703} 1704 1705SDValue 1706MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 1707 if (Subtarget->hasMips64()) 1708 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2()); 1709 1710 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2()); 1711} 1712 1713static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1714 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1715 DebugLoc DL = Op.getDebugLoc(); 1716 1717 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1718 // to i32. 1719 SDValue X = (Op.getValueType() == MVT::f32) ? 1720 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1721 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1722 Const1); 1723 1724 // Clear MSB. 1725 if (HasR2) 1726 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, 1727 DAG.getRegister(Mips::ZERO, MVT::i32), 1728 DAG.getConstant(31, MVT::i32), Const1, X); 1729 else { 1730 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1731 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1732 } 1733 1734 if (Op.getValueType() == MVT::f32) 1735 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res); 1736 1737 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1738 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1739 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1740} 1741 1742static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1743 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1744 DebugLoc DL = Op.getDebugLoc(); 1745 1746 // Bitcast to integer node. 1747 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0)); 1748 1749 // Clear MSB. 1750 if (HasR2) 1751 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64, 1752 DAG.getRegister(Mips::ZERO_64, MVT::i64), 1753 DAG.getConstant(63, MVT::i32), Const1, X); 1754 else { 1755 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1); 1756 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1); 1757 } 1758 1759 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res); 1760} 1761 1762SDValue 1763MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { 1764 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64)) 1765 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2()); 1766 1767 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2()); 1768} 1769 1770SDValue MipsTargetLowering:: 1771lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 1772 // check the depth 1773 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1774 "Frame address can only be determined for current frame."); 1775 1776 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1777 MFI->setFrameAddressIsTaken(true); 1778 EVT VT = Op.getValueType(); 1779 DebugLoc DL = Op.getDebugLoc(); 1780 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, 1781 IsN64 ? Mips::FP_64 : Mips::FP, VT); 1782 return FrameAddr; 1783} 1784 1785SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 1786 SelectionDAG &DAG) const { 1787 // check the depth 1788 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1789 "Return address can be determined only for current frame."); 1790 1791 MachineFunction &MF = DAG.getMachineFunction(); 1792 MachineFrameInfo *MFI = MF.getFrameInfo(); 1793 MVT VT = Op.getSimpleValueType(); 1794 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA; 1795 MFI->setReturnAddressIsTaken(true); 1796 1797 // Return RA, which contains the return address. Mark it an implicit live-in. 1798 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 1799 return DAG.getCopyFromReg(DAG.getEntryNode(), Op.getDebugLoc(), Reg, VT); 1800} 1801 1802// An EH_RETURN is the result of lowering llvm.eh.return which in turn is 1803// generated from __builtin_eh_return (offset, handler) 1804// The effect of this is to adjust the stack pointer by "offset" 1805// and then branch to "handler". 1806SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 1807 const { 1808 MachineFunction &MF = DAG.getMachineFunction(); 1809 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 1810 1811 MipsFI->setCallsEhReturn(); 1812 SDValue Chain = Op.getOperand(0); 1813 SDValue Offset = Op.getOperand(1); 1814 SDValue Handler = Op.getOperand(2); 1815 DebugLoc DL = Op.getDebugLoc(); 1816 EVT Ty = IsN64 ? MVT::i64 : MVT::i32; 1817 1818 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 1819 // EH_RETURN nodes, so that instructions are emitted back-to-back. 1820 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1; 1821 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0; 1822 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 1823 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 1824 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 1825 DAG.getRegister(OffsetReg, Ty), 1826 DAG.getRegister(AddrReg, getPointerTy()), 1827 Chain.getValue(1)); 1828} 1829 1830// TODO: set SType according to the desired memory barrier behavior. 1831SDValue 1832MipsTargetLowering::lowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const { 1833 unsigned SType = 0; 1834 DebugLoc DL = Op.getDebugLoc(); 1835 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 1836 DAG.getConstant(SType, MVT::i32)); 1837} 1838 1839SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 1840 SelectionDAG &DAG) const { 1841 // FIXME: Need pseudo-fence for 'singlethread' fences 1842 // FIXME: Set SType for weaker fences where supported/appropriate. 1843 unsigned SType = 0; 1844 DebugLoc DL = Op.getDebugLoc(); 1845 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 1846 DAG.getConstant(SType, MVT::i32)); 1847} 1848 1849SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 1850 SelectionDAG &DAG) const { 1851 DebugLoc DL = Op.getDebugLoc(); 1852 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1853 SDValue Shamt = Op.getOperand(2); 1854 1855 // if shamt < 32: 1856 // lo = (shl lo, shamt) 1857 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 1858 // else: 1859 // lo = 0 1860 // hi = (shl lo, shamt[4:0]) 1861 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1862 DAG.getConstant(-1, MVT::i32)); 1863 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, 1864 DAG.getConstant(1, MVT::i32)); 1865 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo, 1866 Not); 1867 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt); 1868 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1869 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt); 1870 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1871 DAG.getConstant(0x20, MVT::i32)); 1872 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1873 DAG.getConstant(0, MVT::i32), ShiftLeftLo); 1874 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or); 1875 1876 SDValue Ops[2] = {Lo, Hi}; 1877 return DAG.getMergeValues(Ops, 2, DL); 1878} 1879 1880SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 1881 bool IsSRA) const { 1882 DebugLoc DL = Op.getDebugLoc(); 1883 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1884 SDValue Shamt = Op.getOperand(2); 1885 1886 // if shamt < 32: 1887 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 1888 // if isSRA: 1889 // hi = (sra hi, shamt) 1890 // else: 1891 // hi = (srl hi, shamt) 1892 // else: 1893 // if isSRA: 1894 // lo = (sra hi, shamt[4:0]) 1895 // hi = (sra hi, 31) 1896 // else: 1897 // lo = (srl hi, shamt[4:0]) 1898 // hi = 0 1899 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1900 DAG.getConstant(-1, MVT::i32)); 1901 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, 1902 DAG.getConstant(1, MVT::i32)); 1903 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not); 1904 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt); 1905 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1906 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32, 1907 Hi, Shamt); 1908 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1909 DAG.getConstant(0x20, MVT::i32)); 1910 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi, 1911 DAG.getConstant(31, MVT::i32)); 1912 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or); 1913 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1914 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32), 1915 ShiftRightHi); 1916 1917 SDValue Ops[2] = {Lo, Hi}; 1918 return DAG.getMergeValues(Ops, 2, DL); 1919} 1920 1921static SDValue CreateLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 1922 SDValue Chain, SDValue Src, unsigned Offset) { 1923 SDValue Ptr = LD->getBasePtr(); 1924 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 1925 EVT BasePtrVT = Ptr.getValueType(); 1926 DebugLoc DL = LD->getDebugLoc(); 1927 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 1928 1929 if (Offset) 1930 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 1931 DAG.getConstant(Offset, BasePtrVT)); 1932 1933 SDValue Ops[] = { Chain, Ptr, Src }; 1934 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 1935 LD->getMemOperand()); 1936} 1937 1938// Expand an unaligned 32 or 64-bit integer load node. 1939SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 1940 LoadSDNode *LD = cast<LoadSDNode>(Op); 1941 EVT MemVT = LD->getMemoryVT(); 1942 1943 // Return if load is aligned or if MemVT is neither i32 nor i64. 1944 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 1945 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 1946 return SDValue(); 1947 1948 bool IsLittle = Subtarget->isLittle(); 1949 EVT VT = Op.getValueType(); 1950 ISD::LoadExtType ExtType = LD->getExtensionType(); 1951 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 1952 1953 assert((VT == MVT::i32) || (VT == MVT::i64)); 1954 1955 // Expand 1956 // (set dst, (i64 (load baseptr))) 1957 // to 1958 // (set tmp, (ldl (add baseptr, 7), undef)) 1959 // (set dst, (ldr baseptr, tmp)) 1960 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 1961 SDValue LDL = CreateLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 1962 IsLittle ? 7 : 0); 1963 return CreateLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 1964 IsLittle ? 0 : 7); 1965 } 1966 1967 SDValue LWL = CreateLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 1968 IsLittle ? 3 : 0); 1969 SDValue LWR = CreateLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 1970 IsLittle ? 0 : 3); 1971 1972 // Expand 1973 // (set dst, (i32 (load baseptr))) or 1974 // (set dst, (i64 (sextload baseptr))) or 1975 // (set dst, (i64 (extload baseptr))) 1976 // to 1977 // (set tmp, (lwl (add baseptr, 3), undef)) 1978 // (set dst, (lwr baseptr, tmp)) 1979 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 1980 (ExtType == ISD::EXTLOAD)) 1981 return LWR; 1982 1983 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 1984 1985 // Expand 1986 // (set dst, (i64 (zextload baseptr))) 1987 // to 1988 // (set tmp0, (lwl (add baseptr, 3), undef)) 1989 // (set tmp1, (lwr baseptr, tmp0)) 1990 // (set tmp2, (shl tmp1, 32)) 1991 // (set dst, (srl tmp2, 32)) 1992 DebugLoc DL = LD->getDebugLoc(); 1993 SDValue Const32 = DAG.getConstant(32, MVT::i32); 1994 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 1995 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 1996 SDValue Ops[] = { SRL, LWR.getValue(1) }; 1997 return DAG.getMergeValues(Ops, 2, DL); 1998} 1999 2000static SDValue CreateStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2001 SDValue Chain, unsigned Offset) { 2002 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2003 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2004 DebugLoc DL = SD->getDebugLoc(); 2005 SDVTList VTList = DAG.getVTList(MVT::Other); 2006 2007 if (Offset) 2008 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2009 DAG.getConstant(Offset, BasePtrVT)); 2010 2011 SDValue Ops[] = { Chain, Value, Ptr }; 2012 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 2013 SD->getMemOperand()); 2014} 2015 2016// Expand an unaligned 32 or 64-bit integer store node. 2017SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2018 StoreSDNode *SD = cast<StoreSDNode>(Op); 2019 EVT MemVT = SD->getMemoryVT(); 2020 2021 // Return if store is aligned or if MemVT is neither i32 nor i64. 2022 if ((SD->getAlignment() >= MemVT.getSizeInBits() / 8) || 2023 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 2024 return SDValue(); 2025 2026 bool IsLittle = Subtarget->isLittle(); 2027 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2028 EVT VT = Value.getValueType(); 2029 2030 // Expand 2031 // (store val, baseptr) or 2032 // (truncstore val, baseptr) 2033 // to 2034 // (swl val, (add baseptr, 3)) 2035 // (swr val, baseptr) 2036 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2037 SDValue SWL = CreateStoreLR(MipsISD::SWL, DAG, SD, Chain, 2038 IsLittle ? 3 : 0); 2039 return CreateStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2040 } 2041 2042 assert(VT == MVT::i64); 2043 2044 // Expand 2045 // (store val, baseptr) 2046 // to 2047 // (sdl val, (add baseptr, 7)) 2048 // (sdr val, baseptr) 2049 SDValue SDL = CreateStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2050 return CreateStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2051} 2052 2053static SDValue initAccumulator(SDValue In, DebugLoc DL, SelectionDAG &DAG) { 2054 SDValue InLo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In, 2055 DAG.getConstant(0, MVT::i32)); 2056 SDValue InHi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, In, 2057 DAG.getConstant(1, MVT::i32)); 2058 return DAG.getNode(MipsISD::InsertLOHI, DL, MVT::Untyped, InLo, InHi); 2059} 2060 2061static SDValue extractLOHI(SDValue Op, DebugLoc DL, SelectionDAG &DAG) { 2062 SDValue Lo = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op, 2063 DAG.getConstant(Mips::sub_lo, MVT::i32)); 2064 SDValue Hi = DAG.getNode(MipsISD::ExtractLOHI, DL, MVT::i32, Op, 2065 DAG.getConstant(Mips::sub_hi, MVT::i32)); 2066 return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi); 2067} 2068 2069// This function expands mips intrinsic nodes which have 64-bit input operands 2070// or output values. 2071// 2072// out64 = intrinsic-node in64 2073// => 2074// lo = copy (extract-element (in64, 0)) 2075// hi = copy (extract-element (in64, 1)) 2076// mips-specific-node 2077// v0 = copy lo 2078// v1 = copy hi 2079// out64 = merge-values (v0, v1) 2080// 2081static SDValue lowerDSPIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) { 2082 DebugLoc DL = Op.getDebugLoc(); 2083 bool HasChainIn = Op->getOperand(0).getValueType() == MVT::Other; 2084 SmallVector<SDValue, 3> Ops; 2085 unsigned OpNo = 0; 2086 2087 // See if Op has a chain input. 2088 if (HasChainIn) 2089 Ops.push_back(Op->getOperand(OpNo++)); 2090 2091 // The next operand is the intrinsic opcode. 2092 assert(Op->getOperand(OpNo).getOpcode() == ISD::TargetConstant); 2093 2094 // See if the next operand has type i64. 2095 SDValue Opnd = Op->getOperand(++OpNo), In64; 2096 2097 if (Opnd.getValueType() == MVT::i64) 2098 In64 = initAccumulator(Opnd, DL, DAG); 2099 else 2100 Ops.push_back(Opnd); 2101 2102 // Push the remaining operands. 2103 for (++OpNo ; OpNo < Op->getNumOperands(); ++OpNo) 2104 Ops.push_back(Op->getOperand(OpNo)); 2105 2106 // Add In64 to the end of the list. 2107 if (In64.getNode()) 2108 Ops.push_back(In64); 2109 2110 // Scan output. 2111 SmallVector<EVT, 2> ResTys; 2112 2113 for (SDNode::value_iterator I = Op->value_begin(), E = Op->value_end(); 2114 I != E; ++I) 2115 ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I); 2116 2117 // Create node. 2118 SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size()); 2119 SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val; 2120 2121 if (!HasChainIn) 2122 return Out; 2123 2124 assert(Val->getValueType(1) == MVT::Other); 2125 SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) }; 2126 return DAG.getMergeValues(Vals, 2, DL); 2127} 2128 2129SDValue MipsTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op, 2130 SelectionDAG &DAG) const { 2131 switch (cast<ConstantSDNode>(Op->getOperand(0))->getZExtValue()) { 2132 default: 2133 return SDValue(); 2134 case Intrinsic::mips_shilo: 2135 return lowerDSPIntr(Op, DAG, MipsISD::SHILO); 2136 case Intrinsic::mips_dpau_h_qbl: 2137 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBL); 2138 case Intrinsic::mips_dpau_h_qbr: 2139 return lowerDSPIntr(Op, DAG, MipsISD::DPAU_H_QBR); 2140 case Intrinsic::mips_dpsu_h_qbl: 2141 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBL); 2142 case Intrinsic::mips_dpsu_h_qbr: 2143 return lowerDSPIntr(Op, DAG, MipsISD::DPSU_H_QBR); 2144 case Intrinsic::mips_dpa_w_ph: 2145 return lowerDSPIntr(Op, DAG, MipsISD::DPA_W_PH); 2146 case Intrinsic::mips_dps_w_ph: 2147 return lowerDSPIntr(Op, DAG, MipsISD::DPS_W_PH); 2148 case Intrinsic::mips_dpax_w_ph: 2149 return lowerDSPIntr(Op, DAG, MipsISD::DPAX_W_PH); 2150 case Intrinsic::mips_dpsx_w_ph: 2151 return lowerDSPIntr(Op, DAG, MipsISD::DPSX_W_PH); 2152 case Intrinsic::mips_mulsa_w_ph: 2153 return lowerDSPIntr(Op, DAG, MipsISD::MULSA_W_PH); 2154 case Intrinsic::mips_mult: 2155 return lowerDSPIntr(Op, DAG, MipsISD::Mult); 2156 case Intrinsic::mips_multu: 2157 return lowerDSPIntr(Op, DAG, MipsISD::Multu); 2158 case Intrinsic::mips_madd: 2159 return lowerDSPIntr(Op, DAG, MipsISD::MAdd); 2160 case Intrinsic::mips_maddu: 2161 return lowerDSPIntr(Op, DAG, MipsISD::MAddu); 2162 case Intrinsic::mips_msub: 2163 return lowerDSPIntr(Op, DAG, MipsISD::MSub); 2164 case Intrinsic::mips_msubu: 2165 return lowerDSPIntr(Op, DAG, MipsISD::MSubu); 2166 } 2167} 2168 2169SDValue MipsTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op, 2170 SelectionDAG &DAG) const { 2171 switch (cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue()) { 2172 default: 2173 return SDValue(); 2174 case Intrinsic::mips_extp: 2175 return lowerDSPIntr(Op, DAG, MipsISD::EXTP); 2176 case Intrinsic::mips_extpdp: 2177 return lowerDSPIntr(Op, DAG, MipsISD::EXTPDP); 2178 case Intrinsic::mips_extr_w: 2179 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_W); 2180 case Intrinsic::mips_extr_r_w: 2181 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_R_W); 2182 case Intrinsic::mips_extr_rs_w: 2183 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_RS_W); 2184 case Intrinsic::mips_extr_s_h: 2185 return lowerDSPIntr(Op, DAG, MipsISD::EXTR_S_H); 2186 case Intrinsic::mips_mthlip: 2187 return lowerDSPIntr(Op, DAG, MipsISD::MTHLIP); 2188 case Intrinsic::mips_mulsaq_s_w_ph: 2189 return lowerDSPIntr(Op, DAG, MipsISD::MULSAQ_S_W_PH); 2190 case Intrinsic::mips_maq_s_w_phl: 2191 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHL); 2192 case Intrinsic::mips_maq_s_w_phr: 2193 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_S_W_PHR); 2194 case Intrinsic::mips_maq_sa_w_phl: 2195 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHL); 2196 case Intrinsic::mips_maq_sa_w_phr: 2197 return lowerDSPIntr(Op, DAG, MipsISD::MAQ_SA_W_PHR); 2198 case Intrinsic::mips_dpaq_s_w_ph: 2199 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_S_W_PH); 2200 case Intrinsic::mips_dpsq_s_w_ph: 2201 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_S_W_PH); 2202 case Intrinsic::mips_dpaq_sa_l_w: 2203 return lowerDSPIntr(Op, DAG, MipsISD::DPAQ_SA_L_W); 2204 case Intrinsic::mips_dpsq_sa_l_w: 2205 return lowerDSPIntr(Op, DAG, MipsISD::DPSQ_SA_L_W); 2206 case Intrinsic::mips_dpaqx_s_w_ph: 2207 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_S_W_PH); 2208 case Intrinsic::mips_dpaqx_sa_w_ph: 2209 return lowerDSPIntr(Op, DAG, MipsISD::DPAQX_SA_W_PH); 2210 case Intrinsic::mips_dpsqx_s_w_ph: 2211 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_S_W_PH); 2212 case Intrinsic::mips_dpsqx_sa_w_ph: 2213 return lowerDSPIntr(Op, DAG, MipsISD::DPSQX_SA_W_PH); 2214 } 2215} 2216 2217SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const { 2218 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR 2219 || cast<ConstantSDNode> 2220 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0 2221 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET) 2222 return SDValue(); 2223 2224 // The pattern 2225 // (add (frameaddr 0), (frame_to_args_offset)) 2226 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to 2227 // (add FrameObject, 0) 2228 // where FrameObject is a fixed StackObject with offset 0 which points to 2229 // the old stack pointer. 2230 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2231 EVT ValTy = Op->getValueType(0); 2232 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2233 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy); 2234 return DAG.getNode(ISD::ADD, Op->getDebugLoc(), ValTy, InArgsAddr, 2235 DAG.getConstant(0, ValTy)); 2236} 2237 2238//===----------------------------------------------------------------------===// 2239// Calling Convention Implementation 2240//===----------------------------------------------------------------------===// 2241 2242//===----------------------------------------------------------------------===// 2243// TODO: Implement a generic logic using tblgen that can support this. 2244// Mips O32 ABI rules: 2245// --- 2246// i32 - Passed in A0, A1, A2, A3 and stack 2247// f32 - Only passed in f32 registers if no int reg has been used yet to hold 2248// an argument. Otherwise, passed in A1, A2, A3 and stack. 2249// f64 - Only passed in two aliased f32 registers if no int reg has been used 2250// yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2251// not used, it must be shadowed. If only A3 is avaiable, shadow it and 2252// go to stack. 2253// 2254// For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2255//===----------------------------------------------------------------------===// 2256 2257static bool CC_MipsO32(unsigned ValNo, MVT ValVT, 2258 MVT LocVT, CCValAssign::LocInfo LocInfo, 2259 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2260 2261 static const unsigned IntRegsSize=4, FloatRegsSize=2; 2262 2263 static const uint16_t IntRegs[] = { 2264 Mips::A0, Mips::A1, Mips::A2, Mips::A3 2265 }; 2266 static const uint16_t F32Regs[] = { 2267 Mips::F12, Mips::F14 2268 }; 2269 static const uint16_t F64Regs[] = { 2270 Mips::D6, Mips::D7 2271 }; 2272 2273 // Do not process byval args here. 2274 if (ArgFlags.isByVal()) 2275 return true; 2276 2277 // Promote i8 and i16 2278 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2279 LocVT = MVT::i32; 2280 if (ArgFlags.isSExt()) 2281 LocInfo = CCValAssign::SExt; 2282 else if (ArgFlags.isZExt()) 2283 LocInfo = CCValAssign::ZExt; 2284 else 2285 LocInfo = CCValAssign::AExt; 2286 } 2287 2288 unsigned Reg; 2289 2290 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2291 // is true: function is vararg, argument is 3rd or higher, there is previous 2292 // argument which is not f32 or f64. 2293 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 2294 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo; 2295 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2296 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2297 2298 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2299 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2300 // If this is the first part of an i64 arg, 2301 // the allocated register must be either A0 or A2. 2302 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2303 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2304 LocVT = MVT::i32; 2305 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2306 // Allocate int register and shadow next int register. If first 2307 // available register is Mips::A1 or Mips::A3, shadow it too. 2308 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2309 if (Reg == Mips::A1 || Reg == Mips::A3) 2310 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2311 State.AllocateReg(IntRegs, IntRegsSize); 2312 LocVT = MVT::i32; 2313 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2314 // we are guaranteed to find an available float register 2315 if (ValVT == MVT::f32) { 2316 Reg = State.AllocateReg(F32Regs, FloatRegsSize); 2317 // Shadow int register 2318 State.AllocateReg(IntRegs, IntRegsSize); 2319 } else { 2320 Reg = State.AllocateReg(F64Regs, FloatRegsSize); 2321 // Shadow int registers 2322 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize); 2323 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2324 State.AllocateReg(IntRegs, IntRegsSize); 2325 State.AllocateReg(IntRegs, IntRegsSize); 2326 } 2327 } else 2328 llvm_unreachable("Cannot handle this ValVT."); 2329 2330 if (!Reg) { 2331 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3, 2332 OrigAlign); 2333 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2334 } else 2335 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2336 2337 return false; 2338} 2339 2340#include "MipsGenCallingConv.inc" 2341 2342//===----------------------------------------------------------------------===// 2343// Call Calling Convention Implementation 2344//===----------------------------------------------------------------------===// 2345 2346static const unsigned O32IntRegsSize = 4; 2347 2348// Return next O32 integer argument register. 2349static unsigned getNextIntArgReg(unsigned Reg) { 2350 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 2351 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 2352} 2353 2354SDValue 2355MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 2356 SDValue Chain, SDValue Arg, DebugLoc DL, 2357 bool IsTailCall, SelectionDAG &DAG) const { 2358 if (!IsTailCall) { 2359 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, 2360 DAG.getIntPtrConstant(Offset)); 2361 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false, 2362 false, 0); 2363 } 2364 2365 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2366 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 2367 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2368 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 2369 /*isVolatile=*/ true, false, 0); 2370} 2371 2372void MipsTargetLowering:: 2373getOpndList(SmallVectorImpl<SDValue> &Ops, 2374 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 2375 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 2376 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const { 2377 // Insert node "GP copy globalreg" before call to function. 2378 // 2379 // R_MIPS_CALL* operators (emitted when non-internal functions are called 2380 // in PIC mode) allow symbols to be resolved via lazy binding. 2381 // The lazy binding stub requires GP to point to the GOT. 2382 if (IsPICCall && !InternalLinkage) { 2383 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP; 2384 EVT Ty = IsN64 ? MVT::i64 : MVT::i32; 2385 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 2386 } 2387 2388 // Build a sequence of copy-to-reg nodes chained together with token 2389 // chain and flag operands which copy the outgoing args into registers. 2390 // The InFlag in necessary since all emitted instructions must be 2391 // stuck together. 2392 SDValue InFlag; 2393 2394 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 2395 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 2396 RegsToPass[i].second, InFlag); 2397 InFlag = Chain.getValue(1); 2398 } 2399 2400 // Add argument registers to the end of the list so that they are 2401 // known live into the call. 2402 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 2403 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 2404 RegsToPass[i].second.getValueType())); 2405 2406 // Add a register mask operand representing the call-preserved registers. 2407 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 2408 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv); 2409 assert(Mask && "Missing call preserved mask for calling convention"); 2410 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 2411 2412 if (InFlag.getNode()) 2413 Ops.push_back(InFlag); 2414} 2415 2416/// LowerCall - functions arguments are copied from virtual regs to 2417/// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 2418SDValue 2419MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 2420 SmallVectorImpl<SDValue> &InVals) const { 2421 SelectionDAG &DAG = CLI.DAG; 2422 DebugLoc &DL = CLI.DL; 2423 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; 2424 SmallVector<SDValue, 32> &OutVals = CLI.OutVals; 2425 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; 2426 SDValue Chain = CLI.Chain; 2427 SDValue Callee = CLI.Callee; 2428 bool &IsTailCall = CLI.IsTailCall; 2429 CallingConv::ID CallConv = CLI.CallConv; 2430 bool IsVarArg = CLI.IsVarArg; 2431 2432 MachineFunction &MF = DAG.getMachineFunction(); 2433 MachineFrameInfo *MFI = MF.getFrameInfo(); 2434 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering(); 2435 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_; 2436 2437 // Analyze operands of the call, assigning locations to each operand. 2438 SmallVector<CCValAssign, 16> ArgLocs; 2439 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2440 getTargetMachine(), ArgLocs, *DAG.getContext()); 2441 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2442 2443 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg, 2444 getTargetMachine().Options.UseSoftFloat, 2445 Callee.getNode(), CLI.Args); 2446 2447 // Get a count of how many bytes are to be pushed on the stack. 2448 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2449 2450 // Check if it's really possible to do a tail call. 2451 if (IsTailCall) 2452 IsTailCall = 2453 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset, 2454 *MF.getInfo<MipsFunctionInfo>()); 2455 2456 if (IsTailCall) 2457 ++NumTailCalls; 2458 2459 // Chain is the output chain of the last Load/Store or CopyToReg node. 2460 // ByValChain is the output chain of the last Memcpy node created for copying 2461 // byval arguments to the stack. 2462 unsigned StackAlignment = TFL->getStackAlignment(); 2463 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment); 2464 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true); 2465 2466 if (!IsTailCall) 2467 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal); 2468 2469 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, 2470 IsN64 ? Mips::SP_64 : Mips::SP, 2471 getPointerTy()); 2472 2473 // With EABI is it possible to have 16 args on registers. 2474 std::deque< std::pair<unsigned, SDValue> > RegsToPass; 2475 SmallVector<SDValue, 8> MemOpChains; 2476 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2477 2478 // Walk the register/memloc assignments, inserting copies/loads. 2479 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2480 SDValue Arg = OutVals[i]; 2481 CCValAssign &VA = ArgLocs[i]; 2482 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 2483 ISD::ArgFlagsTy Flags = Outs[i].Flags; 2484 2485 // ByVal Arg. 2486 if (Flags.isByVal()) { 2487 assert(Flags.getByValSize() && 2488 "ByVal args of size 0 should have been ignored by front-end."); 2489 assert(ByValArg != MipsCCInfo.byval_end()); 2490 assert(!IsTailCall && 2491 "Do not tail-call optimize if there is a byval argument."); 2492 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 2493 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle()); 2494 ++ByValArg; 2495 continue; 2496 } 2497 2498 // Promote the value if needed. 2499 switch (VA.getLocInfo()) { 2500 default: llvm_unreachable("Unknown loc info!"); 2501 case CCValAssign::Full: 2502 if (VA.isRegLoc()) { 2503 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 2504 (ValVT == MVT::f64 && LocVT == MVT::i64) || 2505 (ValVT == MVT::i64 && LocVT == MVT::f64)) 2506 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 2507 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 2508 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2509 Arg, DAG.getConstant(0, MVT::i32)); 2510 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2511 Arg, DAG.getConstant(1, MVT::i32)); 2512 if (!Subtarget->isLittle()) 2513 std::swap(Lo, Hi); 2514 unsigned LocRegLo = VA.getLocReg(); 2515 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 2516 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 2517 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 2518 continue; 2519 } 2520 } 2521 break; 2522 case CCValAssign::SExt: 2523 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 2524 break; 2525 case CCValAssign::ZExt: 2526 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 2527 break; 2528 case CCValAssign::AExt: 2529 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 2530 break; 2531 } 2532 2533 // Arguments that can be passed on register must be kept at 2534 // RegsToPass vector 2535 if (VA.isRegLoc()) { 2536 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 2537 continue; 2538 } 2539 2540 // Register can't get to this point... 2541 assert(VA.isMemLoc()); 2542 2543 // emit ISD::STORE whichs stores the 2544 // parameter value to a stack Location 2545 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 2546 Chain, Arg, DL, IsTailCall, DAG)); 2547 } 2548 2549 // Transform all store nodes into one single node because all store 2550 // nodes are independent of each other. 2551 if (!MemOpChains.empty()) 2552 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2553 &MemOpChains[0], MemOpChains.size()); 2554 2555 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 2556 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 2557 // node so that legalize doesn't hack it. 2558 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25 2559 bool GlobalOrExternal = false, InternalLinkage = false; 2560 SDValue CalleeLo; 2561 2562 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2563 if (IsPICCall) { 2564 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 2565 2566 if (InternalLinkage) 2567 Callee = getAddrLocal(Callee, DAG, HasMips64); 2568 else if (LargeGOT) 2569 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2570 MipsII::MO_CALL_LO16); 2571 else 2572 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2573 } else 2574 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0, 2575 MipsII::MO_NO_FLAG); 2576 GlobalOrExternal = true; 2577 } 2578 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 2579 if (!IsN64 && !IsPIC) // !N64 && static 2580 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(), 2581 MipsII::MO_NO_FLAG); 2582 else if (LargeGOT) 2583 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2584 MipsII::MO_CALL_LO16); 2585 else // N64 || PIC 2586 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2587 2588 GlobalOrExternal = true; 2589 } 2590 2591 SmallVector<SDValue, 8> Ops(1, Chain); 2592 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 2593 2594 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage, 2595 CLI, Callee, Chain); 2596 2597 if (IsTailCall) 2598 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size()); 2599 2600 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size()); 2601 SDValue InFlag = Chain.getValue(1); 2602 2603 // Create the CALLSEQ_END node. 2604 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 2605 DAG.getIntPtrConstant(0, true), InFlag); 2606 InFlag = Chain.getValue(1); 2607 2608 // Handle result values, copying them out of physregs into vregs that we 2609 // return. 2610 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, 2611 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy); 2612} 2613 2614/// LowerCallResult - Lower the result values of a call into the 2615/// appropriate copies out of appropriate physical registers. 2616SDValue 2617MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 2618 CallingConv::ID CallConv, bool IsVarArg, 2619 const SmallVectorImpl<ISD::InputArg> &Ins, 2620 DebugLoc DL, SelectionDAG &DAG, 2621 SmallVectorImpl<SDValue> &InVals, 2622 const SDNode *CallNode, 2623 const Type *RetTy) const { 2624 // Assign locations to each value returned by this call. 2625 SmallVector<CCValAssign, 16> RVLocs; 2626 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2627 getTargetMachine(), RVLocs, *DAG.getContext()); 2628 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2629 2630 MipsCCInfo.analyzeCallResult(Ins, getTargetMachine().Options.UseSoftFloat, 2631 CallNode, RetTy); 2632 2633 // Copy all of the result registers out of their specified physreg. 2634 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2635 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 2636 RVLocs[i].getLocVT(), InFlag); 2637 Chain = Val.getValue(1); 2638 InFlag = Val.getValue(2); 2639 2640 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2641 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val); 2642 2643 InVals.push_back(Val); 2644 } 2645 2646 return Chain; 2647} 2648 2649//===----------------------------------------------------------------------===// 2650// Formal Arguments Calling Convention Implementation 2651//===----------------------------------------------------------------------===// 2652/// LowerFormalArguments - transform physical registers into virtual registers 2653/// and generate load operations for arguments places on the stack. 2654SDValue 2655MipsTargetLowering::LowerFormalArguments(SDValue Chain, 2656 CallingConv::ID CallConv, 2657 bool IsVarArg, 2658 const SmallVectorImpl<ISD::InputArg> &Ins, 2659 DebugLoc DL, SelectionDAG &DAG, 2660 SmallVectorImpl<SDValue> &InVals) 2661 const { 2662 MachineFunction &MF = DAG.getMachineFunction(); 2663 MachineFrameInfo *MFI = MF.getFrameInfo(); 2664 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2665 2666 MipsFI->setVarArgsFrameIndex(0); 2667 2668 // Used with vargs to acumulate store chains. 2669 std::vector<SDValue> OutChains; 2670 2671 // Assign locations to all of the incoming arguments. 2672 SmallVector<CCValAssign, 16> ArgLocs; 2673 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2674 getTargetMachine(), ArgLocs, *DAG.getContext()); 2675 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2676 Function::const_arg_iterator FuncArg = 2677 DAG.getMachineFunction().getFunction()->arg_begin(); 2678 bool UseSoftFloat = getTargetMachine().Options.UseSoftFloat; 2679 2680 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg); 2681 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 2682 MipsCCInfo.hasByValArg()); 2683 2684 unsigned CurArgIdx = 0; 2685 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2686 2687 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2688 CCValAssign &VA = ArgLocs[i]; 2689 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx); 2690 CurArgIdx = Ins[i].OrigArgIndex; 2691 EVT ValVT = VA.getValVT(); 2692 ISD::ArgFlagsTy Flags = Ins[i].Flags; 2693 bool IsRegLoc = VA.isRegLoc(); 2694 2695 if (Flags.isByVal()) { 2696 assert(Flags.getByValSize() && 2697 "ByVal args of size 0 should have been ignored by front-end."); 2698 assert(ByValArg != MipsCCInfo.byval_end()); 2699 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 2700 MipsCCInfo, *ByValArg); 2701 ++ByValArg; 2702 continue; 2703 } 2704 2705 // Arguments stored on registers 2706 if (IsRegLoc) { 2707 EVT RegVT = VA.getLocVT(); 2708 unsigned ArgReg = VA.getLocReg(); 2709 const TargetRegisterClass *RC; 2710 2711 if (RegVT == MVT::i32) 2712 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass : 2713 &Mips::CPURegsRegClass; 2714 else if (RegVT == MVT::i64) 2715 RC = &Mips::CPU64RegsRegClass; 2716 else if (RegVT == MVT::f32) 2717 RC = &Mips::FGR32RegClass; 2718 else if (RegVT == MVT::f64) 2719 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass; 2720 else 2721 llvm_unreachable("RegVT not supported by FormalArguments Lowering"); 2722 2723 // Transform the arguments stored on 2724 // physical registers into virtual ones 2725 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 2726 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 2727 2728 // If this is an 8 or 16-bit value, it has been passed promoted 2729 // to 32 bits. Insert an assert[sz]ext to capture this, then 2730 // truncate to the right size. 2731 if (VA.getLocInfo() != CCValAssign::Full) { 2732 unsigned Opcode = 0; 2733 if (VA.getLocInfo() == CCValAssign::SExt) 2734 Opcode = ISD::AssertSext; 2735 else if (VA.getLocInfo() == CCValAssign::ZExt) 2736 Opcode = ISD::AssertZext; 2737 if (Opcode) 2738 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue, 2739 DAG.getValueType(ValVT)); 2740 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue); 2741 } 2742 2743 // Handle floating point arguments passed in integer registers and 2744 // long double arguments passed in floating point registers. 2745 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 2746 (RegVT == MVT::i64 && ValVT == MVT::f64) || 2747 (RegVT == MVT::f64 && ValVT == MVT::i64)) 2748 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 2749 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) { 2750 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 2751 getNextIntArgReg(ArgReg), RC); 2752 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 2753 if (!Subtarget->isLittle()) 2754 std::swap(ArgValue, ArgValue2); 2755 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 2756 ArgValue, ArgValue2); 2757 } 2758 2759 InVals.push_back(ArgValue); 2760 } else { // VA.isRegLoc() 2761 2762 // sanity check 2763 assert(VA.isMemLoc()); 2764 2765 // The stack pointer offset is relative to the caller stack frame. 2766 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, 2767 VA.getLocMemOffset(), true); 2768 2769 // Create load nodes to retrieve arguments from the stack 2770 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2771 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN, 2772 MachinePointerInfo::getFixedStack(FI), 2773 false, false, false, 0)); 2774 } 2775 } 2776 2777 // The mips ABIs for returning structs by value requires that we copy 2778 // the sret argument into $v0 for the return. Save the argument into 2779 // a virtual register so that we can access it from the return points. 2780 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) { 2781 unsigned Reg = MipsFI->getSRetReturnReg(); 2782 if (!Reg) { 2783 Reg = MF.getRegInfo(). 2784 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32)); 2785 MipsFI->setSRetReturnReg(Reg); 2786 } 2787 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]); 2788 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 2789 } 2790 2791 if (IsVarArg) 2792 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG); 2793 2794 // All stores are grouped in one node to allow the matching between 2795 // the size of Ins and InVals. This only happens when on varg functions 2796 if (!OutChains.empty()) { 2797 OutChains.push_back(Chain); 2798 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2799 &OutChains[0], OutChains.size()); 2800 } 2801 2802 return Chain; 2803} 2804 2805//===----------------------------------------------------------------------===// 2806// Return Value Calling Convention Implementation 2807//===----------------------------------------------------------------------===// 2808 2809bool 2810MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 2811 MachineFunction &MF, bool IsVarArg, 2812 const SmallVectorImpl<ISD::OutputArg> &Outs, 2813 LLVMContext &Context) const { 2814 SmallVector<CCValAssign, 16> RVLocs; 2815 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), 2816 RVLocs, Context); 2817 return CCInfo.CheckReturn(Outs, RetCC_Mips); 2818} 2819 2820SDValue 2821MipsTargetLowering::LowerReturn(SDValue Chain, 2822 CallingConv::ID CallConv, bool IsVarArg, 2823 const SmallVectorImpl<ISD::OutputArg> &Outs, 2824 const SmallVectorImpl<SDValue> &OutVals, 2825 DebugLoc DL, SelectionDAG &DAG) const { 2826 // CCValAssign - represent the assignment of 2827 // the return value to a location 2828 SmallVector<CCValAssign, 16> RVLocs; 2829 MachineFunction &MF = DAG.getMachineFunction(); 2830 2831 // CCState - Info about the registers and stack slot. 2832 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs, 2833 *DAG.getContext()); 2834 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2835 2836 // Analyze return values. 2837 MipsCCInfo.analyzeReturn(Outs, getTargetMachine().Options.UseSoftFloat, 2838 MF.getFunction()->getReturnType()); 2839 2840 SDValue Flag; 2841 SmallVector<SDValue, 4> RetOps(1, Chain); 2842 2843 // Copy the result values into the output registers. 2844 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2845 SDValue Val = OutVals[i]; 2846 CCValAssign &VA = RVLocs[i]; 2847 assert(VA.isRegLoc() && "Can only return in registers!"); 2848 2849 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2850 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val); 2851 2852 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 2853 2854 // Guarantee that all emitted copies are stuck together with flags. 2855 Flag = Chain.getValue(1); 2856 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2857 } 2858 2859 // The mips ABIs for returning structs by value requires that we copy 2860 // the sret argument into $v0 for the return. We saved the argument into 2861 // a virtual register in the entry block, so now we copy the value out 2862 // and into $v0. 2863 if (MF.getFunction()->hasStructRetAttr()) { 2864 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2865 unsigned Reg = MipsFI->getSRetReturnReg(); 2866 2867 if (!Reg) 2868 llvm_unreachable("sret virtual register not created in the entry block"); 2869 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy()); 2870 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0; 2871 2872 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 2873 Flag = Chain.getValue(1); 2874 RetOps.push_back(DAG.getRegister(V0, getPointerTy())); 2875 } 2876 2877 RetOps[0] = Chain; // Update chain. 2878 2879 // Add the flag if we have it. 2880 if (Flag.getNode()) 2881 RetOps.push_back(Flag); 2882 2883 // Return on Mips is always a "jr $ra" 2884 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size()); 2885} 2886 2887//===----------------------------------------------------------------------===// 2888// Mips Inline Assembly Support 2889//===----------------------------------------------------------------------===// 2890 2891/// getConstraintType - Given a constraint letter, return the type of 2892/// constraint it is for this target. 2893MipsTargetLowering::ConstraintType MipsTargetLowering:: 2894getConstraintType(const std::string &Constraint) const 2895{ 2896 // Mips specific constrainy 2897 // GCC config/mips/constraints.md 2898 // 2899 // 'd' : An address register. Equivalent to r 2900 // unless generating MIPS16 code. 2901 // 'y' : Equivalent to r; retained for 2902 // backwards compatibility. 2903 // 'c' : A register suitable for use in an indirect 2904 // jump. This will always be $25 for -mabicalls. 2905 // 'l' : The lo register. 1 word storage. 2906 // 'x' : The hilo register pair. Double word storage. 2907 if (Constraint.size() == 1) { 2908 switch (Constraint[0]) { 2909 default : break; 2910 case 'd': 2911 case 'y': 2912 case 'f': 2913 case 'c': 2914 case 'l': 2915 case 'x': 2916 return C_RegisterClass; 2917 case 'R': 2918 return C_Memory; 2919 } 2920 } 2921 return TargetLowering::getConstraintType(Constraint); 2922} 2923 2924/// Examine constraint type and operand type and determine a weight value. 2925/// This object must already have been set up with the operand type 2926/// and the current alternative constraint selected. 2927TargetLowering::ConstraintWeight 2928MipsTargetLowering::getSingleConstraintMatchWeight( 2929 AsmOperandInfo &info, const char *constraint) const { 2930 ConstraintWeight weight = CW_Invalid; 2931 Value *CallOperandVal = info.CallOperandVal; 2932 // If we don't have a value, we can't do a match, 2933 // but allow it at the lowest weight. 2934 if (CallOperandVal == NULL) 2935 return CW_Default; 2936 Type *type = CallOperandVal->getType(); 2937 // Look at the constraint type. 2938 switch (*constraint) { 2939 default: 2940 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 2941 break; 2942 case 'd': 2943 case 'y': 2944 if (type->isIntegerTy()) 2945 weight = CW_Register; 2946 break; 2947 case 'f': 2948 if (type->isFloatTy()) 2949 weight = CW_Register; 2950 break; 2951 case 'c': // $25 for indirect jumps 2952 case 'l': // lo register 2953 case 'x': // hilo register pair 2954 if (type->isIntegerTy()) 2955 weight = CW_SpecificReg; 2956 break; 2957 case 'I': // signed 16 bit immediate 2958 case 'J': // integer zero 2959 case 'K': // unsigned 16 bit immediate 2960 case 'L': // signed 32 bit immediate where lower 16 bits are 0 2961 case 'N': // immediate in the range of -65535 to -1 (inclusive) 2962 case 'O': // signed 15 bit immediate (+- 16383) 2963 case 'P': // immediate in the range of 65535 to 1 (inclusive) 2964 if (isa<ConstantInt>(CallOperandVal)) 2965 weight = CW_Constant; 2966 break; 2967 case 'R': 2968 weight = CW_Memory; 2969 break; 2970 } 2971 return weight; 2972} 2973 2974/// Given a register class constraint, like 'r', if this corresponds directly 2975/// to an LLVM register class, return a register of 0 and the register class 2976/// pointer. 2977std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering:: 2978getRegForInlineAsmConstraint(const std::string &Constraint, EVT VT) const 2979{ 2980 if (Constraint.size() == 1) { 2981 switch (Constraint[0]) { 2982 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 2983 case 'y': // Same as 'r'. Exists for compatibility. 2984 case 'r': 2985 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 2986 if (Subtarget->inMips16Mode()) 2987 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 2988 return std::make_pair(0U, &Mips::CPURegsRegClass); 2989 } 2990 if (VT == MVT::i64 && !HasMips64) 2991 return std::make_pair(0U, &Mips::CPURegsRegClass); 2992 if (VT == MVT::i64 && HasMips64) 2993 return std::make_pair(0U, &Mips::CPU64RegsRegClass); 2994 // This will generate an error message 2995 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2996 case 'f': 2997 if (VT == MVT::f32) 2998 return std::make_pair(0U, &Mips::FGR32RegClass); 2999 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) { 3000 if (Subtarget->isFP64bit()) 3001 return std::make_pair(0U, &Mips::FGR64RegClass); 3002 return std::make_pair(0U, &Mips::AFGR64RegClass); 3003 } 3004 break; 3005 case 'c': // register suitable for indirect jump 3006 if (VT == MVT::i32) 3007 return std::make_pair((unsigned)Mips::T9, &Mips::CPURegsRegClass); 3008 assert(VT == MVT::i64 && "Unexpected type."); 3009 return std::make_pair((unsigned)Mips::T9_64, &Mips::CPU64RegsRegClass); 3010 case 'l': // register suitable for indirect jump 3011 if (VT == MVT::i32) 3012 return std::make_pair((unsigned)Mips::LO, &Mips::HILORegClass); 3013 return std::make_pair((unsigned)Mips::LO64, &Mips::HILO64RegClass); 3014 case 'x': // register suitable for indirect jump 3015 // Fixme: Not triggering the use of both hi and low 3016 // This will generate an error message 3017 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 3018 } 3019 } 3020 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3021} 3022 3023/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3024/// vector. If it is invalid, don't add anything to Ops. 3025void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3026 std::string &Constraint, 3027 std::vector<SDValue>&Ops, 3028 SelectionDAG &DAG) const { 3029 SDValue Result(0, 0); 3030 3031 // Only support length 1 constraints for now. 3032 if (Constraint.length() > 1) return; 3033 3034 char ConstraintLetter = Constraint[0]; 3035 switch (ConstraintLetter) { 3036 default: break; // This will fall through to the generic implementation 3037 case 'I': // Signed 16 bit constant 3038 // If this fails, the parent routine will give an error 3039 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3040 EVT Type = Op.getValueType(); 3041 int64_t Val = C->getSExtValue(); 3042 if (isInt<16>(Val)) { 3043 Result = DAG.getTargetConstant(Val, Type); 3044 break; 3045 } 3046 } 3047 return; 3048 case 'J': // integer zero 3049 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3050 EVT Type = Op.getValueType(); 3051 int64_t Val = C->getZExtValue(); 3052 if (Val == 0) { 3053 Result = DAG.getTargetConstant(0, Type); 3054 break; 3055 } 3056 } 3057 return; 3058 case 'K': // unsigned 16 bit immediate 3059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3060 EVT Type = Op.getValueType(); 3061 uint64_t Val = (uint64_t)C->getZExtValue(); 3062 if (isUInt<16>(Val)) { 3063 Result = DAG.getTargetConstant(Val, Type); 3064 break; 3065 } 3066 } 3067 return; 3068 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3069 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3070 EVT Type = Op.getValueType(); 3071 int64_t Val = C->getSExtValue(); 3072 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3073 Result = DAG.getTargetConstant(Val, Type); 3074 break; 3075 } 3076 } 3077 return; 3078 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3079 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3080 EVT Type = Op.getValueType(); 3081 int64_t Val = C->getSExtValue(); 3082 if ((Val >= -65535) && (Val <= -1)) { 3083 Result = DAG.getTargetConstant(Val, Type); 3084 break; 3085 } 3086 } 3087 return; 3088 case 'O': // signed 15 bit immediate 3089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3090 EVT Type = Op.getValueType(); 3091 int64_t Val = C->getSExtValue(); 3092 if ((isInt<15>(Val))) { 3093 Result = DAG.getTargetConstant(Val, Type); 3094 break; 3095 } 3096 } 3097 return; 3098 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3099 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3100 EVT Type = Op.getValueType(); 3101 int64_t Val = C->getSExtValue(); 3102 if ((Val <= 65535) && (Val >= 1)) { 3103 Result = DAG.getTargetConstant(Val, Type); 3104 break; 3105 } 3106 } 3107 return; 3108 } 3109 3110 if (Result.getNode()) { 3111 Ops.push_back(Result); 3112 return; 3113 } 3114 3115 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 3116} 3117 3118bool 3119MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const { 3120 // No global is ever allowed as a base. 3121 if (AM.BaseGV) 3122 return false; 3123 3124 switch (AM.Scale) { 3125 case 0: // "r+i" or just "i", depending on HasBaseReg. 3126 break; 3127 case 1: 3128 if (!AM.HasBaseReg) // allow "r+i". 3129 break; 3130 return false; // disallow "r+r" or "r+r+i". 3131 default: 3132 return false; 3133 } 3134 3135 return true; 3136} 3137 3138bool 3139MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 3140 // The Mips target isn't yet aware of offsets. 3141 return false; 3142} 3143 3144EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 3145 unsigned SrcAlign, 3146 bool IsMemset, bool ZeroMemset, 3147 bool MemcpyStrSrc, 3148 MachineFunction &MF) const { 3149 if (Subtarget->hasMips64()) 3150 return MVT::i64; 3151 3152 return MVT::i32; 3153} 3154 3155bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 3156 if (VT != MVT::f32 && VT != MVT::f64) 3157 return false; 3158 if (Imm.isNegZero()) 3159 return false; 3160 return Imm.isZero(); 3161} 3162 3163unsigned MipsTargetLowering::getJumpTableEncoding() const { 3164 if (IsN64) 3165 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 3166 3167 return TargetLowering::getJumpTableEncoding(); 3168} 3169 3170/// This function returns true if CallSym is a long double emulation routine. 3171static bool isF128SoftLibCall(const char *CallSym) { 3172 const char *const LibCalls[] = 3173 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2", 3174 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi", 3175 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf", 3176 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2", 3177 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3", 3178 "__trunctfdf2", "__trunctfsf2", "__unordtf2", 3179 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl", 3180 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl", 3181 "truncl"}; 3182 3183 const char * const *End = LibCalls + array_lengthof(LibCalls); 3184 3185 // Check that LibCalls is sorted alphabetically. 3186 MipsTargetLowering::LTStr Comp; 3187 3188#ifndef NDEBUG 3189 for (const char * const *I = LibCalls; I < End - 1; ++I) 3190 assert(Comp(*I, *(I + 1))); 3191#endif 3192 3193 return std::binary_search(LibCalls, End, CallSym, Comp); 3194} 3195 3196/// This function returns true if Ty is fp128 or i128 which was originally a 3197/// fp128. 3198static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) { 3199 if (Ty->isFP128Ty()) 3200 return true; 3201 3202 const ExternalSymbolSDNode *ES = 3203 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode); 3204 3205 // If the Ty is i128 and the function being called is a long double emulation 3206 // routine, then the original type is f128. 3207 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol())); 3208} 3209 3210MipsTargetLowering::MipsCC::MipsCC(CallingConv::ID CC, bool IsO32_, 3211 CCState &Info) 3212 : CCInfo(Info), CallConv(CC), IsO32(IsO32_) { 3213 // Pre-allocate reserved argument area. 3214 CCInfo.AllocateStack(reservedArgArea(), 1); 3215} 3216 3217void MipsTargetLowering::MipsCC:: 3218analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args, 3219 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode, 3220 std::vector<ArgListEntry> &FuncArgs) { 3221 assert((CallConv != CallingConv::Fast || !IsVarArg) && 3222 "CallingConv::Fast shouldn't be used for vararg functions."); 3223 3224 unsigned NumOpnds = Args.size(); 3225 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn(); 3226 3227 for (unsigned I = 0; I != NumOpnds; ++I) { 3228 MVT ArgVT = Args[I].VT; 3229 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3230 bool R; 3231 3232 if (ArgFlags.isByVal()) { 3233 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3234 continue; 3235 } 3236 3237 if (IsVarArg && !Args[I].IsFixed) 3238 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo); 3239 else { 3240 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode, 3241 IsSoftFloat); 3242 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo); 3243 } 3244 3245 if (R) { 3246#ifndef NDEBUG 3247 dbgs() << "Call operand #" << I << " has unhandled type " 3248 << EVT(ArgVT).getEVTString(); 3249#endif 3250 llvm_unreachable(0); 3251 } 3252 } 3253} 3254 3255void MipsTargetLowering::MipsCC:: 3256analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args, 3257 bool IsSoftFloat, Function::const_arg_iterator FuncArg) { 3258 unsigned NumArgs = Args.size(); 3259 llvm::CCAssignFn *FixedFn = fixedArgFn(); 3260 unsigned CurArgIdx = 0; 3261 3262 for (unsigned I = 0; I != NumArgs; ++I) { 3263 MVT ArgVT = Args[I].VT; 3264 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3265 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx); 3266 CurArgIdx = Args[I].OrigArgIndex; 3267 3268 if (ArgFlags.isByVal()) { 3269 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3270 continue; 3271 } 3272 3273 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat); 3274 3275 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo)) 3276 continue; 3277 3278#ifndef NDEBUG 3279 dbgs() << "Formal Arg #" << I << " has unhandled type " 3280 << EVT(ArgVT).getEVTString(); 3281#endif 3282 llvm_unreachable(0); 3283 } 3284} 3285 3286template<typename Ty> 3287void MipsTargetLowering::MipsCC:: 3288analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat, 3289 const SDNode *CallNode, const Type *RetTy) const { 3290 CCAssignFn *Fn; 3291 3292 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode)) 3293 Fn = RetCC_F128Soft; 3294 else 3295 Fn = RetCC_Mips; 3296 3297 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) { 3298 MVT VT = RetVals[I].VT; 3299 ISD::ArgFlagsTy Flags = RetVals[I].Flags; 3300 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat); 3301 3302 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) { 3303#ifndef NDEBUG 3304 dbgs() << "Call result #" << I << " has unhandled type " 3305 << EVT(VT).getEVTString() << '\n'; 3306#endif 3307 llvm_unreachable(0); 3308 } 3309 } 3310} 3311 3312void MipsTargetLowering::MipsCC:: 3313analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat, 3314 const SDNode *CallNode, const Type *RetTy) const { 3315 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy); 3316} 3317 3318void MipsTargetLowering::MipsCC:: 3319analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat, 3320 const Type *RetTy) const { 3321 analyzeReturn(Outs, IsSoftFloat, 0, RetTy); 3322} 3323 3324void 3325MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT, 3326 MVT LocVT, 3327 CCValAssign::LocInfo LocInfo, 3328 ISD::ArgFlagsTy ArgFlags) { 3329 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0."); 3330 3331 struct ByValArgInfo ByVal; 3332 unsigned RegSize = regSize(); 3333 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize); 3334 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize), 3335 RegSize * 2); 3336 3337 if (useRegsForByval()) 3338 allocateRegs(ByVal, ByValSize, Align); 3339 3340 // Allocate space on caller's stack. 3341 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs, 3342 Align); 3343 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT, 3344 LocInfo)); 3345 ByValArgs.push_back(ByVal); 3346} 3347 3348unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const { 3349 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs); 3350} 3351 3352unsigned MipsTargetLowering::MipsCC::reservedArgArea() const { 3353 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0; 3354} 3355 3356const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const { 3357 return IsO32 ? O32IntRegs : Mips64IntRegs; 3358} 3359 3360llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const { 3361 if (CallConv == CallingConv::Fast) 3362 return CC_Mips_FastCC; 3363 3364 return IsO32 ? CC_MipsO32 : CC_MipsN; 3365} 3366 3367llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const { 3368 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg; 3369} 3370 3371const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const { 3372 return IsO32 ? O32IntRegs : Mips64DPRegs; 3373} 3374 3375void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal, 3376 unsigned ByValSize, 3377 unsigned Align) { 3378 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs(); 3379 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs(); 3380 assert(!(ByValSize % RegSize) && !(Align % RegSize) && 3381 "Byval argument's size and alignment should be a multiple of" 3382 "RegSize."); 3383 3384 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs); 3385 3386 // If Align > RegSize, the first arg register must be even. 3387 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) { 3388 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]); 3389 ++ByVal.FirstIdx; 3390 } 3391 3392 // Mark the registers allocated. 3393 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs); 3394 ByValSize -= RegSize, ++I, ++ByVal.NumRegs) 3395 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]); 3396} 3397 3398MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy, 3399 const SDNode *CallNode, 3400 bool IsSoftFloat) const { 3401 if (IsSoftFloat || IsO32) 3402 return VT; 3403 3404 // Check if the original type was fp128. 3405 if (originalTypeIsF128(OrigTy, CallNode)) { 3406 assert(VT == MVT::i64); 3407 return MVT::f64; 3408 } 3409 3410 return VT; 3411} 3412 3413void MipsTargetLowering:: 3414copyByValRegs(SDValue Chain, DebugLoc DL, std::vector<SDValue> &OutChains, 3415 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 3416 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 3417 const MipsCC &CC, const ByValArgInfo &ByVal) const { 3418 MachineFunction &MF = DAG.getMachineFunction(); 3419 MachineFrameInfo *MFI = MF.getFrameInfo(); 3420 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize(); 3421 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 3422 int FrameObjOffset; 3423 3424 if (RegAreaSize) 3425 FrameObjOffset = (int)CC.reservedArgArea() - 3426 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize()); 3427 else 3428 FrameObjOffset = ByVal.Address; 3429 3430 // Create frame object. 3431 EVT PtrTy = getPointerTy(); 3432 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true); 3433 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 3434 InVals.push_back(FIN); 3435 3436 if (!ByVal.NumRegs) 3437 return; 3438 3439 // Copy arg registers. 3440 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8); 3441 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3442 3443 for (unsigned I = 0; I < ByVal.NumRegs; ++I) { 3444 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I]; 3445 unsigned VReg = addLiveIn(MF, ArgReg, RC); 3446 unsigned Offset = I * CC.regSize(); 3447 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 3448 DAG.getConstant(Offset, PtrTy)); 3449 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 3450 StorePtr, MachinePointerInfo(FuncArg, Offset), 3451 false, false, 0); 3452 OutChains.push_back(Store); 3453 } 3454} 3455 3456// Copy byVal arg to registers and stack. 3457void MipsTargetLowering:: 3458passByValArg(SDValue Chain, DebugLoc DL, 3459 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 3460 SmallVector<SDValue, 8> &MemOpChains, SDValue StackPtr, 3461 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg, 3462 const MipsCC &CC, const ByValArgInfo &ByVal, 3463 const ISD::ArgFlagsTy &Flags, bool isLittle) const { 3464 unsigned ByValSize = Flags.getByValSize(); 3465 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct. 3466 unsigned RegSize = CC.regSize(); 3467 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize); 3468 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8); 3469 3470 if (ByVal.NumRegs) { 3471 const uint16_t *ArgRegs = CC.intArgRegs(); 3472 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize); 3473 unsigned I = 0; 3474 3475 // Copy words to registers. 3476 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) { 3477 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3478 DAG.getConstant(Offset, PtrTy)); 3479 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 3480 MachinePointerInfo(), false, false, false, 3481 Alignment); 3482 MemOpChains.push_back(LoadVal.getValue(1)); 3483 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3484 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 3485 } 3486 3487 // Return if the struct has been fully copied. 3488 if (ByValSize == Offset) 3489 return; 3490 3491 // Copy the remainder of the byval argument with sub-word loads and shifts. 3492 if (LeftoverBytes) { 3493 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) && 3494 "Size of the remainder should be smaller than RegSize."); 3495 SDValue Val; 3496 3497 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0; 3498 Offset < ByValSize; LoadSize /= 2) { 3499 unsigned RemSize = ByValSize - Offset; 3500 3501 if (RemSize < LoadSize) 3502 continue; 3503 3504 // Load subword. 3505 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3506 DAG.getConstant(Offset, PtrTy)); 3507 SDValue LoadVal = 3508 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, 3509 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8), 3510 false, false, Alignment); 3511 MemOpChains.push_back(LoadVal.getValue(1)); 3512 3513 // Shift the loaded value. 3514 unsigned Shamt; 3515 3516 if (isLittle) 3517 Shamt = TotalSizeLoaded; 3518 else 3519 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8; 3520 3521 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 3522 DAG.getConstant(Shamt, MVT::i32)); 3523 3524 if (Val.getNode()) 3525 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 3526 else 3527 Val = Shift; 3528 3529 Offset += LoadSize; 3530 TotalSizeLoaded += LoadSize; 3531 Alignment = std::min(Alignment, LoadSize); 3532 } 3533 3534 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3535 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 3536 return; 3537 } 3538 } 3539 3540 // Copy remainder of byval arg to it with memcpy. 3541 unsigned MemCpySize = ByValSize - Offset; 3542 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3543 DAG.getConstant(Offset, PtrTy)); 3544 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 3545 DAG.getIntPtrConstant(ByVal.Address)); 3546 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 3547 DAG.getConstant(MemCpySize, PtrTy), Alignment, 3548 /*isVolatile=*/false, /*AlwaysInline=*/false, 3549 MachinePointerInfo(0), MachinePointerInfo(0)); 3550 MemOpChains.push_back(Chain); 3551} 3552 3553void 3554MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 3555 const MipsCC &CC, SDValue Chain, 3556 DebugLoc DL, SelectionDAG &DAG) const { 3557 unsigned NumRegs = CC.numIntArgRegs(); 3558 const uint16_t *ArgRegs = CC.intArgRegs(); 3559 const CCState &CCInfo = CC.getCCInfo(); 3560 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs); 3561 unsigned RegSize = CC.regSize(); 3562 MVT RegTy = MVT::getIntegerVT(RegSize * 8); 3563 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3564 MachineFunction &MF = DAG.getMachineFunction(); 3565 MachineFrameInfo *MFI = MF.getFrameInfo(); 3566 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3567 3568 // Offset of the first variable argument from stack pointer. 3569 int VaArgOffset; 3570 3571 if (NumRegs == Idx) 3572 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize); 3573 else 3574 VaArgOffset = 3575 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx)); 3576 3577 // Record the frame index of the first variable argument 3578 // which is a value necessary to VASTART. 3579 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3580 MipsFI->setVarArgsFrameIndex(FI); 3581 3582 // Copy the integer registers that have not been used for argument passing 3583 // to the argument register save area. For O32, the save area is allocated 3584 // in the caller's stack frame, while for N32/64, it is allocated in the 3585 // callee's stack frame. 3586 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) { 3587 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 3588 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 3589 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3590 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy()); 3591 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, 3592 MachinePointerInfo(), false, false, 0); 3593 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0); 3594 OutChains.push_back(Store); 3595 } 3596} 3597