| 1 | //===-- Mips16ISelLowering.h - Mips16 DAG Lowering Interface ----*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // Subclass of MipsTargetLowering specialized for mips16. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | #include "Mips16ISelLowering.h" |
| 13 | #include "MCTargetDesc/MipsBaseInfo.h" |
| 14 | #include "Mips16HardFloatInfo.h" |
| 15 | #include "MipsMachineFunction.h" |
| 16 | #include "MipsRegisterInfo.h" |
| 17 | #include "MipsTargetMachine.h" |
| 18 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 19 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 20 | #include "llvm/Support/CommandLine.h" |
| 21 | |
| 22 | using namespace llvm; |
| 23 | |
| 24 | #define DEBUG_TYPE "mips-lower" |
| 25 | |
| 26 | static cl::opt<bool> DontExpandCondPseudos16( |
| 27 | "mips16-dont-expand-cond-pseudo" , |
| 28 | cl::init(Val: false), |
| 29 | cl::desc("Don't expand conditional move related " |
| 30 | "pseudos for Mips 16" ), |
| 31 | cl::Hidden); |
| 32 | |
| 33 | namespace { |
| 34 | struct Mips16IntrinsicHelperType{ |
| 35 | const char* Name; |
| 36 | const char* Helper; |
| 37 | |
| 38 | bool operator<(const Mips16IntrinsicHelperType &RHS) const { |
| 39 | return std::strcmp(s1: Name, s2: RHS.Name) < 0; |
| 40 | } |
| 41 | bool operator==(const Mips16IntrinsicHelperType &RHS) const { |
| 42 | return std::strcmp(s1: Name, s2: RHS.Name) == 0; |
| 43 | } |
| 44 | }; |
| 45 | } // namespace |
| 46 | |
| 47 | static const Mips16IntrinsicHelperType Mips16IntrinsicHelper[] = { |
| 48 | {.Name: "__fixunsdfsi" , .Helper: "__mips16_call_stub_2" }, |
| 49 | {.Name: "ceil" , .Helper: "__mips16_call_stub_df_2" }, |
| 50 | {.Name: "ceilf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 51 | {.Name: "copysign" , .Helper: "__mips16_call_stub_df_10" }, |
| 52 | {.Name: "copysignf" , .Helper: "__mips16_call_stub_sf_5" }, |
| 53 | {.Name: "cos" , .Helper: "__mips16_call_stub_df_2" }, |
| 54 | {.Name: "cosf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 55 | {.Name: "exp2" , .Helper: "__mips16_call_stub_df_2" }, |
| 56 | {.Name: "exp2f" , .Helper: "__mips16_call_stub_sf_1" }, |
| 57 | {.Name: "floor" , .Helper: "__mips16_call_stub_df_2" }, |
| 58 | {.Name: "floorf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 59 | {.Name: "log2" , .Helper: "__mips16_call_stub_df_2" }, |
| 60 | {.Name: "log2f" , .Helper: "__mips16_call_stub_sf_1" }, |
| 61 | {.Name: "nearbyint" , .Helper: "__mips16_call_stub_df_2" }, |
| 62 | {.Name: "nearbyintf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 63 | {.Name: "rint" , .Helper: "__mips16_call_stub_df_2" }, |
| 64 | {.Name: "rintf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 65 | {.Name: "sin" , .Helper: "__mips16_call_stub_df_2" }, |
| 66 | {.Name: "sinf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 67 | {.Name: "sqrt" , .Helper: "__mips16_call_stub_df_2" }, |
| 68 | {.Name: "sqrtf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 69 | {.Name: "trunc" , .Helper: "__mips16_call_stub_df_2" }, |
| 70 | {.Name: "truncf" , .Helper: "__mips16_call_stub_sf_1" }, |
| 71 | }; |
| 72 | |
| 73 | Mips16TargetLowering::Mips16TargetLowering(const MipsTargetMachine &TM, |
| 74 | const MipsSubtarget &STI) |
| 75 | : MipsTargetLowering(TM, STI) { |
| 76 | |
| 77 | // Set up the register classes |
| 78 | addRegisterClass(VT: MVT::i32, RC: &Mips::CPU16RegsRegClass); |
| 79 | |
| 80 | setOperationAction(Op: ISD::ATOMIC_FENCE, VT: MVT::Other, Action: LibCall); |
| 81 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP, VT: MVT::i32, Action: LibCall); |
| 82 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT: MVT::i32, Action: LibCall); |
| 83 | setOperationAction(Op: ISD::ATOMIC_LOAD_ADD, VT: MVT::i32, Action: LibCall); |
| 84 | setOperationAction(Op: ISD::ATOMIC_LOAD_SUB, VT: MVT::i32, Action: LibCall); |
| 85 | setOperationAction(Op: ISD::ATOMIC_LOAD_AND, VT: MVT::i32, Action: LibCall); |
| 86 | setOperationAction(Op: ISD::ATOMIC_LOAD_OR, VT: MVT::i32, Action: LibCall); |
| 87 | setOperationAction(Op: ISD::ATOMIC_LOAD_XOR, VT: MVT::i32, Action: LibCall); |
| 88 | setOperationAction(Op: ISD::ATOMIC_LOAD_NAND, VT: MVT::i32, Action: LibCall); |
| 89 | setOperationAction(Op: ISD::ATOMIC_LOAD_MIN, VT: MVT::i32, Action: LibCall); |
| 90 | setOperationAction(Op: ISD::ATOMIC_LOAD_MAX, VT: MVT::i32, Action: LibCall); |
| 91 | setOperationAction(Op: ISD::ATOMIC_LOAD_UMIN, VT: MVT::i32, Action: LibCall); |
| 92 | setOperationAction(Op: ISD::ATOMIC_LOAD_UMAX, VT: MVT::i32, Action: LibCall); |
| 93 | |
| 94 | setOperationAction(Op: ISD::ROTR, VT: MVT::i32, Action: Expand); |
| 95 | setOperationAction(Op: ISD::ROTR, VT: MVT::i64, Action: Expand); |
| 96 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Expand); |
| 97 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i64, Action: Expand); |
| 98 | |
| 99 | computeRegisterProperties(TRI: STI.getRegisterInfo()); |
| 100 | } |
| 101 | |
| 102 | const MipsTargetLowering * |
| 103 | llvm::createMips16TargetLowering(const MipsTargetMachine &TM, |
| 104 | const MipsSubtarget &STI) { |
| 105 | return new Mips16TargetLowering(TM, STI); |
| 106 | } |
| 107 | |
| 108 | bool Mips16TargetLowering::allowsMisalignedMemoryAccesses( |
| 109 | EVT VT, unsigned, Align, MachineMemOperand::Flags, unsigned *Fast) const { |
| 110 | return false; |
| 111 | } |
| 112 | |
| 113 | MachineBasicBlock * |
| 114 | Mips16TargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
| 115 | MachineBasicBlock *BB) const { |
| 116 | switch (MI.getOpcode()) { |
| 117 | default: |
| 118 | return MipsTargetLowering::EmitInstrWithCustomInserter(MI, MBB: BB); |
| 119 | case Mips::SelBeqZ: |
| 120 | return emitSel16(Opc: Mips::BeqzRxImm16, MI, BB); |
| 121 | case Mips::SelBneZ: |
| 122 | return emitSel16(Opc: Mips::BnezRxImm16, MI, BB); |
| 123 | case Mips::SelTBteqZCmpi: |
| 124 | return emitSeliT16(Opc1: Mips::Bteqz16, Opc2: Mips::CmpiRxImmX16, MI, BB); |
| 125 | case Mips::SelTBteqZSlti: |
| 126 | return emitSeliT16(Opc1: Mips::Bteqz16, Opc2: Mips::SltiRxImmX16, MI, BB); |
| 127 | case Mips::SelTBteqZSltiu: |
| 128 | return emitSeliT16(Opc1: Mips::Bteqz16, Opc2: Mips::SltiuRxImmX16, MI, BB); |
| 129 | case Mips::SelTBtneZCmpi: |
| 130 | return emitSeliT16(Opc1: Mips::Btnez16, Opc2: Mips::CmpiRxImmX16, MI, BB); |
| 131 | case Mips::SelTBtneZSlti: |
| 132 | return emitSeliT16(Opc1: Mips::Btnez16, Opc2: Mips::SltiRxImmX16, MI, BB); |
| 133 | case Mips::SelTBtneZSltiu: |
| 134 | return emitSeliT16(Opc1: Mips::Btnez16, Opc2: Mips::SltiuRxImmX16, MI, BB); |
| 135 | case Mips::SelTBteqZCmp: |
| 136 | return emitSelT16(Opc1: Mips::Bteqz16, Opc2: Mips::CmpRxRy16, MI, BB); |
| 137 | case Mips::SelTBteqZSlt: |
| 138 | return emitSelT16(Opc1: Mips::Bteqz16, Opc2: Mips::SltRxRy16, MI, BB); |
| 139 | case Mips::SelTBteqZSltu: |
| 140 | return emitSelT16(Opc1: Mips::Bteqz16, Opc2: Mips::SltuRxRy16, MI, BB); |
| 141 | case Mips::SelTBtneZCmp: |
| 142 | return emitSelT16(Opc1: Mips::Btnez16, Opc2: Mips::CmpRxRy16, MI, BB); |
| 143 | case Mips::SelTBtneZSlt: |
| 144 | return emitSelT16(Opc1: Mips::Btnez16, Opc2: Mips::SltRxRy16, MI, BB); |
| 145 | case Mips::SelTBtneZSltu: |
| 146 | return emitSelT16(Opc1: Mips::Btnez16, Opc2: Mips::SltuRxRy16, MI, BB); |
| 147 | case Mips::BteqzT8CmpX16: |
| 148 | return emitFEXT_T8I816_ins(BtOpc: Mips::Bteqz16, CmpOpc: Mips::CmpRxRy16, MI, BB); |
| 149 | case Mips::BteqzT8SltX16: |
| 150 | return emitFEXT_T8I816_ins(BtOpc: Mips::Bteqz16, CmpOpc: Mips::SltRxRy16, MI, BB); |
| 151 | case Mips::BteqzT8SltuX16: |
| 152 | // TBD: figure out a way to get this or remove the instruction |
| 153 | // altogether. |
| 154 | return emitFEXT_T8I816_ins(BtOpc: Mips::Bteqz16, CmpOpc: Mips::SltuRxRy16, MI, BB); |
| 155 | case Mips::BtnezT8CmpX16: |
| 156 | return emitFEXT_T8I816_ins(BtOpc: Mips::Btnez16, CmpOpc: Mips::CmpRxRy16, MI, BB); |
| 157 | case Mips::BtnezT8SltX16: |
| 158 | return emitFEXT_T8I816_ins(BtOpc: Mips::Btnez16, CmpOpc: Mips::SltRxRy16, MI, BB); |
| 159 | case Mips::BtnezT8SltuX16: |
| 160 | // TBD: figure out a way to get this or remove the instruction |
| 161 | // altogether. |
| 162 | return emitFEXT_T8I816_ins(BtOpc: Mips::Btnez16, CmpOpc: Mips::SltuRxRy16, MI, BB); |
| 163 | case Mips::BteqzT8CmpiX16: return emitFEXT_T8I8I16_ins( |
| 164 | BtOpc: Mips::Bteqz16, CmpiOpc: Mips::CmpiRxImm16, CmpiXOpc: Mips::CmpiRxImmX16, ImmSigned: false, MI, BB); |
| 165 | case Mips::BteqzT8SltiX16: return emitFEXT_T8I8I16_ins( |
| 166 | BtOpc: Mips::Bteqz16, CmpiOpc: Mips::SltiRxImm16, CmpiXOpc: Mips::SltiRxImmX16, ImmSigned: true, MI, BB); |
| 167 | case Mips::BteqzT8SltiuX16: return emitFEXT_T8I8I16_ins( |
| 168 | BtOpc: Mips::Bteqz16, CmpiOpc: Mips::SltiuRxImm16, CmpiXOpc: Mips::SltiuRxImmX16, ImmSigned: false, MI, BB); |
| 169 | case Mips::BtnezT8CmpiX16: return emitFEXT_T8I8I16_ins( |
| 170 | BtOpc: Mips::Btnez16, CmpiOpc: Mips::CmpiRxImm16, CmpiXOpc: Mips::CmpiRxImmX16, ImmSigned: false, MI, BB); |
| 171 | case Mips::BtnezT8SltiX16: return emitFEXT_T8I8I16_ins( |
| 172 | BtOpc: Mips::Btnez16, CmpiOpc: Mips::SltiRxImm16, CmpiXOpc: Mips::SltiRxImmX16, ImmSigned: true, MI, BB); |
| 173 | case Mips::BtnezT8SltiuX16: return emitFEXT_T8I8I16_ins( |
| 174 | BtOpc: Mips::Btnez16, CmpiOpc: Mips::SltiuRxImm16, CmpiXOpc: Mips::SltiuRxImmX16, ImmSigned: false, MI, BB); |
| 175 | break; |
| 176 | case Mips::SltCCRxRy16: |
| 177 | return emitFEXT_CCRX16_ins(SltOpc: Mips::SltRxRy16, MI, BB); |
| 178 | break; |
| 179 | case Mips::SltiCCRxImmX16: |
| 180 | return emitFEXT_CCRXI16_ins |
| 181 | (SltiOpc: Mips::SltiRxImm16, SltiXOpc: Mips::SltiRxImmX16, MI, BB); |
| 182 | case Mips::SltiuCCRxImmX16: |
| 183 | return emitFEXT_CCRXI16_ins |
| 184 | (SltiOpc: Mips::SltiuRxImm16, SltiXOpc: Mips::SltiuRxImmX16, MI, BB); |
| 185 | case Mips::SltuCCRxRy16: |
| 186 | return emitFEXT_CCRX16_ins |
| 187 | (SltOpc: Mips::SltuRxRy16, MI, BB); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | bool Mips16TargetLowering::isEligibleForTailCallOptimization( |
| 192 | const CCState &CCInfo, unsigned NextStackOffset, |
| 193 | const MipsFunctionInfo &FI) const { |
| 194 | // No tail call optimization for mips16. |
| 195 | return false; |
| 196 | } |
| 197 | |
| 198 | // |
| 199 | // The Mips16 hard float is a crazy quilt inherited from gcc. I have a much |
| 200 | // cleaner way to do all of this but it will have to wait until the traditional |
| 201 | // gcc mechanism is completed. |
| 202 | // |
| 203 | // For Pic, in order for Mips16 code to call Mips32 code which according the abi |
| 204 | // have either arguments or returned values placed in floating point registers, |
| 205 | // we use a set of helper functions. (This includes functions which return type |
| 206 | // complex which on Mips are returned in a pair of floating point registers). |
| 207 | // |
| 208 | // This is an encoding that we inherited from gcc. |
| 209 | // In Mips traditional O32, N32 ABI, floating point numbers are passed in |
| 210 | // floating point argument registers 1,2 only when the first and optionally |
| 211 | // the second arguments are float (sf) or double (df). |
| 212 | // For Mips16 we are only concerned with the situations where floating point |
| 213 | // arguments are being passed in floating point registers by the ABI, because |
| 214 | // Mips16 mode code cannot execute floating point instructions to load those |
| 215 | // values and hence helper functions are needed. |
| 216 | // The possibilities are (), (sf), (sf, sf), (sf, df), (df), (df, sf), (df, df) |
| 217 | // the helper function suffixs for these are: |
| 218 | // 0, 1, 5, 9, 2, 6, 10 |
| 219 | // this suffix can then be calculated as follows: |
| 220 | // for a given argument Arg: |
| 221 | // Arg1x, Arg2x = 1 : Arg is sf |
| 222 | // 2 : Arg is df |
| 223 | // 0: Arg is neither sf or df |
| 224 | // So this stub is the string for number Arg1x + Arg2x*4. |
| 225 | // However not all numbers between 0 and 10 are possible, we check anyway and |
| 226 | // assert if the impossible exists. |
| 227 | // |
| 228 | |
| 229 | unsigned int Mips16TargetLowering::getMips16HelperFunctionStubNumber |
| 230 | (ArgListTy &Args) const { |
| 231 | unsigned int resultNum = 0; |
| 232 | if (Args.size() >= 1) { |
| 233 | Type *t = Args[0].Ty; |
| 234 | if (t->isFloatTy()) { |
| 235 | resultNum = 1; |
| 236 | } |
| 237 | else if (t->isDoubleTy()) { |
| 238 | resultNum = 2; |
| 239 | } |
| 240 | } |
| 241 | if (resultNum) { |
| 242 | if (Args.size() >=2) { |
| 243 | Type *t = Args[1].Ty; |
| 244 | if (t->isFloatTy()) { |
| 245 | resultNum += 4; |
| 246 | } |
| 247 | else if (t->isDoubleTy()) { |
| 248 | resultNum += 8; |
| 249 | } |
| 250 | } |
| 251 | } |
| 252 | return resultNum; |
| 253 | } |
| 254 | |
| 255 | // |
| 256 | // Prefixes are attached to stub numbers depending on the return type. |
| 257 | // return type: float sf_ |
| 258 | // double df_ |
| 259 | // single complex sc_ |
| 260 | // double complext dc_ |
| 261 | // others NO PREFIX |
| 262 | // |
| 263 | // |
| 264 | // The full name of a helper function is__mips16_call_stub + |
| 265 | // return type dependent prefix + stub number |
| 266 | // |
| 267 | // FIXME: This is something that probably should be in a different source file |
| 268 | // and perhaps done differently but my main purpose is to not waste runtime |
| 269 | // on something that we can enumerate in the source. Another possibility is |
| 270 | // to have a python script to generate these mapping tables. This will do |
| 271 | // for now. There are a whole series of helper function mapping arrays, one |
| 272 | // for each return type class as outlined above. There there are 11 possible |
| 273 | // entries. Ones with 0 are ones which should never be selected. |
| 274 | // |
| 275 | // All the arrays are similar except for ones which return neither |
| 276 | // sf, df, sc, dc, in which we only care about ones which have sf or df as a |
| 277 | // first parameter. |
| 278 | // |
| 279 | #define P_ "__mips16_call_stub_" |
| 280 | #define MAX_STUB_NUMBER 10 |
| 281 | #define T1 P "1", P "2", 0, 0, P "5", P "6", 0, 0, P "9", P "10" |
| 282 | #define T P "0" , T1 |
| 283 | #define P P_ |
| 284 | static char const * vMips16Helper[MAX_STUB_NUMBER+1] = |
| 285 | {nullptr, T1 }; |
| 286 | #undef P |
| 287 | #define P P_ "sf_" |
| 288 | static char const * sfMips16Helper[MAX_STUB_NUMBER+1] = |
| 289 | { T }; |
| 290 | #undef P |
| 291 | #define P P_ "df_" |
| 292 | static char const * dfMips16Helper[MAX_STUB_NUMBER+1] = |
| 293 | { T }; |
| 294 | #undef P |
| 295 | #define P P_ "sc_" |
| 296 | static char const * scMips16Helper[MAX_STUB_NUMBER+1] = |
| 297 | { T }; |
| 298 | #undef P |
| 299 | #define P P_ "dc_" |
| 300 | static char const * dcMips16Helper[MAX_STUB_NUMBER+1] = |
| 301 | { T }; |
| 302 | #undef P |
| 303 | #undef P_ |
| 304 | |
| 305 | |
| 306 | const char* Mips16TargetLowering:: |
| 307 | getMips16HelperFunction |
| 308 | (Type* RetTy, ArgListTy &Args, bool &needHelper) const { |
| 309 | const unsigned int stubNum = getMips16HelperFunctionStubNumber(Args); |
| 310 | #ifndef NDEBUG |
| 311 | const unsigned int maxStubNum = 10; |
| 312 | assert(stubNum <= maxStubNum); |
| 313 | const bool validStubNum[maxStubNum+1] = |
| 314 | {true, true, true, false, false, true, true, false, false, true, true}; |
| 315 | assert(validStubNum[stubNum]); |
| 316 | #endif |
| 317 | const char *result; |
| 318 | if (RetTy->isFloatTy()) { |
| 319 | result = sfMips16Helper[stubNum]; |
| 320 | } |
| 321 | else if (RetTy ->isDoubleTy()) { |
| 322 | result = dfMips16Helper[stubNum]; |
| 323 | } else if (StructType *SRetTy = dyn_cast<StructType>(Val: RetTy)) { |
| 324 | // check if it's complex |
| 325 | if (SRetTy->getNumElements() == 2) { |
| 326 | if ((SRetTy->getElementType(N: 0)->isFloatTy()) && |
| 327 | (SRetTy->getElementType(N: 1)->isFloatTy())) { |
| 328 | result = scMips16Helper[stubNum]; |
| 329 | } else if ((SRetTy->getElementType(N: 0)->isDoubleTy()) && |
| 330 | (SRetTy->getElementType(N: 1)->isDoubleTy())) { |
| 331 | result = dcMips16Helper[stubNum]; |
| 332 | } else { |
| 333 | llvm_unreachable("Uncovered condition" ); |
| 334 | } |
| 335 | } else { |
| 336 | llvm_unreachable("Uncovered condition" ); |
| 337 | } |
| 338 | } else { |
| 339 | if (stubNum == 0) { |
| 340 | needHelper = false; |
| 341 | return "" ; |
| 342 | } |
| 343 | result = vMips16Helper[stubNum]; |
| 344 | } |
| 345 | needHelper = true; |
| 346 | return result; |
| 347 | } |
| 348 | |
| 349 | static bool isMips16HardFloatLibcall(StringRef Name) { |
| 350 | // FIXME: Use getSupportedLibcallImpl instead of blindly parsing the name. |
| 351 | iota_range<RTLIB::LibcallImpl> ParsedLibcalls = |
| 352 | RTLIB::RuntimeLibcallsInfo::lookupLibcallImplName(Name); |
| 353 | return !ParsedLibcalls.empty() && |
| 354 | binary_search(Range: MipsSubtarget::HardFloatLibCalls, |
| 355 | Value: *ParsedLibcalls.begin()); |
| 356 | } |
| 357 | |
| 358 | void Mips16TargetLowering:: |
| 359 | getOpndList(SmallVectorImpl<SDValue> &Ops, |
| 360 | std::deque< std::pair<unsigned, SDValue> > &RegsToPass, |
| 361 | bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, |
| 362 | bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, |
| 363 | SDValue Chain) const { |
| 364 | SelectionDAG &DAG = CLI.DAG; |
| 365 | MachineFunction &MF = DAG.getMachineFunction(); |
| 366 | MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); |
| 367 | const char* Mips16HelperFunction = nullptr; |
| 368 | bool NeedMips16Helper = false; |
| 369 | |
| 370 | if (Subtarget.inMips16HardFloat()) { |
| 371 | // |
| 372 | // currently we don't have symbols tagged with the mips16 or mips32 |
| 373 | // qualifier so we will assume that we don't know what kind it is. |
| 374 | // and generate the helper |
| 375 | // |
| 376 | bool LookupHelper = true; |
| 377 | if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Val&: CLI.Callee)) { |
| 378 | if (isMips16HardFloatLibcall(Name: S->getSymbol())) |
| 379 | LookupHelper = false; |
| 380 | else { |
| 381 | const char *Symbol = S->getSymbol(); |
| 382 | Mips16IntrinsicHelperType IntrinsicFind = { .Name: Symbol, .Helper: "" }; |
| 383 | const Mips16HardFloatInfo::FuncSignature *Signature = |
| 384 | Mips16HardFloatInfo::findFuncSignature(name: Symbol); |
| 385 | if (!IsPICCall && Signature && |
| 386 | FuncInfo->StubsNeeded.try_emplace(k: Symbol, args&: Signature).second) { |
| 387 | // |
| 388 | // S2 is normally saved if the stub is for a function which |
| 389 | // returns a float or double value and is not otherwise. This is |
| 390 | // because more work is required after the function the stub |
| 391 | // is calling completes, and so the stub cannot directly return |
| 392 | // and the stub has no stack space to store the return address so |
| 393 | // S2 is used for that purpose. |
| 394 | // In order to take advantage of not saving S2, we need to also |
| 395 | // optimize the call in the stub and this requires some further |
| 396 | // functionality in MipsAsmPrinter which we don't have yet. |
| 397 | // So for now we always save S2. The optimization will be done |
| 398 | // in a follow-on patch. |
| 399 | // |
| 400 | if (true || (Signature->RetSig != Mips16HardFloatInfo::NoFPRet)) |
| 401 | FuncInfo->setSaveS2(); |
| 402 | } |
| 403 | // one more look at list of intrinsics |
| 404 | const Mips16IntrinsicHelperType *Helper = |
| 405 | llvm::lower_bound(Range: Mips16IntrinsicHelper, Value&: IntrinsicFind); |
| 406 | if (Helper != std::end(arr: Mips16IntrinsicHelper) && |
| 407 | *Helper == IntrinsicFind) { |
| 408 | Mips16HelperFunction = Helper->Helper; |
| 409 | NeedMips16Helper = true; |
| 410 | LookupHelper = false; |
| 411 | } |
| 412 | |
| 413 | } |
| 414 | } else if (GlobalAddressSDNode *G = |
| 415 | dyn_cast<GlobalAddressSDNode>(Val&: CLI.Callee)) { |
| 416 | |
| 417 | if (isMips16HardFloatLibcall(Name: G->getGlobal()->getName())) |
| 418 | LookupHelper = false; |
| 419 | } |
| 420 | if (LookupHelper) |
| 421 | Mips16HelperFunction = |
| 422 | getMips16HelperFunction(RetTy: CLI.RetTy, Args&: CLI.getArgs(), needHelper&: NeedMips16Helper); |
| 423 | } |
| 424 | |
| 425 | SDValue JumpTarget = Callee; |
| 426 | |
| 427 | // T9 should contain the address of the callee function if |
| 428 | // -relocation-model=pic or it is an indirect call. |
| 429 | if (IsPICCall || !GlobalOrExternal) { |
| 430 | unsigned V0Reg = Mips::V0; |
| 431 | if (NeedMips16Helper) { |
| 432 | RegsToPass.push_front(x: std::make_pair(x&: V0Reg, y&: Callee)); |
| 433 | JumpTarget = DAG.getExternalSymbol(Sym: Mips16HelperFunction, |
| 434 | VT: getPointerTy(DL: DAG.getDataLayout())); |
| 435 | ExternalSymbolSDNode *S = cast<ExternalSymbolSDNode>(Val&: JumpTarget); |
| 436 | JumpTarget = getAddrGlobal(N: S, DL: CLI.DL, Ty: JumpTarget.getValueType(), DAG, |
| 437 | Flag: MipsII::MO_GOT, Chain, |
| 438 | PtrInfo: FuncInfo->callPtrInfo(MF, ES: S->getSymbol())); |
| 439 | } else |
| 440 | RegsToPass.push_front(x: std::make_pair(x: (unsigned)Mips::T9, y&: Callee)); |
| 441 | } |
| 442 | |
| 443 | Ops.push_back(Elt: JumpTarget); |
| 444 | |
| 445 | MipsTargetLowering::getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, |
| 446 | InternalLinkage, IsCallReloc, CLI, Callee, |
| 447 | Chain); |
| 448 | } |
| 449 | |
| 450 | MachineBasicBlock * |
| 451 | Mips16TargetLowering::emitSel16(unsigned Opc, MachineInstr &MI, |
| 452 | MachineBasicBlock *BB) const { |
| 453 | if (DontExpandCondPseudos16) |
| 454 | return BB; |
| 455 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 456 | DebugLoc DL = MI.getDebugLoc(); |
| 457 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
| 458 | // diamond control-flow pattern. The incoming instruction knows the |
| 459 | // destination vreg to set, the condition code register to branch on, the |
| 460 | // true/false values to select between, and a branch opcode to use. |
| 461 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 462 | MachineFunction::iterator It = ++BB->getIterator(); |
| 463 | |
| 464 | // thisMBB: |
| 465 | // ... |
| 466 | // TrueVal = ... |
| 467 | // setcc r1, r2, r3 |
| 468 | // bNE r1, r0, copy1MBB |
| 469 | // fallthrough --> copy0MBB |
| 470 | MachineBasicBlock *thisMBB = BB; |
| 471 | MachineFunction *F = BB->getParent(); |
| 472 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 473 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 474 | F->insert(MBBI: It, MBB: copy0MBB); |
| 475 | F->insert(MBBI: It, MBB: sinkMBB); |
| 476 | |
| 477 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
| 478 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
| 479 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 480 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 481 | |
| 482 | // Next, add the true and fallthrough blocks as its successors. |
| 483 | BB->addSuccessor(Succ: copy0MBB); |
| 484 | BB->addSuccessor(Succ: sinkMBB); |
| 485 | |
| 486 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc)) |
| 487 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
| 488 | .addMBB(MBB: sinkMBB); |
| 489 | |
| 490 | // copy0MBB: |
| 491 | // %FalseValue = ... |
| 492 | // # fallthrough to sinkMBB |
| 493 | BB = copy0MBB; |
| 494 | |
| 495 | // Update machine-CFG edges |
| 496 | BB->addSuccessor(Succ: sinkMBB); |
| 497 | |
| 498 | // sinkMBB: |
| 499 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
| 500 | // ... |
| 501 | BB = sinkMBB; |
| 502 | |
| 503 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
| 504 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 505 | .addMBB(MBB: thisMBB) |
| 506 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 507 | .addMBB(MBB: copy0MBB); |
| 508 | |
| 509 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 510 | return BB; |
| 511 | } |
| 512 | |
| 513 | MachineBasicBlock * |
| 514 | Mips16TargetLowering::emitSelT16(unsigned Opc1, unsigned Opc2, MachineInstr &MI, |
| 515 | MachineBasicBlock *BB) const { |
| 516 | if (DontExpandCondPseudos16) |
| 517 | return BB; |
| 518 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 519 | DebugLoc DL = MI.getDebugLoc(); |
| 520 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
| 521 | // diamond control-flow pattern. The incoming instruction knows the |
| 522 | // destination vreg to set, the condition code register to branch on, the |
| 523 | // true/false values to select between, and a branch opcode to use. |
| 524 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 525 | MachineFunction::iterator It = ++BB->getIterator(); |
| 526 | |
| 527 | // thisMBB: |
| 528 | // ... |
| 529 | // TrueVal = ... |
| 530 | // setcc r1, r2, r3 |
| 531 | // bNE r1, r0, copy1MBB |
| 532 | // fallthrough --> copy0MBB |
| 533 | MachineBasicBlock *thisMBB = BB; |
| 534 | MachineFunction *F = BB->getParent(); |
| 535 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 536 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 537 | F->insert(MBBI: It, MBB: copy0MBB); |
| 538 | F->insert(MBBI: It, MBB: sinkMBB); |
| 539 | |
| 540 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
| 541 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
| 542 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 543 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 544 | |
| 545 | // Next, add the true and fallthrough blocks as its successors. |
| 546 | BB->addSuccessor(Succ: copy0MBB); |
| 547 | BB->addSuccessor(Succ: sinkMBB); |
| 548 | |
| 549 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc2)) |
| 550 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
| 551 | .addReg(RegNo: MI.getOperand(i: 4).getReg()); |
| 552 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc1)).addMBB(MBB: sinkMBB); |
| 553 | |
| 554 | // copy0MBB: |
| 555 | // %FalseValue = ... |
| 556 | // # fallthrough to sinkMBB |
| 557 | BB = copy0MBB; |
| 558 | |
| 559 | // Update machine-CFG edges |
| 560 | BB->addSuccessor(Succ: sinkMBB); |
| 561 | |
| 562 | // sinkMBB: |
| 563 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
| 564 | // ... |
| 565 | BB = sinkMBB; |
| 566 | |
| 567 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
| 568 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 569 | .addMBB(MBB: thisMBB) |
| 570 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 571 | .addMBB(MBB: copy0MBB); |
| 572 | |
| 573 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 574 | return BB; |
| 575 | |
| 576 | } |
| 577 | |
| 578 | MachineBasicBlock * |
| 579 | Mips16TargetLowering::emitSeliT16(unsigned Opc1, unsigned Opc2, |
| 580 | MachineInstr &MI, |
| 581 | MachineBasicBlock *BB) const { |
| 582 | if (DontExpandCondPseudos16) |
| 583 | return BB; |
| 584 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 585 | DebugLoc DL = MI.getDebugLoc(); |
| 586 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
| 587 | // diamond control-flow pattern. The incoming instruction knows the |
| 588 | // destination vreg to set, the condition code register to branch on, the |
| 589 | // true/false values to select between, and a branch opcode to use. |
| 590 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 591 | MachineFunction::iterator It = ++BB->getIterator(); |
| 592 | |
| 593 | // thisMBB: |
| 594 | // ... |
| 595 | // TrueVal = ... |
| 596 | // setcc r1, r2, r3 |
| 597 | // bNE r1, r0, copy1MBB |
| 598 | // fallthrough --> copy0MBB |
| 599 | MachineBasicBlock *thisMBB = BB; |
| 600 | MachineFunction *F = BB->getParent(); |
| 601 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 602 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 603 | F->insert(MBBI: It, MBB: copy0MBB); |
| 604 | F->insert(MBBI: It, MBB: sinkMBB); |
| 605 | |
| 606 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
| 607 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
| 608 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 609 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 610 | |
| 611 | // Next, add the true and fallthrough blocks as its successors. |
| 612 | BB->addSuccessor(Succ: copy0MBB); |
| 613 | BB->addSuccessor(Succ: sinkMBB); |
| 614 | |
| 615 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc2)) |
| 616 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
| 617 | .addImm(Val: MI.getOperand(i: 4).getImm()); |
| 618 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc1)).addMBB(MBB: sinkMBB); |
| 619 | |
| 620 | // copy0MBB: |
| 621 | // %FalseValue = ... |
| 622 | // # fallthrough to sinkMBB |
| 623 | BB = copy0MBB; |
| 624 | |
| 625 | // Update machine-CFG edges |
| 626 | BB->addSuccessor(Succ: sinkMBB); |
| 627 | |
| 628 | // sinkMBB: |
| 629 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
| 630 | // ... |
| 631 | BB = sinkMBB; |
| 632 | |
| 633 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
| 634 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 635 | .addMBB(MBB: thisMBB) |
| 636 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 637 | .addMBB(MBB: copy0MBB); |
| 638 | |
| 639 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 640 | return BB; |
| 641 | |
| 642 | } |
| 643 | |
| 644 | MachineBasicBlock * |
| 645 | Mips16TargetLowering::emitFEXT_T8I816_ins(unsigned BtOpc, unsigned CmpOpc, |
| 646 | MachineInstr &MI, |
| 647 | MachineBasicBlock *BB) const { |
| 648 | if (DontExpandCondPseudos16) |
| 649 | return BB; |
| 650 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 651 | Register regX = MI.getOperand(i: 0).getReg(); |
| 652 | Register regY = MI.getOperand(i: 1).getReg(); |
| 653 | MachineBasicBlock *target = MI.getOperand(i: 2).getMBB(); |
| 654 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: CmpOpc)) |
| 655 | .addReg(RegNo: regX) |
| 656 | .addReg(RegNo: regY); |
| 657 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: BtOpc)).addMBB(MBB: target); |
| 658 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 659 | return BB; |
| 660 | } |
| 661 | |
| 662 | MachineBasicBlock *Mips16TargetLowering::emitFEXT_T8I8I16_ins( |
| 663 | unsigned BtOpc, unsigned CmpiOpc, unsigned CmpiXOpc, bool ImmSigned, |
| 664 | MachineInstr &MI, MachineBasicBlock *BB) const { |
| 665 | if (DontExpandCondPseudos16) |
| 666 | return BB; |
| 667 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 668 | Register regX = MI.getOperand(i: 0).getReg(); |
| 669 | int64_t imm = MI.getOperand(i: 1).getImm(); |
| 670 | MachineBasicBlock *target = MI.getOperand(i: 2).getMBB(); |
| 671 | unsigned CmpOpc; |
| 672 | if (isUInt<8>(x: imm)) |
| 673 | CmpOpc = CmpiOpc; |
| 674 | else if ((!ImmSigned && isUInt<16>(x: imm)) || |
| 675 | (ImmSigned && isInt<16>(x: imm))) |
| 676 | CmpOpc = CmpiXOpc; |
| 677 | else |
| 678 | llvm_unreachable("immediate field not usable" ); |
| 679 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: CmpOpc)).addReg(RegNo: regX).addImm(Val: imm); |
| 680 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: BtOpc)).addMBB(MBB: target); |
| 681 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 682 | return BB; |
| 683 | } |
| 684 | |
| 685 | static unsigned Mips16WhichOp8uOr16simm |
| 686 | (unsigned shortOp, unsigned longOp, int64_t Imm) { |
| 687 | if (isUInt<8>(x: Imm)) |
| 688 | return shortOp; |
| 689 | else if (isInt<16>(x: Imm)) |
| 690 | return longOp; |
| 691 | else |
| 692 | llvm_unreachable("immediate field not usable" ); |
| 693 | } |
| 694 | |
| 695 | MachineBasicBlock * |
| 696 | Mips16TargetLowering::emitFEXT_CCRX16_ins(unsigned SltOpc, MachineInstr &MI, |
| 697 | MachineBasicBlock *BB) const { |
| 698 | if (DontExpandCondPseudos16) |
| 699 | return BB; |
| 700 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 701 | Register CC = MI.getOperand(i: 0).getReg(); |
| 702 | Register regX = MI.getOperand(i: 1).getReg(); |
| 703 | Register regY = MI.getOperand(i: 2).getReg(); |
| 704 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: SltOpc)) |
| 705 | .addReg(RegNo: regX) |
| 706 | .addReg(RegNo: regY); |
| 707 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: Mips::MoveR3216), DestReg: CC) |
| 708 | .addReg(RegNo: Mips::T8); |
| 709 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 710 | return BB; |
| 711 | } |
| 712 | |
| 713 | MachineBasicBlock * |
| 714 | Mips16TargetLowering::emitFEXT_CCRXI16_ins(unsigned SltiOpc, unsigned SltiXOpc, |
| 715 | MachineInstr &MI, |
| 716 | MachineBasicBlock *BB) const { |
| 717 | if (DontExpandCondPseudos16) |
| 718 | return BB; |
| 719 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 720 | Register CC = MI.getOperand(i: 0).getReg(); |
| 721 | Register regX = MI.getOperand(i: 1).getReg(); |
| 722 | int64_t Imm = MI.getOperand(i: 2).getImm(); |
| 723 | unsigned SltOpc = Mips16WhichOp8uOr16simm(shortOp: SltiOpc, longOp: SltiXOpc, Imm); |
| 724 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: SltOpc)).addReg(RegNo: regX).addImm(Val: Imm); |
| 725 | BuildMI(BB&: *BB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: Mips::MoveR3216), DestReg: CC) |
| 726 | .addReg(RegNo: Mips::T8); |
| 727 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 728 | return BB; |
| 729 | |
| 730 | } |
| 731 | |