MipsISelLowering.cpp revision bfb07b1054b653661306848e695b34e79289a15b
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/Support/CommandLine.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/ErrorHandling.h" 36#include "llvm/Support/raw_ostream.h" 37#include <cctype> 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 cl::opt<bool> 48NoZeroDivCheck("mno-check-zero-division", cl::Hidden, 49 cl::desc("MIPS: Don't trap on integer division by zero."), 50 cl::init(false)); 51 52static const uint16_t O32IntRegs[4] = { 53 Mips::A0, Mips::A1, Mips::A2, Mips::A3 54}; 55 56static const uint16_t Mips64IntRegs[8] = { 57 Mips::A0_64, Mips::A1_64, Mips::A2_64, Mips::A3_64, 58 Mips::T0_64, Mips::T1_64, Mips::T2_64, Mips::T3_64 59}; 60 61static const uint16_t Mips64DPRegs[8] = { 62 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, 63 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 64}; 65 66// If I is a shifted mask, set the size (Size) and the first bit of the 67// mask (Pos), and return true. 68// For example, if I is 0x003ff800, (Pos, Size) = (11, 11). 69static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) { 70 if (!isShiftedMask_64(I)) 71 return false; 72 73 Size = CountPopulation_64(I); 74 Pos = countTrailingZeros(I); 75 return true; 76} 77 78SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { 79 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>(); 80 return DAG.getRegister(FI->getGlobalBaseReg(), Ty); 81} 82 83static SDValue getTargetNode(SDValue Op, SelectionDAG &DAG, unsigned Flag) { 84 EVT Ty = Op.getValueType(); 85 86 if (GlobalAddressSDNode *N = dyn_cast<GlobalAddressSDNode>(Op)) 87 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(Op), Ty, 0, 88 Flag); 89 if (ExternalSymbolSDNode *N = dyn_cast<ExternalSymbolSDNode>(Op)) 90 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag); 91 if (BlockAddressSDNode *N = dyn_cast<BlockAddressSDNode>(Op)) 92 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 93 if (JumpTableSDNode *N = dyn_cast<JumpTableSDNode>(Op)) 94 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 95 if (ConstantPoolSDNode *N = dyn_cast<ConstantPoolSDNode>(Op)) 96 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(), 97 N->getOffset(), Flag); 98 99 llvm_unreachable("Unexpected node type."); 100 return SDValue(); 101} 102 103static SDValue getAddrNonPIC(SDValue Op, SelectionDAG &DAG) { 104 SDLoc DL(Op); 105 EVT Ty = Op.getValueType(); 106 SDValue Hi = getTargetNode(Op, DAG, MipsII::MO_ABS_HI); 107 SDValue Lo = getTargetNode(Op, DAG, MipsII::MO_ABS_LO); 108 return DAG.getNode(ISD::ADD, DL, Ty, 109 DAG.getNode(MipsISD::Hi, DL, Ty, Hi), 110 DAG.getNode(MipsISD::Lo, DL, Ty, Lo)); 111} 112 113SDValue MipsTargetLowering::getAddrLocal(SDValue Op, SelectionDAG &DAG, 114 bool HasMips64) const { 115 SDLoc DL(Op); 116 EVT Ty = Op.getValueType(); 117 unsigned GOTFlag = HasMips64 ? MipsII::MO_GOT_PAGE : MipsII::MO_GOT; 118 SDValue GOT = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty), 119 getTargetNode(Op, DAG, GOTFlag)); 120 SDValue Load = DAG.getLoad(Ty, DL, DAG.getEntryNode(), GOT, 121 MachinePointerInfo::getGOT(), false, false, false, 122 0); 123 unsigned LoFlag = HasMips64 ? MipsII::MO_GOT_OFST : MipsII::MO_ABS_LO; 124 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, Ty, getTargetNode(Op, DAG, LoFlag)); 125 return DAG.getNode(ISD::ADD, DL, Ty, Load, Lo); 126} 127 128SDValue MipsTargetLowering::getAddrGlobal(SDValue Op, SelectionDAG &DAG, 129 unsigned Flag) const { 130 SDLoc DL(Op); 131 EVT Ty = Op.getValueType(); 132 SDValue Tgt = DAG.getNode(MipsISD::Wrapper, DL, Ty, getGlobalReg(DAG, Ty), 133 getTargetNode(Op, DAG, Flag)); 134 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Tgt, 135 MachinePointerInfo::getGOT(), false, false, false, 0); 136} 137 138SDValue MipsTargetLowering::getAddrGlobalLargeGOT(SDValue Op, SelectionDAG &DAG, 139 unsigned HiFlag, 140 unsigned LoFlag) const { 141 SDLoc DL(Op); 142 EVT Ty = Op.getValueType(); 143 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, Ty, getTargetNode(Op, DAG, HiFlag)); 144 Hi = DAG.getNode(ISD::ADD, DL, Ty, Hi, getGlobalReg(DAG, Ty)); 145 SDValue Wrapper = DAG.getNode(MipsISD::Wrapper, DL, Ty, Hi, 146 getTargetNode(Op, DAG, LoFlag)); 147 return DAG.getLoad(Ty, DL, DAG.getEntryNode(), Wrapper, 148 MachinePointerInfo::getGOT(), false, false, false, 0); 149} 150 151const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { 152 switch (Opcode) { 153 case MipsISD::JmpLink: return "MipsISD::JmpLink"; 154 case MipsISD::TailCall: return "MipsISD::TailCall"; 155 case MipsISD::Hi: return "MipsISD::Hi"; 156 case MipsISD::Lo: return "MipsISD::Lo"; 157 case MipsISD::GPRel: return "MipsISD::GPRel"; 158 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer"; 159 case MipsISD::Ret: return "MipsISD::Ret"; 160 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN"; 161 case MipsISD::FPBrcond: return "MipsISD::FPBrcond"; 162 case MipsISD::FPCmp: return "MipsISD::FPCmp"; 163 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T"; 164 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F"; 165 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP"; 166 case MipsISD::ExtractLOHI: return "MipsISD::ExtractLOHI"; 167 case MipsISD::InsertLOHI: return "MipsISD::InsertLOHI"; 168 case MipsISD::Mult: return "MipsISD::Mult"; 169 case MipsISD::Multu: return "MipsISD::Multu"; 170 case MipsISD::MAdd: return "MipsISD::MAdd"; 171 case MipsISD::MAddu: return "MipsISD::MAddu"; 172 case MipsISD::MSub: return "MipsISD::MSub"; 173 case MipsISD::MSubu: return "MipsISD::MSubu"; 174 case MipsISD::DivRem: return "MipsISD::DivRem"; 175 case MipsISD::DivRemU: return "MipsISD::DivRemU"; 176 case MipsISD::DivRem16: return "MipsISD::DivRem16"; 177 case MipsISD::DivRemU16: return "MipsISD::DivRemU16"; 178 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64"; 179 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64"; 180 case MipsISD::Wrapper: return "MipsISD::Wrapper"; 181 case MipsISD::Sync: return "MipsISD::Sync"; 182 case MipsISD::Ext: return "MipsISD::Ext"; 183 case MipsISD::Ins: return "MipsISD::Ins"; 184 case MipsISD::LWL: return "MipsISD::LWL"; 185 case MipsISD::LWR: return "MipsISD::LWR"; 186 case MipsISD::SWL: return "MipsISD::SWL"; 187 case MipsISD::SWR: return "MipsISD::SWR"; 188 case MipsISD::LDL: return "MipsISD::LDL"; 189 case MipsISD::LDR: return "MipsISD::LDR"; 190 case MipsISD::SDL: return "MipsISD::SDL"; 191 case MipsISD::SDR: return "MipsISD::SDR"; 192 case MipsISD::EXTP: return "MipsISD::EXTP"; 193 case MipsISD::EXTPDP: return "MipsISD::EXTPDP"; 194 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H"; 195 case MipsISD::EXTR_W: return "MipsISD::EXTR_W"; 196 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W"; 197 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W"; 198 case MipsISD::SHILO: return "MipsISD::SHILO"; 199 case MipsISD::MTHLIP: return "MipsISD::MTHLIP"; 200 case MipsISD::MULT: return "MipsISD::MULT"; 201 case MipsISD::MULTU: return "MipsISD::MULTU"; 202 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP"; 203 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP"; 204 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP"; 205 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP"; 206 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP"; 207 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP"; 208 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP"; 209 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP"; 210 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP"; 211 default: return NULL; 212 } 213} 214 215MipsTargetLowering:: 216MipsTargetLowering(MipsTargetMachine &TM) 217 : TargetLowering(TM, new MipsTargetObjectFile()), 218 Subtarget(&TM.getSubtarget<MipsSubtarget>()), 219 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()), 220 IsO32(Subtarget->isABI_O32()) { 221 // Mips does not have i1 type, so use i32 for 222 // setcc operations results (slt, sgt, ...). 223 setBooleanContents(ZeroOrOneBooleanContent); 224 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 225 226 // Load extented operations for i1 types must be promoted 227 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 228 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 229 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 230 231 // MIPS doesn't have extending float->double load/store 232 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 233 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 234 235 // Used by legalize types to correctly generate the setcc result. 236 // Without this, every float setcc comes with a AND/OR with the result, 237 // we don't want this, since the fpcmp result goes to a flag register, 238 // which is used implicitly by brcond and select operations. 239 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 240 241 // Mips Custom Operations 242 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 243 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 244 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 245 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 246 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 247 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 248 setOperationAction(ISD::SELECT, MVT::f32, Custom); 249 setOperationAction(ISD::SELECT, MVT::f64, Custom); 250 setOperationAction(ISD::SELECT, MVT::i32, Custom); 251 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 252 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 253 setOperationAction(ISD::SETCC, MVT::f32, Custom); 254 setOperationAction(ISD::SETCC, MVT::f64, Custom); 255 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 256 setOperationAction(ISD::VASTART, MVT::Other, Custom); 257 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 258 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 259 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 260 261 if (!TM.Options.NoNaNsFPMath) { 262 setOperationAction(ISD::FABS, MVT::f32, Custom); 263 setOperationAction(ISD::FABS, MVT::f64, Custom); 264 } 265 266 if (HasMips64) { 267 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 268 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 269 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 270 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 271 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 272 setOperationAction(ISD::SELECT, MVT::i64, Custom); 273 setOperationAction(ISD::LOAD, MVT::i64, Custom); 274 setOperationAction(ISD::STORE, MVT::i64, Custom); 275 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 276 } 277 278 if (!HasMips64) { 279 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 280 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 281 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 282 } 283 284 setOperationAction(ISD::ADD, MVT::i32, Custom); 285 if (HasMips64) 286 setOperationAction(ISD::ADD, MVT::i64, Custom); 287 288 setOperationAction(ISD::SDIV, MVT::i32, Expand); 289 setOperationAction(ISD::SREM, MVT::i32, Expand); 290 setOperationAction(ISD::UDIV, MVT::i32, Expand); 291 setOperationAction(ISD::UREM, MVT::i32, Expand); 292 setOperationAction(ISD::SDIV, MVT::i64, Expand); 293 setOperationAction(ISD::SREM, MVT::i64, Expand); 294 setOperationAction(ISD::UDIV, MVT::i64, Expand); 295 setOperationAction(ISD::UREM, MVT::i64, Expand); 296 297 // Operations not directly supported by Mips. 298 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 299 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 300 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 301 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 302 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); 303 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 304 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 305 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 306 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 308 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 309 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 310 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 311 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 312 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 313 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); 314 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 315 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand); 316 setOperationAction(ISD::ROTL, MVT::i32, Expand); 317 setOperationAction(ISD::ROTL, MVT::i64, Expand); 318 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 319 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 320 321 if (!Subtarget->hasMips32r2()) 322 setOperationAction(ISD::ROTR, MVT::i32, Expand); 323 324 if (!Subtarget->hasMips64r2()) 325 setOperationAction(ISD::ROTR, MVT::i64, Expand); 326 327 setOperationAction(ISD::FSIN, MVT::f32, Expand); 328 setOperationAction(ISD::FSIN, MVT::f64, Expand); 329 setOperationAction(ISD::FCOS, MVT::f32, Expand); 330 setOperationAction(ISD::FCOS, MVT::f64, Expand); 331 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 332 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 333 setOperationAction(ISD::FPOWI, MVT::f32, Expand); 334 setOperationAction(ISD::FPOW, MVT::f32, Expand); 335 setOperationAction(ISD::FPOW, MVT::f64, Expand); 336 setOperationAction(ISD::FLOG, MVT::f32, Expand); 337 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 338 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 339 setOperationAction(ISD::FEXP, MVT::f32, Expand); 340 setOperationAction(ISD::FMA, MVT::f32, Expand); 341 setOperationAction(ISD::FMA, MVT::f64, Expand); 342 setOperationAction(ISD::FREM, MVT::f32, Expand); 343 setOperationAction(ISD::FREM, MVT::f64, Expand); 344 345 if (!TM.Options.NoNaNsFPMath) { 346 setOperationAction(ISD::FNEG, MVT::f32, Expand); 347 setOperationAction(ISD::FNEG, MVT::f64, Expand); 348 } 349 350 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 351 352 setOperationAction(ISD::VAARG, MVT::Other, Expand); 353 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 354 setOperationAction(ISD::VAEND, MVT::Other, Expand); 355 356 // Use the default for now 357 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 358 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 359 360 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand); 361 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 362 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand); 363 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 364 365 setInsertFencesForAtomic(true); 366 367 if (!Subtarget->hasSEInReg()) { 368 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 369 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 370 } 371 372 if (!Subtarget->hasBitCount()) { 373 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 374 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 375 } 376 377 if (!Subtarget->hasSwap()) { 378 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 379 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 380 } 381 382 if (HasMips64) { 383 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom); 384 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom); 385 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom); 386 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 387 } 388 389 setOperationAction(ISD::TRAP, MVT::Other, Legal); 390 391 setTargetDAGCombine(ISD::SDIVREM); 392 setTargetDAGCombine(ISD::UDIVREM); 393 setTargetDAGCombine(ISD::SELECT); 394 setTargetDAGCombine(ISD::AND); 395 setTargetDAGCombine(ISD::OR); 396 setTargetDAGCombine(ISD::ADD); 397 398 setMinFunctionAlignment(HasMips64 ? 3 : 2); 399 400 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP); 401 402 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0); 403 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1); 404 405 MaxStoresPerMemcpy = 16; 406} 407 408const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) { 409 if (TM.getSubtargetImpl()->inMips16Mode()) 410 return llvm::createMips16TargetLowering(TM); 411 412 return llvm::createMipsSETargetLowering(TM); 413} 414 415EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const { 416 if (!VT.isVector()) 417 return MVT::i32; 418 return VT.changeVectorElementTypeToInteger(); 419} 420 421static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 422 TargetLowering::DAGCombinerInfo &DCI, 423 const MipsSubtarget *Subtarget) { 424 if (DCI.isBeforeLegalizeOps()) 425 return SDValue(); 426 427 EVT Ty = N->getValueType(0); 428 unsigned LO = (Ty == MVT::i32) ? Mips::LO : Mips::LO64; 429 unsigned HI = (Ty == MVT::i32) ? Mips::HI : Mips::HI64; 430 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 431 MipsISD::DivRemU16; 432 SDLoc DL(N); 433 434 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 435 N->getOperand(0), N->getOperand(1)); 436 SDValue InChain = DAG.getEntryNode(); 437 SDValue InGlue = DivRem; 438 439 // insert MFLO 440 if (N->hasAnyUseOfValue(0)) { 441 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 442 InGlue); 443 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 444 InChain = CopyFromLo.getValue(1); 445 InGlue = CopyFromLo.getValue(2); 446 } 447 448 // insert MFHI 449 if (N->hasAnyUseOfValue(1)) { 450 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 451 HI, Ty, InGlue); 452 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 453 } 454 455 return SDValue(); 456} 457 458static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 459 switch (CC) { 460 default: llvm_unreachable("Unknown fp condition code!"); 461 case ISD::SETEQ: 462 case ISD::SETOEQ: return Mips::FCOND_OEQ; 463 case ISD::SETUNE: return Mips::FCOND_UNE; 464 case ISD::SETLT: 465 case ISD::SETOLT: return Mips::FCOND_OLT; 466 case ISD::SETGT: 467 case ISD::SETOGT: return Mips::FCOND_OGT; 468 case ISD::SETLE: 469 case ISD::SETOLE: return Mips::FCOND_OLE; 470 case ISD::SETGE: 471 case ISD::SETOGE: return Mips::FCOND_OGE; 472 case ISD::SETULT: return Mips::FCOND_ULT; 473 case ISD::SETULE: return Mips::FCOND_ULE; 474 case ISD::SETUGT: return Mips::FCOND_UGT; 475 case ISD::SETUGE: return Mips::FCOND_UGE; 476 case ISD::SETUO: return Mips::FCOND_UN; 477 case ISD::SETO: return Mips::FCOND_OR; 478 case ISD::SETNE: 479 case ISD::SETONE: return Mips::FCOND_ONE; 480 case ISD::SETUEQ: return Mips::FCOND_UEQ; 481 } 482} 483 484 485/// This function returns true if the floating point conditional branches and 486/// conditional moves which use condition code CC should be inverted. 487static bool invertFPCondCodeUser(Mips::CondCode CC) { 488 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 489 return false; 490 491 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 492 "Illegal Condition Code"); 493 494 return true; 495} 496 497// Creates and returns an FPCmp node from a setcc node. 498// Returns Op if setcc is not a floating point comparison. 499static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 500 // must be a SETCC node 501 if (Op.getOpcode() != ISD::SETCC) 502 return Op; 503 504 SDValue LHS = Op.getOperand(0); 505 506 if (!LHS.getValueType().isFloatingPoint()) 507 return Op; 508 509 SDValue RHS = Op.getOperand(1); 510 SDLoc DL(Op); 511 512 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 513 // node if necessary. 514 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 515 516 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 517 DAG.getConstant(condCodeToFCC(CC), MVT::i32)); 518} 519 520// Creates and returns a CMovFPT/F node. 521static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 522 SDValue False, SDLoc DL) { 523 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 524 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 525 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 526 527 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 528 True.getValueType(), True, FCC0, False, Cond); 529} 530 531static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 532 TargetLowering::DAGCombinerInfo &DCI, 533 const MipsSubtarget *Subtarget) { 534 if (DCI.isBeforeLegalizeOps()) 535 return SDValue(); 536 537 SDValue SetCC = N->getOperand(0); 538 539 if ((SetCC.getOpcode() != ISD::SETCC) || 540 !SetCC.getOperand(0).getValueType().isInteger()) 541 return SDValue(); 542 543 SDValue False = N->getOperand(2); 544 EVT FalseTy = False.getValueType(); 545 546 if (!FalseTy.isInteger()) 547 return SDValue(); 548 549 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False); 550 551 if (!CN || CN->getZExtValue()) 552 return SDValue(); 553 554 const SDLoc DL(N); 555 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 556 SDValue True = N->getOperand(1); 557 558 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 559 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 560 561 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 562} 563 564static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 565 TargetLowering::DAGCombinerInfo &DCI, 566 const MipsSubtarget *Subtarget) { 567 // Pattern match EXT. 568 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 569 // => ext $dst, $src, size, pos 570 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 571 return SDValue(); 572 573 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1); 574 unsigned ShiftRightOpc = ShiftRight.getOpcode(); 575 576 // Op's first operand must be a shift right. 577 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL) 578 return SDValue(); 579 580 // The second operand of the shift must be an immediate. 581 ConstantSDNode *CN; 582 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1)))) 583 return SDValue(); 584 585 uint64_t Pos = CN->getZExtValue(); 586 uint64_t SMPos, SMSize; 587 588 // Op's second operand must be a shifted mask. 589 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 590 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 591 return SDValue(); 592 593 // Return if the shifted mask does not start at bit 0 or the sum of its size 594 // and Pos exceeds the word's size. 595 EVT ValTy = N->getValueType(0); 596 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 597 return SDValue(); 598 599 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy, 600 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32), 601 DAG.getConstant(SMSize, MVT::i32)); 602} 603 604static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 605 TargetLowering::DAGCombinerInfo &DCI, 606 const MipsSubtarget *Subtarget) { 607 // Pattern match INS. 608 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 609 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 610 // => ins $dst, $src, size, pos, $src1 611 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 612 return SDValue(); 613 614 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 615 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 616 ConstantSDNode *CN; 617 618 // See if Op's first operand matches (and $src1 , mask0). 619 if (And0.getOpcode() != ISD::AND) 620 return SDValue(); 621 622 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 623 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 624 return SDValue(); 625 626 // See if Op's second operand matches (and (shl $src, pos), mask1). 627 if (And1.getOpcode() != ISD::AND) 628 return SDValue(); 629 630 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 631 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 632 return SDValue(); 633 634 // The shift masks must have the same position and size. 635 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 636 return SDValue(); 637 638 SDValue Shl = And1.getOperand(0); 639 if (Shl.getOpcode() != ISD::SHL) 640 return SDValue(); 641 642 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 643 return SDValue(); 644 645 unsigned Shamt = CN->getZExtValue(); 646 647 // Return if the shift amount and the first bit position of mask are not the 648 // same. 649 EVT ValTy = N->getValueType(0); 650 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 651 return SDValue(); 652 653 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0), 654 DAG.getConstant(SMPos0, MVT::i32), 655 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0)); 656} 657 658static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 659 TargetLowering::DAGCombinerInfo &DCI, 660 const MipsSubtarget *Subtarget) { 661 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 662 663 if (DCI.isBeforeLegalizeOps()) 664 return SDValue(); 665 666 SDValue Add = N->getOperand(1); 667 668 if (Add.getOpcode() != ISD::ADD) 669 return SDValue(); 670 671 SDValue Lo = Add.getOperand(1); 672 673 if ((Lo.getOpcode() != MipsISD::Lo) || 674 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 675 return SDValue(); 676 677 EVT ValTy = N->getValueType(0); 678 SDLoc DL(N); 679 680 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 681 Add.getOperand(0)); 682 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 683} 684 685SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 686 const { 687 SelectionDAG &DAG = DCI.DAG; 688 unsigned Opc = N->getOpcode(); 689 690 switch (Opc) { 691 default: break; 692 case ISD::SDIVREM: 693 case ISD::UDIVREM: 694 return performDivRemCombine(N, DAG, DCI, Subtarget); 695 case ISD::SELECT: 696 return performSELECTCombine(N, DAG, DCI, Subtarget); 697 case ISD::AND: 698 return performANDCombine(N, DAG, DCI, Subtarget); 699 case ISD::OR: 700 return performORCombine(N, DAG, DCI, Subtarget); 701 case ISD::ADD: 702 return performADDCombine(N, DAG, DCI, Subtarget); 703 } 704 705 return SDValue(); 706} 707 708void 709MipsTargetLowering::LowerOperationWrapper(SDNode *N, 710 SmallVectorImpl<SDValue> &Results, 711 SelectionDAG &DAG) const { 712 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 713 714 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 715 Results.push_back(Res.getValue(I)); 716} 717 718void 719MipsTargetLowering::ReplaceNodeResults(SDNode *N, 720 SmallVectorImpl<SDValue> &Results, 721 SelectionDAG &DAG) const { 722 return LowerOperationWrapper(N, Results, DAG); 723} 724 725SDValue MipsTargetLowering:: 726LowerOperation(SDValue Op, SelectionDAG &DAG) const 727{ 728 switch (Op.getOpcode()) 729 { 730 case ISD::BR_JT: return lowerBR_JT(Op, DAG); 731 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 732 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 733 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 734 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 735 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 736 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 737 case ISD::SELECT: return lowerSELECT(Op, DAG); 738 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG); 739 case ISD::SETCC: return lowerSETCC(Op, DAG); 740 case ISD::VASTART: return lowerVASTART(Op, DAG); 741 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 742 case ISD::FABS: return lowerFABS(Op, DAG); 743 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 744 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 745 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 746 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 747 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 748 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 749 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 750 case ISD::LOAD: return lowerLOAD(Op, DAG); 751 case ISD::STORE: return lowerSTORE(Op, DAG); 752 case ISD::ADD: return lowerADD(Op, DAG); 753 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 754 } 755 return SDValue(); 756} 757 758//===----------------------------------------------------------------------===// 759// Lower helper functions 760//===----------------------------------------------------------------------===// 761 762// addLiveIn - This helper function adds the specified physical register to the 763// MachineFunction as a live in value. It also creates a corresponding 764// virtual register for it. 765static unsigned 766addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 767{ 768 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 769 MF.getRegInfo().addLiveIn(PReg, VReg); 770 return VReg; 771} 772 773static MachineBasicBlock *expandPseudoDIV(MachineInstr *MI, 774 MachineBasicBlock &MBB, 775 const TargetInstrInfo &TII, 776 bool Is64Bit) { 777 if (NoZeroDivCheck) 778 return &MBB; 779 780 // Insert instruction "teq $divisor_reg, $zero, 7". 781 MachineBasicBlock::iterator I(MI); 782 MachineInstrBuilder MIB; 783 MIB = BuildMI(MBB, llvm::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ)) 784 .addOperand(MI->getOperand(2)).addReg(Mips::ZERO).addImm(7); 785 786 // Use the 32-bit sub-register if this is a 64-bit division. 787 if (Is64Bit) 788 MIB->getOperand(0).setSubReg(Mips::sub_32); 789 790 return &MBB; 791} 792 793MachineBasicBlock * 794MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 795 MachineBasicBlock *BB) const { 796 switch (MI->getOpcode()) { 797 default: 798 llvm_unreachable("Unexpected instr type to insert"); 799 case Mips::ATOMIC_LOAD_ADD_I8: 800 case Mips::ATOMIC_LOAD_ADD_I8_P8: 801 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 802 case Mips::ATOMIC_LOAD_ADD_I16: 803 case Mips::ATOMIC_LOAD_ADD_I16_P8: 804 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 805 case Mips::ATOMIC_LOAD_ADD_I32: 806 case Mips::ATOMIC_LOAD_ADD_I32_P8: 807 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 808 case Mips::ATOMIC_LOAD_ADD_I64: 809 case Mips::ATOMIC_LOAD_ADD_I64_P8: 810 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 811 812 case Mips::ATOMIC_LOAD_AND_I8: 813 case Mips::ATOMIC_LOAD_AND_I8_P8: 814 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 815 case Mips::ATOMIC_LOAD_AND_I16: 816 case Mips::ATOMIC_LOAD_AND_I16_P8: 817 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 818 case Mips::ATOMIC_LOAD_AND_I32: 819 case Mips::ATOMIC_LOAD_AND_I32_P8: 820 return emitAtomicBinary(MI, BB, 4, Mips::AND); 821 case Mips::ATOMIC_LOAD_AND_I64: 822 case Mips::ATOMIC_LOAD_AND_I64_P8: 823 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 824 825 case Mips::ATOMIC_LOAD_OR_I8: 826 case Mips::ATOMIC_LOAD_OR_I8_P8: 827 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 828 case Mips::ATOMIC_LOAD_OR_I16: 829 case Mips::ATOMIC_LOAD_OR_I16_P8: 830 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 831 case Mips::ATOMIC_LOAD_OR_I32: 832 case Mips::ATOMIC_LOAD_OR_I32_P8: 833 return emitAtomicBinary(MI, BB, 4, Mips::OR); 834 case Mips::ATOMIC_LOAD_OR_I64: 835 case Mips::ATOMIC_LOAD_OR_I64_P8: 836 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 837 838 case Mips::ATOMIC_LOAD_XOR_I8: 839 case Mips::ATOMIC_LOAD_XOR_I8_P8: 840 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 841 case Mips::ATOMIC_LOAD_XOR_I16: 842 case Mips::ATOMIC_LOAD_XOR_I16_P8: 843 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 844 case Mips::ATOMIC_LOAD_XOR_I32: 845 case Mips::ATOMIC_LOAD_XOR_I32_P8: 846 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 847 case Mips::ATOMIC_LOAD_XOR_I64: 848 case Mips::ATOMIC_LOAD_XOR_I64_P8: 849 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 850 851 case Mips::ATOMIC_LOAD_NAND_I8: 852 case Mips::ATOMIC_LOAD_NAND_I8_P8: 853 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 854 case Mips::ATOMIC_LOAD_NAND_I16: 855 case Mips::ATOMIC_LOAD_NAND_I16_P8: 856 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 857 case Mips::ATOMIC_LOAD_NAND_I32: 858 case Mips::ATOMIC_LOAD_NAND_I32_P8: 859 return emitAtomicBinary(MI, BB, 4, 0, true); 860 case Mips::ATOMIC_LOAD_NAND_I64: 861 case Mips::ATOMIC_LOAD_NAND_I64_P8: 862 return emitAtomicBinary(MI, BB, 8, 0, true); 863 864 case Mips::ATOMIC_LOAD_SUB_I8: 865 case Mips::ATOMIC_LOAD_SUB_I8_P8: 866 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 867 case Mips::ATOMIC_LOAD_SUB_I16: 868 case Mips::ATOMIC_LOAD_SUB_I16_P8: 869 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 870 case Mips::ATOMIC_LOAD_SUB_I32: 871 case Mips::ATOMIC_LOAD_SUB_I32_P8: 872 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 873 case Mips::ATOMIC_LOAD_SUB_I64: 874 case Mips::ATOMIC_LOAD_SUB_I64_P8: 875 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 876 877 case Mips::ATOMIC_SWAP_I8: 878 case Mips::ATOMIC_SWAP_I8_P8: 879 return emitAtomicBinaryPartword(MI, BB, 1, 0); 880 case Mips::ATOMIC_SWAP_I16: 881 case Mips::ATOMIC_SWAP_I16_P8: 882 return emitAtomicBinaryPartword(MI, BB, 2, 0); 883 case Mips::ATOMIC_SWAP_I32: 884 case Mips::ATOMIC_SWAP_I32_P8: 885 return emitAtomicBinary(MI, BB, 4, 0); 886 case Mips::ATOMIC_SWAP_I64: 887 case Mips::ATOMIC_SWAP_I64_P8: 888 return emitAtomicBinary(MI, BB, 8, 0); 889 890 case Mips::ATOMIC_CMP_SWAP_I8: 891 case Mips::ATOMIC_CMP_SWAP_I8_P8: 892 return emitAtomicCmpSwapPartword(MI, BB, 1); 893 case Mips::ATOMIC_CMP_SWAP_I16: 894 case Mips::ATOMIC_CMP_SWAP_I16_P8: 895 return emitAtomicCmpSwapPartword(MI, BB, 2); 896 case Mips::ATOMIC_CMP_SWAP_I32: 897 case Mips::ATOMIC_CMP_SWAP_I32_P8: 898 return emitAtomicCmpSwap(MI, BB, 4); 899 case Mips::ATOMIC_CMP_SWAP_I64: 900 case Mips::ATOMIC_CMP_SWAP_I64_P8: 901 return emitAtomicCmpSwap(MI, BB, 8); 902 case Mips::PseudoSDIV: 903 case Mips::PseudoUDIV: 904 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), false); 905 case Mips::PseudoDSDIV: 906 case Mips::PseudoDUDIV: 907 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), true); 908 } 909} 910 911// This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 912// Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 913MachineBasicBlock * 914MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 915 unsigned Size, unsigned BinOpcode, 916 bool Nand) const { 917 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 918 919 MachineFunction *MF = BB->getParent(); 920 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 921 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 922 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 923 DebugLoc DL = MI->getDebugLoc(); 924 unsigned LL, SC, AND, NOR, ZERO, BEQ; 925 926 if (Size == 4) { 927 LL = IsN64 ? Mips::LL_P8 : Mips::LL; 928 SC = IsN64 ? Mips::SC_P8 : Mips::SC; 929 AND = Mips::AND; 930 NOR = Mips::NOR; 931 ZERO = Mips::ZERO; 932 BEQ = Mips::BEQ; 933 } 934 else { 935 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD; 936 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD; 937 AND = Mips::AND64; 938 NOR = Mips::NOR64; 939 ZERO = Mips::ZERO_64; 940 BEQ = Mips::BEQ64; 941 } 942 943 unsigned OldVal = MI->getOperand(0).getReg(); 944 unsigned Ptr = MI->getOperand(1).getReg(); 945 unsigned Incr = MI->getOperand(2).getReg(); 946 947 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 948 unsigned AndRes = RegInfo.createVirtualRegister(RC); 949 unsigned Success = RegInfo.createVirtualRegister(RC); 950 951 // insert new blocks after the current block 952 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 953 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 954 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 955 MachineFunction::iterator It = BB; 956 ++It; 957 MF->insert(It, loopMBB); 958 MF->insert(It, exitMBB); 959 960 // Transfer the remainder of BB and its successor edges to exitMBB. 961 exitMBB->splice(exitMBB->begin(), BB, 962 llvm::next(MachineBasicBlock::iterator(MI)), 963 BB->end()); 964 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 965 966 // thisMBB: 967 // ... 968 // fallthrough --> loopMBB 969 BB->addSuccessor(loopMBB); 970 loopMBB->addSuccessor(loopMBB); 971 loopMBB->addSuccessor(exitMBB); 972 973 // loopMBB: 974 // ll oldval, 0(ptr) 975 // <binop> storeval, oldval, incr 976 // sc success, storeval, 0(ptr) 977 // beq success, $0, loopMBB 978 BB = loopMBB; 979 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 980 if (Nand) { 981 // and andres, oldval, incr 982 // nor storeval, $0, andres 983 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 984 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 985 } else if (BinOpcode) { 986 // <binop> storeval, oldval, incr 987 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 988 } else { 989 StoreVal = Incr; 990 } 991 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 992 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 993 994 MI->eraseFromParent(); // The instruction is gone now. 995 996 return exitMBB; 997} 998 999MachineBasicBlock * 1000MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI, 1001 MachineBasicBlock *BB, 1002 unsigned Size, unsigned BinOpcode, 1003 bool Nand) const { 1004 assert((Size == 1 || Size == 2) && 1005 "Unsupported size for EmitAtomicBinaryPartial."); 1006 1007 MachineFunction *MF = BB->getParent(); 1008 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1009 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1010 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1011 DebugLoc DL = MI->getDebugLoc(); 1012 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL; 1013 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC; 1014 1015 unsigned Dest = MI->getOperand(0).getReg(); 1016 unsigned Ptr = MI->getOperand(1).getReg(); 1017 unsigned Incr = MI->getOperand(2).getReg(); 1018 1019 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1020 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1021 unsigned Mask = RegInfo.createVirtualRegister(RC); 1022 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1023 unsigned NewVal = RegInfo.createVirtualRegister(RC); 1024 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1025 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 1026 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1027 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1028 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1029 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1030 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1031 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1032 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1033 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1034 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1035 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1036 unsigned Success = RegInfo.createVirtualRegister(RC); 1037 1038 // insert new blocks after the current block 1039 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1040 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1041 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1042 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1043 MachineFunction::iterator It = BB; 1044 ++It; 1045 MF->insert(It, loopMBB); 1046 MF->insert(It, sinkMBB); 1047 MF->insert(It, exitMBB); 1048 1049 // Transfer the remainder of BB and its successor edges to exitMBB. 1050 exitMBB->splice(exitMBB->begin(), BB, 1051 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1052 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1053 1054 BB->addSuccessor(loopMBB); 1055 loopMBB->addSuccessor(loopMBB); 1056 loopMBB->addSuccessor(sinkMBB); 1057 sinkMBB->addSuccessor(exitMBB); 1058 1059 // thisMBB: 1060 // addiu masklsb2,$0,-4 # 0xfffffffc 1061 // and alignedaddr,ptr,masklsb2 1062 // andi ptrlsb2,ptr,3 1063 // sll shiftamt,ptrlsb2,3 1064 // ori maskupper,$0,255 # 0xff 1065 // sll mask,maskupper,shiftamt 1066 // nor mask2,$0,mask 1067 // sll incr2,incr,shiftamt 1068 1069 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1070 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1071 .addReg(Mips::ZERO).addImm(-4); 1072 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1073 .addReg(Ptr).addReg(MaskLSB2); 1074 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1075 if (Subtarget->isLittle()) { 1076 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1077 } else { 1078 unsigned Off = RegInfo.createVirtualRegister(RC); 1079 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1080 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1081 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1082 } 1083 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1084 .addReg(Mips::ZERO).addImm(MaskImm); 1085 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1086 .addReg(MaskUpper).addReg(ShiftAmt); 1087 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1088 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1089 1090 // atomic.load.binop 1091 // loopMBB: 1092 // ll oldval,0(alignedaddr) 1093 // binop binopres,oldval,incr2 1094 // and newval,binopres,mask 1095 // and maskedoldval0,oldval,mask2 1096 // or storeval,maskedoldval0,newval 1097 // sc success,storeval,0(alignedaddr) 1098 // beq success,$0,loopMBB 1099 1100 // atomic.swap 1101 // loopMBB: 1102 // ll oldval,0(alignedaddr) 1103 // and newval,incr2,mask 1104 // and maskedoldval0,oldval,mask2 1105 // or storeval,maskedoldval0,newval 1106 // sc success,storeval,0(alignedaddr) 1107 // beq success,$0,loopMBB 1108 1109 BB = loopMBB; 1110 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1111 if (Nand) { 1112 // and andres, oldval, incr2 1113 // nor binopres, $0, andres 1114 // and newval, binopres, mask 1115 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1116 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1117 .addReg(Mips::ZERO).addReg(AndRes); 1118 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1119 } else if (BinOpcode) { 1120 // <binop> binopres, oldval, incr2 1121 // and newval, binopres, mask 1122 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1123 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1124 } else {// atomic.swap 1125 // and newval, incr2, mask 1126 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1127 } 1128 1129 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1130 .addReg(OldVal).addReg(Mask2); 1131 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1132 .addReg(MaskedOldVal0).addReg(NewVal); 1133 BuildMI(BB, DL, TII->get(SC), Success) 1134 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1135 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1136 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1137 1138 // sinkMBB: 1139 // and maskedoldval1,oldval,mask 1140 // srl srlres,maskedoldval1,shiftamt 1141 // sll sllres,srlres,24 1142 // sra dest,sllres,24 1143 BB = sinkMBB; 1144 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1145 1146 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1147 .addReg(OldVal).addReg(Mask); 1148 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1149 .addReg(MaskedOldVal1).addReg(ShiftAmt); 1150 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1151 .addReg(SrlRes).addImm(ShiftImm); 1152 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1153 .addReg(SllRes).addImm(ShiftImm); 1154 1155 MI->eraseFromParent(); // The instruction is gone now. 1156 1157 return exitMBB; 1158} 1159 1160MachineBasicBlock * 1161MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI, 1162 MachineBasicBlock *BB, 1163 unsigned Size) const { 1164 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1165 1166 MachineFunction *MF = BB->getParent(); 1167 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1168 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1169 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1170 DebugLoc DL = MI->getDebugLoc(); 1171 unsigned LL, SC, ZERO, BNE, BEQ; 1172 1173 if (Size == 4) { 1174 LL = IsN64 ? Mips::LL_P8 : Mips::LL; 1175 SC = IsN64 ? Mips::SC_P8 : Mips::SC; 1176 ZERO = Mips::ZERO; 1177 BNE = Mips::BNE; 1178 BEQ = Mips::BEQ; 1179 } 1180 else { 1181 LL = IsN64 ? Mips::LLD_P8 : Mips::LLD; 1182 SC = IsN64 ? Mips::SCD_P8 : Mips::SCD; 1183 ZERO = Mips::ZERO_64; 1184 BNE = Mips::BNE64; 1185 BEQ = Mips::BEQ64; 1186 } 1187 1188 unsigned Dest = MI->getOperand(0).getReg(); 1189 unsigned Ptr = MI->getOperand(1).getReg(); 1190 unsigned OldVal = MI->getOperand(2).getReg(); 1191 unsigned NewVal = MI->getOperand(3).getReg(); 1192 1193 unsigned Success = RegInfo.createVirtualRegister(RC); 1194 1195 // insert new blocks after the current block 1196 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1197 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1198 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1199 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1200 MachineFunction::iterator It = BB; 1201 ++It; 1202 MF->insert(It, loop1MBB); 1203 MF->insert(It, loop2MBB); 1204 MF->insert(It, exitMBB); 1205 1206 // Transfer the remainder of BB and its successor edges to exitMBB. 1207 exitMBB->splice(exitMBB->begin(), BB, 1208 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1209 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1210 1211 // thisMBB: 1212 // ... 1213 // fallthrough --> loop1MBB 1214 BB->addSuccessor(loop1MBB); 1215 loop1MBB->addSuccessor(exitMBB); 1216 loop1MBB->addSuccessor(loop2MBB); 1217 loop2MBB->addSuccessor(loop1MBB); 1218 loop2MBB->addSuccessor(exitMBB); 1219 1220 // loop1MBB: 1221 // ll dest, 0(ptr) 1222 // bne dest, oldval, exitMBB 1223 BB = loop1MBB; 1224 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1225 BuildMI(BB, DL, TII->get(BNE)) 1226 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1227 1228 // loop2MBB: 1229 // sc success, newval, 0(ptr) 1230 // beq success, $0, loop1MBB 1231 BB = loop2MBB; 1232 BuildMI(BB, DL, TII->get(SC), Success) 1233 .addReg(NewVal).addReg(Ptr).addImm(0); 1234 BuildMI(BB, DL, TII->get(BEQ)) 1235 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1236 1237 MI->eraseFromParent(); // The instruction is gone now. 1238 1239 return exitMBB; 1240} 1241 1242MachineBasicBlock * 1243MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI, 1244 MachineBasicBlock *BB, 1245 unsigned Size) const { 1246 assert((Size == 1 || Size == 2) && 1247 "Unsupported size for EmitAtomicCmpSwapPartial."); 1248 1249 MachineFunction *MF = BB->getParent(); 1250 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1251 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1252 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1253 DebugLoc DL = MI->getDebugLoc(); 1254 unsigned LL = IsN64 ? Mips::LL_P8 : Mips::LL; 1255 unsigned SC = IsN64 ? Mips::SC_P8 : Mips::SC; 1256 1257 unsigned Dest = MI->getOperand(0).getReg(); 1258 unsigned Ptr = MI->getOperand(1).getReg(); 1259 unsigned CmpVal = MI->getOperand(2).getReg(); 1260 unsigned NewVal = MI->getOperand(3).getReg(); 1261 1262 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1263 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1264 unsigned Mask = RegInfo.createVirtualRegister(RC); 1265 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1266 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1267 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1268 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1269 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1270 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1271 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1272 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1273 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1274 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1275 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1276 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1277 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1278 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1279 unsigned Success = RegInfo.createVirtualRegister(RC); 1280 1281 // insert new blocks after the current block 1282 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1283 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1284 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1285 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1286 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1287 MachineFunction::iterator It = BB; 1288 ++It; 1289 MF->insert(It, loop1MBB); 1290 MF->insert(It, loop2MBB); 1291 MF->insert(It, sinkMBB); 1292 MF->insert(It, exitMBB); 1293 1294 // Transfer the remainder of BB and its successor edges to exitMBB. 1295 exitMBB->splice(exitMBB->begin(), BB, 1296 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1297 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1298 1299 BB->addSuccessor(loop1MBB); 1300 loop1MBB->addSuccessor(sinkMBB); 1301 loop1MBB->addSuccessor(loop2MBB); 1302 loop2MBB->addSuccessor(loop1MBB); 1303 loop2MBB->addSuccessor(sinkMBB); 1304 sinkMBB->addSuccessor(exitMBB); 1305 1306 // FIXME: computation of newval2 can be moved to loop2MBB. 1307 // thisMBB: 1308 // addiu masklsb2,$0,-4 # 0xfffffffc 1309 // and alignedaddr,ptr,masklsb2 1310 // andi ptrlsb2,ptr,3 1311 // sll shiftamt,ptrlsb2,3 1312 // ori maskupper,$0,255 # 0xff 1313 // sll mask,maskupper,shiftamt 1314 // nor mask2,$0,mask 1315 // andi maskedcmpval,cmpval,255 1316 // sll shiftedcmpval,maskedcmpval,shiftamt 1317 // andi maskednewval,newval,255 1318 // sll shiftednewval,maskednewval,shiftamt 1319 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1320 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1321 .addReg(Mips::ZERO).addImm(-4); 1322 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1323 .addReg(Ptr).addReg(MaskLSB2); 1324 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1325 if (Subtarget->isLittle()) { 1326 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1327 } else { 1328 unsigned Off = RegInfo.createVirtualRegister(RC); 1329 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1330 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1331 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1332 } 1333 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1334 .addReg(Mips::ZERO).addImm(MaskImm); 1335 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1336 .addReg(MaskUpper).addReg(ShiftAmt); 1337 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1338 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1339 .addReg(CmpVal).addImm(MaskImm); 1340 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1341 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1342 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1343 .addReg(NewVal).addImm(MaskImm); 1344 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1345 .addReg(MaskedNewVal).addReg(ShiftAmt); 1346 1347 // loop1MBB: 1348 // ll oldval,0(alginedaddr) 1349 // and maskedoldval0,oldval,mask 1350 // bne maskedoldval0,shiftedcmpval,sinkMBB 1351 BB = loop1MBB; 1352 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0); 1353 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1354 .addReg(OldVal).addReg(Mask); 1355 BuildMI(BB, DL, TII->get(Mips::BNE)) 1356 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1357 1358 // loop2MBB: 1359 // and maskedoldval1,oldval,mask2 1360 // or storeval,maskedoldval1,shiftednewval 1361 // sc success,storeval,0(alignedaddr) 1362 // beq success,$0,loop1MBB 1363 BB = loop2MBB; 1364 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1365 .addReg(OldVal).addReg(Mask2); 1366 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1367 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1368 BuildMI(BB, DL, TII->get(SC), Success) 1369 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1370 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1371 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1372 1373 // sinkMBB: 1374 // srl srlres,maskedoldval0,shiftamt 1375 // sll sllres,srlres,24 1376 // sra dest,sllres,24 1377 BB = sinkMBB; 1378 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1379 1380 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1381 .addReg(MaskedOldVal0).addReg(ShiftAmt); 1382 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1383 .addReg(SrlRes).addImm(ShiftImm); 1384 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1385 .addReg(SllRes).addImm(ShiftImm); 1386 1387 MI->eraseFromParent(); // The instruction is gone now. 1388 1389 return exitMBB; 1390} 1391 1392//===----------------------------------------------------------------------===// 1393// Misc Lower Operation implementation 1394//===----------------------------------------------------------------------===// 1395SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 1396 SDValue Chain = Op.getOperand(0); 1397 SDValue Table = Op.getOperand(1); 1398 SDValue Index = Op.getOperand(2); 1399 SDLoc DL(Op); 1400 EVT PTy = getPointerTy(); 1401 unsigned EntrySize = 1402 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout()); 1403 1404 Index = DAG.getNode(ISD::MUL, DL, PTy, Index, 1405 DAG.getConstant(EntrySize, PTy)); 1406 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table); 1407 1408 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); 1409 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr, 1410 MachinePointerInfo::getJumpTable(), MemVT, false, false, 1411 0); 1412 Chain = Addr.getValue(1); 1413 1414 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) { 1415 // For PIC, the sequence is: 1416 // BRIND(load(Jumptable + index) + RelocBase) 1417 // RelocBase can be JumpTable, GOT or some sort of global base. 1418 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr, 1419 getPICJumpTableRelocBase(Table, DAG)); 1420 } 1421 1422 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr); 1423} 1424 1425SDValue MipsTargetLowering:: 1426lowerBRCOND(SDValue Op, SelectionDAG &DAG) const 1427{ 1428 // The first operand is the chain, the second is the condition, the third is 1429 // the block to branch to if the condition is true. 1430 SDValue Chain = Op.getOperand(0); 1431 SDValue Dest = Op.getOperand(2); 1432 SDLoc DL(Op); 1433 1434 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1435 1436 // Return if flag is not set by a floating point comparison. 1437 if (CondRes.getOpcode() != MipsISD::FPCmp) 1438 return Op; 1439 1440 SDValue CCNode = CondRes.getOperand(2); 1441 Mips::CondCode CC = 1442 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1443 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1444 SDValue BrCode = DAG.getConstant(Opc, MVT::i32); 1445 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 1446 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1447 FCC0, Dest, CondRes); 1448} 1449 1450SDValue MipsTargetLowering:: 1451lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1452{ 1453 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1454 1455 // Return if flag is not set by a floating point comparison. 1456 if (Cond.getOpcode() != MipsISD::FPCmp) 1457 return Op; 1458 1459 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1460 SDLoc(Op)); 1461} 1462 1463SDValue MipsTargetLowering:: 1464lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 1465{ 1466 SDLoc DL(Op); 1467 EVT Ty = Op.getOperand(0).getValueType(); 1468 SDValue Cond = DAG.getNode(ISD::SETCC, DL, 1469 getSetCCResultType(*DAG.getContext(), Ty), 1470 Op.getOperand(0), Op.getOperand(1), 1471 Op.getOperand(4)); 1472 1473 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2), 1474 Op.getOperand(3)); 1475} 1476 1477SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 1478 SDValue Cond = createFPCmp(DAG, Op); 1479 1480 assert(Cond.getOpcode() == MipsISD::FPCmp && 1481 "Floating point operand expected."); 1482 1483 SDValue True = DAG.getConstant(1, MVT::i32); 1484 SDValue False = DAG.getConstant(0, MVT::i32); 1485 1486 return createCMovFP(DAG, Cond, True, False, SDLoc(Op)); 1487} 1488 1489SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 1490 SelectionDAG &DAG) const { 1491 // FIXME there isn't actually debug info here 1492 SDLoc DL(Op); 1493 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1494 1495 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) { 1496 const MipsTargetObjectFile &TLOF = 1497 (const MipsTargetObjectFile&)getObjFileLowering(); 1498 1499 // %gp_rel relocation 1500 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) { 1501 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0, 1502 MipsII::MO_GPREL); 1503 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL, 1504 DAG.getVTList(MVT::i32), &GA, 1); 1505 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32); 1506 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode); 1507 } 1508 1509 // %hi/%lo relocation 1510 return getAddrNonPIC(Op, DAG); 1511 } 1512 1513 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))) 1514 return getAddrLocal(Op, DAG, HasMips64); 1515 1516 if (LargeGOT) 1517 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16, 1518 MipsII::MO_GOT_LO16); 1519 1520 return getAddrGlobal(Op, DAG, 1521 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16); 1522} 1523 1524SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 1525 SelectionDAG &DAG) const { 1526 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1527 return getAddrNonPIC(Op, DAG); 1528 1529 return getAddrLocal(Op, DAG, HasMips64); 1530} 1531 1532SDValue MipsTargetLowering:: 1533lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 1534{ 1535 // If the relocation model is PIC, use the General Dynamic TLS Model or 1536 // Local Dynamic TLS model, otherwise use the Initial Exec or 1537 // Local Exec TLS Model. 1538 1539 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1540 SDLoc DL(GA); 1541 const GlobalValue *GV = GA->getGlobal(); 1542 EVT PtrVT = getPointerTy(); 1543 1544 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 1545 1546 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 1547 // General Dynamic and Local Dynamic TLS Model. 1548 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 1549 : MipsII::MO_TLSGD; 1550 1551 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 1552 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 1553 getGlobalReg(DAG, PtrVT), TGA); 1554 unsigned PtrSize = PtrVT.getSizeInBits(); 1555 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 1556 1557 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 1558 1559 ArgListTy Args; 1560 ArgListEntry Entry; 1561 Entry.Node = Argument; 1562 Entry.Ty = PtrTy; 1563 Args.push_back(Entry); 1564 1565 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy, 1566 false, false, false, false, 0, CallingConv::C, 1567 /*IsTailCall=*/false, /*doesNotRet=*/false, 1568 /*isReturnValueUsed=*/true, 1569 TlsGetAddr, Args, DAG, DL); 1570 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 1571 1572 SDValue Ret = CallResult.first; 1573 1574 if (model != TLSModel::LocalDynamic) 1575 return Ret; 1576 1577 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1578 MipsII::MO_DTPREL_HI); 1579 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1580 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1581 MipsII::MO_DTPREL_LO); 1582 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1583 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 1584 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 1585 } 1586 1587 SDValue Offset; 1588 if (model == TLSModel::InitialExec) { 1589 // Initial Exec TLS Model 1590 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1591 MipsII::MO_GOTTPREL); 1592 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 1593 TGA); 1594 Offset = DAG.getLoad(PtrVT, DL, 1595 DAG.getEntryNode(), TGA, MachinePointerInfo(), 1596 false, false, false, 0); 1597 } else { 1598 // Local Exec TLS Model 1599 assert(model == TLSModel::LocalExec); 1600 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1601 MipsII::MO_TPREL_HI); 1602 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1603 MipsII::MO_TPREL_LO); 1604 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1605 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1606 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 1607 } 1608 1609 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 1610 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 1611} 1612 1613SDValue MipsTargetLowering:: 1614lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 1615{ 1616 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1617 return getAddrNonPIC(Op, DAG); 1618 1619 return getAddrLocal(Op, DAG, HasMips64); 1620} 1621 1622SDValue MipsTargetLowering:: 1623lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 1624{ 1625 // gp_rel relocation 1626 // FIXME: we should reference the constant pool using small data sections, 1627 // but the asm printer currently doesn't support this feature without 1628 // hacking it. This feature should come soon so we can uncomment the 1629 // stuff below. 1630 //if (IsInSmallSection(C->getType())) { 1631 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP); 1632 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32); 1633 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode); 1634 1635 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1636 return getAddrNonPIC(Op, DAG); 1637 1638 return getAddrLocal(Op, DAG, HasMips64); 1639} 1640 1641SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 1642 MachineFunction &MF = DAG.getMachineFunction(); 1643 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 1644 1645 SDLoc DL(Op); 1646 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 1647 getPointerTy()); 1648 1649 // vastart just stores the address of the VarArgsFrameIndex slot into the 1650 // memory location argument. 1651 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1652 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 1653 MachinePointerInfo(SV), false, false, 0); 1654} 1655 1656static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1657 EVT TyX = Op.getOperand(0).getValueType(); 1658 EVT TyY = Op.getOperand(1).getValueType(); 1659 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1660 SDValue Const31 = DAG.getConstant(31, MVT::i32); 1661 SDLoc DL(Op); 1662 SDValue Res; 1663 1664 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1665 // to i32. 1666 SDValue X = (TyX == MVT::f32) ? 1667 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1668 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1669 Const1); 1670 SDValue Y = (TyY == MVT::f32) ? 1671 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 1672 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 1673 Const1); 1674 1675 if (HasR2) { 1676 // ext E, Y, 31, 1 ; extract bit31 of Y 1677 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 1678 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 1679 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 1680 } else { 1681 // sll SllX, X, 1 1682 // srl SrlX, SllX, 1 1683 // srl SrlY, Y, 31 1684 // sll SllY, SrlX, 31 1685 // or Or, SrlX, SllY 1686 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1687 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1688 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 1689 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 1690 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 1691 } 1692 1693 if (TyX == MVT::f32) 1694 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 1695 1696 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1697 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1698 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1699} 1700 1701static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1702 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 1703 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 1704 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 1705 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1706 SDLoc DL(Op); 1707 1708 // Bitcast to integer nodes. 1709 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 1710 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 1711 1712 if (HasR2) { 1713 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 1714 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 1715 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 1716 DAG.getConstant(WidthY - 1, MVT::i32), Const1); 1717 1718 if (WidthX > WidthY) 1719 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 1720 else if (WidthY > WidthX) 1721 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 1722 1723 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 1724 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X); 1725 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 1726 } 1727 1728 // (d)sll SllX, X, 1 1729 // (d)srl SrlX, SllX, 1 1730 // (d)srl SrlY, Y, width(Y)-1 1731 // (d)sll SllY, SrlX, width(Y)-1 1732 // or Or, SrlX, SllY 1733 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 1734 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 1735 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 1736 DAG.getConstant(WidthY - 1, MVT::i32)); 1737 1738 if (WidthX > WidthY) 1739 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 1740 else if (WidthY > WidthX) 1741 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 1742 1743 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 1744 DAG.getConstant(WidthX - 1, MVT::i32)); 1745 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 1746 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 1747} 1748 1749SDValue 1750MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 1751 if (Subtarget->hasMips64()) 1752 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2()); 1753 1754 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2()); 1755} 1756 1757static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1758 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1759 SDLoc DL(Op); 1760 1761 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1762 // to i32. 1763 SDValue X = (Op.getValueType() == MVT::f32) ? 1764 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1765 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1766 Const1); 1767 1768 // Clear MSB. 1769 if (HasR2) 1770 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, 1771 DAG.getRegister(Mips::ZERO, MVT::i32), 1772 DAG.getConstant(31, MVT::i32), Const1, X); 1773 else { 1774 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1775 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1776 } 1777 1778 if (Op.getValueType() == MVT::f32) 1779 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res); 1780 1781 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1782 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1783 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1784} 1785 1786static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1787 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1788 SDLoc DL(Op); 1789 1790 // Bitcast to integer node. 1791 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0)); 1792 1793 // Clear MSB. 1794 if (HasR2) 1795 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64, 1796 DAG.getRegister(Mips::ZERO_64, MVT::i64), 1797 DAG.getConstant(63, MVT::i32), Const1, X); 1798 else { 1799 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1); 1800 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1); 1801 } 1802 1803 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res); 1804} 1805 1806SDValue 1807MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { 1808 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64)) 1809 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2()); 1810 1811 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2()); 1812} 1813 1814SDValue MipsTargetLowering:: 1815lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 1816 // check the depth 1817 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1818 "Frame address can only be determined for current frame."); 1819 1820 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1821 MFI->setFrameAddressIsTaken(true); 1822 EVT VT = Op.getValueType(); 1823 SDLoc DL(Op); 1824 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, 1825 IsN64 ? Mips::FP_64 : Mips::FP, VT); 1826 return FrameAddr; 1827} 1828 1829SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 1830 SelectionDAG &DAG) const { 1831 // check the depth 1832 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1833 "Return address can be determined only for current frame."); 1834 1835 MachineFunction &MF = DAG.getMachineFunction(); 1836 MachineFrameInfo *MFI = MF.getFrameInfo(); 1837 MVT VT = Op.getSimpleValueType(); 1838 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA; 1839 MFI->setReturnAddressIsTaken(true); 1840 1841 // Return RA, which contains the return address. Mark it an implicit live-in. 1842 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 1843 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 1844} 1845 1846// An EH_RETURN is the result of lowering llvm.eh.return which in turn is 1847// generated from __builtin_eh_return (offset, handler) 1848// The effect of this is to adjust the stack pointer by "offset" 1849// and then branch to "handler". 1850SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 1851 const { 1852 MachineFunction &MF = DAG.getMachineFunction(); 1853 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 1854 1855 MipsFI->setCallsEhReturn(); 1856 SDValue Chain = Op.getOperand(0); 1857 SDValue Offset = Op.getOperand(1); 1858 SDValue Handler = Op.getOperand(2); 1859 SDLoc DL(Op); 1860 EVT Ty = IsN64 ? MVT::i64 : MVT::i32; 1861 1862 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 1863 // EH_RETURN nodes, so that instructions are emitted back-to-back. 1864 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1; 1865 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0; 1866 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 1867 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 1868 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 1869 DAG.getRegister(OffsetReg, Ty), 1870 DAG.getRegister(AddrReg, getPointerTy()), 1871 Chain.getValue(1)); 1872} 1873 1874SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 1875 SelectionDAG &DAG) const { 1876 // FIXME: Need pseudo-fence for 'singlethread' fences 1877 // FIXME: Set SType for weaker fences where supported/appropriate. 1878 unsigned SType = 0; 1879 SDLoc DL(Op); 1880 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 1881 DAG.getConstant(SType, MVT::i32)); 1882} 1883 1884SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 1885 SelectionDAG &DAG) const { 1886 SDLoc DL(Op); 1887 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1888 SDValue Shamt = Op.getOperand(2); 1889 1890 // if shamt < 32: 1891 // lo = (shl lo, shamt) 1892 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 1893 // else: 1894 // lo = 0 1895 // hi = (shl lo, shamt[4:0]) 1896 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1897 DAG.getConstant(-1, MVT::i32)); 1898 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, 1899 DAG.getConstant(1, MVT::i32)); 1900 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo, 1901 Not); 1902 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt); 1903 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1904 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt); 1905 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1906 DAG.getConstant(0x20, MVT::i32)); 1907 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1908 DAG.getConstant(0, MVT::i32), ShiftLeftLo); 1909 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or); 1910 1911 SDValue Ops[2] = {Lo, Hi}; 1912 return DAG.getMergeValues(Ops, 2, DL); 1913} 1914 1915SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 1916 bool IsSRA) const { 1917 SDLoc DL(Op); 1918 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1919 SDValue Shamt = Op.getOperand(2); 1920 1921 // if shamt < 32: 1922 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 1923 // if isSRA: 1924 // hi = (sra hi, shamt) 1925 // else: 1926 // hi = (srl hi, shamt) 1927 // else: 1928 // if isSRA: 1929 // lo = (sra hi, shamt[4:0]) 1930 // hi = (sra hi, 31) 1931 // else: 1932 // lo = (srl hi, shamt[4:0]) 1933 // hi = 0 1934 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1935 DAG.getConstant(-1, MVT::i32)); 1936 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, 1937 DAG.getConstant(1, MVT::i32)); 1938 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not); 1939 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt); 1940 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1941 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32, 1942 Hi, Shamt); 1943 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1944 DAG.getConstant(0x20, MVT::i32)); 1945 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi, 1946 DAG.getConstant(31, MVT::i32)); 1947 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or); 1948 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1949 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32), 1950 ShiftRightHi); 1951 1952 SDValue Ops[2] = {Lo, Hi}; 1953 return DAG.getMergeValues(Ops, 2, DL); 1954} 1955 1956static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 1957 SDValue Chain, SDValue Src, unsigned Offset) { 1958 SDValue Ptr = LD->getBasePtr(); 1959 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 1960 EVT BasePtrVT = Ptr.getValueType(); 1961 SDLoc DL(LD); 1962 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 1963 1964 if (Offset) 1965 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 1966 DAG.getConstant(Offset, BasePtrVT)); 1967 1968 SDValue Ops[] = { Chain, Ptr, Src }; 1969 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 1970 LD->getMemOperand()); 1971} 1972 1973// Expand an unaligned 32 or 64-bit integer load node. 1974SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 1975 LoadSDNode *LD = cast<LoadSDNode>(Op); 1976 EVT MemVT = LD->getMemoryVT(); 1977 1978 // Return if load is aligned or if MemVT is neither i32 nor i64. 1979 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 1980 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 1981 return SDValue(); 1982 1983 bool IsLittle = Subtarget->isLittle(); 1984 EVT VT = Op.getValueType(); 1985 ISD::LoadExtType ExtType = LD->getExtensionType(); 1986 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 1987 1988 assert((VT == MVT::i32) || (VT == MVT::i64)); 1989 1990 // Expand 1991 // (set dst, (i64 (load baseptr))) 1992 // to 1993 // (set tmp, (ldl (add baseptr, 7), undef)) 1994 // (set dst, (ldr baseptr, tmp)) 1995 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 1996 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 1997 IsLittle ? 7 : 0); 1998 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 1999 IsLittle ? 0 : 7); 2000 } 2001 2002 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 2003 IsLittle ? 3 : 0); 2004 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 2005 IsLittle ? 0 : 3); 2006 2007 // Expand 2008 // (set dst, (i32 (load baseptr))) or 2009 // (set dst, (i64 (sextload baseptr))) or 2010 // (set dst, (i64 (extload baseptr))) 2011 // to 2012 // (set tmp, (lwl (add baseptr, 3), undef)) 2013 // (set dst, (lwr baseptr, tmp)) 2014 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2015 (ExtType == ISD::EXTLOAD)) 2016 return LWR; 2017 2018 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2019 2020 // Expand 2021 // (set dst, (i64 (zextload baseptr))) 2022 // to 2023 // (set tmp0, (lwl (add baseptr, 3), undef)) 2024 // (set tmp1, (lwr baseptr, tmp0)) 2025 // (set tmp2, (shl tmp1, 32)) 2026 // (set dst, (srl tmp2, 32)) 2027 SDLoc DL(LD); 2028 SDValue Const32 = DAG.getConstant(32, MVT::i32); 2029 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2030 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2031 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2032 return DAG.getMergeValues(Ops, 2, DL); 2033} 2034 2035static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2036 SDValue Chain, unsigned Offset) { 2037 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2038 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2039 SDLoc DL(SD); 2040 SDVTList VTList = DAG.getVTList(MVT::Other); 2041 2042 if (Offset) 2043 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2044 DAG.getConstant(Offset, BasePtrVT)); 2045 2046 SDValue Ops[] = { Chain, Value, Ptr }; 2047 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 2048 SD->getMemOperand()); 2049} 2050 2051// Expand an unaligned 32 or 64-bit integer store node. 2052static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2053 bool IsLittle) { 2054 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2055 EVT VT = Value.getValueType(); 2056 2057 // Expand 2058 // (store val, baseptr) or 2059 // (truncstore val, baseptr) 2060 // to 2061 // (swl val, (add baseptr, 3)) 2062 // (swr val, baseptr) 2063 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2064 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2065 IsLittle ? 3 : 0); 2066 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2067 } 2068 2069 assert(VT == MVT::i64); 2070 2071 // Expand 2072 // (store val, baseptr) 2073 // to 2074 // (sdl val, (add baseptr, 7)) 2075 // (sdr val, baseptr) 2076 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2077 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2078} 2079 2080// Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2081static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) { 2082 SDValue Val = SD->getValue(); 2083 2084 if (Val.getOpcode() != ISD::FP_TO_SINT) 2085 return SDValue(); 2086 2087 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2088 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2089 Val.getOperand(0)); 2090 2091 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2092 SD->getPointerInfo(), SD->isVolatile(), 2093 SD->isNonTemporal(), SD->getAlignment()); 2094} 2095 2096SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2097 StoreSDNode *SD = cast<StoreSDNode>(Op); 2098 EVT MemVT = SD->getMemoryVT(); 2099 2100 // Lower unaligned integer stores. 2101 if ((SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2102 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2103 return lowerUnalignedIntStore(SD, DAG, Subtarget->isLittle()); 2104 2105 return lowerFP_TO_SINT_STORE(SD, DAG); 2106} 2107 2108SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const { 2109 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR 2110 || cast<ConstantSDNode> 2111 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0 2112 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET) 2113 return SDValue(); 2114 2115 // The pattern 2116 // (add (frameaddr 0), (frame_to_args_offset)) 2117 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to 2118 // (add FrameObject, 0) 2119 // where FrameObject is a fixed StackObject with offset 0 which points to 2120 // the old stack pointer. 2121 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2122 EVT ValTy = Op->getValueType(0); 2123 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2124 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy); 2125 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr, 2126 DAG.getConstant(0, ValTy)); 2127} 2128 2129SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2130 SelectionDAG &DAG) const { 2131 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2132 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2133 Op.getOperand(0)); 2134 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2135} 2136 2137//===----------------------------------------------------------------------===// 2138// Calling Convention Implementation 2139//===----------------------------------------------------------------------===// 2140 2141//===----------------------------------------------------------------------===// 2142// TODO: Implement a generic logic using tblgen that can support this. 2143// Mips O32 ABI rules: 2144// --- 2145// i32 - Passed in A0, A1, A2, A3 and stack 2146// f32 - Only passed in f32 registers if no int reg has been used yet to hold 2147// an argument. Otherwise, passed in A1, A2, A3 and stack. 2148// f64 - Only passed in two aliased f32 registers if no int reg has been used 2149// yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2150// not used, it must be shadowed. If only A3 is avaiable, shadow it and 2151// go to stack. 2152// 2153// For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2154//===----------------------------------------------------------------------===// 2155 2156static bool CC_MipsO32(unsigned ValNo, MVT ValVT, 2157 MVT LocVT, CCValAssign::LocInfo LocInfo, 2158 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2159 2160 static const unsigned IntRegsSize=4, FloatRegsSize=2; 2161 2162 static const uint16_t IntRegs[] = { 2163 Mips::A0, Mips::A1, Mips::A2, Mips::A3 2164 }; 2165 static const uint16_t F32Regs[] = { 2166 Mips::F12, Mips::F14 2167 }; 2168 static const uint16_t F64Regs[] = { 2169 Mips::D6, Mips::D7 2170 }; 2171 2172 // Do not process byval args here. 2173 if (ArgFlags.isByVal()) 2174 return true; 2175 2176 // Promote i8 and i16 2177 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2178 LocVT = MVT::i32; 2179 if (ArgFlags.isSExt()) 2180 LocInfo = CCValAssign::SExt; 2181 else if (ArgFlags.isZExt()) 2182 LocInfo = CCValAssign::ZExt; 2183 else 2184 LocInfo = CCValAssign::AExt; 2185 } 2186 2187 unsigned Reg; 2188 2189 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2190 // is true: function is vararg, argument is 3rd or higher, there is previous 2191 // argument which is not f32 or f64. 2192 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 2193 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo; 2194 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2195 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2196 2197 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2198 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2199 // If this is the first part of an i64 arg, 2200 // the allocated register must be either A0 or A2. 2201 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2202 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2203 LocVT = MVT::i32; 2204 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2205 // Allocate int register and shadow next int register. If first 2206 // available register is Mips::A1 or Mips::A3, shadow it too. 2207 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2208 if (Reg == Mips::A1 || Reg == Mips::A3) 2209 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2210 State.AllocateReg(IntRegs, IntRegsSize); 2211 LocVT = MVT::i32; 2212 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2213 // we are guaranteed to find an available float register 2214 if (ValVT == MVT::f32) { 2215 Reg = State.AllocateReg(F32Regs, FloatRegsSize); 2216 // Shadow int register 2217 State.AllocateReg(IntRegs, IntRegsSize); 2218 } else { 2219 Reg = State.AllocateReg(F64Regs, FloatRegsSize); 2220 // Shadow int registers 2221 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize); 2222 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2223 State.AllocateReg(IntRegs, IntRegsSize); 2224 State.AllocateReg(IntRegs, IntRegsSize); 2225 } 2226 } else 2227 llvm_unreachable("Cannot handle this ValVT."); 2228 2229 if (!Reg) { 2230 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3, 2231 OrigAlign); 2232 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2233 } else 2234 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2235 2236 return false; 2237} 2238 2239#include "MipsGenCallingConv.inc" 2240 2241//===----------------------------------------------------------------------===// 2242// Call Calling Convention Implementation 2243//===----------------------------------------------------------------------===// 2244 2245static const unsigned O32IntRegsSize = 4; 2246 2247// Return next O32 integer argument register. 2248static unsigned getNextIntArgReg(unsigned Reg) { 2249 assert((Reg == Mips::A0) || (Reg == Mips::A2)); 2250 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3; 2251} 2252 2253SDValue 2254MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, 2255 SDValue Chain, SDValue Arg, SDLoc DL, 2256 bool IsTailCall, SelectionDAG &DAG) const { 2257 if (!IsTailCall) { 2258 SDValue PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, 2259 DAG.getIntPtrConstant(Offset)); 2260 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false, 2261 false, 0); 2262 } 2263 2264 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2265 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false); 2266 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2267 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(), 2268 /*isVolatile=*/ true, false, 0); 2269} 2270 2271void MipsTargetLowering:: 2272getOpndList(SmallVectorImpl<SDValue> &Ops, 2273 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 2274 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, 2275 CallLoweringInfo &CLI, SDValue Callee, SDValue Chain) const { 2276 // Insert node "GP copy globalreg" before call to function. 2277 // 2278 // R_MIPS_CALL* operators (emitted when non-internal functions are called 2279 // in PIC mode) allow symbols to be resolved via lazy binding. 2280 // The lazy binding stub requires GP to point to the GOT. 2281 if (IsPICCall && !InternalLinkage) { 2282 unsigned GPReg = IsN64 ? Mips::GP_64 : Mips::GP; 2283 EVT Ty = IsN64 ? MVT::i64 : MVT::i32; 2284 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty))); 2285 } 2286 2287 // Build a sequence of copy-to-reg nodes chained together with token 2288 // chain and flag operands which copy the outgoing args into registers. 2289 // The InFlag in necessary since all emitted instructions must be 2290 // stuck together. 2291 SDValue InFlag; 2292 2293 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 2294 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first, 2295 RegsToPass[i].second, InFlag); 2296 InFlag = Chain.getValue(1); 2297 } 2298 2299 // Add argument registers to the end of the list so that they are 2300 // known live into the call. 2301 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 2302 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first, 2303 RegsToPass[i].second.getValueType())); 2304 2305 // Add a register mask operand representing the call-preserved registers. 2306 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 2307 const uint32_t *Mask = TRI->getCallPreservedMask(CLI.CallConv); 2308 assert(Mask && "Missing call preserved mask for calling convention"); 2309 if (Subtarget->inMips16HardFloat()) { 2310 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) { 2311 llvm::StringRef Sym = G->getGlobal()->getName(); 2312 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 2313 if (F->hasFnAttribute("__Mips16RetHelper")) { 2314 Mask = MipsRegisterInfo::getMips16RetHelperMask(); 2315 } 2316 } 2317 } 2318 Ops.push_back(CLI.DAG.getRegisterMask(Mask)); 2319 2320 if (InFlag.getNode()) 2321 Ops.push_back(InFlag); 2322} 2323 2324/// LowerCall - functions arguments are copied from virtual regs to 2325/// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. 2326SDValue 2327MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, 2328 SmallVectorImpl<SDValue> &InVals) const { 2329 SelectionDAG &DAG = CLI.DAG; 2330 SDLoc DL = CLI.DL; 2331 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; 2332 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; 2333 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; 2334 SDValue Chain = CLI.Chain; 2335 SDValue Callee = CLI.Callee; 2336 bool &IsTailCall = CLI.IsTailCall; 2337 CallingConv::ID CallConv = CLI.CallConv; 2338 bool IsVarArg = CLI.IsVarArg; 2339 2340 MachineFunction &MF = DAG.getMachineFunction(); 2341 MachineFrameInfo *MFI = MF.getFrameInfo(); 2342 const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering(); 2343 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_; 2344 2345 // Analyze operands of the call, assigning locations to each operand. 2346 SmallVector<CCValAssign, 16> ArgLocs; 2347 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2348 getTargetMachine(), ArgLocs, *DAG.getContext()); 2349 MipsCC::SpecialCallingConvType SpecialCallingConv = 2350 getSpecialCallingConv(Callee); 2351 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo, SpecialCallingConv); 2352 2353 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg, 2354 getTargetMachine().Options.UseSoftFloat, 2355 Callee.getNode(), CLI.Args); 2356 2357 // Get a count of how many bytes are to be pushed on the stack. 2358 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2359 2360 // Check if it's really possible to do a tail call. 2361 if (IsTailCall) 2362 IsTailCall = 2363 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset, 2364 *MF.getInfo<MipsFunctionInfo>()); 2365 2366 if (IsTailCall) 2367 ++NumTailCalls; 2368 2369 // Chain is the output chain of the last Load/Store or CopyToReg node. 2370 // ByValChain is the output chain of the last Memcpy node created for copying 2371 // byval arguments to the stack. 2372 unsigned StackAlignment = TFL->getStackAlignment(); 2373 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment); 2374 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true); 2375 2376 if (!IsTailCall) 2377 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL); 2378 2379 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, 2380 IsN64 ? Mips::SP_64 : Mips::SP, 2381 getPointerTy()); 2382 2383 // With EABI is it possible to have 16 args on registers. 2384 std::deque< std::pair<unsigned, SDValue> > RegsToPass; 2385 SmallVector<SDValue, 8> MemOpChains; 2386 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2387 2388 // Walk the register/memloc assignments, inserting copies/loads. 2389 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2390 SDValue Arg = OutVals[i]; 2391 CCValAssign &VA = ArgLocs[i]; 2392 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 2393 ISD::ArgFlagsTy Flags = Outs[i].Flags; 2394 2395 // ByVal Arg. 2396 if (Flags.isByVal()) { 2397 assert(Flags.getByValSize() && 2398 "ByVal args of size 0 should have been ignored by front-end."); 2399 assert(ByValArg != MipsCCInfo.byval_end()); 2400 assert(!IsTailCall && 2401 "Do not tail-call optimize if there is a byval argument."); 2402 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 2403 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle()); 2404 ++ByValArg; 2405 continue; 2406 } 2407 2408 // Promote the value if needed. 2409 switch (VA.getLocInfo()) { 2410 default: llvm_unreachable("Unknown loc info!"); 2411 case CCValAssign::Full: 2412 if (VA.isRegLoc()) { 2413 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 2414 (ValVT == MVT::f64 && LocVT == MVT::i64) || 2415 (ValVT == MVT::i64 && LocVT == MVT::f64)) 2416 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 2417 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 2418 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2419 Arg, DAG.getConstant(0, MVT::i32)); 2420 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2421 Arg, DAG.getConstant(1, MVT::i32)); 2422 if (!Subtarget->isLittle()) 2423 std::swap(Lo, Hi); 2424 unsigned LocRegLo = VA.getLocReg(); 2425 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 2426 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 2427 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 2428 continue; 2429 } 2430 } 2431 break; 2432 case CCValAssign::SExt: 2433 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 2434 break; 2435 case CCValAssign::ZExt: 2436 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 2437 break; 2438 case CCValAssign::AExt: 2439 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 2440 break; 2441 } 2442 2443 // Arguments that can be passed on register must be kept at 2444 // RegsToPass vector 2445 if (VA.isRegLoc()) { 2446 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 2447 continue; 2448 } 2449 2450 // Register can't get to this point... 2451 assert(VA.isMemLoc()); 2452 2453 // emit ISD::STORE whichs stores the 2454 // parameter value to a stack Location 2455 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 2456 Chain, Arg, DL, IsTailCall, DAG)); 2457 } 2458 2459 // Transform all store nodes into one single node because all store 2460 // nodes are independent of each other. 2461 if (!MemOpChains.empty()) 2462 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2463 &MemOpChains[0], MemOpChains.size()); 2464 2465 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 2466 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 2467 // node so that legalize doesn't hack it. 2468 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25 2469 bool GlobalOrExternal = false, InternalLinkage = false; 2470 SDValue CalleeLo; 2471 2472 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2473 if (IsPICCall) { 2474 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 2475 2476 if (InternalLinkage) 2477 Callee = getAddrLocal(Callee, DAG, HasMips64); 2478 else if (LargeGOT) 2479 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2480 MipsII::MO_CALL_LO16); 2481 else 2482 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2483 } else 2484 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0, 2485 MipsII::MO_NO_FLAG); 2486 GlobalOrExternal = true; 2487 } 2488 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 2489 if (!IsN64 && !IsPIC) // !N64 && static 2490 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(), 2491 MipsII::MO_NO_FLAG); 2492 else if (LargeGOT) 2493 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2494 MipsII::MO_CALL_LO16); 2495 else // N64 || PIC 2496 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2497 2498 GlobalOrExternal = true; 2499 } 2500 2501 SmallVector<SDValue, 8> Ops(1, Chain); 2502 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 2503 2504 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage, 2505 CLI, Callee, Chain); 2506 2507 if (IsTailCall) 2508 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size()); 2509 2510 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size()); 2511 SDValue InFlag = Chain.getValue(1); 2512 2513 // Create the CALLSEQ_END node. 2514 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 2515 DAG.getIntPtrConstant(0, true), InFlag, DL); 2516 InFlag = Chain.getValue(1); 2517 2518 // Handle result values, copying them out of physregs into vregs that we 2519 // return. 2520 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, 2521 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy); 2522} 2523 2524/// LowerCallResult - Lower the result values of a call into the 2525/// appropriate copies out of appropriate physical registers. 2526SDValue 2527MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 2528 CallingConv::ID CallConv, bool IsVarArg, 2529 const SmallVectorImpl<ISD::InputArg> &Ins, 2530 SDLoc DL, SelectionDAG &DAG, 2531 SmallVectorImpl<SDValue> &InVals, 2532 const SDNode *CallNode, 2533 const Type *RetTy) const { 2534 // Assign locations to each value returned by this call. 2535 SmallVector<CCValAssign, 16> RVLocs; 2536 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2537 getTargetMachine(), RVLocs, *DAG.getContext()); 2538 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2539 2540 MipsCCInfo.analyzeCallResult(Ins, getTargetMachine().Options.UseSoftFloat, 2541 CallNode, RetTy); 2542 2543 // Copy all of the result registers out of their specified physreg. 2544 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2545 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 2546 RVLocs[i].getLocVT(), InFlag); 2547 Chain = Val.getValue(1); 2548 InFlag = Val.getValue(2); 2549 2550 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2551 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val); 2552 2553 InVals.push_back(Val); 2554 } 2555 2556 return Chain; 2557} 2558 2559//===----------------------------------------------------------------------===// 2560// Formal Arguments Calling Convention Implementation 2561//===----------------------------------------------------------------------===// 2562/// LowerFormalArguments - transform physical registers into virtual registers 2563/// and generate load operations for arguments places on the stack. 2564SDValue 2565MipsTargetLowering::LowerFormalArguments(SDValue Chain, 2566 CallingConv::ID CallConv, 2567 bool IsVarArg, 2568 const SmallVectorImpl<ISD::InputArg> &Ins, 2569 SDLoc DL, SelectionDAG &DAG, 2570 SmallVectorImpl<SDValue> &InVals) 2571 const { 2572 MachineFunction &MF = DAG.getMachineFunction(); 2573 MachineFrameInfo *MFI = MF.getFrameInfo(); 2574 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2575 2576 MipsFI->setVarArgsFrameIndex(0); 2577 2578 // Used with vargs to acumulate store chains. 2579 std::vector<SDValue> OutChains; 2580 2581 // Assign locations to all of the incoming arguments. 2582 SmallVector<CCValAssign, 16> ArgLocs; 2583 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2584 getTargetMachine(), ArgLocs, *DAG.getContext()); 2585 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2586 Function::const_arg_iterator FuncArg = 2587 DAG.getMachineFunction().getFunction()->arg_begin(); 2588 bool UseSoftFloat = getTargetMachine().Options.UseSoftFloat; 2589 2590 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg); 2591 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 2592 MipsCCInfo.hasByValArg()); 2593 2594 unsigned CurArgIdx = 0; 2595 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2596 2597 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2598 CCValAssign &VA = ArgLocs[i]; 2599 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx); 2600 CurArgIdx = Ins[i].OrigArgIndex; 2601 EVT ValVT = VA.getValVT(); 2602 ISD::ArgFlagsTy Flags = Ins[i].Flags; 2603 bool IsRegLoc = VA.isRegLoc(); 2604 2605 if (Flags.isByVal()) { 2606 assert(Flags.getByValSize() && 2607 "ByVal args of size 0 should have been ignored by front-end."); 2608 assert(ByValArg != MipsCCInfo.byval_end()); 2609 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 2610 MipsCCInfo, *ByValArg); 2611 ++ByValArg; 2612 continue; 2613 } 2614 2615 // Arguments stored on registers 2616 if (IsRegLoc) { 2617 EVT RegVT = VA.getLocVT(); 2618 unsigned ArgReg = VA.getLocReg(); 2619 const TargetRegisterClass *RC; 2620 2621 if (RegVT == MVT::i32) 2622 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass : 2623 &Mips::GPR32RegClass; 2624 else if (RegVT == MVT::i64) 2625 RC = &Mips::GPR64RegClass; 2626 else if (RegVT == MVT::f32) 2627 RC = &Mips::FGR32RegClass; 2628 else if (RegVT == MVT::f64) 2629 RC = HasMips64 ? &Mips::FGR64RegClass : &Mips::AFGR64RegClass; 2630 else 2631 llvm_unreachable("RegVT not supported by FormalArguments Lowering"); 2632 2633 // Transform the arguments stored on 2634 // physical registers into virtual ones 2635 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 2636 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 2637 2638 // If this is an 8 or 16-bit value, it has been passed promoted 2639 // to 32 bits. Insert an assert[sz]ext to capture this, then 2640 // truncate to the right size. 2641 if (VA.getLocInfo() != CCValAssign::Full) { 2642 unsigned Opcode = 0; 2643 if (VA.getLocInfo() == CCValAssign::SExt) 2644 Opcode = ISD::AssertSext; 2645 else if (VA.getLocInfo() == CCValAssign::ZExt) 2646 Opcode = ISD::AssertZext; 2647 if (Opcode) 2648 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue, 2649 DAG.getValueType(ValVT)); 2650 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue); 2651 } 2652 2653 // Handle floating point arguments passed in integer registers and 2654 // long double arguments passed in floating point registers. 2655 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 2656 (RegVT == MVT::i64 && ValVT == MVT::f64) || 2657 (RegVT == MVT::f64 && ValVT == MVT::i64)) 2658 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 2659 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) { 2660 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 2661 getNextIntArgReg(ArgReg), RC); 2662 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 2663 if (!Subtarget->isLittle()) 2664 std::swap(ArgValue, ArgValue2); 2665 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 2666 ArgValue, ArgValue2); 2667 } 2668 2669 InVals.push_back(ArgValue); 2670 } else { // VA.isRegLoc() 2671 2672 // sanity check 2673 assert(VA.isMemLoc()); 2674 2675 // The stack pointer offset is relative to the caller stack frame. 2676 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, 2677 VA.getLocMemOffset(), true); 2678 2679 // Create load nodes to retrieve arguments from the stack 2680 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2681 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN, 2682 MachinePointerInfo::getFixedStack(FI), 2683 false, false, false, 0)); 2684 } 2685 } 2686 2687 // The mips ABIs for returning structs by value requires that we copy 2688 // the sret argument into $v0 for the return. Save the argument into 2689 // a virtual register so that we can access it from the return points. 2690 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) { 2691 unsigned Reg = MipsFI->getSRetReturnReg(); 2692 if (!Reg) { 2693 Reg = MF.getRegInfo(). 2694 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32)); 2695 MipsFI->setSRetReturnReg(Reg); 2696 } 2697 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]); 2698 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 2699 } 2700 2701 if (IsVarArg) 2702 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG); 2703 2704 // All stores are grouped in one node to allow the matching between 2705 // the size of Ins and InVals. This only happens when on varg functions 2706 if (!OutChains.empty()) { 2707 OutChains.push_back(Chain); 2708 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2709 &OutChains[0], OutChains.size()); 2710 } 2711 2712 return Chain; 2713} 2714 2715//===----------------------------------------------------------------------===// 2716// Return Value Calling Convention Implementation 2717//===----------------------------------------------------------------------===// 2718 2719bool 2720MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 2721 MachineFunction &MF, bool IsVarArg, 2722 const SmallVectorImpl<ISD::OutputArg> &Outs, 2723 LLVMContext &Context) const { 2724 SmallVector<CCValAssign, 16> RVLocs; 2725 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), 2726 RVLocs, Context); 2727 return CCInfo.CheckReturn(Outs, RetCC_Mips); 2728} 2729 2730SDValue 2731MipsTargetLowering::LowerReturn(SDValue Chain, 2732 CallingConv::ID CallConv, bool IsVarArg, 2733 const SmallVectorImpl<ISD::OutputArg> &Outs, 2734 const SmallVectorImpl<SDValue> &OutVals, 2735 SDLoc DL, SelectionDAG &DAG) const { 2736 // CCValAssign - represent the assignment of 2737 // the return value to a location 2738 SmallVector<CCValAssign, 16> RVLocs; 2739 MachineFunction &MF = DAG.getMachineFunction(); 2740 2741 // CCState - Info about the registers and stack slot. 2742 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs, 2743 *DAG.getContext()); 2744 MipsCC MipsCCInfo(CallConv, IsO32, CCInfo); 2745 2746 // Analyze return values. 2747 MipsCCInfo.analyzeReturn(Outs, getTargetMachine().Options.UseSoftFloat, 2748 MF.getFunction()->getReturnType()); 2749 2750 SDValue Flag; 2751 SmallVector<SDValue, 4> RetOps(1, Chain); 2752 2753 // Copy the result values into the output registers. 2754 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2755 SDValue Val = OutVals[i]; 2756 CCValAssign &VA = RVLocs[i]; 2757 assert(VA.isRegLoc() && "Can only return in registers!"); 2758 2759 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2760 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val); 2761 2762 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 2763 2764 // Guarantee that all emitted copies are stuck together with flags. 2765 Flag = Chain.getValue(1); 2766 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2767 } 2768 2769 // The mips ABIs for returning structs by value requires that we copy 2770 // the sret argument into $v0 for the return. We saved the argument into 2771 // a virtual register in the entry block, so now we copy the value out 2772 // and into $v0. 2773 if (MF.getFunction()->hasStructRetAttr()) { 2774 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2775 unsigned Reg = MipsFI->getSRetReturnReg(); 2776 2777 if (!Reg) 2778 llvm_unreachable("sret virtual register not created in the entry block"); 2779 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy()); 2780 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0; 2781 2782 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 2783 Flag = Chain.getValue(1); 2784 RetOps.push_back(DAG.getRegister(V0, getPointerTy())); 2785 } 2786 2787 RetOps[0] = Chain; // Update chain. 2788 2789 // Add the flag if we have it. 2790 if (Flag.getNode()) 2791 RetOps.push_back(Flag); 2792 2793 // Return on Mips is always a "jr $ra" 2794 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size()); 2795} 2796 2797//===----------------------------------------------------------------------===// 2798// Mips Inline Assembly Support 2799//===----------------------------------------------------------------------===// 2800 2801/// getConstraintType - Given a constraint letter, return the type of 2802/// constraint it is for this target. 2803MipsTargetLowering::ConstraintType MipsTargetLowering:: 2804getConstraintType(const std::string &Constraint) const 2805{ 2806 // Mips specific constrainy 2807 // GCC config/mips/constraints.md 2808 // 2809 // 'd' : An address register. Equivalent to r 2810 // unless generating MIPS16 code. 2811 // 'y' : Equivalent to r; retained for 2812 // backwards compatibility. 2813 // 'c' : A register suitable for use in an indirect 2814 // jump. This will always be $25 for -mabicalls. 2815 // 'l' : The lo register. 1 word storage. 2816 // 'x' : The hilo register pair. Double word storage. 2817 if (Constraint.size() == 1) { 2818 switch (Constraint[0]) { 2819 default : break; 2820 case 'd': 2821 case 'y': 2822 case 'f': 2823 case 'c': 2824 case 'l': 2825 case 'x': 2826 return C_RegisterClass; 2827 case 'R': 2828 return C_Memory; 2829 } 2830 } 2831 return TargetLowering::getConstraintType(Constraint); 2832} 2833 2834/// Examine constraint type and operand type and determine a weight value. 2835/// This object must already have been set up with the operand type 2836/// and the current alternative constraint selected. 2837TargetLowering::ConstraintWeight 2838MipsTargetLowering::getSingleConstraintMatchWeight( 2839 AsmOperandInfo &info, const char *constraint) const { 2840 ConstraintWeight weight = CW_Invalid; 2841 Value *CallOperandVal = info.CallOperandVal; 2842 // If we don't have a value, we can't do a match, 2843 // but allow it at the lowest weight. 2844 if (CallOperandVal == NULL) 2845 return CW_Default; 2846 Type *type = CallOperandVal->getType(); 2847 // Look at the constraint type. 2848 switch (*constraint) { 2849 default: 2850 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 2851 break; 2852 case 'd': 2853 case 'y': 2854 if (type->isIntegerTy()) 2855 weight = CW_Register; 2856 break; 2857 case 'f': 2858 if (type->isFloatTy()) 2859 weight = CW_Register; 2860 break; 2861 case 'c': // $25 for indirect jumps 2862 case 'l': // lo register 2863 case 'x': // hilo register pair 2864 if (type->isIntegerTy()) 2865 weight = CW_SpecificReg; 2866 break; 2867 case 'I': // signed 16 bit immediate 2868 case 'J': // integer zero 2869 case 'K': // unsigned 16 bit immediate 2870 case 'L': // signed 32 bit immediate where lower 16 bits are 0 2871 case 'N': // immediate in the range of -65535 to -1 (inclusive) 2872 case 'O': // signed 15 bit immediate (+- 16383) 2873 case 'P': // immediate in the range of 65535 to 1 (inclusive) 2874 if (isa<ConstantInt>(CallOperandVal)) 2875 weight = CW_Constant; 2876 break; 2877 case 'R': 2878 weight = CW_Memory; 2879 break; 2880 } 2881 return weight; 2882} 2883 2884/// This is a helper function to parse a physical register string and split it 2885/// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 2886/// that is returned indicates whether parsing was successful. The second flag 2887/// is true if the numeric part exists. 2888static std::pair<bool, bool> 2889parsePhysicalReg(const StringRef &C, std::string &Prefix, 2890 unsigned long long &Reg) { 2891 if (C.front() != '{' || C.back() != '}') 2892 return std::make_pair(false, false); 2893 2894 // Search for the first numeric character. 2895 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 2896 I = std::find_if(B, E, std::ptr_fun(isdigit)); 2897 2898 Prefix.assign(B, I - B); 2899 2900 // The second flag is set to false if no numeric characters were found. 2901 if (I == E) 2902 return std::make_pair(true, false); 2903 2904 // Parse the numeric characters. 2905 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 2906 true); 2907} 2908 2909std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 2910parseRegForInlineAsmConstraint(const StringRef &C, MVT VT) const { 2911 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 2912 const TargetRegisterClass *RC; 2913 std::string Prefix; 2914 unsigned long long Reg; 2915 2916 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 2917 2918 if (!R.first) 2919 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2920 2921 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 2922 // No numeric characters follow "hi" or "lo". 2923 if (R.second) 2924 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2925 2926 RC = TRI->getRegClass(Prefix == "hi" ? 2927 Mips::HIRegsRegClassID : Mips::LORegsRegClassID); 2928 return std::make_pair(*(RC->begin()), RC); 2929 } 2930 2931 if (!R.second) 2932 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2933 2934 if (Prefix == "$f") { // Parse $f0-$f31. 2935 // If the size of FP registers is 64-bit or Reg is an even number, select 2936 // the 64-bit register class. Otherwise, select the 32-bit register class. 2937 if (VT == MVT::Other) 2938 VT = (Subtarget->isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 2939 2940 RC= getRegClassFor(VT); 2941 2942 if (RC == &Mips::AFGR64RegClass) { 2943 assert(Reg % 2 == 0); 2944 Reg >>= 1; 2945 } 2946 } else if (Prefix == "$fcc") { // Parse $fcc0-$fcc7. 2947 RC = TRI->getRegClass(Mips::FCCRegClassID); 2948 } else { // Parse $0-$31. 2949 assert(Prefix == "$"); 2950 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 2951 } 2952 2953 assert(Reg < RC->getNumRegs()); 2954 return std::make_pair(*(RC->begin() + Reg), RC); 2955} 2956 2957/// Given a register class constraint, like 'r', if this corresponds directly 2958/// to an LLVM register class, return a register of 0 and the register class 2959/// pointer. 2960std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering:: 2961getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const 2962{ 2963 if (Constraint.size() == 1) { 2964 switch (Constraint[0]) { 2965 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 2966 case 'y': // Same as 'r'. Exists for compatibility. 2967 case 'r': 2968 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 2969 if (Subtarget->inMips16Mode()) 2970 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 2971 return std::make_pair(0U, &Mips::GPR32RegClass); 2972 } 2973 if (VT == MVT::i64 && !HasMips64) 2974 return std::make_pair(0U, &Mips::GPR32RegClass); 2975 if (VT == MVT::i64 && HasMips64) 2976 return std::make_pair(0U, &Mips::GPR64RegClass); 2977 // This will generate an error message 2978 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2979 case 'f': 2980 if (VT == MVT::f32) 2981 return std::make_pair(0U, &Mips::FGR32RegClass); 2982 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) { 2983 if (Subtarget->isFP64bit()) 2984 return std::make_pair(0U, &Mips::FGR64RegClass); 2985 return std::make_pair(0U, &Mips::AFGR64RegClass); 2986 } 2987 break; 2988 case 'c': // register suitable for indirect jump 2989 if (VT == MVT::i32) 2990 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 2991 assert(VT == MVT::i64 && "Unexpected type."); 2992 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 2993 case 'l': // register suitable for indirect jump 2994 if (VT == MVT::i32) 2995 return std::make_pair((unsigned)Mips::LO, &Mips::LORegsRegClass); 2996 return std::make_pair((unsigned)Mips::LO64, &Mips::LORegs64RegClass); 2997 case 'x': // register suitable for indirect jump 2998 // Fixme: Not triggering the use of both hi and low 2999 // This will generate an error message 3000 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 3001 } 3002 } 3003 3004 std::pair<unsigned, const TargetRegisterClass *> R; 3005 R = parseRegForInlineAsmConstraint(Constraint, VT); 3006 3007 if (R.second) 3008 return R; 3009 3010 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3011} 3012 3013/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3014/// vector. If it is invalid, don't add anything to Ops. 3015void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3016 std::string &Constraint, 3017 std::vector<SDValue>&Ops, 3018 SelectionDAG &DAG) const { 3019 SDValue Result(0, 0); 3020 3021 // Only support length 1 constraints for now. 3022 if (Constraint.length() > 1) return; 3023 3024 char ConstraintLetter = Constraint[0]; 3025 switch (ConstraintLetter) { 3026 default: break; // This will fall through to the generic implementation 3027 case 'I': // Signed 16 bit constant 3028 // If this fails, the parent routine will give an error 3029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3030 EVT Type = Op.getValueType(); 3031 int64_t Val = C->getSExtValue(); 3032 if (isInt<16>(Val)) { 3033 Result = DAG.getTargetConstant(Val, Type); 3034 break; 3035 } 3036 } 3037 return; 3038 case 'J': // integer zero 3039 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3040 EVT Type = Op.getValueType(); 3041 int64_t Val = C->getZExtValue(); 3042 if (Val == 0) { 3043 Result = DAG.getTargetConstant(0, Type); 3044 break; 3045 } 3046 } 3047 return; 3048 case 'K': // unsigned 16 bit immediate 3049 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3050 EVT Type = Op.getValueType(); 3051 uint64_t Val = (uint64_t)C->getZExtValue(); 3052 if (isUInt<16>(Val)) { 3053 Result = DAG.getTargetConstant(Val, Type); 3054 break; 3055 } 3056 } 3057 return; 3058 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3060 EVT Type = Op.getValueType(); 3061 int64_t Val = C->getSExtValue(); 3062 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3063 Result = DAG.getTargetConstant(Val, Type); 3064 break; 3065 } 3066 } 3067 return; 3068 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3069 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3070 EVT Type = Op.getValueType(); 3071 int64_t Val = C->getSExtValue(); 3072 if ((Val >= -65535) && (Val <= -1)) { 3073 Result = DAG.getTargetConstant(Val, Type); 3074 break; 3075 } 3076 } 3077 return; 3078 case 'O': // signed 15 bit immediate 3079 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3080 EVT Type = Op.getValueType(); 3081 int64_t Val = C->getSExtValue(); 3082 if ((isInt<15>(Val))) { 3083 Result = DAG.getTargetConstant(Val, Type); 3084 break; 3085 } 3086 } 3087 return; 3088 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3089 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3090 EVT Type = Op.getValueType(); 3091 int64_t Val = C->getSExtValue(); 3092 if ((Val <= 65535) && (Val >= 1)) { 3093 Result = DAG.getTargetConstant(Val, Type); 3094 break; 3095 } 3096 } 3097 return; 3098 } 3099 3100 if (Result.getNode()) { 3101 Ops.push_back(Result); 3102 return; 3103 } 3104 3105 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 3106} 3107 3108bool 3109MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const { 3110 // No global is ever allowed as a base. 3111 if (AM.BaseGV) 3112 return false; 3113 3114 switch (AM.Scale) { 3115 case 0: // "r+i" or just "i", depending on HasBaseReg. 3116 break; 3117 case 1: 3118 if (!AM.HasBaseReg) // allow "r+i". 3119 break; 3120 return false; // disallow "r+r" or "r+r+i". 3121 default: 3122 return false; 3123 } 3124 3125 return true; 3126} 3127 3128bool 3129MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 3130 // The Mips target isn't yet aware of offsets. 3131 return false; 3132} 3133 3134EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 3135 unsigned SrcAlign, 3136 bool IsMemset, bool ZeroMemset, 3137 bool MemcpyStrSrc, 3138 MachineFunction &MF) const { 3139 if (Subtarget->hasMips64()) 3140 return MVT::i64; 3141 3142 return MVT::i32; 3143} 3144 3145bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 3146 if (VT != MVT::f32 && VT != MVT::f64) 3147 return false; 3148 if (Imm.isNegZero()) 3149 return false; 3150 return Imm.isZero(); 3151} 3152 3153unsigned MipsTargetLowering::getJumpTableEncoding() const { 3154 if (IsN64) 3155 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 3156 3157 return TargetLowering::getJumpTableEncoding(); 3158} 3159 3160/// This function returns true if CallSym is a long double emulation routine. 3161static bool isF128SoftLibCall(const char *CallSym) { 3162 const char *const LibCalls[] = 3163 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2", 3164 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi", 3165 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf", 3166 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2", 3167 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3", 3168 "__trunctfdf2", "__trunctfsf2", "__unordtf2", 3169 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl", 3170 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl", 3171 "truncl"}; 3172 3173 const char * const *End = LibCalls + array_lengthof(LibCalls); 3174 3175 // Check that LibCalls is sorted alphabetically. 3176 MipsTargetLowering::LTStr Comp; 3177 3178#ifndef NDEBUG 3179 for (const char * const *I = LibCalls; I < End - 1; ++I) 3180 assert(Comp(*I, *(I + 1))); 3181#endif 3182 3183 return std::binary_search(LibCalls, End, CallSym, Comp); 3184} 3185 3186/// This function returns true if Ty is fp128 or i128 which was originally a 3187/// fp128. 3188static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) { 3189 if (Ty->isFP128Ty()) 3190 return true; 3191 3192 const ExternalSymbolSDNode *ES = 3193 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode); 3194 3195 // If the Ty is i128 and the function being called is a long double emulation 3196 // routine, then the original type is f128. 3197 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol())); 3198} 3199 3200MipsTargetLowering::MipsCC::SpecialCallingConvType 3201 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const { 3202 MipsCC::SpecialCallingConvType SpecialCallingConv = 3203 MipsCC::NoSpecialCallingConv;; 3204 if (Subtarget->inMips16HardFloat()) { 3205 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3206 llvm::StringRef Sym = G->getGlobal()->getName(); 3207 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 3208 if (F->hasFnAttribute("__Mips16RetHelper")) { 3209 SpecialCallingConv = MipsCC::Mips16RetHelperConv; 3210 } 3211 } 3212 } 3213 return SpecialCallingConv; 3214} 3215 3216MipsTargetLowering::MipsCC::MipsCC( 3217 CallingConv::ID CC, bool IsO32_, CCState &Info, 3218 MipsCC::SpecialCallingConvType SpecialCallingConv_) 3219 : CCInfo(Info), CallConv(CC), IsO32(IsO32_), 3220 SpecialCallingConv(SpecialCallingConv_){ 3221 // Pre-allocate reserved argument area. 3222 CCInfo.AllocateStack(reservedArgArea(), 1); 3223} 3224 3225 3226void MipsTargetLowering::MipsCC:: 3227analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args, 3228 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode, 3229 std::vector<ArgListEntry> &FuncArgs) { 3230 assert((CallConv != CallingConv::Fast || !IsVarArg) && 3231 "CallingConv::Fast shouldn't be used for vararg functions."); 3232 3233 unsigned NumOpnds = Args.size(); 3234 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn(); 3235 3236 for (unsigned I = 0; I != NumOpnds; ++I) { 3237 MVT ArgVT = Args[I].VT; 3238 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3239 bool R; 3240 3241 if (ArgFlags.isByVal()) { 3242 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3243 continue; 3244 } 3245 3246 if (IsVarArg && !Args[I].IsFixed) 3247 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo); 3248 else { 3249 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode, 3250 IsSoftFloat); 3251 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo); 3252 } 3253 3254 if (R) { 3255#ifndef NDEBUG 3256 dbgs() << "Call operand #" << I << " has unhandled type " 3257 << EVT(ArgVT).getEVTString(); 3258#endif 3259 llvm_unreachable(0); 3260 } 3261 } 3262} 3263 3264void MipsTargetLowering::MipsCC:: 3265analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args, 3266 bool IsSoftFloat, Function::const_arg_iterator FuncArg) { 3267 unsigned NumArgs = Args.size(); 3268 llvm::CCAssignFn *FixedFn = fixedArgFn(); 3269 unsigned CurArgIdx = 0; 3270 3271 for (unsigned I = 0; I != NumArgs; ++I) { 3272 MVT ArgVT = Args[I].VT; 3273 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3274 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx); 3275 CurArgIdx = Args[I].OrigArgIndex; 3276 3277 if (ArgFlags.isByVal()) { 3278 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3279 continue; 3280 } 3281 3282 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat); 3283 3284 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo)) 3285 continue; 3286 3287#ifndef NDEBUG 3288 dbgs() << "Formal Arg #" << I << " has unhandled type " 3289 << EVT(ArgVT).getEVTString(); 3290#endif 3291 llvm_unreachable(0); 3292 } 3293} 3294 3295template<typename Ty> 3296void MipsTargetLowering::MipsCC:: 3297analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat, 3298 const SDNode *CallNode, const Type *RetTy) const { 3299 CCAssignFn *Fn; 3300 3301 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode)) 3302 Fn = RetCC_F128Soft; 3303 else 3304 Fn = RetCC_Mips; 3305 3306 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) { 3307 MVT VT = RetVals[I].VT; 3308 ISD::ArgFlagsTy Flags = RetVals[I].Flags; 3309 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat); 3310 3311 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) { 3312#ifndef NDEBUG 3313 dbgs() << "Call result #" << I << " has unhandled type " 3314 << EVT(VT).getEVTString() << '\n'; 3315#endif 3316 llvm_unreachable(0); 3317 } 3318 } 3319} 3320 3321void MipsTargetLowering::MipsCC:: 3322analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat, 3323 const SDNode *CallNode, const Type *RetTy) const { 3324 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy); 3325} 3326 3327void MipsTargetLowering::MipsCC:: 3328analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat, 3329 const Type *RetTy) const { 3330 analyzeReturn(Outs, IsSoftFloat, 0, RetTy); 3331} 3332 3333void 3334MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT, 3335 MVT LocVT, 3336 CCValAssign::LocInfo LocInfo, 3337 ISD::ArgFlagsTy ArgFlags) { 3338 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0."); 3339 3340 struct ByValArgInfo ByVal; 3341 unsigned RegSize = regSize(); 3342 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize); 3343 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize), 3344 RegSize * 2); 3345 3346 if (useRegsForByval()) 3347 allocateRegs(ByVal, ByValSize, Align); 3348 3349 // Allocate space on caller's stack. 3350 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs, 3351 Align); 3352 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT, 3353 LocInfo)); 3354 ByValArgs.push_back(ByVal); 3355} 3356 3357unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const { 3358 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs); 3359} 3360 3361unsigned MipsTargetLowering::MipsCC::reservedArgArea() const { 3362 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0; 3363} 3364 3365const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const { 3366 return IsO32 ? O32IntRegs : Mips64IntRegs; 3367} 3368 3369llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const { 3370 if (CallConv == CallingConv::Fast) 3371 return CC_Mips_FastCC; 3372 3373 if (SpecialCallingConv == Mips16RetHelperConv) 3374 return CC_Mips16RetHelper; 3375 return IsO32 ? CC_MipsO32 : CC_MipsN; 3376} 3377 3378llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const { 3379 return IsO32 ? CC_MipsO32 : CC_MipsN_VarArg; 3380} 3381 3382const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const { 3383 return IsO32 ? O32IntRegs : Mips64DPRegs; 3384} 3385 3386void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal, 3387 unsigned ByValSize, 3388 unsigned Align) { 3389 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs(); 3390 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs(); 3391 assert(!(ByValSize % RegSize) && !(Align % RegSize) && 3392 "Byval argument's size and alignment should be a multiple of" 3393 "RegSize."); 3394 3395 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs); 3396 3397 // If Align > RegSize, the first arg register must be even. 3398 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) { 3399 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]); 3400 ++ByVal.FirstIdx; 3401 } 3402 3403 // Mark the registers allocated. 3404 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs); 3405 ByValSize -= RegSize, ++I, ++ByVal.NumRegs) 3406 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]); 3407} 3408 3409MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy, 3410 const SDNode *CallNode, 3411 bool IsSoftFloat) const { 3412 if (IsSoftFloat || IsO32) 3413 return VT; 3414 3415 // Check if the original type was fp128. 3416 if (originalTypeIsF128(OrigTy, CallNode)) { 3417 assert(VT == MVT::i64); 3418 return MVT::f64; 3419 } 3420 3421 return VT; 3422} 3423 3424void MipsTargetLowering:: 3425copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains, 3426 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 3427 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 3428 const MipsCC &CC, const ByValArgInfo &ByVal) const { 3429 MachineFunction &MF = DAG.getMachineFunction(); 3430 MachineFrameInfo *MFI = MF.getFrameInfo(); 3431 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize(); 3432 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 3433 int FrameObjOffset; 3434 3435 if (RegAreaSize) 3436 FrameObjOffset = (int)CC.reservedArgArea() - 3437 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize()); 3438 else 3439 FrameObjOffset = ByVal.Address; 3440 3441 // Create frame object. 3442 EVT PtrTy = getPointerTy(); 3443 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true); 3444 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 3445 InVals.push_back(FIN); 3446 3447 if (!ByVal.NumRegs) 3448 return; 3449 3450 // Copy arg registers. 3451 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8); 3452 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3453 3454 for (unsigned I = 0; I < ByVal.NumRegs; ++I) { 3455 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I]; 3456 unsigned VReg = addLiveIn(MF, ArgReg, RC); 3457 unsigned Offset = I * CC.regSize(); 3458 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 3459 DAG.getConstant(Offset, PtrTy)); 3460 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 3461 StorePtr, MachinePointerInfo(FuncArg, Offset), 3462 false, false, 0); 3463 OutChains.push_back(Store); 3464 } 3465} 3466 3467// Copy byVal arg to registers and stack. 3468void MipsTargetLowering:: 3469passByValArg(SDValue Chain, SDLoc DL, 3470 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 3471 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 3472 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg, 3473 const MipsCC &CC, const ByValArgInfo &ByVal, 3474 const ISD::ArgFlagsTy &Flags, bool isLittle) const { 3475 unsigned ByValSize = Flags.getByValSize(); 3476 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct. 3477 unsigned RegSize = CC.regSize(); 3478 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize); 3479 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8); 3480 3481 if (ByVal.NumRegs) { 3482 const uint16_t *ArgRegs = CC.intArgRegs(); 3483 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize); 3484 unsigned I = 0; 3485 3486 // Copy words to registers. 3487 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) { 3488 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3489 DAG.getConstant(Offset, PtrTy)); 3490 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 3491 MachinePointerInfo(), false, false, false, 3492 Alignment); 3493 MemOpChains.push_back(LoadVal.getValue(1)); 3494 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3495 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 3496 } 3497 3498 // Return if the struct has been fully copied. 3499 if (ByValSize == Offset) 3500 return; 3501 3502 // Copy the remainder of the byval argument with sub-word loads and shifts. 3503 if (LeftoverBytes) { 3504 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) && 3505 "Size of the remainder should be smaller than RegSize."); 3506 SDValue Val; 3507 3508 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0; 3509 Offset < ByValSize; LoadSize /= 2) { 3510 unsigned RemSize = ByValSize - Offset; 3511 3512 if (RemSize < LoadSize) 3513 continue; 3514 3515 // Load subword. 3516 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3517 DAG.getConstant(Offset, PtrTy)); 3518 SDValue LoadVal = 3519 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, 3520 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8), 3521 false, false, Alignment); 3522 MemOpChains.push_back(LoadVal.getValue(1)); 3523 3524 // Shift the loaded value. 3525 unsigned Shamt; 3526 3527 if (isLittle) 3528 Shamt = TotalSizeLoaded; 3529 else 3530 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8; 3531 3532 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 3533 DAG.getConstant(Shamt, MVT::i32)); 3534 3535 if (Val.getNode()) 3536 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 3537 else 3538 Val = Shift; 3539 3540 Offset += LoadSize; 3541 TotalSizeLoaded += LoadSize; 3542 Alignment = std::min(Alignment, LoadSize); 3543 } 3544 3545 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3546 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 3547 return; 3548 } 3549 } 3550 3551 // Copy remainder of byval arg to it with memcpy. 3552 unsigned MemCpySize = ByValSize - Offset; 3553 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3554 DAG.getConstant(Offset, PtrTy)); 3555 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 3556 DAG.getIntPtrConstant(ByVal.Address)); 3557 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 3558 DAG.getConstant(MemCpySize, PtrTy), Alignment, 3559 /*isVolatile=*/false, /*AlwaysInline=*/false, 3560 MachinePointerInfo(0), MachinePointerInfo(0)); 3561 MemOpChains.push_back(Chain); 3562} 3563 3564void 3565MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 3566 const MipsCC &CC, SDValue Chain, 3567 SDLoc DL, SelectionDAG &DAG) const { 3568 unsigned NumRegs = CC.numIntArgRegs(); 3569 const uint16_t *ArgRegs = CC.intArgRegs(); 3570 const CCState &CCInfo = CC.getCCInfo(); 3571 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs); 3572 unsigned RegSize = CC.regSize(); 3573 MVT RegTy = MVT::getIntegerVT(RegSize * 8); 3574 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3575 MachineFunction &MF = DAG.getMachineFunction(); 3576 MachineFrameInfo *MFI = MF.getFrameInfo(); 3577 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3578 3579 // Offset of the first variable argument from stack pointer. 3580 int VaArgOffset; 3581 3582 if (NumRegs == Idx) 3583 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize); 3584 else 3585 VaArgOffset = 3586 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx)); 3587 3588 // Record the frame index of the first variable argument 3589 // which is a value necessary to VASTART. 3590 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3591 MipsFI->setVarArgsFrameIndex(FI); 3592 3593 // Copy the integer registers that have not been used for argument passing 3594 // to the argument register save area. For O32, the save area is allocated 3595 // in the caller's stack frame, while for N32/64, it is allocated in the 3596 // callee's stack frame. 3597 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) { 3598 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 3599 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 3600 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3601 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy()); 3602 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, 3603 MachinePointerInfo(), false, false, 0); 3604 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0); 3605 OutChains.push_back(Store); 3606 } 3607} 3608