MipsISelLowering.cpp revision 3706eda52c4565016959902a3f5aaf7271516286
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 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO"; 212 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO"; 213 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO"; 214 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO"; 215 case MipsISD::VCEQ: return "MipsISD::VCEQ"; 216 case MipsISD::VCLE_S: return "MipsISD::VCLE_S"; 217 case MipsISD::VCLE_U: return "MipsISD::VCLE_U"; 218 case MipsISD::VCLT_S: return "MipsISD::VCLT_S"; 219 case MipsISD::VCLT_U: return "MipsISD::VCLT_U"; 220 case MipsISD::VSMAX: return "MipsISD::VSMAX"; 221 case MipsISD::VSMIN: return "MipsISD::VSMIN"; 222 case MipsISD::VUMAX: return "MipsISD::VUMAX"; 223 case MipsISD::VUMIN: return "MipsISD::VUMIN"; 224 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT"; 225 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT"; 226 case MipsISD::VNOR: return "MipsISD::VNOR"; 227 case MipsISD::VSHF: return "MipsISD::VSHF"; 228 case MipsISD::SHF: return "MipsISD::SHF"; 229 case MipsISD::ILVEV: return "MipsISD::ILVEV"; 230 case MipsISD::ILVOD: return "MipsISD::ILVOD"; 231 case MipsISD::ILVL: return "MipsISD::ILVL"; 232 case MipsISD::ILVR: return "MipsISD::ILVR"; 233 case MipsISD::PCKEV: return "MipsISD::PCKEV"; 234 case MipsISD::PCKOD: return "MipsISD::PCKOD"; 235 default: return NULL; 236 } 237} 238 239MipsTargetLowering:: 240MipsTargetLowering(MipsTargetMachine &TM) 241 : TargetLowering(TM, new MipsTargetObjectFile()), 242 Subtarget(&TM.getSubtarget<MipsSubtarget>()), 243 HasMips64(Subtarget->hasMips64()), IsN64(Subtarget->isABI_N64()), 244 IsO32(Subtarget->isABI_O32()) { 245 // Mips does not have i1 type, so use i32 for 246 // setcc operations results (slt, sgt, ...). 247 setBooleanContents(ZeroOrOneBooleanContent); 248 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 249 250 // Load extented operations for i1 types must be promoted 251 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote); 252 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote); 253 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 254 255 // MIPS doesn't have extending float->double load/store 256 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 257 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 258 259 // Used by legalize types to correctly generate the setcc result. 260 // Without this, every float setcc comes with a AND/OR with the result, 261 // we don't want this, since the fpcmp result goes to a flag register, 262 // which is used implicitly by brcond and select operations. 263 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); 264 265 // Mips Custom Operations 266 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 267 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 268 setOperationAction(ISD::BlockAddress, MVT::i32, Custom); 269 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 270 setOperationAction(ISD::JumpTable, MVT::i32, Custom); 271 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 272 setOperationAction(ISD::SELECT, MVT::f32, Custom); 273 setOperationAction(ISD::SELECT, MVT::f64, Custom); 274 setOperationAction(ISD::SELECT, MVT::i32, Custom); 275 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 276 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 277 setOperationAction(ISD::SETCC, MVT::f32, Custom); 278 setOperationAction(ISD::SETCC, MVT::f64, Custom); 279 setOperationAction(ISD::BRCOND, MVT::Other, Custom); 280 setOperationAction(ISD::VASTART, MVT::Other, Custom); 281 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 282 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 283 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 284 285 if (!TM.Options.NoNaNsFPMath) { 286 setOperationAction(ISD::FABS, MVT::f32, Custom); 287 setOperationAction(ISD::FABS, MVT::f64, Custom); 288 } 289 290 if (HasMips64) { 291 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 292 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 293 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 294 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 295 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 296 setOperationAction(ISD::SELECT, MVT::i64, Custom); 297 setOperationAction(ISD::LOAD, MVT::i64, Custom); 298 setOperationAction(ISD::STORE, MVT::i64, Custom); 299 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 300 } 301 302 if (!HasMips64) { 303 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); 304 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); 305 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); 306 } 307 308 setOperationAction(ISD::ADD, MVT::i32, Custom); 309 if (HasMips64) 310 setOperationAction(ISD::ADD, MVT::i64, Custom); 311 312 setOperationAction(ISD::SDIV, MVT::i32, Expand); 313 setOperationAction(ISD::SREM, MVT::i32, Expand); 314 setOperationAction(ISD::UDIV, MVT::i32, Expand); 315 setOperationAction(ISD::UREM, MVT::i32, Expand); 316 setOperationAction(ISD::SDIV, MVT::i64, Expand); 317 setOperationAction(ISD::SREM, MVT::i64, Expand); 318 setOperationAction(ISD::UDIV, MVT::i64, Expand); 319 setOperationAction(ISD::UREM, MVT::i64, Expand); 320 321 // Operations not directly supported by Mips. 322 setOperationAction(ISD::BR_CC, MVT::f32, Expand); 323 setOperationAction(ISD::BR_CC, MVT::f64, Expand); 324 setOperationAction(ISD::BR_CC, MVT::i32, Expand); 325 setOperationAction(ISD::BR_CC, MVT::i64, Expand); 326 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand); 327 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); 328 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); 329 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); 330 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); 331 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 332 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 333 setOperationAction(ISD::CTPOP, MVT::i64, Expand); 334 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 335 setOperationAction(ISD::CTTZ, MVT::i64, Expand); 336 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand); 337 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand); 338 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand); 339 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand); 340 setOperationAction(ISD::ROTL, MVT::i32, Expand); 341 setOperationAction(ISD::ROTL, MVT::i64, Expand); 342 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 343 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 344 345 if (!Subtarget->hasMips32r2()) 346 setOperationAction(ISD::ROTR, MVT::i32, Expand); 347 348 if (!Subtarget->hasMips64r2()) 349 setOperationAction(ISD::ROTR, MVT::i64, Expand); 350 351 setOperationAction(ISD::FSIN, MVT::f32, Expand); 352 setOperationAction(ISD::FSIN, MVT::f64, Expand); 353 setOperationAction(ISD::FCOS, MVT::f32, Expand); 354 setOperationAction(ISD::FCOS, MVT::f64, Expand); 355 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 356 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 357 setOperationAction(ISD::FPOWI, MVT::f32, Expand); 358 setOperationAction(ISD::FPOW, MVT::f32, Expand); 359 setOperationAction(ISD::FPOW, MVT::f64, Expand); 360 setOperationAction(ISD::FLOG, MVT::f32, Expand); 361 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 362 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 363 setOperationAction(ISD::FEXP, MVT::f32, Expand); 364 setOperationAction(ISD::FMA, MVT::f32, Expand); 365 setOperationAction(ISD::FMA, MVT::f64, Expand); 366 setOperationAction(ISD::FREM, MVT::f32, Expand); 367 setOperationAction(ISD::FREM, MVT::f64, Expand); 368 369 if (!TM.Options.NoNaNsFPMath) { 370 setOperationAction(ISD::FNEG, MVT::f32, Expand); 371 setOperationAction(ISD::FNEG, MVT::f64, Expand); 372 } 373 374 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); 375 376 setOperationAction(ISD::VAARG, MVT::Other, Expand); 377 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 378 setOperationAction(ISD::VAEND, MVT::Other, Expand); 379 380 // Use the default for now 381 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 382 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 383 384 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand); 385 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); 386 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand); 387 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); 388 389 setInsertFencesForAtomic(true); 390 391 if (!Subtarget->hasSEInReg()) { 392 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 393 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 394 } 395 396 if (!Subtarget->hasBitCount()) { 397 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 398 setOperationAction(ISD::CTLZ, MVT::i64, Expand); 399 } 400 401 if (!Subtarget->hasSwap()) { 402 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 403 setOperationAction(ISD::BSWAP, MVT::i64, Expand); 404 } 405 406 if (HasMips64) { 407 setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Custom); 408 setLoadExtAction(ISD::ZEXTLOAD, MVT::i32, Custom); 409 setLoadExtAction(ISD::EXTLOAD, MVT::i32, Custom); 410 setTruncStoreAction(MVT::i64, MVT::i32, Custom); 411 } 412 413 setOperationAction(ISD::TRAP, MVT::Other, Legal); 414 415 setTargetDAGCombine(ISD::SDIVREM); 416 setTargetDAGCombine(ISD::UDIVREM); 417 setTargetDAGCombine(ISD::SELECT); 418 setTargetDAGCombine(ISD::AND); 419 setTargetDAGCombine(ISD::OR); 420 setTargetDAGCombine(ISD::ADD); 421 422 setMinFunctionAlignment(HasMips64 ? 3 : 2); 423 424 setStackPointerRegisterToSaveRestore(IsN64 ? Mips::SP_64 : Mips::SP); 425 426 setExceptionPointerRegister(IsN64 ? Mips::A0_64 : Mips::A0); 427 setExceptionSelectorRegister(IsN64 ? Mips::A1_64 : Mips::A1); 428 429 MaxStoresPerMemcpy = 16; 430} 431 432const MipsTargetLowering *MipsTargetLowering::create(MipsTargetMachine &TM) { 433 if (TM.getSubtargetImpl()->inMips16Mode()) 434 return llvm::createMips16TargetLowering(TM); 435 436 return llvm::createMipsSETargetLowering(TM); 437} 438 439EVT MipsTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const { 440 if (!VT.isVector()) 441 return MVT::i32; 442 return VT.changeVectorElementTypeToInteger(); 443} 444 445static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, 446 TargetLowering::DAGCombinerInfo &DCI, 447 const MipsSubtarget *Subtarget) { 448 if (DCI.isBeforeLegalizeOps()) 449 return SDValue(); 450 451 EVT Ty = N->getValueType(0); 452 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; 453 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; 454 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : 455 MipsISD::DivRemU16; 456 SDLoc DL(N); 457 458 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue, 459 N->getOperand(0), N->getOperand(1)); 460 SDValue InChain = DAG.getEntryNode(); 461 SDValue InGlue = DivRem; 462 463 // insert MFLO 464 if (N->hasAnyUseOfValue(0)) { 465 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty, 466 InGlue); 467 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); 468 InChain = CopyFromLo.getValue(1); 469 InGlue = CopyFromLo.getValue(2); 470 } 471 472 // insert MFHI 473 if (N->hasAnyUseOfValue(1)) { 474 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL, 475 HI, Ty, InGlue); 476 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); 477 } 478 479 return SDValue(); 480} 481 482static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { 483 switch (CC) { 484 default: llvm_unreachable("Unknown fp condition code!"); 485 case ISD::SETEQ: 486 case ISD::SETOEQ: return Mips::FCOND_OEQ; 487 case ISD::SETUNE: return Mips::FCOND_UNE; 488 case ISD::SETLT: 489 case ISD::SETOLT: return Mips::FCOND_OLT; 490 case ISD::SETGT: 491 case ISD::SETOGT: return Mips::FCOND_OGT; 492 case ISD::SETLE: 493 case ISD::SETOLE: return Mips::FCOND_OLE; 494 case ISD::SETGE: 495 case ISD::SETOGE: return Mips::FCOND_OGE; 496 case ISD::SETULT: return Mips::FCOND_ULT; 497 case ISD::SETULE: return Mips::FCOND_ULE; 498 case ISD::SETUGT: return Mips::FCOND_UGT; 499 case ISD::SETUGE: return Mips::FCOND_UGE; 500 case ISD::SETUO: return Mips::FCOND_UN; 501 case ISD::SETO: return Mips::FCOND_OR; 502 case ISD::SETNE: 503 case ISD::SETONE: return Mips::FCOND_ONE; 504 case ISD::SETUEQ: return Mips::FCOND_UEQ; 505 } 506} 507 508 509/// This function returns true if the floating point conditional branches and 510/// conditional moves which use condition code CC should be inverted. 511static bool invertFPCondCodeUser(Mips::CondCode CC) { 512 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) 513 return false; 514 515 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && 516 "Illegal Condition Code"); 517 518 return true; 519} 520 521// Creates and returns an FPCmp node from a setcc node. 522// Returns Op if setcc is not a floating point comparison. 523static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { 524 // must be a SETCC node 525 if (Op.getOpcode() != ISD::SETCC) 526 return Op; 527 528 SDValue LHS = Op.getOperand(0); 529 530 if (!LHS.getValueType().isFloatingPoint()) 531 return Op; 532 533 SDValue RHS = Op.getOperand(1); 534 SDLoc DL(Op); 535 536 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of 537 // node if necessary. 538 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 539 540 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS, 541 DAG.getConstant(condCodeToFCC(CC), MVT::i32)); 542} 543 544// Creates and returns a CMovFPT/F node. 545static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, 546 SDValue False, SDLoc DL) { 547 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2)); 548 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue()); 549 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 550 551 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, 552 True.getValueType(), True, FCC0, False, Cond); 553} 554 555static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, 556 TargetLowering::DAGCombinerInfo &DCI, 557 const MipsSubtarget *Subtarget) { 558 if (DCI.isBeforeLegalizeOps()) 559 return SDValue(); 560 561 SDValue SetCC = N->getOperand(0); 562 563 if ((SetCC.getOpcode() != ISD::SETCC) || 564 !SetCC.getOperand(0).getValueType().isInteger()) 565 return SDValue(); 566 567 SDValue False = N->getOperand(2); 568 EVT FalseTy = False.getValueType(); 569 570 if (!FalseTy.isInteger()) 571 return SDValue(); 572 573 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(False); 574 575 if (!CN || CN->getZExtValue()) 576 return SDValue(); 577 578 const SDLoc DL(N); 579 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get(); 580 SDValue True = N->getOperand(1); 581 582 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0), 583 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true)); 584 585 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True); 586} 587 588static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, 589 TargetLowering::DAGCombinerInfo &DCI, 590 const MipsSubtarget *Subtarget) { 591 // Pattern match EXT. 592 // $dst = and ((sra or srl) $src , pos), (2**size - 1) 593 // => ext $dst, $src, size, pos 594 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 595 return SDValue(); 596 597 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1); 598 unsigned ShiftRightOpc = ShiftRight.getOpcode(); 599 600 // Op's first operand must be a shift right. 601 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL) 602 return SDValue(); 603 604 // The second operand of the shift must be an immediate. 605 ConstantSDNode *CN; 606 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1)))) 607 return SDValue(); 608 609 uint64_t Pos = CN->getZExtValue(); 610 uint64_t SMPos, SMSize; 611 612 // Op's second operand must be a shifted mask. 613 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) || 614 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize)) 615 return SDValue(); 616 617 // Return if the shifted mask does not start at bit 0 or the sum of its size 618 // and Pos exceeds the word's size. 619 EVT ValTy = N->getValueType(0); 620 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) 621 return SDValue(); 622 623 return DAG.getNode(MipsISD::Ext, SDLoc(N), ValTy, 624 ShiftRight.getOperand(0), DAG.getConstant(Pos, MVT::i32), 625 DAG.getConstant(SMSize, MVT::i32)); 626} 627 628static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, 629 TargetLowering::DAGCombinerInfo &DCI, 630 const MipsSubtarget *Subtarget) { 631 // Pattern match INS. 632 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), 633 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 634 // => ins $dst, $src, size, pos, $src1 635 if (DCI.isBeforeLegalizeOps() || !Subtarget->hasMips32r2()) 636 return SDValue(); 637 638 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1); 639 uint64_t SMPos0, SMSize0, SMPos1, SMSize1; 640 ConstantSDNode *CN; 641 642 // See if Op's first operand matches (and $src1 , mask0). 643 if (And0.getOpcode() != ISD::AND) 644 return SDValue(); 645 646 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) || 647 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0)) 648 return SDValue(); 649 650 // See if Op's second operand matches (and (shl $src, pos), mask1). 651 if (And1.getOpcode() != ISD::AND) 652 return SDValue(); 653 654 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) || 655 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1)) 656 return SDValue(); 657 658 // The shift masks must have the same position and size. 659 if (SMPos0 != SMPos1 || SMSize0 != SMSize1) 660 return SDValue(); 661 662 SDValue Shl = And1.getOperand(0); 663 if (Shl.getOpcode() != ISD::SHL) 664 return SDValue(); 665 666 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1)))) 667 return SDValue(); 668 669 unsigned Shamt = CN->getZExtValue(); 670 671 // Return if the shift amount and the first bit position of mask are not the 672 // same. 673 EVT ValTy = N->getValueType(0); 674 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) 675 return SDValue(); 676 677 return DAG.getNode(MipsISD::Ins, SDLoc(N), ValTy, Shl.getOperand(0), 678 DAG.getConstant(SMPos0, MVT::i32), 679 DAG.getConstant(SMSize0, MVT::i32), And0.getOperand(0)); 680} 681 682static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, 683 TargetLowering::DAGCombinerInfo &DCI, 684 const MipsSubtarget *Subtarget) { 685 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) 686 687 if (DCI.isBeforeLegalizeOps()) 688 return SDValue(); 689 690 SDValue Add = N->getOperand(1); 691 692 if (Add.getOpcode() != ISD::ADD) 693 return SDValue(); 694 695 SDValue Lo = Add.getOperand(1); 696 697 if ((Lo.getOpcode() != MipsISD::Lo) || 698 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable)) 699 return SDValue(); 700 701 EVT ValTy = N->getValueType(0); 702 SDLoc DL(N); 703 704 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0), 705 Add.getOperand(0)); 706 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo); 707} 708 709SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) 710 const { 711 SelectionDAG &DAG = DCI.DAG; 712 unsigned Opc = N->getOpcode(); 713 714 switch (Opc) { 715 default: break; 716 case ISD::SDIVREM: 717 case ISD::UDIVREM: 718 return performDivRemCombine(N, DAG, DCI, Subtarget); 719 case ISD::SELECT: 720 return performSELECTCombine(N, DAG, DCI, Subtarget); 721 case ISD::AND: 722 return performANDCombine(N, DAG, DCI, Subtarget); 723 case ISD::OR: 724 return performORCombine(N, DAG, DCI, Subtarget); 725 case ISD::ADD: 726 return performADDCombine(N, DAG, DCI, Subtarget); 727 } 728 729 return SDValue(); 730} 731 732void 733MipsTargetLowering::LowerOperationWrapper(SDNode *N, 734 SmallVectorImpl<SDValue> &Results, 735 SelectionDAG &DAG) const { 736 SDValue Res = LowerOperation(SDValue(N, 0), DAG); 737 738 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) 739 Results.push_back(Res.getValue(I)); 740} 741 742void 743MipsTargetLowering::ReplaceNodeResults(SDNode *N, 744 SmallVectorImpl<SDValue> &Results, 745 SelectionDAG &DAG) const { 746 return LowerOperationWrapper(N, Results, DAG); 747} 748 749SDValue MipsTargetLowering:: 750LowerOperation(SDValue Op, SelectionDAG &DAG) const 751{ 752 switch (Op.getOpcode()) 753 { 754 case ISD::BR_JT: return lowerBR_JT(Op, DAG); 755 case ISD::BRCOND: return lowerBRCOND(Op, DAG); 756 case ISD::ConstantPool: return lowerConstantPool(Op, DAG); 757 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); 758 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); 759 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); 760 case ISD::JumpTable: return lowerJumpTable(Op, DAG); 761 case ISD::SELECT: return lowerSELECT(Op, DAG); 762 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG); 763 case ISD::SETCC: return lowerSETCC(Op, DAG); 764 case ISD::VASTART: return lowerVASTART(Op, DAG); 765 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); 766 case ISD::FABS: return lowerFABS(Op, DAG); 767 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); 768 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); 769 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); 770 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); 771 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); 772 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true); 773 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false); 774 case ISD::LOAD: return lowerLOAD(Op, DAG); 775 case ISD::STORE: return lowerSTORE(Op, DAG); 776 case ISD::ADD: return lowerADD(Op, DAG); 777 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); 778 } 779 return SDValue(); 780} 781 782//===----------------------------------------------------------------------===// 783// Lower helper functions 784//===----------------------------------------------------------------------===// 785 786// addLiveIn - This helper function adds the specified physical register to the 787// MachineFunction as a live in value. It also creates a corresponding 788// virtual register for it. 789static unsigned 790addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) 791{ 792 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC); 793 MF.getRegInfo().addLiveIn(PReg, VReg); 794 return VReg; 795} 796 797static MachineBasicBlock *expandPseudoDIV(MachineInstr *MI, 798 MachineBasicBlock &MBB, 799 const TargetInstrInfo &TII, 800 bool Is64Bit) { 801 if (NoZeroDivCheck) 802 return &MBB; 803 804 // Insert instruction "teq $divisor_reg, $zero, 7". 805 MachineBasicBlock::iterator I(MI); 806 MachineInstrBuilder MIB; 807 MIB = BuildMI(MBB, llvm::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ)) 808 .addOperand(MI->getOperand(2)).addReg(Mips::ZERO).addImm(7); 809 810 // Use the 32-bit sub-register if this is a 64-bit division. 811 if (Is64Bit) 812 MIB->getOperand(0).setSubReg(Mips::sub_32); 813 814 return &MBB; 815} 816 817MachineBasicBlock * 818MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 819 MachineBasicBlock *BB) const { 820 switch (MI->getOpcode()) { 821 default: 822 llvm_unreachable("Unexpected instr type to insert"); 823 case Mips::ATOMIC_LOAD_ADD_I8: 824 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu); 825 case Mips::ATOMIC_LOAD_ADD_I16: 826 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu); 827 case Mips::ATOMIC_LOAD_ADD_I32: 828 return emitAtomicBinary(MI, BB, 4, Mips::ADDu); 829 case Mips::ATOMIC_LOAD_ADD_I64: 830 return emitAtomicBinary(MI, BB, 8, Mips::DADDu); 831 832 case Mips::ATOMIC_LOAD_AND_I8: 833 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND); 834 case Mips::ATOMIC_LOAD_AND_I16: 835 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND); 836 case Mips::ATOMIC_LOAD_AND_I32: 837 return emitAtomicBinary(MI, BB, 4, Mips::AND); 838 case Mips::ATOMIC_LOAD_AND_I64: 839 return emitAtomicBinary(MI, BB, 8, Mips::AND64); 840 841 case Mips::ATOMIC_LOAD_OR_I8: 842 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR); 843 case Mips::ATOMIC_LOAD_OR_I16: 844 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR); 845 case Mips::ATOMIC_LOAD_OR_I32: 846 return emitAtomicBinary(MI, BB, 4, Mips::OR); 847 case Mips::ATOMIC_LOAD_OR_I64: 848 return emitAtomicBinary(MI, BB, 8, Mips::OR64); 849 850 case Mips::ATOMIC_LOAD_XOR_I8: 851 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR); 852 case Mips::ATOMIC_LOAD_XOR_I16: 853 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR); 854 case Mips::ATOMIC_LOAD_XOR_I32: 855 return emitAtomicBinary(MI, BB, 4, Mips::XOR); 856 case Mips::ATOMIC_LOAD_XOR_I64: 857 return emitAtomicBinary(MI, BB, 8, Mips::XOR64); 858 859 case Mips::ATOMIC_LOAD_NAND_I8: 860 return emitAtomicBinaryPartword(MI, BB, 1, 0, true); 861 case Mips::ATOMIC_LOAD_NAND_I16: 862 return emitAtomicBinaryPartword(MI, BB, 2, 0, true); 863 case Mips::ATOMIC_LOAD_NAND_I32: 864 return emitAtomicBinary(MI, BB, 4, 0, true); 865 case Mips::ATOMIC_LOAD_NAND_I64: 866 return emitAtomicBinary(MI, BB, 8, 0, true); 867 868 case Mips::ATOMIC_LOAD_SUB_I8: 869 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu); 870 case Mips::ATOMIC_LOAD_SUB_I16: 871 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu); 872 case Mips::ATOMIC_LOAD_SUB_I32: 873 return emitAtomicBinary(MI, BB, 4, Mips::SUBu); 874 case Mips::ATOMIC_LOAD_SUB_I64: 875 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu); 876 877 case Mips::ATOMIC_SWAP_I8: 878 return emitAtomicBinaryPartword(MI, BB, 1, 0); 879 case Mips::ATOMIC_SWAP_I16: 880 return emitAtomicBinaryPartword(MI, BB, 2, 0); 881 case Mips::ATOMIC_SWAP_I32: 882 return emitAtomicBinary(MI, BB, 4, 0); 883 case Mips::ATOMIC_SWAP_I64: 884 return emitAtomicBinary(MI, BB, 8, 0); 885 886 case Mips::ATOMIC_CMP_SWAP_I8: 887 return emitAtomicCmpSwapPartword(MI, BB, 1); 888 case Mips::ATOMIC_CMP_SWAP_I16: 889 return emitAtomicCmpSwapPartword(MI, BB, 2); 890 case Mips::ATOMIC_CMP_SWAP_I32: 891 return emitAtomicCmpSwap(MI, BB, 4); 892 case Mips::ATOMIC_CMP_SWAP_I64: 893 return emitAtomicCmpSwap(MI, BB, 8); 894 case Mips::PseudoSDIV: 895 case Mips::PseudoUDIV: 896 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), false); 897 case Mips::PseudoDSDIV: 898 case Mips::PseudoDUDIV: 899 return expandPseudoDIV(MI, *BB, *getTargetMachine().getInstrInfo(), true); 900 } 901} 902 903// This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and 904// Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) 905MachineBasicBlock * 906MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB, 907 unsigned Size, unsigned BinOpcode, 908 bool Nand) const { 909 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary."); 910 911 MachineFunction *MF = BB->getParent(); 912 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 913 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 914 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 915 DebugLoc DL = MI->getDebugLoc(); 916 unsigned LL, SC, AND, NOR, ZERO, BEQ; 917 918 if (Size == 4) { 919 LL = Mips::LL; 920 SC = Mips::SC; 921 AND = Mips::AND; 922 NOR = Mips::NOR; 923 ZERO = Mips::ZERO; 924 BEQ = Mips::BEQ; 925 } 926 else { 927 LL = Mips::LLD; 928 SC = Mips::SCD; 929 AND = Mips::AND64; 930 NOR = Mips::NOR64; 931 ZERO = Mips::ZERO_64; 932 BEQ = Mips::BEQ64; 933 } 934 935 unsigned OldVal = MI->getOperand(0).getReg(); 936 unsigned Ptr = MI->getOperand(1).getReg(); 937 unsigned Incr = MI->getOperand(2).getReg(); 938 939 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 940 unsigned AndRes = RegInfo.createVirtualRegister(RC); 941 unsigned Success = RegInfo.createVirtualRegister(RC); 942 943 // insert new blocks after the current block 944 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 945 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 946 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 947 MachineFunction::iterator It = BB; 948 ++It; 949 MF->insert(It, loopMBB); 950 MF->insert(It, exitMBB); 951 952 // Transfer the remainder of BB and its successor edges to exitMBB. 953 exitMBB->splice(exitMBB->begin(), BB, 954 llvm::next(MachineBasicBlock::iterator(MI)), 955 BB->end()); 956 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 957 958 // thisMBB: 959 // ... 960 // fallthrough --> loopMBB 961 BB->addSuccessor(loopMBB); 962 loopMBB->addSuccessor(loopMBB); 963 loopMBB->addSuccessor(exitMBB); 964 965 // loopMBB: 966 // ll oldval, 0(ptr) 967 // <binop> storeval, oldval, incr 968 // sc success, storeval, 0(ptr) 969 // beq success, $0, loopMBB 970 BB = loopMBB; 971 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0); 972 if (Nand) { 973 // and andres, oldval, incr 974 // nor storeval, $0, andres 975 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr); 976 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes); 977 } else if (BinOpcode) { 978 // <binop> storeval, oldval, incr 979 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr); 980 } else { 981 StoreVal = Incr; 982 } 983 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0); 984 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB); 985 986 MI->eraseFromParent(); // The instruction is gone now. 987 988 return exitMBB; 989} 990 991MachineBasicBlock * 992MipsTargetLowering::emitAtomicBinaryPartword(MachineInstr *MI, 993 MachineBasicBlock *BB, 994 unsigned Size, unsigned BinOpcode, 995 bool Nand) const { 996 assert((Size == 1 || Size == 2) && 997 "Unsupported size for EmitAtomicBinaryPartial."); 998 999 MachineFunction *MF = BB->getParent(); 1000 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1001 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1002 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1003 DebugLoc DL = MI->getDebugLoc(); 1004 1005 unsigned Dest = MI->getOperand(0).getReg(); 1006 unsigned Ptr = MI->getOperand(1).getReg(); 1007 unsigned Incr = MI->getOperand(2).getReg(); 1008 1009 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1010 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1011 unsigned Mask = RegInfo.createVirtualRegister(RC); 1012 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1013 unsigned NewVal = RegInfo.createVirtualRegister(RC); 1014 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1015 unsigned Incr2 = RegInfo.createVirtualRegister(RC); 1016 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1017 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1018 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1019 unsigned AndRes = RegInfo.createVirtualRegister(RC); 1020 unsigned BinOpRes = RegInfo.createVirtualRegister(RC); 1021 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1022 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1023 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1024 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1025 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1026 unsigned Success = RegInfo.createVirtualRegister(RC); 1027 1028 // insert new blocks after the current block 1029 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1030 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1031 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1032 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1033 MachineFunction::iterator It = BB; 1034 ++It; 1035 MF->insert(It, loopMBB); 1036 MF->insert(It, sinkMBB); 1037 MF->insert(It, exitMBB); 1038 1039 // Transfer the remainder of BB and its successor edges to exitMBB. 1040 exitMBB->splice(exitMBB->begin(), BB, 1041 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1042 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1043 1044 BB->addSuccessor(loopMBB); 1045 loopMBB->addSuccessor(loopMBB); 1046 loopMBB->addSuccessor(sinkMBB); 1047 sinkMBB->addSuccessor(exitMBB); 1048 1049 // thisMBB: 1050 // addiu masklsb2,$0,-4 # 0xfffffffc 1051 // and alignedaddr,ptr,masklsb2 1052 // andi ptrlsb2,ptr,3 1053 // sll shiftamt,ptrlsb2,3 1054 // ori maskupper,$0,255 # 0xff 1055 // sll mask,maskupper,shiftamt 1056 // nor mask2,$0,mask 1057 // sll incr2,incr,shiftamt 1058 1059 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1060 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1061 .addReg(Mips::ZERO).addImm(-4); 1062 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1063 .addReg(Ptr).addReg(MaskLSB2); 1064 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1065 if (Subtarget->isLittle()) { 1066 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1067 } else { 1068 unsigned Off = RegInfo.createVirtualRegister(RC); 1069 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1070 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1071 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1072 } 1073 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1074 .addReg(Mips::ZERO).addImm(MaskImm); 1075 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1076 .addReg(MaskUpper).addReg(ShiftAmt); 1077 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1078 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt); 1079 1080 // atomic.load.binop 1081 // loopMBB: 1082 // ll oldval,0(alignedaddr) 1083 // binop binopres,oldval,incr2 1084 // and newval,binopres,mask 1085 // and maskedoldval0,oldval,mask2 1086 // or storeval,maskedoldval0,newval 1087 // sc success,storeval,0(alignedaddr) 1088 // beq success,$0,loopMBB 1089 1090 // atomic.swap 1091 // loopMBB: 1092 // ll oldval,0(alignedaddr) 1093 // and newval,incr2,mask 1094 // and maskedoldval0,oldval,mask2 1095 // or storeval,maskedoldval0,newval 1096 // sc success,storeval,0(alignedaddr) 1097 // beq success,$0,loopMBB 1098 1099 BB = loopMBB; 1100 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0); 1101 if (Nand) { 1102 // and andres, oldval, incr2 1103 // nor binopres, $0, andres 1104 // and newval, binopres, mask 1105 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2); 1106 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes) 1107 .addReg(Mips::ZERO).addReg(AndRes); 1108 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1109 } else if (BinOpcode) { 1110 // <binop> binopres, oldval, incr2 1111 // and newval, binopres, mask 1112 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2); 1113 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask); 1114 } else {// atomic.swap 1115 // and newval, incr2, mask 1116 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask); 1117 } 1118 1119 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1120 .addReg(OldVal).addReg(Mask2); 1121 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1122 .addReg(MaskedOldVal0).addReg(NewVal); 1123 BuildMI(BB, DL, TII->get(Mips::SC), Success) 1124 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1125 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1126 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB); 1127 1128 // sinkMBB: 1129 // and maskedoldval1,oldval,mask 1130 // srl srlres,maskedoldval1,shiftamt 1131 // sll sllres,srlres,24 1132 // sra dest,sllres,24 1133 BB = sinkMBB; 1134 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1135 1136 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1137 .addReg(OldVal).addReg(Mask); 1138 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1139 .addReg(MaskedOldVal1).addReg(ShiftAmt); 1140 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1141 .addReg(SrlRes).addImm(ShiftImm); 1142 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1143 .addReg(SllRes).addImm(ShiftImm); 1144 1145 MI->eraseFromParent(); // The instruction is gone now. 1146 1147 return exitMBB; 1148} 1149 1150MachineBasicBlock * 1151MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI, 1152 MachineBasicBlock *BB, 1153 unsigned Size) const { 1154 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap."); 1155 1156 MachineFunction *MF = BB->getParent(); 1157 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1158 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8)); 1159 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1160 DebugLoc DL = MI->getDebugLoc(); 1161 unsigned LL, SC, ZERO, BNE, BEQ; 1162 1163 if (Size == 4) { 1164 LL = Mips::LL; 1165 SC = Mips::SC; 1166 ZERO = Mips::ZERO; 1167 BNE = Mips::BNE; 1168 BEQ = Mips::BEQ; 1169 } 1170 else { 1171 LL = Mips::LLD; 1172 SC = Mips::SCD; 1173 ZERO = Mips::ZERO_64; 1174 BNE = Mips::BNE64; 1175 BEQ = Mips::BEQ64; 1176 } 1177 1178 unsigned Dest = MI->getOperand(0).getReg(); 1179 unsigned Ptr = MI->getOperand(1).getReg(); 1180 unsigned OldVal = MI->getOperand(2).getReg(); 1181 unsigned NewVal = MI->getOperand(3).getReg(); 1182 1183 unsigned Success = RegInfo.createVirtualRegister(RC); 1184 1185 // insert new blocks after the current block 1186 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1187 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1188 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1189 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1190 MachineFunction::iterator It = BB; 1191 ++It; 1192 MF->insert(It, loop1MBB); 1193 MF->insert(It, loop2MBB); 1194 MF->insert(It, exitMBB); 1195 1196 // Transfer the remainder of BB and its successor edges to exitMBB. 1197 exitMBB->splice(exitMBB->begin(), BB, 1198 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1199 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1200 1201 // thisMBB: 1202 // ... 1203 // fallthrough --> loop1MBB 1204 BB->addSuccessor(loop1MBB); 1205 loop1MBB->addSuccessor(exitMBB); 1206 loop1MBB->addSuccessor(loop2MBB); 1207 loop2MBB->addSuccessor(loop1MBB); 1208 loop2MBB->addSuccessor(exitMBB); 1209 1210 // loop1MBB: 1211 // ll dest, 0(ptr) 1212 // bne dest, oldval, exitMBB 1213 BB = loop1MBB; 1214 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0); 1215 BuildMI(BB, DL, TII->get(BNE)) 1216 .addReg(Dest).addReg(OldVal).addMBB(exitMBB); 1217 1218 // loop2MBB: 1219 // sc success, newval, 0(ptr) 1220 // beq success, $0, loop1MBB 1221 BB = loop2MBB; 1222 BuildMI(BB, DL, TII->get(SC), Success) 1223 .addReg(NewVal).addReg(Ptr).addImm(0); 1224 BuildMI(BB, DL, TII->get(BEQ)) 1225 .addReg(Success).addReg(ZERO).addMBB(loop1MBB); 1226 1227 MI->eraseFromParent(); // The instruction is gone now. 1228 1229 return exitMBB; 1230} 1231 1232MachineBasicBlock * 1233MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI, 1234 MachineBasicBlock *BB, 1235 unsigned Size) const { 1236 assert((Size == 1 || Size == 2) && 1237 "Unsupported size for EmitAtomicCmpSwapPartial."); 1238 1239 MachineFunction *MF = BB->getParent(); 1240 MachineRegisterInfo &RegInfo = MF->getRegInfo(); 1241 const TargetRegisterClass *RC = getRegClassFor(MVT::i32); 1242 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 1243 DebugLoc DL = MI->getDebugLoc(); 1244 1245 unsigned Dest = MI->getOperand(0).getReg(); 1246 unsigned Ptr = MI->getOperand(1).getReg(); 1247 unsigned CmpVal = MI->getOperand(2).getReg(); 1248 unsigned NewVal = MI->getOperand(3).getReg(); 1249 1250 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC); 1251 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC); 1252 unsigned Mask = RegInfo.createVirtualRegister(RC); 1253 unsigned Mask2 = RegInfo.createVirtualRegister(RC); 1254 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC); 1255 unsigned OldVal = RegInfo.createVirtualRegister(RC); 1256 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC); 1257 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC); 1258 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC); 1259 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC); 1260 unsigned MaskUpper = RegInfo.createVirtualRegister(RC); 1261 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC); 1262 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC); 1263 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC); 1264 unsigned StoreVal = RegInfo.createVirtualRegister(RC); 1265 unsigned SrlRes = RegInfo.createVirtualRegister(RC); 1266 unsigned SllRes = RegInfo.createVirtualRegister(RC); 1267 unsigned Success = RegInfo.createVirtualRegister(RC); 1268 1269 // insert new blocks after the current block 1270 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 1271 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1272 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB); 1273 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1274 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB); 1275 MachineFunction::iterator It = BB; 1276 ++It; 1277 MF->insert(It, loop1MBB); 1278 MF->insert(It, loop2MBB); 1279 MF->insert(It, sinkMBB); 1280 MF->insert(It, exitMBB); 1281 1282 // Transfer the remainder of BB and its successor edges to exitMBB. 1283 exitMBB->splice(exitMBB->begin(), BB, 1284 llvm::next(MachineBasicBlock::iterator(MI)), BB->end()); 1285 exitMBB->transferSuccessorsAndUpdatePHIs(BB); 1286 1287 BB->addSuccessor(loop1MBB); 1288 loop1MBB->addSuccessor(sinkMBB); 1289 loop1MBB->addSuccessor(loop2MBB); 1290 loop2MBB->addSuccessor(loop1MBB); 1291 loop2MBB->addSuccessor(sinkMBB); 1292 sinkMBB->addSuccessor(exitMBB); 1293 1294 // FIXME: computation of newval2 can be moved to loop2MBB. 1295 // thisMBB: 1296 // addiu masklsb2,$0,-4 # 0xfffffffc 1297 // and alignedaddr,ptr,masklsb2 1298 // andi ptrlsb2,ptr,3 1299 // sll shiftamt,ptrlsb2,3 1300 // ori maskupper,$0,255 # 0xff 1301 // sll mask,maskupper,shiftamt 1302 // nor mask2,$0,mask 1303 // andi maskedcmpval,cmpval,255 1304 // sll shiftedcmpval,maskedcmpval,shiftamt 1305 // andi maskednewval,newval,255 1306 // sll shiftednewval,maskednewval,shiftamt 1307 int64_t MaskImm = (Size == 1) ? 255 : 65535; 1308 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2) 1309 .addReg(Mips::ZERO).addImm(-4); 1310 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr) 1311 .addReg(Ptr).addReg(MaskLSB2); 1312 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3); 1313 if (Subtarget->isLittle()) { 1314 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3); 1315 } else { 1316 unsigned Off = RegInfo.createVirtualRegister(RC); 1317 BuildMI(BB, DL, TII->get(Mips::XORi), Off) 1318 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2); 1319 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3); 1320 } 1321 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper) 1322 .addReg(Mips::ZERO).addImm(MaskImm); 1323 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask) 1324 .addReg(MaskUpper).addReg(ShiftAmt); 1325 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask); 1326 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal) 1327 .addReg(CmpVal).addImm(MaskImm); 1328 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal) 1329 .addReg(MaskedCmpVal).addReg(ShiftAmt); 1330 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal) 1331 .addReg(NewVal).addImm(MaskImm); 1332 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal) 1333 .addReg(MaskedNewVal).addReg(ShiftAmt); 1334 1335 // loop1MBB: 1336 // ll oldval,0(alginedaddr) 1337 // and maskedoldval0,oldval,mask 1338 // bne maskedoldval0,shiftedcmpval,sinkMBB 1339 BB = loop1MBB; 1340 BuildMI(BB, DL, TII->get(Mips::LL), OldVal).addReg(AlignedAddr).addImm(0); 1341 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0) 1342 .addReg(OldVal).addReg(Mask); 1343 BuildMI(BB, DL, TII->get(Mips::BNE)) 1344 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB); 1345 1346 // loop2MBB: 1347 // and maskedoldval1,oldval,mask2 1348 // or storeval,maskedoldval1,shiftednewval 1349 // sc success,storeval,0(alignedaddr) 1350 // beq success,$0,loop1MBB 1351 BB = loop2MBB; 1352 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1) 1353 .addReg(OldVal).addReg(Mask2); 1354 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal) 1355 .addReg(MaskedOldVal1).addReg(ShiftedNewVal); 1356 BuildMI(BB, DL, TII->get(Mips::SC), Success) 1357 .addReg(StoreVal).addReg(AlignedAddr).addImm(0); 1358 BuildMI(BB, DL, TII->get(Mips::BEQ)) 1359 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB); 1360 1361 // sinkMBB: 1362 // srl srlres,maskedoldval0,shiftamt 1363 // sll sllres,srlres,24 1364 // sra dest,sllres,24 1365 BB = sinkMBB; 1366 int64_t ShiftImm = (Size == 1) ? 24 : 16; 1367 1368 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes) 1369 .addReg(MaskedOldVal0).addReg(ShiftAmt); 1370 BuildMI(BB, DL, TII->get(Mips::SLL), SllRes) 1371 .addReg(SrlRes).addImm(ShiftImm); 1372 BuildMI(BB, DL, TII->get(Mips::SRA), Dest) 1373 .addReg(SllRes).addImm(ShiftImm); 1374 1375 MI->eraseFromParent(); // The instruction is gone now. 1376 1377 return exitMBB; 1378} 1379 1380//===----------------------------------------------------------------------===// 1381// Misc Lower Operation implementation 1382//===----------------------------------------------------------------------===// 1383SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const { 1384 SDValue Chain = Op.getOperand(0); 1385 SDValue Table = Op.getOperand(1); 1386 SDValue Index = Op.getOperand(2); 1387 SDLoc DL(Op); 1388 EVT PTy = getPointerTy(); 1389 unsigned EntrySize = 1390 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(*getDataLayout()); 1391 1392 Index = DAG.getNode(ISD::MUL, DL, PTy, Index, 1393 DAG.getConstant(EntrySize, PTy)); 1394 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table); 1395 1396 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); 1397 Addr = DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr, 1398 MachinePointerInfo::getJumpTable(), MemVT, false, false, 1399 0); 1400 Chain = Addr.getValue(1); 1401 1402 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || IsN64) { 1403 // For PIC, the sequence is: 1404 // BRIND(load(Jumptable + index) + RelocBase) 1405 // RelocBase can be JumpTable, GOT or some sort of global base. 1406 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr, 1407 getPICJumpTableRelocBase(Table, DAG)); 1408 } 1409 1410 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr); 1411} 1412 1413SDValue MipsTargetLowering:: 1414lowerBRCOND(SDValue Op, SelectionDAG &DAG) const 1415{ 1416 // The first operand is the chain, the second is the condition, the third is 1417 // the block to branch to if the condition is true. 1418 SDValue Chain = Op.getOperand(0); 1419 SDValue Dest = Op.getOperand(2); 1420 SDLoc DL(Op); 1421 1422 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1)); 1423 1424 // Return if flag is not set by a floating point comparison. 1425 if (CondRes.getOpcode() != MipsISD::FPCmp) 1426 return Op; 1427 1428 SDValue CCNode = CondRes.getOperand(2); 1429 Mips::CondCode CC = 1430 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue(); 1431 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; 1432 SDValue BrCode = DAG.getConstant(Opc, MVT::i32); 1433 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32); 1434 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode, 1435 FCC0, Dest, CondRes); 1436} 1437 1438SDValue MipsTargetLowering:: 1439lowerSELECT(SDValue Op, SelectionDAG &DAG) const 1440{ 1441 SDValue Cond = createFPCmp(DAG, Op.getOperand(0)); 1442 1443 // Return if flag is not set by a floating point comparison. 1444 if (Cond.getOpcode() != MipsISD::FPCmp) 1445 return Op; 1446 1447 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2), 1448 SDLoc(Op)); 1449} 1450 1451SDValue MipsTargetLowering:: 1452lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const 1453{ 1454 SDLoc DL(Op); 1455 EVT Ty = Op.getOperand(0).getValueType(); 1456 SDValue Cond = DAG.getNode(ISD::SETCC, DL, 1457 getSetCCResultType(*DAG.getContext(), Ty), 1458 Op.getOperand(0), Op.getOperand(1), 1459 Op.getOperand(4)); 1460 1461 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2), 1462 Op.getOperand(3)); 1463} 1464 1465SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { 1466 SDValue Cond = createFPCmp(DAG, Op); 1467 1468 assert(Cond.getOpcode() == MipsISD::FPCmp && 1469 "Floating point operand expected."); 1470 1471 SDValue True = DAG.getConstant(1, MVT::i32); 1472 SDValue False = DAG.getConstant(0, MVT::i32); 1473 1474 return createCMovFP(DAG, Cond, True, False, SDLoc(Op)); 1475} 1476 1477SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, 1478 SelectionDAG &DAG) const { 1479 // FIXME there isn't actually debug info here 1480 SDLoc DL(Op); 1481 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1482 1483 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) { 1484 const MipsTargetObjectFile &TLOF = 1485 (const MipsTargetObjectFile&)getObjFileLowering(); 1486 1487 // %gp_rel relocation 1488 if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) { 1489 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i32, 0, 1490 MipsII::MO_GPREL); 1491 SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, DL, 1492 DAG.getVTList(MVT::i32), &GA, 1); 1493 SDValue GPReg = DAG.getRegister(Mips::GP, MVT::i32); 1494 return DAG.getNode(ISD::ADD, DL, MVT::i32, GPReg, GPRelNode); 1495 } 1496 1497 // %hi/%lo relocation 1498 return getAddrNonPIC(Op, DAG); 1499 } 1500 1501 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))) 1502 return getAddrLocal(Op, DAG, HasMips64); 1503 1504 if (LargeGOT) 1505 return getAddrGlobalLargeGOT(Op, DAG, MipsII::MO_GOT_HI16, 1506 MipsII::MO_GOT_LO16); 1507 1508 return getAddrGlobal(Op, DAG, 1509 HasMips64 ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16); 1510} 1511 1512SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, 1513 SelectionDAG &DAG) const { 1514 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1515 return getAddrNonPIC(Op, DAG); 1516 1517 return getAddrLocal(Op, DAG, HasMips64); 1518} 1519 1520SDValue MipsTargetLowering:: 1521lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const 1522{ 1523 // If the relocation model is PIC, use the General Dynamic TLS Model or 1524 // Local Dynamic TLS model, otherwise use the Initial Exec or 1525 // Local Exec TLS Model. 1526 1527 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1528 SDLoc DL(GA); 1529 const GlobalValue *GV = GA->getGlobal(); 1530 EVT PtrVT = getPointerTy(); 1531 1532 TLSModel::Model model = getTargetMachine().getTLSModel(GV); 1533 1534 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { 1535 // General Dynamic and Local Dynamic TLS Model. 1536 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM 1537 : MipsII::MO_TLSGD; 1538 1539 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag); 1540 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, 1541 getGlobalReg(DAG, PtrVT), TGA); 1542 unsigned PtrSize = PtrVT.getSizeInBits(); 1543 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize); 1544 1545 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT); 1546 1547 ArgListTy Args; 1548 ArgListEntry Entry; 1549 Entry.Node = Argument; 1550 Entry.Ty = PtrTy; 1551 Args.push_back(Entry); 1552 1553 TargetLowering::CallLoweringInfo CLI(DAG.getEntryNode(), PtrTy, 1554 false, false, false, false, 0, CallingConv::C, 1555 /*IsTailCall=*/false, /*doesNotRet=*/false, 1556 /*isReturnValueUsed=*/true, 1557 TlsGetAddr, Args, DAG, DL); 1558 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 1559 1560 SDValue Ret = CallResult.first; 1561 1562 if (model != TLSModel::LocalDynamic) 1563 return Ret; 1564 1565 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1566 MipsII::MO_DTPREL_HI); 1567 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1568 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1569 MipsII::MO_DTPREL_LO); 1570 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1571 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret); 1572 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo); 1573 } 1574 1575 SDValue Offset; 1576 if (model == TLSModel::InitialExec) { 1577 // Initial Exec TLS Model 1578 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1579 MipsII::MO_GOTTPREL); 1580 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT), 1581 TGA); 1582 Offset = DAG.getLoad(PtrVT, DL, 1583 DAG.getEntryNode(), TGA, MachinePointerInfo(), 1584 false, false, false, 0); 1585 } else { 1586 // Local Exec TLS Model 1587 assert(model == TLSModel::LocalExec); 1588 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1589 MipsII::MO_TPREL_HI); 1590 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 1591 MipsII::MO_TPREL_LO); 1592 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi); 1593 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo); 1594 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); 1595 } 1596 1597 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT); 1598 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset); 1599} 1600 1601SDValue MipsTargetLowering:: 1602lowerJumpTable(SDValue Op, SelectionDAG &DAG) const 1603{ 1604 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1605 return getAddrNonPIC(Op, DAG); 1606 1607 return getAddrLocal(Op, DAG, HasMips64); 1608} 1609 1610SDValue MipsTargetLowering:: 1611lowerConstantPool(SDValue Op, SelectionDAG &DAG) const 1612{ 1613 // gp_rel relocation 1614 // FIXME: we should reference the constant pool using small data sections, 1615 // but the asm printer currently doesn't support this feature without 1616 // hacking it. This feature should come soon so we can uncomment the 1617 // stuff below. 1618 //if (IsInSmallSection(C->getType())) { 1619 // SDValue GPRelNode = DAG.getNode(MipsISD::GPRel, MVT::i32, CP); 1620 // SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32); 1621 // ResNode = DAG.getNode(ISD::ADD, MVT::i32, GOT, GPRelNode); 1622 1623 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !IsN64) 1624 return getAddrNonPIC(Op, DAG); 1625 1626 return getAddrLocal(Op, DAG, HasMips64); 1627} 1628 1629SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { 1630 MachineFunction &MF = DAG.getMachineFunction(); 1631 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); 1632 1633 SDLoc DL(Op); 1634 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), 1635 getPointerTy()); 1636 1637 // vastart just stores the address of the VarArgsFrameIndex slot into the 1638 // memory location argument. 1639 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1640 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), 1641 MachinePointerInfo(SV), false, false, 0); 1642} 1643 1644static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1645 EVT TyX = Op.getOperand(0).getValueType(); 1646 EVT TyY = Op.getOperand(1).getValueType(); 1647 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1648 SDValue Const31 = DAG.getConstant(31, MVT::i32); 1649 SDLoc DL(Op); 1650 SDValue Res; 1651 1652 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1653 // to i32. 1654 SDValue X = (TyX == MVT::f32) ? 1655 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1656 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1657 Const1); 1658 SDValue Y = (TyY == MVT::f32) ? 1659 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) : 1660 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1), 1661 Const1); 1662 1663 if (HasR2) { 1664 // ext E, Y, 31, 1 ; extract bit31 of Y 1665 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X 1666 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1); 1667 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X); 1668 } else { 1669 // sll SllX, X, 1 1670 // srl SrlX, SllX, 1 1671 // srl SrlY, Y, 31 1672 // sll SllY, SrlX, 31 1673 // or Or, SrlX, SllY 1674 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1675 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1676 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31); 1677 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31); 1678 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY); 1679 } 1680 1681 if (TyX == MVT::f32) 1682 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res); 1683 1684 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1685 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1686 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1687} 1688 1689static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1690 unsigned WidthX = Op.getOperand(0).getValueSizeInBits(); 1691 unsigned WidthY = Op.getOperand(1).getValueSizeInBits(); 1692 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY); 1693 SDValue Const1 = DAG.getConstant(1, MVT::i32); 1694 SDLoc DL(Op); 1695 1696 // Bitcast to integer nodes. 1697 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0)); 1698 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1)); 1699 1700 if (HasR2) { 1701 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y 1702 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X 1703 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y, 1704 DAG.getConstant(WidthY - 1, MVT::i32), Const1); 1705 1706 if (WidthX > WidthY) 1707 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E); 1708 else if (WidthY > WidthX) 1709 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E); 1710 1711 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E, 1712 DAG.getConstant(WidthX - 1, MVT::i32), Const1, X); 1713 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I); 1714 } 1715 1716 // (d)sll SllX, X, 1 1717 // (d)srl SrlX, SllX, 1 1718 // (d)srl SrlY, Y, width(Y)-1 1719 // (d)sll SllY, SrlX, width(Y)-1 1720 // or Or, SrlX, SllY 1721 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1); 1722 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1); 1723 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y, 1724 DAG.getConstant(WidthY - 1, MVT::i32)); 1725 1726 if (WidthX > WidthY) 1727 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY); 1728 else if (WidthY > WidthX) 1729 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY); 1730 1731 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY, 1732 DAG.getConstant(WidthX - 1, MVT::i32)); 1733 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY); 1734 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or); 1735} 1736 1737SDValue 1738MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { 1739 if (Subtarget->hasMips64()) 1740 return lowerFCOPYSIGN64(Op, DAG, Subtarget->hasMips32r2()); 1741 1742 return lowerFCOPYSIGN32(Op, DAG, Subtarget->hasMips32r2()); 1743} 1744 1745static SDValue lowerFABS32(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1746 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1747 SDLoc DL(Op); 1748 1749 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it 1750 // to i32. 1751 SDValue X = (Op.getValueType() == MVT::f32) ? 1752 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) : 1753 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0), 1754 Const1); 1755 1756 // Clear MSB. 1757 if (HasR2) 1758 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, 1759 DAG.getRegister(Mips::ZERO, MVT::i32), 1760 DAG.getConstant(31, MVT::i32), Const1, X); 1761 else { 1762 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1); 1763 Res = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1); 1764 } 1765 1766 if (Op.getValueType() == MVT::f32) 1767 return DAG.getNode(ISD::BITCAST, DL, MVT::f32, Res); 1768 1769 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 1770 Op.getOperand(0), DAG.getConstant(0, MVT::i32)); 1771 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res); 1772} 1773 1774static SDValue lowerFABS64(SDValue Op, SelectionDAG &DAG, bool HasR2) { 1775 SDValue Res, Const1 = DAG.getConstant(1, MVT::i32); 1776 SDLoc DL(Op); 1777 1778 // Bitcast to integer node. 1779 SDValue X = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Op.getOperand(0)); 1780 1781 // Clear MSB. 1782 if (HasR2) 1783 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i64, 1784 DAG.getRegister(Mips::ZERO_64, MVT::i64), 1785 DAG.getConstant(63, MVT::i32), Const1, X); 1786 else { 1787 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i64, X, Const1); 1788 Res = DAG.getNode(ISD::SRL, DL, MVT::i64, SllX, Const1); 1789 } 1790 1791 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, Res); 1792} 1793 1794SDValue 1795MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { 1796 if (Subtarget->hasMips64() && (Op.getValueType() == MVT::f64)) 1797 return lowerFABS64(Op, DAG, Subtarget->hasMips32r2()); 1798 1799 return lowerFABS32(Op, DAG, Subtarget->hasMips32r2()); 1800} 1801 1802SDValue MipsTargetLowering:: 1803lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { 1804 // check the depth 1805 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1806 "Frame address can only be determined for current frame."); 1807 1808 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1809 MFI->setFrameAddressIsTaken(true); 1810 EVT VT = Op.getValueType(); 1811 SDLoc DL(Op); 1812 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, 1813 IsN64 ? Mips::FP_64 : Mips::FP, VT); 1814 return FrameAddr; 1815} 1816 1817SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, 1818 SelectionDAG &DAG) const { 1819 // check the depth 1820 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) && 1821 "Return address can be determined only for current frame."); 1822 1823 MachineFunction &MF = DAG.getMachineFunction(); 1824 MachineFrameInfo *MFI = MF.getFrameInfo(); 1825 MVT VT = Op.getSimpleValueType(); 1826 unsigned RA = IsN64 ? Mips::RA_64 : Mips::RA; 1827 MFI->setReturnAddressIsTaken(true); 1828 1829 // Return RA, which contains the return address. Mark it an implicit live-in. 1830 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT)); 1831 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT); 1832} 1833 1834// An EH_RETURN is the result of lowering llvm.eh.return which in turn is 1835// generated from __builtin_eh_return (offset, handler) 1836// The effect of this is to adjust the stack pointer by "offset" 1837// and then branch to "handler". 1838SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) 1839 const { 1840 MachineFunction &MF = DAG.getMachineFunction(); 1841 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 1842 1843 MipsFI->setCallsEhReturn(); 1844 SDValue Chain = Op.getOperand(0); 1845 SDValue Offset = Op.getOperand(1); 1846 SDValue Handler = Op.getOperand(2); 1847 SDLoc DL(Op); 1848 EVT Ty = IsN64 ? MVT::i64 : MVT::i32; 1849 1850 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and 1851 // EH_RETURN nodes, so that instructions are emitted back-to-back. 1852 unsigned OffsetReg = IsN64 ? Mips::V1_64 : Mips::V1; 1853 unsigned AddrReg = IsN64 ? Mips::V0_64 : Mips::V0; 1854 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue()); 1855 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1)); 1856 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain, 1857 DAG.getRegister(OffsetReg, Ty), 1858 DAG.getRegister(AddrReg, getPointerTy()), 1859 Chain.getValue(1)); 1860} 1861 1862SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, 1863 SelectionDAG &DAG) const { 1864 // FIXME: Need pseudo-fence for 'singlethread' fences 1865 // FIXME: Set SType for weaker fences where supported/appropriate. 1866 unsigned SType = 0; 1867 SDLoc DL(Op); 1868 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0), 1869 DAG.getConstant(SType, MVT::i32)); 1870} 1871 1872SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, 1873 SelectionDAG &DAG) const { 1874 SDLoc DL(Op); 1875 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1876 SDValue Shamt = Op.getOperand(2); 1877 1878 // if shamt < 32: 1879 // lo = (shl lo, shamt) 1880 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt)) 1881 // else: 1882 // lo = 0 1883 // hi = (shl lo, shamt[4:0]) 1884 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1885 DAG.getConstant(-1, MVT::i32)); 1886 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, 1887 DAG.getConstant(1, MVT::i32)); 1888 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, ShiftRight1Lo, 1889 Not); 1890 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, Shamt); 1891 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1892 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, MVT::i32, Lo, Shamt); 1893 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1894 DAG.getConstant(0x20, MVT::i32)); 1895 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1896 DAG.getConstant(0, MVT::i32), ShiftLeftLo); 1897 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftLeftLo, Or); 1898 1899 SDValue Ops[2] = {Lo, Hi}; 1900 return DAG.getMergeValues(Ops, 2, DL); 1901} 1902 1903SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, 1904 bool IsSRA) const { 1905 SDLoc DL(Op); 1906 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1); 1907 SDValue Shamt = Op.getOperand(2); 1908 1909 // if shamt < 32: 1910 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt)) 1911 // if isSRA: 1912 // hi = (sra hi, shamt) 1913 // else: 1914 // hi = (srl hi, shamt) 1915 // else: 1916 // if isSRA: 1917 // lo = (sra hi, shamt[4:0]) 1918 // hi = (sra hi, 31) 1919 // else: 1920 // lo = (srl hi, shamt[4:0]) 1921 // hi = 0 1922 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt, 1923 DAG.getConstant(-1, MVT::i32)); 1924 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, MVT::i32, Hi, 1925 DAG.getConstant(1, MVT::i32)); 1926 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, MVT::i32, ShiftLeft1Hi, Not); 1927 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, MVT::i32, Lo, Shamt); 1928 SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, ShiftLeftHi, ShiftRightLo); 1929 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL, DL, MVT::i32, 1930 Hi, Shamt); 1931 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt, 1932 DAG.getConstant(0x20, MVT::i32)); 1933 SDValue Shift31 = DAG.getNode(ISD::SRA, DL, MVT::i32, Hi, 1934 DAG.getConstant(31, MVT::i32)); 1935 Lo = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, ShiftRightHi, Or); 1936 Hi = DAG.getNode(ISD::SELECT, DL, MVT::i32, Cond, 1937 IsSRA ? Shift31 : DAG.getConstant(0, MVT::i32), 1938 ShiftRightHi); 1939 1940 SDValue Ops[2] = {Lo, Hi}; 1941 return DAG.getMergeValues(Ops, 2, DL); 1942} 1943 1944static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, 1945 SDValue Chain, SDValue Src, unsigned Offset) { 1946 SDValue Ptr = LD->getBasePtr(); 1947 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT(); 1948 EVT BasePtrVT = Ptr.getValueType(); 1949 SDLoc DL(LD); 1950 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 1951 1952 if (Offset) 1953 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 1954 DAG.getConstant(Offset, BasePtrVT)); 1955 1956 SDValue Ops[] = { Chain, Ptr, Src }; 1957 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 1958 LD->getMemOperand()); 1959} 1960 1961// Expand an unaligned 32 or 64-bit integer load node. 1962SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { 1963 LoadSDNode *LD = cast<LoadSDNode>(Op); 1964 EVT MemVT = LD->getMemoryVT(); 1965 1966 // Return if load is aligned or if MemVT is neither i32 nor i64. 1967 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) || 1968 ((MemVT != MVT::i32) && (MemVT != MVT::i64))) 1969 return SDValue(); 1970 1971 bool IsLittle = Subtarget->isLittle(); 1972 EVT VT = Op.getValueType(); 1973 ISD::LoadExtType ExtType = LD->getExtensionType(); 1974 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); 1975 1976 assert((VT == MVT::i32) || (VT == MVT::i64)); 1977 1978 // Expand 1979 // (set dst, (i64 (load baseptr))) 1980 // to 1981 // (set tmp, (ldl (add baseptr, 7), undef)) 1982 // (set dst, (ldr baseptr, tmp)) 1983 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { 1984 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef, 1985 IsLittle ? 7 : 0); 1986 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL, 1987 IsLittle ? 0 : 7); 1988 } 1989 1990 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef, 1991 IsLittle ? 3 : 0); 1992 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL, 1993 IsLittle ? 0 : 3); 1994 1995 // Expand 1996 // (set dst, (i32 (load baseptr))) or 1997 // (set dst, (i64 (sextload baseptr))) or 1998 // (set dst, (i64 (extload baseptr))) 1999 // to 2000 // (set tmp, (lwl (add baseptr, 3), undef)) 2001 // (set dst, (lwr baseptr, tmp)) 2002 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || 2003 (ExtType == ISD::EXTLOAD)) 2004 return LWR; 2005 2006 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); 2007 2008 // Expand 2009 // (set dst, (i64 (zextload baseptr))) 2010 // to 2011 // (set tmp0, (lwl (add baseptr, 3), undef)) 2012 // (set tmp1, (lwr baseptr, tmp0)) 2013 // (set tmp2, (shl tmp1, 32)) 2014 // (set dst, (srl tmp2, 32)) 2015 SDLoc DL(LD); 2016 SDValue Const32 = DAG.getConstant(32, MVT::i32); 2017 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32); 2018 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32); 2019 SDValue Ops[] = { SRL, LWR.getValue(1) }; 2020 return DAG.getMergeValues(Ops, 2, DL); 2021} 2022 2023static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, 2024 SDValue Chain, unsigned Offset) { 2025 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); 2026 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); 2027 SDLoc DL(SD); 2028 SDVTList VTList = DAG.getVTList(MVT::Other); 2029 2030 if (Offset) 2031 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr, 2032 DAG.getConstant(Offset, BasePtrVT)); 2033 2034 SDValue Ops[] = { Chain, Value, Ptr }; 2035 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, 3, MemVT, 2036 SD->getMemOperand()); 2037} 2038 2039// Expand an unaligned 32 or 64-bit integer store node. 2040static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, 2041 bool IsLittle) { 2042 SDValue Value = SD->getValue(), Chain = SD->getChain(); 2043 EVT VT = Value.getValueType(); 2044 2045 // Expand 2046 // (store val, baseptr) or 2047 // (truncstore val, baseptr) 2048 // to 2049 // (swl val, (add baseptr, 3)) 2050 // (swr val, baseptr) 2051 if ((VT == MVT::i32) || SD->isTruncatingStore()) { 2052 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain, 2053 IsLittle ? 3 : 0); 2054 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3); 2055 } 2056 2057 assert(VT == MVT::i64); 2058 2059 // Expand 2060 // (store val, baseptr) 2061 // to 2062 // (sdl val, (add baseptr, 7)) 2063 // (sdr val, baseptr) 2064 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0); 2065 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7); 2066} 2067 2068// Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). 2069static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) { 2070 SDValue Val = SD->getValue(); 2071 2072 if (Val.getOpcode() != ISD::FP_TO_SINT) 2073 return SDValue(); 2074 2075 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits()); 2076 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy, 2077 Val.getOperand(0)); 2078 2079 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(), 2080 SD->getPointerInfo(), SD->isVolatile(), 2081 SD->isNonTemporal(), SD->getAlignment()); 2082} 2083 2084SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { 2085 StoreSDNode *SD = cast<StoreSDNode>(Op); 2086 EVT MemVT = SD->getMemoryVT(); 2087 2088 // Lower unaligned integer stores. 2089 if ((SD->getAlignment() < MemVT.getSizeInBits() / 8) && 2090 ((MemVT == MVT::i32) || (MemVT == MVT::i64))) 2091 return lowerUnalignedIntStore(SD, DAG, Subtarget->isLittle()); 2092 2093 return lowerFP_TO_SINT_STORE(SD, DAG); 2094} 2095 2096SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const { 2097 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR 2098 || cast<ConstantSDNode> 2099 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0 2100 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET) 2101 return SDValue(); 2102 2103 // The pattern 2104 // (add (frameaddr 0), (frame_to_args_offset)) 2105 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to 2106 // (add FrameObject, 0) 2107 // where FrameObject is a fixed StackObject with offset 0 which points to 2108 // the old stack pointer. 2109 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 2110 EVT ValTy = Op->getValueType(0); 2111 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false); 2112 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy); 2113 return DAG.getNode(ISD::ADD, SDLoc(Op), ValTy, InArgsAddr, 2114 DAG.getConstant(0, ValTy)); 2115} 2116 2117SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, 2118 SelectionDAG &DAG) const { 2119 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); 2120 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy, 2121 Op.getOperand(0)); 2122 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc); 2123} 2124 2125//===----------------------------------------------------------------------===// 2126// Calling Convention Implementation 2127//===----------------------------------------------------------------------===// 2128 2129//===----------------------------------------------------------------------===// 2130// TODO: Implement a generic logic using tblgen that can support this. 2131// Mips O32 ABI rules: 2132// --- 2133// i32 - Passed in A0, A1, A2, A3 and stack 2134// f32 - Only passed in f32 registers if no int reg has been used yet to hold 2135// an argument. Otherwise, passed in A1, A2, A3 and stack. 2136// f64 - Only passed in two aliased f32 registers if no int reg has been used 2137// yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is 2138// not used, it must be shadowed. If only A3 is avaiable, shadow it and 2139// go to stack. 2140// 2141// For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. 2142//===----------------------------------------------------------------------===// 2143 2144static bool CC_MipsO32(unsigned ValNo, MVT ValVT, 2145 MVT LocVT, CCValAssign::LocInfo LocInfo, 2146 ISD::ArgFlagsTy ArgFlags, CCState &State, 2147 const uint16_t *F64Regs) { 2148 2149 static const unsigned IntRegsSize=4, FloatRegsSize=2; 2150 2151 static const uint16_t IntRegs[] = { 2152 Mips::A0, Mips::A1, Mips::A2, Mips::A3 2153 }; 2154 static const uint16_t F32Regs[] = { 2155 Mips::F12, Mips::F14 2156 }; 2157 2158 // Do not process byval args here. 2159 if (ArgFlags.isByVal()) 2160 return true; 2161 2162 // Promote i8 and i16 2163 if (LocVT == MVT::i8 || LocVT == MVT::i16) { 2164 LocVT = MVT::i32; 2165 if (ArgFlags.isSExt()) 2166 LocInfo = CCValAssign::SExt; 2167 else if (ArgFlags.isZExt()) 2168 LocInfo = CCValAssign::ZExt; 2169 else 2170 LocInfo = CCValAssign::AExt; 2171 } 2172 2173 unsigned Reg; 2174 2175 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following 2176 // is true: function is vararg, argument is 3rd or higher, there is previous 2177 // argument which is not f32 or f64. 2178 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 2179 || State.getFirstUnallocated(F32Regs, FloatRegsSize) != ValNo; 2180 unsigned OrigAlign = ArgFlags.getOrigAlign(); 2181 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8); 2182 2183 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { 2184 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2185 // If this is the first part of an i64 arg, 2186 // the allocated register must be either A0 or A2. 2187 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) 2188 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2189 LocVT = MVT::i32; 2190 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { 2191 // Allocate int register and shadow next int register. If first 2192 // available register is Mips::A1 or Mips::A3, shadow it too. 2193 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2194 if (Reg == Mips::A1 || Reg == Mips::A3) 2195 Reg = State.AllocateReg(IntRegs, IntRegsSize); 2196 State.AllocateReg(IntRegs, IntRegsSize); 2197 LocVT = MVT::i32; 2198 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { 2199 // we are guaranteed to find an available float register 2200 if (ValVT == MVT::f32) { 2201 Reg = State.AllocateReg(F32Regs, FloatRegsSize); 2202 // Shadow int register 2203 State.AllocateReg(IntRegs, IntRegsSize); 2204 } else { 2205 Reg = State.AllocateReg(F64Regs, FloatRegsSize); 2206 // Shadow int registers 2207 unsigned Reg2 = State.AllocateReg(IntRegs, IntRegsSize); 2208 if (Reg2 == Mips::A1 || Reg2 == Mips::A3) 2209 State.AllocateReg(IntRegs, IntRegsSize); 2210 State.AllocateReg(IntRegs, IntRegsSize); 2211 } 2212 } else 2213 llvm_unreachable("Cannot handle this ValVT."); 2214 2215 if (!Reg) { 2216 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3, 2217 OrigAlign); 2218 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); 2219 } else 2220 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 2221 2222 return false; 2223} 2224 2225static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, 2226 MVT LocVT, CCValAssign::LocInfo LocInfo, 2227 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2228 static const uint16_t F64Regs[] = { Mips::D6, Mips::D7 }; 2229 2230 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2231} 2232 2233static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, 2234 MVT LocVT, CCValAssign::LocInfo LocInfo, 2235 ISD::ArgFlagsTy ArgFlags, CCState &State) { 2236 static const uint16_t F64Regs[] = { Mips::D12_64, Mips::D12_64 }; 2237 2238 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); 2239} 2240 2241#include "MipsGenCallingConv.inc" 2242 2243//===----------------------------------------------------------------------===// 2244// Call Calling Convention Implementation 2245//===----------------------------------------------------------------------===// 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, Subtarget->isFP64bit(), CCInfo, 2352 SpecialCallingConv); 2353 2354 MipsCCInfo.analyzeCallOperands(Outs, IsVarArg, 2355 Subtarget->mipsSEUsesSoftFloat(), 2356 Callee.getNode(), CLI.Args); 2357 2358 // Get a count of how many bytes are to be pushed on the stack. 2359 unsigned NextStackOffset = CCInfo.getNextStackOffset(); 2360 2361 // Check if it's really possible to do a tail call. 2362 if (IsTailCall) 2363 IsTailCall = 2364 isEligibleForTailCallOptimization(MipsCCInfo, NextStackOffset, 2365 *MF.getInfo<MipsFunctionInfo>()); 2366 2367 if (IsTailCall) 2368 ++NumTailCalls; 2369 2370 // Chain is the output chain of the last Load/Store or CopyToReg node. 2371 // ByValChain is the output chain of the last Memcpy node created for copying 2372 // byval arguments to the stack. 2373 unsigned StackAlignment = TFL->getStackAlignment(); 2374 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment); 2375 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true); 2376 2377 if (!IsTailCall) 2378 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL); 2379 2380 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, 2381 IsN64 ? Mips::SP_64 : Mips::SP, 2382 getPointerTy()); 2383 2384 // With EABI is it possible to have 16 args on registers. 2385 std::deque< std::pair<unsigned, SDValue> > RegsToPass; 2386 SmallVector<SDValue, 8> MemOpChains; 2387 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2388 2389 // Walk the register/memloc assignments, inserting copies/loads. 2390 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2391 SDValue Arg = OutVals[i]; 2392 CCValAssign &VA = ArgLocs[i]; 2393 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); 2394 ISD::ArgFlagsTy Flags = Outs[i].Flags; 2395 2396 // ByVal Arg. 2397 if (Flags.isByVal()) { 2398 assert(Flags.getByValSize() && 2399 "ByVal args of size 0 should have been ignored by front-end."); 2400 assert(ByValArg != MipsCCInfo.byval_end()); 2401 assert(!IsTailCall && 2402 "Do not tail-call optimize if there is a byval argument."); 2403 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, 2404 MipsCCInfo, *ByValArg, Flags, Subtarget->isLittle()); 2405 ++ByValArg; 2406 continue; 2407 } 2408 2409 // Promote the value if needed. 2410 switch (VA.getLocInfo()) { 2411 default: llvm_unreachable("Unknown loc info!"); 2412 case CCValAssign::Full: 2413 if (VA.isRegLoc()) { 2414 if ((ValVT == MVT::f32 && LocVT == MVT::i32) || 2415 (ValVT == MVT::f64 && LocVT == MVT::i64) || 2416 (ValVT == MVT::i64 && LocVT == MVT::f64)) 2417 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg); 2418 else if (ValVT == MVT::f64 && LocVT == MVT::i32) { 2419 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2420 Arg, DAG.getConstant(0, MVT::i32)); 2421 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, 2422 Arg, DAG.getConstant(1, MVT::i32)); 2423 if (!Subtarget->isLittle()) 2424 std::swap(Lo, Hi); 2425 unsigned LocRegLo = VA.getLocReg(); 2426 unsigned LocRegHigh = getNextIntArgReg(LocRegLo); 2427 RegsToPass.push_back(std::make_pair(LocRegLo, Lo)); 2428 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi)); 2429 continue; 2430 } 2431 } 2432 break; 2433 case CCValAssign::SExt: 2434 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg); 2435 break; 2436 case CCValAssign::ZExt: 2437 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg); 2438 break; 2439 case CCValAssign::AExt: 2440 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg); 2441 break; 2442 } 2443 2444 // Arguments that can be passed on register must be kept at 2445 // RegsToPass vector 2446 if (VA.isRegLoc()) { 2447 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 2448 continue; 2449 } 2450 2451 // Register can't get to this point... 2452 assert(VA.isMemLoc()); 2453 2454 // emit ISD::STORE whichs stores the 2455 // parameter value to a stack Location 2456 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(), 2457 Chain, Arg, DL, IsTailCall, DAG)); 2458 } 2459 2460 // Transform all store nodes into one single node because all store 2461 // nodes are independent of each other. 2462 if (!MemOpChains.empty()) 2463 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2464 &MemOpChains[0], MemOpChains.size()); 2465 2466 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 2467 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 2468 // node so that legalize doesn't hack it. 2469 bool IsPICCall = (IsN64 || IsPIC); // true if calls are translated to jalr $25 2470 bool GlobalOrExternal = false, InternalLinkage = false; 2471 SDValue CalleeLo; 2472 2473 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 2474 if (IsPICCall) { 2475 InternalLinkage = G->getGlobal()->hasInternalLinkage(); 2476 2477 if (InternalLinkage) 2478 Callee = getAddrLocal(Callee, DAG, HasMips64); 2479 else if (LargeGOT) 2480 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2481 MipsII::MO_CALL_LO16); 2482 else 2483 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2484 } else 2485 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, getPointerTy(), 0, 2486 MipsII::MO_NO_FLAG); 2487 GlobalOrExternal = true; 2488 } 2489 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 2490 if (!IsN64 && !IsPIC) // !N64 && static 2491 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(), 2492 MipsII::MO_NO_FLAG); 2493 else if (LargeGOT) 2494 Callee = getAddrGlobalLargeGOT(Callee, DAG, MipsII::MO_CALL_HI16, 2495 MipsII::MO_CALL_LO16); 2496 else // N64 || PIC 2497 Callee = getAddrGlobal(Callee, DAG, MipsII::MO_GOT_CALL); 2498 2499 GlobalOrExternal = true; 2500 } 2501 2502 SmallVector<SDValue, 8> Ops(1, Chain); 2503 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 2504 2505 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage, 2506 CLI, Callee, Chain); 2507 2508 if (IsTailCall) 2509 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, &Ops[0], Ops.size()); 2510 2511 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, &Ops[0], Ops.size()); 2512 SDValue InFlag = Chain.getValue(1); 2513 2514 // Create the CALLSEQ_END node. 2515 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal, 2516 DAG.getIntPtrConstant(0, true), InFlag, DL); 2517 InFlag = Chain.getValue(1); 2518 2519 // Handle result values, copying them out of physregs into vregs that we 2520 // return. 2521 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, 2522 Ins, DL, DAG, InVals, CLI.Callee.getNode(), CLI.RetTy); 2523} 2524 2525/// LowerCallResult - Lower the result values of a call into the 2526/// appropriate copies out of appropriate physical registers. 2527SDValue 2528MipsTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 2529 CallingConv::ID CallConv, bool IsVarArg, 2530 const SmallVectorImpl<ISD::InputArg> &Ins, 2531 SDLoc DL, SelectionDAG &DAG, 2532 SmallVectorImpl<SDValue> &InVals, 2533 const SDNode *CallNode, 2534 const Type *RetTy) const { 2535 // Assign locations to each value returned by this call. 2536 SmallVector<CCValAssign, 16> RVLocs; 2537 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2538 getTargetMachine(), RVLocs, *DAG.getContext()); 2539 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo); 2540 2541 MipsCCInfo.analyzeCallResult(Ins, Subtarget->mipsSEUsesSoftFloat(), 2542 CallNode, RetTy); 2543 2544 // Copy all of the result registers out of their specified physreg. 2545 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2546 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(), 2547 RVLocs[i].getLocVT(), InFlag); 2548 Chain = Val.getValue(1); 2549 InFlag = Val.getValue(2); 2550 2551 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2552 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getValVT(), Val); 2553 2554 InVals.push_back(Val); 2555 } 2556 2557 return Chain; 2558} 2559 2560//===----------------------------------------------------------------------===// 2561// Formal Arguments Calling Convention Implementation 2562//===----------------------------------------------------------------------===// 2563/// LowerFormalArguments - transform physical registers into virtual registers 2564/// and generate load operations for arguments places on the stack. 2565SDValue 2566MipsTargetLowering::LowerFormalArguments(SDValue Chain, 2567 CallingConv::ID CallConv, 2568 bool IsVarArg, 2569 const SmallVectorImpl<ISD::InputArg> &Ins, 2570 SDLoc DL, SelectionDAG &DAG, 2571 SmallVectorImpl<SDValue> &InVals) 2572 const { 2573 MachineFunction &MF = DAG.getMachineFunction(); 2574 MachineFrameInfo *MFI = MF.getFrameInfo(); 2575 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2576 2577 MipsFI->setVarArgsFrameIndex(0); 2578 2579 // Used with vargs to acumulate store chains. 2580 std::vector<SDValue> OutChains; 2581 2582 // Assign locations to all of the incoming arguments. 2583 SmallVector<CCValAssign, 16> ArgLocs; 2584 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), 2585 getTargetMachine(), ArgLocs, *DAG.getContext()); 2586 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo); 2587 Function::const_arg_iterator FuncArg = 2588 DAG.getMachineFunction().getFunction()->arg_begin(); 2589 bool UseSoftFloat = Subtarget->mipsSEUsesSoftFloat(); 2590 2591 MipsCCInfo.analyzeFormalArguments(Ins, UseSoftFloat, FuncArg); 2592 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(), 2593 MipsCCInfo.hasByValArg()); 2594 2595 unsigned CurArgIdx = 0; 2596 MipsCC::byval_iterator ByValArg = MipsCCInfo.byval_begin(); 2597 2598 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2599 CCValAssign &VA = ArgLocs[i]; 2600 std::advance(FuncArg, Ins[i].OrigArgIndex - CurArgIdx); 2601 CurArgIdx = Ins[i].OrigArgIndex; 2602 EVT ValVT = VA.getValVT(); 2603 ISD::ArgFlagsTy Flags = Ins[i].Flags; 2604 bool IsRegLoc = VA.isRegLoc(); 2605 2606 if (Flags.isByVal()) { 2607 assert(Flags.getByValSize() && 2608 "ByVal args of size 0 should have been ignored by front-end."); 2609 assert(ByValArg != MipsCCInfo.byval_end()); 2610 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg, 2611 MipsCCInfo, *ByValArg); 2612 ++ByValArg; 2613 continue; 2614 } 2615 2616 // Arguments stored on registers 2617 if (IsRegLoc) { 2618 EVT RegVT = VA.getLocVT(); 2619 unsigned ArgReg = VA.getLocReg(); 2620 const TargetRegisterClass *RC; 2621 2622 if (RegVT == MVT::i32) 2623 RC = Subtarget->inMips16Mode()? &Mips::CPU16RegsRegClass : 2624 &Mips::GPR32RegClass; 2625 else if (RegVT == MVT::i64) 2626 RC = &Mips::GPR64RegClass; 2627 else if (RegVT == MVT::f32) 2628 RC = &Mips::FGR32RegClass; 2629 else if (RegVT == MVT::f64) 2630 RC = Subtarget->isFP64bit() ? &Mips::FGR64RegClass : 2631 &Mips::AFGR64RegClass; 2632 else 2633 llvm_unreachable("RegVT not supported by FormalArguments Lowering"); 2634 2635 // Transform the arguments stored on 2636 // physical registers into virtual ones 2637 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC); 2638 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 2639 2640 // If this is an 8 or 16-bit value, it has been passed promoted 2641 // to 32 bits. Insert an assert[sz]ext to capture this, then 2642 // truncate to the right size. 2643 if (VA.getLocInfo() != CCValAssign::Full) { 2644 unsigned Opcode = 0; 2645 if (VA.getLocInfo() == CCValAssign::SExt) 2646 Opcode = ISD::AssertSext; 2647 else if (VA.getLocInfo() == CCValAssign::ZExt) 2648 Opcode = ISD::AssertZext; 2649 if (Opcode) 2650 ArgValue = DAG.getNode(Opcode, DL, RegVT, ArgValue, 2651 DAG.getValueType(ValVT)); 2652 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, ValVT, ArgValue); 2653 } 2654 2655 // Handle floating point arguments passed in integer registers and 2656 // long double arguments passed in floating point registers. 2657 if ((RegVT == MVT::i32 && ValVT == MVT::f32) || 2658 (RegVT == MVT::i64 && ValVT == MVT::f64) || 2659 (RegVT == MVT::f64 && ValVT == MVT::i64)) 2660 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue); 2661 else if (IsO32 && RegVT == MVT::i32 && ValVT == MVT::f64) { 2662 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(), 2663 getNextIntArgReg(ArgReg), RC); 2664 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT); 2665 if (!Subtarget->isLittle()) 2666 std::swap(ArgValue, ArgValue2); 2667 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, 2668 ArgValue, ArgValue2); 2669 } 2670 2671 InVals.push_back(ArgValue); 2672 } else { // VA.isRegLoc() 2673 2674 // sanity check 2675 assert(VA.isMemLoc()); 2676 2677 // The stack pointer offset is relative to the caller stack frame. 2678 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, 2679 VA.getLocMemOffset(), true); 2680 2681 // Create load nodes to retrieve arguments from the stack 2682 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 2683 InVals.push_back(DAG.getLoad(ValVT, DL, Chain, FIN, 2684 MachinePointerInfo::getFixedStack(FI), 2685 false, false, false, 0)); 2686 } 2687 } 2688 2689 // The mips ABIs for returning structs by value requires that we copy 2690 // the sret argument into $v0 for the return. Save the argument into 2691 // a virtual register so that we can access it from the return points. 2692 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr()) { 2693 unsigned Reg = MipsFI->getSRetReturnReg(); 2694 if (!Reg) { 2695 Reg = MF.getRegInfo(). 2696 createVirtualRegister(getRegClassFor(IsN64 ? MVT::i64 : MVT::i32)); 2697 MipsFI->setSRetReturnReg(Reg); 2698 } 2699 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]); 2700 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 2701 } 2702 2703 if (IsVarArg) 2704 writeVarArgRegs(OutChains, MipsCCInfo, Chain, DL, DAG); 2705 2706 // All stores are grouped in one node to allow the matching between 2707 // the size of Ins and InVals. This only happens when on varg functions 2708 if (!OutChains.empty()) { 2709 OutChains.push_back(Chain); 2710 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, 2711 &OutChains[0], OutChains.size()); 2712 } 2713 2714 return Chain; 2715} 2716 2717//===----------------------------------------------------------------------===// 2718// Return Value Calling Convention Implementation 2719//===----------------------------------------------------------------------===// 2720 2721bool 2722MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, 2723 MachineFunction &MF, bool IsVarArg, 2724 const SmallVectorImpl<ISD::OutputArg> &Outs, 2725 LLVMContext &Context) const { 2726 SmallVector<CCValAssign, 16> RVLocs; 2727 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), 2728 RVLocs, Context); 2729 return CCInfo.CheckReturn(Outs, RetCC_Mips); 2730} 2731 2732SDValue 2733MipsTargetLowering::LowerReturn(SDValue Chain, 2734 CallingConv::ID CallConv, bool IsVarArg, 2735 const SmallVectorImpl<ISD::OutputArg> &Outs, 2736 const SmallVectorImpl<SDValue> &OutVals, 2737 SDLoc DL, SelectionDAG &DAG) const { 2738 // CCValAssign - represent the assignment of 2739 // the return value to a location 2740 SmallVector<CCValAssign, 16> RVLocs; 2741 MachineFunction &MF = DAG.getMachineFunction(); 2742 2743 // CCState - Info about the registers and stack slot. 2744 CCState CCInfo(CallConv, IsVarArg, MF, getTargetMachine(), RVLocs, 2745 *DAG.getContext()); 2746 MipsCC MipsCCInfo(CallConv, IsO32, Subtarget->isFP64bit(), CCInfo); 2747 2748 // Analyze return values. 2749 MipsCCInfo.analyzeReturn(Outs, Subtarget->mipsSEUsesSoftFloat(), 2750 MF.getFunction()->getReturnType()); 2751 2752 SDValue Flag; 2753 SmallVector<SDValue, 4> RetOps(1, Chain); 2754 2755 // Copy the result values into the output registers. 2756 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2757 SDValue Val = OutVals[i]; 2758 CCValAssign &VA = RVLocs[i]; 2759 assert(VA.isRegLoc() && "Can only return in registers!"); 2760 2761 if (RVLocs[i].getValVT() != RVLocs[i].getLocVT()) 2762 Val = DAG.getNode(ISD::BITCAST, DL, RVLocs[i].getLocVT(), Val); 2763 2764 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag); 2765 2766 // Guarantee that all emitted copies are stuck together with flags. 2767 Flag = Chain.getValue(1); 2768 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); 2769 } 2770 2771 // The mips ABIs for returning structs by value requires that we copy 2772 // the sret argument into $v0 for the return. We saved the argument into 2773 // a virtual register in the entry block, so now we copy the value out 2774 // and into $v0. 2775 if (MF.getFunction()->hasStructRetAttr()) { 2776 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 2777 unsigned Reg = MipsFI->getSRetReturnReg(); 2778 2779 if (!Reg) 2780 llvm_unreachable("sret virtual register not created in the entry block"); 2781 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy()); 2782 unsigned V0 = IsN64 ? Mips::V0_64 : Mips::V0; 2783 2784 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag); 2785 Flag = Chain.getValue(1); 2786 RetOps.push_back(DAG.getRegister(V0, getPointerTy())); 2787 } 2788 2789 RetOps[0] = Chain; // Update chain. 2790 2791 // Add the flag if we have it. 2792 if (Flag.getNode()) 2793 RetOps.push_back(Flag); 2794 2795 // Return on Mips is always a "jr $ra" 2796 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, &RetOps[0], RetOps.size()); 2797} 2798 2799//===----------------------------------------------------------------------===// 2800// Mips Inline Assembly Support 2801//===----------------------------------------------------------------------===// 2802 2803/// getConstraintType - Given a constraint letter, return the type of 2804/// constraint it is for this target. 2805MipsTargetLowering::ConstraintType MipsTargetLowering:: 2806getConstraintType(const std::string &Constraint) const 2807{ 2808 // Mips specific constrainy 2809 // GCC config/mips/constraints.md 2810 // 2811 // 'd' : An address register. Equivalent to r 2812 // unless generating MIPS16 code. 2813 // 'y' : Equivalent to r; retained for 2814 // backwards compatibility. 2815 // 'c' : A register suitable for use in an indirect 2816 // jump. This will always be $25 for -mabicalls. 2817 // 'l' : The lo register. 1 word storage. 2818 // 'x' : The hilo register pair. Double word storage. 2819 if (Constraint.size() == 1) { 2820 switch (Constraint[0]) { 2821 default : break; 2822 case 'd': 2823 case 'y': 2824 case 'f': 2825 case 'c': 2826 case 'l': 2827 case 'x': 2828 return C_RegisterClass; 2829 case 'R': 2830 return C_Memory; 2831 } 2832 } 2833 return TargetLowering::getConstraintType(Constraint); 2834} 2835 2836/// Examine constraint type and operand type and determine a weight value. 2837/// This object must already have been set up with the operand type 2838/// and the current alternative constraint selected. 2839TargetLowering::ConstraintWeight 2840MipsTargetLowering::getSingleConstraintMatchWeight( 2841 AsmOperandInfo &info, const char *constraint) const { 2842 ConstraintWeight weight = CW_Invalid; 2843 Value *CallOperandVal = info.CallOperandVal; 2844 // If we don't have a value, we can't do a match, 2845 // but allow it at the lowest weight. 2846 if (CallOperandVal == NULL) 2847 return CW_Default; 2848 Type *type = CallOperandVal->getType(); 2849 // Look at the constraint type. 2850 switch (*constraint) { 2851 default: 2852 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 2853 break; 2854 case 'd': 2855 case 'y': 2856 if (type->isIntegerTy()) 2857 weight = CW_Register; 2858 break; 2859 case 'f': 2860 if (type->isFloatTy()) 2861 weight = CW_Register; 2862 break; 2863 case 'c': // $25 for indirect jumps 2864 case 'l': // lo register 2865 case 'x': // hilo register pair 2866 if (type->isIntegerTy()) 2867 weight = CW_SpecificReg; 2868 break; 2869 case 'I': // signed 16 bit immediate 2870 case 'J': // integer zero 2871 case 'K': // unsigned 16 bit immediate 2872 case 'L': // signed 32 bit immediate where lower 16 bits are 0 2873 case 'N': // immediate in the range of -65535 to -1 (inclusive) 2874 case 'O': // signed 15 bit immediate (+- 16383) 2875 case 'P': // immediate in the range of 65535 to 1 (inclusive) 2876 if (isa<ConstantInt>(CallOperandVal)) 2877 weight = CW_Constant; 2878 break; 2879 case 'R': 2880 weight = CW_Memory; 2881 break; 2882 } 2883 return weight; 2884} 2885 2886/// This is a helper function to parse a physical register string and split it 2887/// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag 2888/// that is returned indicates whether parsing was successful. The second flag 2889/// is true if the numeric part exists. 2890static std::pair<bool, bool> 2891parsePhysicalReg(const StringRef &C, std::string &Prefix, 2892 unsigned long long &Reg) { 2893 if (C.front() != '{' || C.back() != '}') 2894 return std::make_pair(false, false); 2895 2896 // Search for the first numeric character. 2897 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; 2898 I = std::find_if(B, E, std::ptr_fun(isdigit)); 2899 2900 Prefix.assign(B, I - B); 2901 2902 // The second flag is set to false if no numeric characters were found. 2903 if (I == E) 2904 return std::make_pair(true, false); 2905 2906 // Parse the numeric characters. 2907 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg), 2908 true); 2909} 2910 2911std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: 2912parseRegForInlineAsmConstraint(const StringRef &C, MVT VT) const { 2913 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 2914 const TargetRegisterClass *RC; 2915 std::string Prefix; 2916 unsigned long long Reg; 2917 2918 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); 2919 2920 if (!R.first) 2921 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2922 2923 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo. 2924 // No numeric characters follow "hi" or "lo". 2925 if (R.second) 2926 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2927 2928 RC = TRI->getRegClass(Prefix == "hi" ? 2929 Mips::HI32RegClassID : Mips::LO32RegClassID); 2930 return std::make_pair(*(RC->begin()), RC); 2931 } 2932 2933 if (!R.second) 2934 return std::make_pair((unsigned)0, (const TargetRegisterClass*)0); 2935 2936 if (Prefix == "$f") { // Parse $f0-$f31. 2937 // If the size of FP registers is 64-bit or Reg is an even number, select 2938 // the 64-bit register class. Otherwise, select the 32-bit register class. 2939 if (VT == MVT::Other) 2940 VT = (Subtarget->isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; 2941 2942 RC= getRegClassFor(VT); 2943 2944 if (RC == &Mips::AFGR64RegClass) { 2945 assert(Reg % 2 == 0); 2946 Reg >>= 1; 2947 } 2948 } else if (Prefix == "$fcc") { // Parse $fcc0-$fcc7. 2949 RC = TRI->getRegClass(Mips::FCCRegClassID); 2950 } else { // Parse $0-$31. 2951 assert(Prefix == "$"); 2952 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT); 2953 } 2954 2955 assert(Reg < RC->getNumRegs()); 2956 return std::make_pair(*(RC->begin() + Reg), RC); 2957} 2958 2959/// Given a register class constraint, like 'r', if this corresponds directly 2960/// to an LLVM register class, return a register of 0 and the register class 2961/// pointer. 2962std::pair<unsigned, const TargetRegisterClass*> MipsTargetLowering:: 2963getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const 2964{ 2965 if (Constraint.size() == 1) { 2966 switch (Constraint[0]) { 2967 case 'd': // Address register. Same as 'r' unless generating MIPS16 code. 2968 case 'y': // Same as 'r'. Exists for compatibility. 2969 case 'r': 2970 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { 2971 if (Subtarget->inMips16Mode()) 2972 return std::make_pair(0U, &Mips::CPU16RegsRegClass); 2973 return std::make_pair(0U, &Mips::GPR32RegClass); 2974 } 2975 if (VT == MVT::i64 && !HasMips64) 2976 return std::make_pair(0U, &Mips::GPR32RegClass); 2977 if (VT == MVT::i64 && HasMips64) 2978 return std::make_pair(0U, &Mips::GPR64RegClass); 2979 // This will generate an error message 2980 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2981 case 'f': 2982 if (VT == MVT::f32) 2983 return std::make_pair(0U, &Mips::FGR32RegClass); 2984 if ((VT == MVT::f64) && (!Subtarget->isSingleFloat())) { 2985 if (Subtarget->isFP64bit()) 2986 return std::make_pair(0U, &Mips::FGR64RegClass); 2987 return std::make_pair(0U, &Mips::AFGR64RegClass); 2988 } 2989 break; 2990 case 'c': // register suitable for indirect jump 2991 if (VT == MVT::i32) 2992 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass); 2993 assert(VT == MVT::i64 && "Unexpected type."); 2994 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass); 2995 case 'l': // register suitable for indirect jump 2996 if (VT == MVT::i32) 2997 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass); 2998 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass); 2999 case 'x': // register suitable for indirect jump 3000 // Fixme: Not triggering the use of both hi and low 3001 // This will generate an error message 3002 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 3003 } 3004 } 3005 3006 std::pair<unsigned, const TargetRegisterClass *> R; 3007 R = parseRegForInlineAsmConstraint(Constraint, VT); 3008 3009 if (R.second) 3010 return R; 3011 3012 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3013} 3014 3015/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3016/// vector. If it is invalid, don't add anything to Ops. 3017void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3018 std::string &Constraint, 3019 std::vector<SDValue>&Ops, 3020 SelectionDAG &DAG) const { 3021 SDValue Result(0, 0); 3022 3023 // Only support length 1 constraints for now. 3024 if (Constraint.length() > 1) return; 3025 3026 char ConstraintLetter = Constraint[0]; 3027 switch (ConstraintLetter) { 3028 default: break; // This will fall through to the generic implementation 3029 case 'I': // Signed 16 bit constant 3030 // If this fails, the parent routine will give an error 3031 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3032 EVT Type = Op.getValueType(); 3033 int64_t Val = C->getSExtValue(); 3034 if (isInt<16>(Val)) { 3035 Result = DAG.getTargetConstant(Val, Type); 3036 break; 3037 } 3038 } 3039 return; 3040 case 'J': // integer zero 3041 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3042 EVT Type = Op.getValueType(); 3043 int64_t Val = C->getZExtValue(); 3044 if (Val == 0) { 3045 Result = DAG.getTargetConstant(0, Type); 3046 break; 3047 } 3048 } 3049 return; 3050 case 'K': // unsigned 16 bit immediate 3051 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3052 EVT Type = Op.getValueType(); 3053 uint64_t Val = (uint64_t)C->getZExtValue(); 3054 if (isUInt<16>(Val)) { 3055 Result = DAG.getTargetConstant(Val, Type); 3056 break; 3057 } 3058 } 3059 return; 3060 case 'L': // signed 32 bit immediate where lower 16 bits are 0 3061 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3062 EVT Type = Op.getValueType(); 3063 int64_t Val = C->getSExtValue(); 3064 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){ 3065 Result = DAG.getTargetConstant(Val, Type); 3066 break; 3067 } 3068 } 3069 return; 3070 case 'N': // immediate in the range of -65535 to -1 (inclusive) 3071 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3072 EVT Type = Op.getValueType(); 3073 int64_t Val = C->getSExtValue(); 3074 if ((Val >= -65535) && (Val <= -1)) { 3075 Result = DAG.getTargetConstant(Val, Type); 3076 break; 3077 } 3078 } 3079 return; 3080 case 'O': // signed 15 bit immediate 3081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3082 EVT Type = Op.getValueType(); 3083 int64_t Val = C->getSExtValue(); 3084 if ((isInt<15>(Val))) { 3085 Result = DAG.getTargetConstant(Val, Type); 3086 break; 3087 } 3088 } 3089 return; 3090 case 'P': // immediate in the range of 1 to 65535 (inclusive) 3091 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 3092 EVT Type = Op.getValueType(); 3093 int64_t Val = C->getSExtValue(); 3094 if ((Val <= 65535) && (Val >= 1)) { 3095 Result = DAG.getTargetConstant(Val, Type); 3096 break; 3097 } 3098 } 3099 return; 3100 } 3101 3102 if (Result.getNode()) { 3103 Ops.push_back(Result); 3104 return; 3105 } 3106 3107 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 3108} 3109 3110bool 3111MipsTargetLowering::isLegalAddressingMode(const AddrMode &AM, Type *Ty) const { 3112 // No global is ever allowed as a base. 3113 if (AM.BaseGV) 3114 return false; 3115 3116 switch (AM.Scale) { 3117 case 0: // "r+i" or just "i", depending on HasBaseReg. 3118 break; 3119 case 1: 3120 if (!AM.HasBaseReg) // allow "r+i". 3121 break; 3122 return false; // disallow "r+r" or "r+r+i". 3123 default: 3124 return false; 3125 } 3126 3127 return true; 3128} 3129 3130bool 3131MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 3132 // The Mips target isn't yet aware of offsets. 3133 return false; 3134} 3135 3136EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign, 3137 unsigned SrcAlign, 3138 bool IsMemset, bool ZeroMemset, 3139 bool MemcpyStrSrc, 3140 MachineFunction &MF) const { 3141 if (Subtarget->hasMips64()) 3142 return MVT::i64; 3143 3144 return MVT::i32; 3145} 3146 3147bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { 3148 if (VT != MVT::f32 && VT != MVT::f64) 3149 return false; 3150 if (Imm.isNegZero()) 3151 return false; 3152 return Imm.isZero(); 3153} 3154 3155unsigned MipsTargetLowering::getJumpTableEncoding() const { 3156 if (IsN64) 3157 return MachineJumpTableInfo::EK_GPRel64BlockAddress; 3158 3159 return TargetLowering::getJumpTableEncoding(); 3160} 3161 3162/// This function returns true if CallSym is a long double emulation routine. 3163static bool isF128SoftLibCall(const char *CallSym) { 3164 const char *const LibCalls[] = 3165 {"__addtf3", "__divtf3", "__eqtf2", "__extenddftf2", "__extendsftf2", 3166 "__fixtfdi", "__fixtfsi", "__fixtfti", "__fixunstfdi", "__fixunstfsi", 3167 "__fixunstfti", "__floatditf", "__floatsitf", "__floattitf", 3168 "__floatunditf", "__floatunsitf", "__floatuntitf", "__getf2", "__gttf2", 3169 "__letf2", "__lttf2", "__multf3", "__netf2", "__powitf2", "__subtf3", 3170 "__trunctfdf2", "__trunctfsf2", "__unordtf2", 3171 "ceill", "copysignl", "cosl", "exp2l", "expl", "floorl", "fmal", "fmodl", 3172 "log10l", "log2l", "logl", "nearbyintl", "powl", "rintl", "sinl", "sqrtl", 3173 "truncl"}; 3174 3175 const char * const *End = LibCalls + array_lengthof(LibCalls); 3176 3177 // Check that LibCalls is sorted alphabetically. 3178 MipsTargetLowering::LTStr Comp; 3179 3180#ifndef NDEBUG 3181 for (const char * const *I = LibCalls; I < End - 1; ++I) 3182 assert(Comp(*I, *(I + 1))); 3183#endif 3184 3185 return std::binary_search(LibCalls, End, CallSym, Comp); 3186} 3187 3188/// This function returns true if Ty is fp128 or i128 which was originally a 3189/// fp128. 3190static bool originalTypeIsF128(const Type *Ty, const SDNode *CallNode) { 3191 if (Ty->isFP128Ty()) 3192 return true; 3193 3194 const ExternalSymbolSDNode *ES = 3195 dyn_cast_or_null<const ExternalSymbolSDNode>(CallNode); 3196 3197 // If the Ty is i128 and the function being called is a long double emulation 3198 // routine, then the original type is f128. 3199 return (ES && Ty->isIntegerTy(128) && isF128SoftLibCall(ES->getSymbol())); 3200} 3201 3202MipsTargetLowering::MipsCC::SpecialCallingConvType 3203 MipsTargetLowering::getSpecialCallingConv(SDValue Callee) const { 3204 MipsCC::SpecialCallingConvType SpecialCallingConv = 3205 MipsCC::NoSpecialCallingConv;; 3206 if (Subtarget->inMips16HardFloat()) { 3207 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 3208 llvm::StringRef Sym = G->getGlobal()->getName(); 3209 Function *F = G->getGlobal()->getParent()->getFunction(Sym); 3210 if (F->hasFnAttribute("__Mips16RetHelper")) { 3211 SpecialCallingConv = MipsCC::Mips16RetHelperConv; 3212 } 3213 } 3214 } 3215 return SpecialCallingConv; 3216} 3217 3218MipsTargetLowering::MipsCC::MipsCC( 3219 CallingConv::ID CC, bool IsO32_, bool IsFP64_, CCState &Info, 3220 MipsCC::SpecialCallingConvType SpecialCallingConv_) 3221 : CCInfo(Info), CallConv(CC), IsO32(IsO32_), IsFP64(IsFP64_), 3222 SpecialCallingConv(SpecialCallingConv_){ 3223 // Pre-allocate reserved argument area. 3224 CCInfo.AllocateStack(reservedArgArea(), 1); 3225} 3226 3227 3228void MipsTargetLowering::MipsCC:: 3229analyzeCallOperands(const SmallVectorImpl<ISD::OutputArg> &Args, 3230 bool IsVarArg, bool IsSoftFloat, const SDNode *CallNode, 3231 std::vector<ArgListEntry> &FuncArgs) { 3232 assert((CallConv != CallingConv::Fast || !IsVarArg) && 3233 "CallingConv::Fast shouldn't be used for vararg functions."); 3234 3235 unsigned NumOpnds = Args.size(); 3236 llvm::CCAssignFn *FixedFn = fixedArgFn(), *VarFn = varArgFn(); 3237 3238 for (unsigned I = 0; I != NumOpnds; ++I) { 3239 MVT ArgVT = Args[I].VT; 3240 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3241 bool R; 3242 3243 if (ArgFlags.isByVal()) { 3244 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3245 continue; 3246 } 3247 3248 if (IsVarArg && !Args[I].IsFixed) 3249 R = VarFn(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo); 3250 else { 3251 MVT RegVT = getRegVT(ArgVT, FuncArgs[Args[I].OrigArgIndex].Ty, CallNode, 3252 IsSoftFloat); 3253 R = FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo); 3254 } 3255 3256 if (R) { 3257#ifndef NDEBUG 3258 dbgs() << "Call operand #" << I << " has unhandled type " 3259 << EVT(ArgVT).getEVTString(); 3260#endif 3261 llvm_unreachable(0); 3262 } 3263 } 3264} 3265 3266void MipsTargetLowering::MipsCC:: 3267analyzeFormalArguments(const SmallVectorImpl<ISD::InputArg> &Args, 3268 bool IsSoftFloat, Function::const_arg_iterator FuncArg) { 3269 unsigned NumArgs = Args.size(); 3270 llvm::CCAssignFn *FixedFn = fixedArgFn(); 3271 unsigned CurArgIdx = 0; 3272 3273 for (unsigned I = 0; I != NumArgs; ++I) { 3274 MVT ArgVT = Args[I].VT; 3275 ISD::ArgFlagsTy ArgFlags = Args[I].Flags; 3276 std::advance(FuncArg, Args[I].OrigArgIndex - CurArgIdx); 3277 CurArgIdx = Args[I].OrigArgIndex; 3278 3279 if (ArgFlags.isByVal()) { 3280 handleByValArg(I, ArgVT, ArgVT, CCValAssign::Full, ArgFlags); 3281 continue; 3282 } 3283 3284 MVT RegVT = getRegVT(ArgVT, FuncArg->getType(), 0, IsSoftFloat); 3285 3286 if (!FixedFn(I, ArgVT, RegVT, CCValAssign::Full, ArgFlags, CCInfo)) 3287 continue; 3288 3289#ifndef NDEBUG 3290 dbgs() << "Formal Arg #" << I << " has unhandled type " 3291 << EVT(ArgVT).getEVTString(); 3292#endif 3293 llvm_unreachable(0); 3294 } 3295} 3296 3297template<typename Ty> 3298void MipsTargetLowering::MipsCC:: 3299analyzeReturn(const SmallVectorImpl<Ty> &RetVals, bool IsSoftFloat, 3300 const SDNode *CallNode, const Type *RetTy) const { 3301 CCAssignFn *Fn; 3302 3303 if (IsSoftFloat && originalTypeIsF128(RetTy, CallNode)) 3304 Fn = RetCC_F128Soft; 3305 else 3306 Fn = RetCC_Mips; 3307 3308 for (unsigned I = 0, E = RetVals.size(); I < E; ++I) { 3309 MVT VT = RetVals[I].VT; 3310 ISD::ArgFlagsTy Flags = RetVals[I].Flags; 3311 MVT RegVT = this->getRegVT(VT, RetTy, CallNode, IsSoftFloat); 3312 3313 if (Fn(I, VT, RegVT, CCValAssign::Full, Flags, this->CCInfo)) { 3314#ifndef NDEBUG 3315 dbgs() << "Call result #" << I << " has unhandled type " 3316 << EVT(VT).getEVTString() << '\n'; 3317#endif 3318 llvm_unreachable(0); 3319 } 3320 } 3321} 3322 3323void MipsTargetLowering::MipsCC:: 3324analyzeCallResult(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsSoftFloat, 3325 const SDNode *CallNode, const Type *RetTy) const { 3326 analyzeReturn(Ins, IsSoftFloat, CallNode, RetTy); 3327} 3328 3329void MipsTargetLowering::MipsCC:: 3330analyzeReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsSoftFloat, 3331 const Type *RetTy) const { 3332 analyzeReturn(Outs, IsSoftFloat, 0, RetTy); 3333} 3334 3335void 3336MipsTargetLowering::MipsCC::handleByValArg(unsigned ValNo, MVT ValVT, 3337 MVT LocVT, 3338 CCValAssign::LocInfo LocInfo, 3339 ISD::ArgFlagsTy ArgFlags) { 3340 assert(ArgFlags.getByValSize() && "Byval argument's size shouldn't be 0."); 3341 3342 struct ByValArgInfo ByVal; 3343 unsigned RegSize = regSize(); 3344 unsigned ByValSize = RoundUpToAlignment(ArgFlags.getByValSize(), RegSize); 3345 unsigned Align = std::min(std::max(ArgFlags.getByValAlign(), RegSize), 3346 RegSize * 2); 3347 3348 if (useRegsForByval()) 3349 allocateRegs(ByVal, ByValSize, Align); 3350 3351 // Allocate space on caller's stack. 3352 ByVal.Address = CCInfo.AllocateStack(ByValSize - RegSize * ByVal.NumRegs, 3353 Align); 3354 CCInfo.addLoc(CCValAssign::getMem(ValNo, ValVT, ByVal.Address, LocVT, 3355 LocInfo)); 3356 ByValArgs.push_back(ByVal); 3357} 3358 3359unsigned MipsTargetLowering::MipsCC::numIntArgRegs() const { 3360 return IsO32 ? array_lengthof(O32IntRegs) : array_lengthof(Mips64IntRegs); 3361} 3362 3363unsigned MipsTargetLowering::MipsCC::reservedArgArea() const { 3364 return (IsO32 && (CallConv != CallingConv::Fast)) ? 16 : 0; 3365} 3366 3367const uint16_t *MipsTargetLowering::MipsCC::intArgRegs() const { 3368 return IsO32 ? O32IntRegs : Mips64IntRegs; 3369} 3370 3371llvm::CCAssignFn *MipsTargetLowering::MipsCC::fixedArgFn() const { 3372 if (CallConv == CallingConv::Fast) 3373 return CC_Mips_FastCC; 3374 3375 if (SpecialCallingConv == Mips16RetHelperConv) 3376 return CC_Mips16RetHelper; 3377 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN; 3378} 3379 3380llvm::CCAssignFn *MipsTargetLowering::MipsCC::varArgFn() const { 3381 return IsO32 ? (IsFP64 ? CC_MipsO32_FP64 : CC_MipsO32_FP32) : CC_MipsN_VarArg; 3382} 3383 3384const uint16_t *MipsTargetLowering::MipsCC::shadowRegs() const { 3385 return IsO32 ? O32IntRegs : Mips64DPRegs; 3386} 3387 3388void MipsTargetLowering::MipsCC::allocateRegs(ByValArgInfo &ByVal, 3389 unsigned ByValSize, 3390 unsigned Align) { 3391 unsigned RegSize = regSize(), NumIntArgRegs = numIntArgRegs(); 3392 const uint16_t *IntArgRegs = intArgRegs(), *ShadowRegs = shadowRegs(); 3393 assert(!(ByValSize % RegSize) && !(Align % RegSize) && 3394 "Byval argument's size and alignment should be a multiple of" 3395 "RegSize."); 3396 3397 ByVal.FirstIdx = CCInfo.getFirstUnallocated(IntArgRegs, NumIntArgRegs); 3398 3399 // If Align > RegSize, the first arg register must be even. 3400 if ((Align > RegSize) && (ByVal.FirstIdx % 2)) { 3401 CCInfo.AllocateReg(IntArgRegs[ByVal.FirstIdx], ShadowRegs[ByVal.FirstIdx]); 3402 ++ByVal.FirstIdx; 3403 } 3404 3405 // Mark the registers allocated. 3406 for (unsigned I = ByVal.FirstIdx; ByValSize && (I < NumIntArgRegs); 3407 ByValSize -= RegSize, ++I, ++ByVal.NumRegs) 3408 CCInfo.AllocateReg(IntArgRegs[I], ShadowRegs[I]); 3409} 3410 3411MVT MipsTargetLowering::MipsCC::getRegVT(MVT VT, const Type *OrigTy, 3412 const SDNode *CallNode, 3413 bool IsSoftFloat) const { 3414 if (IsSoftFloat || IsO32) 3415 return VT; 3416 3417 // Check if the original type was fp128. 3418 if (originalTypeIsF128(OrigTy, CallNode)) { 3419 assert(VT == MVT::i64); 3420 return MVT::f64; 3421 } 3422 3423 return VT; 3424} 3425 3426void MipsTargetLowering:: 3427copyByValRegs(SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains, 3428 SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, 3429 SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, 3430 const MipsCC &CC, const ByValArgInfo &ByVal) const { 3431 MachineFunction &MF = DAG.getMachineFunction(); 3432 MachineFrameInfo *MFI = MF.getFrameInfo(); 3433 unsigned RegAreaSize = ByVal.NumRegs * CC.regSize(); 3434 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize); 3435 int FrameObjOffset; 3436 3437 if (RegAreaSize) 3438 FrameObjOffset = (int)CC.reservedArgArea() - 3439 (int)((CC.numIntArgRegs() - ByVal.FirstIdx) * CC.regSize()); 3440 else 3441 FrameObjOffset = ByVal.Address; 3442 3443 // Create frame object. 3444 EVT PtrTy = getPointerTy(); 3445 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true); 3446 SDValue FIN = DAG.getFrameIndex(FI, PtrTy); 3447 InVals.push_back(FIN); 3448 3449 if (!ByVal.NumRegs) 3450 return; 3451 3452 // Copy arg registers. 3453 MVT RegTy = MVT::getIntegerVT(CC.regSize() * 8); 3454 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3455 3456 for (unsigned I = 0; I < ByVal.NumRegs; ++I) { 3457 unsigned ArgReg = CC.intArgRegs()[ByVal.FirstIdx + I]; 3458 unsigned VReg = addLiveIn(MF, ArgReg, RC); 3459 unsigned Offset = I * CC.regSize(); 3460 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN, 3461 DAG.getConstant(Offset, PtrTy)); 3462 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy), 3463 StorePtr, MachinePointerInfo(FuncArg, Offset), 3464 false, false, 0); 3465 OutChains.push_back(Store); 3466 } 3467} 3468 3469// Copy byVal arg to registers and stack. 3470void MipsTargetLowering:: 3471passByValArg(SDValue Chain, SDLoc DL, 3472 std::deque< std::pair<unsigned, SDValue> > &RegsToPass, 3473 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, 3474 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg, 3475 const MipsCC &CC, const ByValArgInfo &ByVal, 3476 const ISD::ArgFlagsTy &Flags, bool isLittle) const { 3477 unsigned ByValSize = Flags.getByValSize(); 3478 unsigned Offset = 0; // Offset in # of bytes from the beginning of struct. 3479 unsigned RegSize = CC.regSize(); 3480 unsigned Alignment = std::min(Flags.getByValAlign(), RegSize); 3481 EVT PtrTy = getPointerTy(), RegTy = MVT::getIntegerVT(RegSize * 8); 3482 3483 if (ByVal.NumRegs) { 3484 const uint16_t *ArgRegs = CC.intArgRegs(); 3485 bool LeftoverBytes = (ByVal.NumRegs * RegSize > ByValSize); 3486 unsigned I = 0; 3487 3488 // Copy words to registers. 3489 for (; I < ByVal.NumRegs - LeftoverBytes; ++I, Offset += RegSize) { 3490 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3491 DAG.getConstant(Offset, PtrTy)); 3492 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr, 3493 MachinePointerInfo(), false, false, false, 3494 Alignment); 3495 MemOpChains.push_back(LoadVal.getValue(1)); 3496 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3497 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal)); 3498 } 3499 3500 // Return if the struct has been fully copied. 3501 if (ByValSize == Offset) 3502 return; 3503 3504 // Copy the remainder of the byval argument with sub-word loads and shifts. 3505 if (LeftoverBytes) { 3506 assert((ByValSize > Offset) && (ByValSize < Offset + RegSize) && 3507 "Size of the remainder should be smaller than RegSize."); 3508 SDValue Val; 3509 3510 for (unsigned LoadSize = RegSize / 2, TotalSizeLoaded = 0; 3511 Offset < ByValSize; LoadSize /= 2) { 3512 unsigned RemSize = ByValSize - Offset; 3513 3514 if (RemSize < LoadSize) 3515 continue; 3516 3517 // Load subword. 3518 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3519 DAG.getConstant(Offset, PtrTy)); 3520 SDValue LoadVal = 3521 DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, 3522 MachinePointerInfo(), MVT::getIntegerVT(LoadSize * 8), 3523 false, false, Alignment); 3524 MemOpChains.push_back(LoadVal.getValue(1)); 3525 3526 // Shift the loaded value. 3527 unsigned Shamt; 3528 3529 if (isLittle) 3530 Shamt = TotalSizeLoaded; 3531 else 3532 Shamt = (RegSize - (TotalSizeLoaded + LoadSize)) * 8; 3533 3534 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal, 3535 DAG.getConstant(Shamt, MVT::i32)); 3536 3537 if (Val.getNode()) 3538 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift); 3539 else 3540 Val = Shift; 3541 3542 Offset += LoadSize; 3543 TotalSizeLoaded += LoadSize; 3544 Alignment = std::min(Alignment, LoadSize); 3545 } 3546 3547 unsigned ArgReg = ArgRegs[ByVal.FirstIdx + I]; 3548 RegsToPass.push_back(std::make_pair(ArgReg, Val)); 3549 return; 3550 } 3551 } 3552 3553 // Copy remainder of byval arg to it with memcpy. 3554 unsigned MemCpySize = ByValSize - Offset; 3555 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg, 3556 DAG.getConstant(Offset, PtrTy)); 3557 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr, 3558 DAG.getIntPtrConstant(ByVal.Address)); 3559 Chain = DAG.getMemcpy(Chain, DL, Dst, Src, 3560 DAG.getConstant(MemCpySize, PtrTy), Alignment, 3561 /*isVolatile=*/false, /*AlwaysInline=*/false, 3562 MachinePointerInfo(0), MachinePointerInfo(0)); 3563 MemOpChains.push_back(Chain); 3564} 3565 3566void 3567MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, 3568 const MipsCC &CC, SDValue Chain, 3569 SDLoc DL, SelectionDAG &DAG) const { 3570 unsigned NumRegs = CC.numIntArgRegs(); 3571 const uint16_t *ArgRegs = CC.intArgRegs(); 3572 const CCState &CCInfo = CC.getCCInfo(); 3573 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs, NumRegs); 3574 unsigned RegSize = CC.regSize(); 3575 MVT RegTy = MVT::getIntegerVT(RegSize * 8); 3576 const TargetRegisterClass *RC = getRegClassFor(RegTy); 3577 MachineFunction &MF = DAG.getMachineFunction(); 3578 MachineFrameInfo *MFI = MF.getFrameInfo(); 3579 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); 3580 3581 // Offset of the first variable argument from stack pointer. 3582 int VaArgOffset; 3583 3584 if (NumRegs == Idx) 3585 VaArgOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), RegSize); 3586 else 3587 VaArgOffset = 3588 (int)CC.reservedArgArea() - (int)(RegSize * (NumRegs - Idx)); 3589 3590 // Record the frame index of the first variable argument 3591 // which is a value necessary to VASTART. 3592 int FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3593 MipsFI->setVarArgsFrameIndex(FI); 3594 3595 // Copy the integer registers that have not been used for argument passing 3596 // to the argument register save area. For O32, the save area is allocated 3597 // in the caller's stack frame, while for N32/64, it is allocated in the 3598 // callee's stack frame. 3599 for (unsigned I = Idx; I < NumRegs; ++I, VaArgOffset += RegSize) { 3600 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC); 3601 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy); 3602 FI = MFI->CreateFixedObject(RegSize, VaArgOffset, true); 3603 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy()); 3604 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, 3605 MachinePointerInfo(), false, false, 0); 3606 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(0); 3607 OutChains.push_back(Store); 3608 } 3609} 3610