| File: | build/source/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp |
| Warning: | line 1504, column 8 Although the value stored to 'Shamt' is used in the enclosing expression, the value is never actually read from 'Shamt' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //=- LoongArchISelLowering.cpp - LoongArch DAG Lowering Implementation ---===// |
| 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 | // This file defines the interfaces that LoongArch uses to lower LLVM code into |
| 10 | // a selection DAG. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "LoongArchISelLowering.h" |
| 15 | #include "LoongArch.h" |
| 16 | #include "LoongArchMachineFunctionInfo.h" |
| 17 | #include "LoongArchRegisterInfo.h" |
| 18 | #include "LoongArchSubtarget.h" |
| 19 | #include "LoongArchTargetMachine.h" |
| 20 | #include "MCTargetDesc/LoongArchBaseInfo.h" |
| 21 | #include "MCTargetDesc/LoongArchMCTargetDesc.h" |
| 22 | #include "llvm/ADT/Statistic.h" |
| 23 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 24 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
| 25 | #include "llvm/IR/IRBuilder.h" |
| 26 | #include "llvm/IR/IntrinsicsLoongArch.h" |
| 27 | #include "llvm/Support/Debug.h" |
| 28 | #include "llvm/Support/KnownBits.h" |
| 29 | #include "llvm/Support/MathExtras.h" |
| 30 | |
| 31 | using namespace llvm; |
| 32 | |
| 33 | #define DEBUG_TYPE"loongarch-isel-lowering" "loongarch-isel-lowering" |
| 34 | |
| 35 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"loongarch-isel-lowering" , "NumTailCalls", "Number of tail calls"}; |
| 36 | |
| 37 | static cl::opt<bool> ZeroDivCheck( |
| 38 | "loongarch-check-zero-division", cl::Hidden, |
| 39 | cl::desc("Trap on integer division by zero."), |
| 40 | cl::init(false)); |
| 41 | |
| 42 | LoongArchTargetLowering::LoongArchTargetLowering(const TargetMachine &TM, |
| 43 | const LoongArchSubtarget &STI) |
| 44 | : TargetLowering(TM), Subtarget(STI) { |
| 45 | |
| 46 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 47 | // Set up the register classes. |
| 48 | addRegisterClass(GRLenVT, &LoongArch::GPRRegClass); |
| 49 | if (Subtarget.hasBasicF()) |
| 50 | addRegisterClass(MVT::f32, &LoongArch::FPR32RegClass); |
| 51 | if (Subtarget.hasBasicD()) |
| 52 | addRegisterClass(MVT::f64, &LoongArch::FPR64RegClass); |
| 53 | |
| 54 | setLoadExtAction({ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}, GRLenVT, |
| 55 | MVT::i1, Promote); |
| 56 | |
| 57 | // TODO: add necessary setOperationAction calls later. |
| 58 | setOperationAction(ISD::SHL_PARTS, GRLenVT, Custom); |
| 59 | setOperationAction(ISD::SRA_PARTS, GRLenVT, Custom); |
| 60 | setOperationAction(ISD::SRL_PARTS, GRLenVT, Custom); |
| 61 | setOperationAction(ISD::FP_TO_SINT, GRLenVT, Custom); |
| 62 | setOperationAction(ISD::ROTL, GRLenVT, Expand); |
| 63 | setOperationAction(ISD::CTPOP, GRLenVT, Expand); |
| 64 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); |
| 65 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
| 66 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
| 67 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
| 68 | |
| 69 | setOperationAction({ISD::GlobalAddress, ISD::BlockAddress, ISD::ConstantPool, |
| 70 | ISD::JumpTable}, |
| 71 | GRLenVT, Custom); |
| 72 | |
| 73 | setOperationAction(ISD::GlobalTLSAddress, GRLenVT, Custom); |
| 74 | |
| 75 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
| 76 | |
| 77 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); |
| 78 | if (Subtarget.is64Bit()) |
| 79 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); |
| 80 | |
| 81 | setOperationAction(ISD::DYNAMIC_STACKALLOC, GRLenVT, Expand); |
| 82 | setOperationAction({ISD::STACKSAVE, ISD::STACKRESTORE}, MVT::Other, Expand); |
| 83 | setOperationAction(ISD::VASTART, MVT::Other, Custom); |
| 84 | setOperationAction({ISD::VAARG, ISD::VACOPY, ISD::VAEND}, MVT::Other, Expand); |
| 85 | |
| 86 | if (Subtarget.is64Bit()) { |
| 87 | setOperationAction(ISD::SHL, MVT::i32, Custom); |
| 88 | setOperationAction(ISD::SRA, MVT::i32, Custom); |
| 89 | setOperationAction(ISD::SRL, MVT::i32, Custom); |
| 90 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); |
| 91 | setOperationAction(ISD::BITCAST, MVT::i32, Custom); |
| 92 | setOperationAction(ISD::ROTR, MVT::i32, Custom); |
| 93 | setOperationAction(ISD::ROTL, MVT::i32, Custom); |
| 94 | setOperationAction(ISD::CTTZ, MVT::i32, Custom); |
| 95 | setOperationAction(ISD::CTLZ, MVT::i32, Custom); |
| 96 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i32, Custom); |
| 97 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i32, Custom); |
| 98 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
| 99 | setOperationAction(ISD::READ_REGISTER, MVT::i32, Custom); |
| 100 | setOperationAction(ISD::WRITE_REGISTER, MVT::i32, Custom); |
| 101 | if (Subtarget.hasBasicF() && !Subtarget.hasBasicD()) |
| 102 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); |
| 103 | if (Subtarget.hasBasicF()) |
| 104 | setOperationAction(ISD::FRINT, MVT::f32, Legal); |
| 105 | if (Subtarget.hasBasicD()) |
| 106 | setOperationAction(ISD::FRINT, MVT::f64, Legal); |
| 107 | } |
| 108 | |
| 109 | // LA32 does not have REVB.2W and REVB.D due to the 64-bit operands, and |
| 110 | // the narrower REVB.W does not exist. But LA32 does have REVB.2H, so i16 |
| 111 | // and i32 could still be byte-swapped relatively cheaply. |
| 112 | setOperationAction(ISD::BSWAP, MVT::i16, Custom); |
| 113 | if (Subtarget.is64Bit()) { |
| 114 | setOperationAction(ISD::BSWAP, MVT::i32, Custom); |
| 115 | } |
| 116 | |
| 117 | // Expand bitreverse.i16 with native-width bitrev and shift for now, before |
| 118 | // we get to know which of sll and revb.2h is faster. |
| 119 | setOperationAction(ISD::BITREVERSE, MVT::i8, Custom); |
| 120 | if (Subtarget.is64Bit()) { |
| 121 | setOperationAction(ISD::BITREVERSE, MVT::i32, Custom); |
| 122 | setOperationAction(ISD::BITREVERSE, MVT::i64, Legal); |
| 123 | } else { |
| 124 | setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); |
| 125 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); |
| 126 | setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom); |
| 127 | setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom); |
| 128 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
| 129 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i64, Custom); |
| 130 | } |
| 131 | |
| 132 | static const ISD::CondCode FPCCToExpand[] = { |
| 133 | ISD::SETOGT, ISD::SETOGE, ISD::SETUGT, ISD::SETUGE, |
| 134 | ISD::SETGE, ISD::SETNE, ISD::SETGT}; |
| 135 | |
| 136 | if (Subtarget.hasBasicF()) { |
| 137 | setCondCodeAction(FPCCToExpand, MVT::f32, Expand); |
| 138 | setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); |
| 139 | setOperationAction(ISD::BR_CC, MVT::f32, Expand); |
| 140 | setOperationAction(ISD::FMA, MVT::f32, Legal); |
| 141 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal); |
| 142 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal); |
| 143 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal); |
| 144 | setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal); |
| 145 | setOperationAction(ISD::FSIN, MVT::f32, Expand); |
| 146 | setOperationAction(ISD::FCOS, MVT::f32, Expand); |
| 147 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); |
| 148 | setOperationAction(ISD::FPOW, MVT::f32, Expand); |
| 149 | setOperationAction(ISD::FREM, MVT::f32, Expand); |
| 150 | } |
| 151 | if (Subtarget.hasBasicD()) { |
| 152 | setCondCodeAction(FPCCToExpand, MVT::f64, Expand); |
| 153 | setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); |
| 154 | setOperationAction(ISD::BR_CC, MVT::f64, Expand); |
| 155 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal); |
| 156 | setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal); |
| 157 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); |
| 158 | setOperationAction(ISD::FMA, MVT::f64, Legal); |
| 159 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal); |
| 160 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal); |
| 161 | setOperationAction(ISD::FSIN, MVT::f64, Expand); |
| 162 | setOperationAction(ISD::FCOS, MVT::f64, Expand); |
| 163 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); |
| 164 | setOperationAction(ISD::FPOW, MVT::f64, Expand); |
| 165 | setOperationAction(ISD::FREM, MVT::f64, Expand); |
| 166 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
| 167 | } |
| 168 | |
| 169 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); |
| 170 | |
| 171 | setOperationAction(ISD::BR_CC, GRLenVT, Expand); |
| 172 | setOperationAction(ISD::SELECT_CC, GRLenVT, Expand); |
| 173 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
| 174 | setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, GRLenVT, Expand); |
| 175 | if (!Subtarget.is64Bit()) |
| 176 | setLibcallName(RTLIB::MUL_I128, nullptr); |
| 177 | |
| 178 | setOperationAction(ISD::FP_TO_UINT, GRLenVT, Custom); |
| 179 | setOperationAction(ISD::UINT_TO_FP, GRLenVT, Expand); |
| 180 | if ((Subtarget.is64Bit() && Subtarget.hasBasicF() && |
| 181 | !Subtarget.hasBasicD())) { |
| 182 | setOperationAction(ISD::SINT_TO_FP, GRLenVT, Custom); |
| 183 | setOperationAction(ISD::UINT_TO_FP, GRLenVT, Custom); |
| 184 | } |
| 185 | |
| 186 | // Compute derived properties from the register classes. |
| 187 | computeRegisterProperties(STI.getRegisterInfo()); |
| 188 | |
| 189 | setStackPointerRegisterToSaveRestore(LoongArch::R3); |
| 190 | |
| 191 | setBooleanContents(ZeroOrOneBooleanContent); |
| 192 | |
| 193 | setMaxAtomicSizeInBitsSupported(Subtarget.getGRLen()); |
| 194 | |
| 195 | setMinCmpXchgSizeInBits(32); |
| 196 | |
| 197 | // Function alignments. |
| 198 | setMinFunctionAlignment(Align(4)); |
| 199 | |
| 200 | setTargetDAGCombine(ISD::AND); |
| 201 | setTargetDAGCombine(ISD::OR); |
| 202 | setTargetDAGCombine(ISD::SRL); |
| 203 | } |
| 204 | |
| 205 | bool LoongArchTargetLowering::isOffsetFoldingLegal( |
| 206 | const GlobalAddressSDNode *GA) const { |
| 207 | // In order to maximise the opportunity for common subexpression elimination, |
| 208 | // keep a separate ADD node for the global address offset instead of folding |
| 209 | // it in the global address node. Later peephole optimisations may choose to |
| 210 | // fold it back in when profitable. |
| 211 | return false; |
| 212 | } |
| 213 | |
| 214 | SDValue LoongArchTargetLowering::LowerOperation(SDValue Op, |
| 215 | SelectionDAG &DAG) const { |
| 216 | switch (Op.getOpcode()) { |
| 217 | case ISD::EH_DWARF_CFA: |
| 218 | return lowerEH_DWARF_CFA(Op, DAG); |
| 219 | case ISD::GlobalAddress: |
| 220 | return lowerGlobalAddress(Op, DAG); |
| 221 | case ISD::GlobalTLSAddress: |
| 222 | return lowerGlobalTLSAddress(Op, DAG); |
| 223 | case ISD::INTRINSIC_WO_CHAIN: |
| 224 | return lowerINTRINSIC_WO_CHAIN(Op, DAG); |
| 225 | case ISD::INTRINSIC_W_CHAIN: |
| 226 | return lowerINTRINSIC_W_CHAIN(Op, DAG); |
| 227 | case ISD::INTRINSIC_VOID: |
| 228 | return lowerINTRINSIC_VOID(Op, DAG); |
| 229 | case ISD::BlockAddress: |
| 230 | return lowerBlockAddress(Op, DAG); |
| 231 | case ISD::JumpTable: |
| 232 | return lowerJumpTable(Op, DAG); |
| 233 | case ISD::SHL_PARTS: |
| 234 | return lowerShiftLeftParts(Op, DAG); |
| 235 | case ISD::SRA_PARTS: |
| 236 | return lowerShiftRightParts(Op, DAG, true); |
| 237 | case ISD::SRL_PARTS: |
| 238 | return lowerShiftRightParts(Op, DAG, false); |
| 239 | case ISD::ConstantPool: |
| 240 | return lowerConstantPool(Op, DAG); |
| 241 | case ISD::FP_TO_SINT: |
| 242 | return lowerFP_TO_SINT(Op, DAG); |
| 243 | case ISD::BITCAST: |
| 244 | return lowerBITCAST(Op, DAG); |
| 245 | case ISD::UINT_TO_FP: |
| 246 | return lowerUINT_TO_FP(Op, DAG); |
| 247 | case ISD::SINT_TO_FP: |
| 248 | return lowerSINT_TO_FP(Op, DAG); |
| 249 | case ISD::VASTART: |
| 250 | return lowerVASTART(Op, DAG); |
| 251 | case ISD::FRAMEADDR: |
| 252 | return lowerFRAMEADDR(Op, DAG); |
| 253 | case ISD::RETURNADDR: |
| 254 | return lowerRETURNADDR(Op, DAG); |
| 255 | case ISD::WRITE_REGISTER: |
| 256 | return lowerWRITE_REGISTER(Op, DAG); |
| 257 | } |
| 258 | return SDValue(); |
| 259 | } |
| 260 | |
| 261 | SDValue LoongArchTargetLowering::lowerWRITE_REGISTER(SDValue Op, |
| 262 | SelectionDAG &DAG) const { |
| 263 | |
| 264 | if (Subtarget.is64Bit() && Op.getOperand(2).getValueType() == MVT::i32) { |
| 265 | DAG.getContext()->emitError( |
| 266 | "On LA64, only 64-bit registers can be written."); |
| 267 | return Op.getOperand(0); |
| 268 | } |
| 269 | |
| 270 | if (!Subtarget.is64Bit() && Op.getOperand(2).getValueType() == MVT::i64) { |
| 271 | DAG.getContext()->emitError( |
| 272 | "On LA32, only 32-bit registers can be written."); |
| 273 | return Op.getOperand(0); |
| 274 | } |
| 275 | |
| 276 | return Op; |
| 277 | } |
| 278 | |
| 279 | SDValue LoongArchTargetLowering::lowerFRAMEADDR(SDValue Op, |
| 280 | SelectionDAG &DAG) const { |
| 281 | if (!isa<ConstantSDNode>(Op.getOperand(0))) { |
| 282 | DAG.getContext()->emitError("argument to '__builtin_frame_address' must " |
| 283 | "be a constant integer"); |
| 284 | return SDValue(); |
| 285 | } |
| 286 | |
| 287 | MachineFunction &MF = DAG.getMachineFunction(); |
| 288 | MF.getFrameInfo().setFrameAddressIsTaken(true); |
| 289 | Register FrameReg = Subtarget.getRegisterInfo()->getFrameRegister(MF); |
| 290 | EVT VT = Op.getValueType(); |
| 291 | SDLoc DL(Op); |
| 292 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT); |
| 293 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
| 294 | int GRLenInBytes = Subtarget.getGRLen() / 8; |
| 295 | |
| 296 | while (Depth--) { |
| 297 | int Offset = -(GRLenInBytes * 2); |
| 298 | SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr, |
| 299 | DAG.getIntPtrConstant(Offset, DL)); |
| 300 | FrameAddr = |
| 301 | DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo()); |
| 302 | } |
| 303 | return FrameAddr; |
| 304 | } |
| 305 | |
| 306 | SDValue LoongArchTargetLowering::lowerRETURNADDR(SDValue Op, |
| 307 | SelectionDAG &DAG) const { |
| 308 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
| 309 | return SDValue(); |
| 310 | |
| 311 | // Currently only support lowering return address for current frame. |
| 312 | if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() != 0) { |
| 313 | DAG.getContext()->emitError( |
| 314 | "return address can only be determined for the current frame"); |
| 315 | return SDValue(); |
| 316 | } |
| 317 | |
| 318 | MachineFunction &MF = DAG.getMachineFunction(); |
| 319 | MF.getFrameInfo().setReturnAddressIsTaken(true); |
| 320 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 321 | |
| 322 | // Return the value of the return address register, marking it an implicit |
| 323 | // live-in. |
| 324 | Register Reg = MF.addLiveIn(Subtarget.getRegisterInfo()->getRARegister(), |
| 325 | getRegClassFor(GRLenVT)); |
| 326 | return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, GRLenVT); |
| 327 | } |
| 328 | |
| 329 | SDValue LoongArchTargetLowering::lowerEH_DWARF_CFA(SDValue Op, |
| 330 | SelectionDAG &DAG) const { |
| 331 | MachineFunction &MF = DAG.getMachineFunction(); |
| 332 | auto Size = Subtarget.getGRLen() / 8; |
| 333 | auto FI = MF.getFrameInfo().CreateFixedObject(Size, 0, false); |
| 334 | return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); |
| 335 | } |
| 336 | |
| 337 | SDValue LoongArchTargetLowering::lowerVASTART(SDValue Op, |
| 338 | SelectionDAG &DAG) const { |
| 339 | MachineFunction &MF = DAG.getMachineFunction(); |
| 340 | auto *FuncInfo = MF.getInfo<LoongArchMachineFunctionInfo>(); |
| 341 | |
| 342 | SDLoc DL(Op); |
| 343 | SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), |
| 344 | getPointerTy(MF.getDataLayout())); |
| 345 | |
| 346 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
| 347 | // memory location argument. |
| 348 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
| 349 | return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), |
| 350 | MachinePointerInfo(SV)); |
| 351 | } |
| 352 | |
| 353 | SDValue LoongArchTargetLowering::lowerUINT_TO_FP(SDValue Op, |
| 354 | SelectionDAG &DAG) const { |
| 355 | assert(Subtarget.is64Bit() && Subtarget.hasBasicF() &&(static_cast <bool> (Subtarget.is64Bit() && Subtarget .hasBasicF() && !Subtarget.hasBasicD() && "unexpected target features" ) ? void (0) : __assert_fail ("Subtarget.is64Bit() && Subtarget.hasBasicF() && !Subtarget.hasBasicD() && \"unexpected target features\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 356, __extension__ __PRETTY_FUNCTION__)) |
| 356 | !Subtarget.hasBasicD() && "unexpected target features")(static_cast <bool> (Subtarget.is64Bit() && Subtarget .hasBasicF() && !Subtarget.hasBasicD() && "unexpected target features" ) ? void (0) : __assert_fail ("Subtarget.is64Bit() && Subtarget.hasBasicF() && !Subtarget.hasBasicD() && \"unexpected target features\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 356, __extension__ __PRETTY_FUNCTION__)); |
| 357 | |
| 358 | SDLoc DL(Op); |
| 359 | SDValue Op0 = Op.getOperand(0); |
| 360 | if (Op0->getOpcode() == ISD::AND) { |
| 361 | auto *C = dyn_cast<ConstantSDNode>(Op0.getOperand(1)); |
| 362 | if (C && C->getZExtValue() < UINT64_C(0xFFFFFFFF)0xFFFFFFFFUL) |
| 363 | return Op; |
| 364 | } |
| 365 | |
| 366 | if (Op0->getOpcode() == LoongArchISD::BSTRPICK && |
| 367 | Op0.getConstantOperandVal(1) < UINT64_C(0X1F)0X1FUL && |
| 368 | Op0.getConstantOperandVal(2) == UINT64_C(0)0UL) |
| 369 | return Op; |
| 370 | |
| 371 | if (Op0.getOpcode() == ISD::AssertZext && |
| 372 | dyn_cast<VTSDNode>(Op0.getOperand(1))->getVT().bitsLT(MVT::i32)) |
| 373 | return Op; |
| 374 | |
| 375 | EVT OpVT = Op0.getValueType(); |
| 376 | EVT RetVT = Op.getValueType(); |
| 377 | RTLIB::Libcall LC = RTLIB::getUINTTOFP(OpVT, RetVT); |
| 378 | MakeLibCallOptions CallOptions; |
| 379 | CallOptions.setTypeListBeforeSoften(OpVT, RetVT, true); |
| 380 | SDValue Chain = SDValue(); |
| 381 | SDValue Result; |
| 382 | std::tie(Result, Chain) = |
| 383 | makeLibCall(DAG, LC, Op.getValueType(), Op0, CallOptions, DL, Chain); |
| 384 | return Result; |
| 385 | } |
| 386 | |
| 387 | SDValue LoongArchTargetLowering::lowerSINT_TO_FP(SDValue Op, |
| 388 | SelectionDAG &DAG) const { |
| 389 | assert(Subtarget.is64Bit() && Subtarget.hasBasicF() &&(static_cast <bool> (Subtarget.is64Bit() && Subtarget .hasBasicF() && !Subtarget.hasBasicD() && "unexpected target features" ) ? void (0) : __assert_fail ("Subtarget.is64Bit() && Subtarget.hasBasicF() && !Subtarget.hasBasicD() && \"unexpected target features\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 390, __extension__ __PRETTY_FUNCTION__)) |
| 390 | !Subtarget.hasBasicD() && "unexpected target features")(static_cast <bool> (Subtarget.is64Bit() && Subtarget .hasBasicF() && !Subtarget.hasBasicD() && "unexpected target features" ) ? void (0) : __assert_fail ("Subtarget.is64Bit() && Subtarget.hasBasicF() && !Subtarget.hasBasicD() && \"unexpected target features\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 390, __extension__ __PRETTY_FUNCTION__)); |
| 391 | |
| 392 | SDLoc DL(Op); |
| 393 | SDValue Op0 = Op.getOperand(0); |
| 394 | |
| 395 | if ((Op0.getOpcode() == ISD::AssertSext || |
| 396 | Op0.getOpcode() == ISD::SIGN_EXTEND_INREG) && |
| 397 | dyn_cast<VTSDNode>(Op0.getOperand(1))->getVT().bitsLE(MVT::i32)) |
| 398 | return Op; |
| 399 | |
| 400 | EVT OpVT = Op0.getValueType(); |
| 401 | EVT RetVT = Op.getValueType(); |
| 402 | RTLIB::Libcall LC = RTLIB::getSINTTOFP(OpVT, RetVT); |
| 403 | MakeLibCallOptions CallOptions; |
| 404 | CallOptions.setTypeListBeforeSoften(OpVT, RetVT, true); |
| 405 | SDValue Chain = SDValue(); |
| 406 | SDValue Result; |
| 407 | std::tie(Result, Chain) = |
| 408 | makeLibCall(DAG, LC, Op.getValueType(), Op0, CallOptions, DL, Chain); |
| 409 | return Result; |
| 410 | } |
| 411 | |
| 412 | SDValue LoongArchTargetLowering::lowerBITCAST(SDValue Op, |
| 413 | SelectionDAG &DAG) const { |
| 414 | |
| 415 | SDLoc DL(Op); |
| 416 | SDValue Op0 = Op.getOperand(0); |
| 417 | |
| 418 | if (Op.getValueType() == MVT::f32 && Op0.getValueType() == MVT::i32 && |
| 419 | Subtarget.is64Bit() && Subtarget.hasBasicF()) { |
| 420 | SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0); |
| 421 | return DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64, DL, MVT::f32, NewOp0); |
| 422 | } |
| 423 | return Op; |
| 424 | } |
| 425 | |
| 426 | SDValue LoongArchTargetLowering::lowerFP_TO_SINT(SDValue Op, |
| 427 | SelectionDAG &DAG) const { |
| 428 | |
| 429 | SDLoc DL(Op); |
| 430 | |
| 431 | if (Op.getValueSizeInBits() > 32 && Subtarget.hasBasicF() && |
| 432 | !Subtarget.hasBasicD()) { |
| 433 | SDValue Dst = |
| 434 | DAG.getNode(LoongArchISD::FTINT, DL, MVT::f32, Op.getOperand(0)); |
| 435 | return DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Dst); |
| 436 | } |
| 437 | |
| 438 | EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits()); |
| 439 | SDValue Trunc = DAG.getNode(LoongArchISD::FTINT, DL, FPTy, Op.getOperand(0)); |
| 440 | return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Trunc); |
| 441 | } |
| 442 | |
| 443 | static SDValue getTargetNode(GlobalAddressSDNode *N, SDLoc DL, EVT Ty, |
| 444 | SelectionDAG &DAG, unsigned Flags) { |
| 445 | return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags); |
| 446 | } |
| 447 | |
| 448 | static SDValue getTargetNode(BlockAddressSDNode *N, SDLoc DL, EVT Ty, |
| 449 | SelectionDAG &DAG, unsigned Flags) { |
| 450 | return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(), |
| 451 | Flags); |
| 452 | } |
| 453 | |
| 454 | static SDValue getTargetNode(ConstantPoolSDNode *N, SDLoc DL, EVT Ty, |
| 455 | SelectionDAG &DAG, unsigned Flags) { |
| 456 | return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(), |
| 457 | N->getOffset(), Flags); |
| 458 | } |
| 459 | |
| 460 | static SDValue getTargetNode(JumpTableSDNode *N, SDLoc DL, EVT Ty, |
| 461 | SelectionDAG &DAG, unsigned Flags) { |
| 462 | return DAG.getTargetJumpTable(N->getIndex(), Ty, Flags); |
| 463 | } |
| 464 | |
| 465 | template <class NodeTy> |
| 466 | SDValue LoongArchTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG, |
| 467 | bool IsLocal) const { |
| 468 | SDLoc DL(N); |
| 469 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
| 470 | SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0); |
| 471 | // TODO: Check CodeModel. |
| 472 | if (IsLocal) |
| 473 | // This generates the pattern (PseudoLA_PCREL sym), which expands to |
| 474 | // (addi.w/d (pcalau12i %pc_hi20(sym)) %pc_lo12(sym)). |
| 475 | return SDValue(DAG.getMachineNode(LoongArch::PseudoLA_PCREL, DL, Ty, Addr), |
| 476 | 0); |
| 477 | |
| 478 | // This generates the pattern (PseudoLA_GOT sym), which expands to (ld.w/d |
| 479 | // (pcalau12i %got_pc_hi20(sym)) %got_pc_lo12(sym)). |
| 480 | return SDValue(DAG.getMachineNode(LoongArch::PseudoLA_GOT, DL, Ty, Addr), 0); |
| 481 | } |
| 482 | |
| 483 | SDValue LoongArchTargetLowering::lowerBlockAddress(SDValue Op, |
| 484 | SelectionDAG &DAG) const { |
| 485 | return getAddr(cast<BlockAddressSDNode>(Op), DAG); |
| 486 | } |
| 487 | |
| 488 | SDValue LoongArchTargetLowering::lowerJumpTable(SDValue Op, |
| 489 | SelectionDAG &DAG) const { |
| 490 | return getAddr(cast<JumpTableSDNode>(Op), DAG); |
| 491 | } |
| 492 | |
| 493 | SDValue LoongArchTargetLowering::lowerConstantPool(SDValue Op, |
| 494 | SelectionDAG &DAG) const { |
| 495 | return getAddr(cast<ConstantPoolSDNode>(Op), DAG); |
| 496 | } |
| 497 | |
| 498 | SDValue LoongArchTargetLowering::lowerGlobalAddress(SDValue Op, |
| 499 | SelectionDAG &DAG) const { |
| 500 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); |
| 501 | assert(N->getOffset() == 0 && "unexpected offset in global node")(static_cast <bool> (N->getOffset() == 0 && "unexpected offset in global node" ) ? void (0) : __assert_fail ("N->getOffset() == 0 && \"unexpected offset in global node\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 501, __extension__ __PRETTY_FUNCTION__)); |
| 502 | return getAddr(N, DAG, N->getGlobal()->isDSOLocal()); |
| 503 | } |
| 504 | |
| 505 | SDValue LoongArchTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N, |
| 506 | SelectionDAG &DAG, |
| 507 | unsigned Opc) const { |
| 508 | SDLoc DL(N); |
| 509 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
| 510 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 511 | |
| 512 | SDValue Addr = DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, 0); |
| 513 | SDValue Offset = SDValue(DAG.getMachineNode(Opc, DL, Ty, Addr), 0); |
| 514 | |
| 515 | // Add the thread pointer. |
| 516 | return DAG.getNode(ISD::ADD, DL, Ty, Offset, |
| 517 | DAG.getRegister(LoongArch::R2, GRLenVT)); |
| 518 | } |
| 519 | |
| 520 | SDValue LoongArchTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N, |
| 521 | SelectionDAG &DAG, |
| 522 | unsigned Opc) const { |
| 523 | SDLoc DL(N); |
| 524 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
| 525 | IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits()); |
| 526 | |
| 527 | // Use a PC-relative addressing mode to access the dynamic GOT address. |
| 528 | SDValue Addr = DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, 0); |
| 529 | SDValue Load = SDValue(DAG.getMachineNode(Opc, DL, Ty, Addr), 0); |
| 530 | |
| 531 | // Prepare argument list to generate call. |
| 532 | ArgListTy Args; |
| 533 | ArgListEntry Entry; |
| 534 | Entry.Node = Load; |
| 535 | Entry.Ty = CallTy; |
| 536 | Args.push_back(Entry); |
| 537 | |
| 538 | // Setup call to __tls_get_addr. |
| 539 | TargetLowering::CallLoweringInfo CLI(DAG); |
| 540 | CLI.setDebugLoc(DL) |
| 541 | .setChain(DAG.getEntryNode()) |
| 542 | .setLibCallee(CallingConv::C, CallTy, |
| 543 | DAG.getExternalSymbol("__tls_get_addr", Ty), |
| 544 | std::move(Args)); |
| 545 | |
| 546 | return LowerCallTo(CLI).first; |
| 547 | } |
| 548 | |
| 549 | SDValue |
| 550 | LoongArchTargetLowering::lowerGlobalTLSAddress(SDValue Op, |
| 551 | SelectionDAG &DAG) const { |
| 552 | if (DAG.getMachineFunction().getFunction().getCallingConv() == |
| 553 | CallingConv::GHC) |
| 554 | report_fatal_error("In GHC calling convention TLS is not supported"); |
| 555 | |
| 556 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); |
| 557 | assert(N->getOffset() == 0 && "unexpected offset in global node")(static_cast <bool> (N->getOffset() == 0 && "unexpected offset in global node" ) ? void (0) : __assert_fail ("N->getOffset() == 0 && \"unexpected offset in global node\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 557, __extension__ __PRETTY_FUNCTION__)); |
| 558 | |
| 559 | SDValue Addr; |
| 560 | switch (getTargetMachine().getTLSModel(N->getGlobal())) { |
| 561 | case TLSModel::GeneralDynamic: |
| 562 | // In this model, application code calls the dynamic linker function |
| 563 | // __tls_get_addr to locate TLS offsets into the dynamic thread vector at |
| 564 | // runtime. |
| 565 | Addr = getDynamicTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_GD); |
| 566 | break; |
| 567 | case TLSModel::LocalDynamic: |
| 568 | // Same as GeneralDynamic, except for assembly modifiers and relocation |
| 569 | // records. |
| 570 | Addr = getDynamicTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_LD); |
| 571 | break; |
| 572 | case TLSModel::InitialExec: |
| 573 | // This model uses the GOT to resolve TLS offsets. |
| 574 | Addr = getStaticTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_IE); |
| 575 | break; |
| 576 | case TLSModel::LocalExec: |
| 577 | // This model is used when static linking as the TLS offsets are resolved |
| 578 | // during program linking. |
| 579 | Addr = getStaticTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_LE); |
| 580 | break; |
| 581 | } |
| 582 | |
| 583 | return Addr; |
| 584 | } |
| 585 | |
| 586 | SDValue |
| 587 | LoongArchTargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op, |
| 588 | SelectionDAG &DAG) const { |
| 589 | switch (Op.getConstantOperandVal(0)) { |
| 590 | default: |
| 591 | return SDValue(); // Don't custom lower most intrinsics. |
| 592 | case Intrinsic::thread_pointer: { |
| 593 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
| 594 | return DAG.getRegister(LoongArch::R2, PtrVT); |
| 595 | } |
| 596 | } |
| 597 | } |
| 598 | |
| 599 | // Helper function that emits error message for intrinsics with chain. |
| 600 | static SDValue emitIntrinsicWithChainErrorMessage(SDValue Op, |
| 601 | StringRef ErrorMsg, |
| 602 | SelectionDAG &DAG) { |
| 603 | |
| 604 | DAG.getContext()->emitError("argument to '" + Op->getOperationName(0) + "' " + |
| 605 | ErrorMsg); |
| 606 | return DAG.getMergeValues({DAG.getUNDEF(Op.getValueType()), Op.getOperand(0)}, |
| 607 | SDLoc(Op)); |
| 608 | } |
| 609 | |
| 610 | SDValue |
| 611 | LoongArchTargetLowering::lowerINTRINSIC_W_CHAIN(SDValue Op, |
| 612 | SelectionDAG &DAG) const { |
| 613 | SDLoc DL(Op); |
| 614 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 615 | SDValue Op0 = Op.getOperand(0); |
| 616 | std::string Name = Op->getOperationName(0); |
| 617 | const StringRef ErrorMsgOOR = "out of range"; |
| 618 | |
| 619 | switch (Op.getConstantOperandVal(1)) { |
| 620 | default: |
| 621 | return Op; |
| 622 | case Intrinsic::loongarch_crc_w_b_w: |
| 623 | case Intrinsic::loongarch_crc_w_h_w: |
| 624 | case Intrinsic::loongarch_crc_w_w_w: |
| 625 | case Intrinsic::loongarch_crc_w_d_w: |
| 626 | case Intrinsic::loongarch_crcc_w_b_w: |
| 627 | case Intrinsic::loongarch_crcc_w_h_w: |
| 628 | case Intrinsic::loongarch_crcc_w_w_w: |
| 629 | case Intrinsic::loongarch_crcc_w_d_w: { |
| 630 | std::string Name = Op->getOperationName(0); |
| 631 | DAG.getContext()->emitError(Name + " requires target: loongarch64"); |
| 632 | return DAG.getMergeValues({DAG.getUNDEF(Op.getValueType()), Op0}, DL); |
| 633 | } |
| 634 | case Intrinsic::loongarch_csrrd_w: |
| 635 | case Intrinsic::loongarch_csrrd_d: { |
| 636 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); |
| 637 | if (!isUInt<14>(Imm)) |
| 638 | return emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG); |
| 639 | return DAG.getMergeValues( |
| 640 | {DAG.getNode(LoongArchISD::CSRRD, DL, GRLenVT, Op0, |
| 641 | DAG.getConstant(Imm, DL, GRLenVT)), |
| 642 | Op0}, |
| 643 | DL); |
| 644 | } |
| 645 | case Intrinsic::loongarch_csrwr_w: |
| 646 | case Intrinsic::loongarch_csrwr_d: { |
| 647 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue(); |
| 648 | if (!isUInt<14>(Imm)) |
| 649 | return emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG); |
| 650 | return DAG.getMergeValues( |
| 651 | {DAG.getNode(LoongArchISD::CSRWR, DL, GRLenVT, Op0, Op.getOperand(2), |
| 652 | DAG.getConstant(Imm, DL, GRLenVT)), |
| 653 | Op0}, |
| 654 | DL); |
| 655 | } |
| 656 | case Intrinsic::loongarch_csrxchg_w: |
| 657 | case Intrinsic::loongarch_csrxchg_d: { |
| 658 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); |
| 659 | if (!isUInt<14>(Imm)) |
| 660 | return emitIntrinsicWithChainErrorMessage(Op, ErrorMsgOOR, DAG); |
| 661 | return DAG.getMergeValues( |
| 662 | {DAG.getNode(LoongArchISD::CSRXCHG, DL, GRLenVT, Op0, Op.getOperand(2), |
| 663 | Op.getOperand(3), DAG.getConstant(Imm, DL, GRLenVT)), |
| 664 | Op0}, |
| 665 | DL); |
| 666 | } |
| 667 | case Intrinsic::loongarch_iocsrrd_d: { |
| 668 | if (Subtarget.is64Bit()) |
| 669 | return DAG.getMergeValues( |
| 670 | {DAG.getNode( |
| 671 | LoongArchISD::IOCSRRD_D, DL, GRLenVT, Op0, |
| 672 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(2))), |
| 673 | Op0}, |
| 674 | DL); |
| 675 | else { |
| 676 | DAG.getContext()->emitError( |
| 677 | "llvm.loongarch.crc.w.d.w requires target: loongarch64"); |
| 678 | return DAG.getMergeValues({DAG.getUNDEF(Op.getValueType()), Op0}, DL); |
| 679 | } |
| 680 | } |
| 681 | #define IOCSRRD_CASE(NAME, NODE) \ |
| 682 | case Intrinsic::loongarch_##NAME: { \ |
| 683 | return DAG.getMergeValues( \ |
| 684 | {DAG.getNode(LoongArchISD::NODE, DL, GRLenVT, Op0, Op.getOperand(2)), \ |
| 685 | Op0}, \ |
| 686 | DL); \ |
| 687 | } |
| 688 | IOCSRRD_CASE(iocsrrd_b, IOCSRRD_B); |
| 689 | IOCSRRD_CASE(iocsrrd_h, IOCSRRD_H); |
| 690 | IOCSRRD_CASE(iocsrrd_w, IOCSRRD_W); |
| 691 | #undef IOCSRRD_CASE |
| 692 | case Intrinsic::loongarch_cpucfg: { |
| 693 | return DAG.getMergeValues( |
| 694 | {DAG.getNode(LoongArchISD::CPUCFG, DL, GRLenVT, Op0, Op.getOperand(2)), |
| 695 | Op0}, |
| 696 | DL); |
| 697 | } |
| 698 | case Intrinsic::loongarch_lddir_d: { |
| 699 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue(); |
| 700 | if (!isUInt<8>(Imm)) { |
| 701 | DAG.getContext()->emitError("argument to '" + Op->getOperationName(0) + |
| 702 | "' out of range"); |
| 703 | return DAG.getMergeValues({DAG.getUNDEF(Op.getValueType()), Op0}, DL); |
| 704 | } |
| 705 | |
| 706 | return Op; |
| 707 | } |
| 708 | case Intrinsic::loongarch_movfcsr2gr: { |
| 709 | if (!Subtarget.hasBasicF()) { |
| 710 | DAG.getContext()->emitError( |
| 711 | "llvm.loongarch.movfcsr2gr expects basic f target feature"); |
| 712 | return DAG.getMergeValues( |
| 713 | {DAG.getUNDEF(Op.getValueType()), Op.getOperand(0)}, SDLoc(Op)); |
| 714 | } |
| 715 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); |
| 716 | if (!isUInt<2>(Imm)) { |
| 717 | DAG.getContext()->emitError("argument to '" + Op->getOperationName(0) + |
| 718 | "' " + ErrorMsgOOR); |
| 719 | return DAG.getMergeValues( |
| 720 | {DAG.getUNDEF(Op.getValueType()), Op.getOperand(0)}, SDLoc(Op)); |
| 721 | } |
| 722 | return DAG.getMergeValues( |
| 723 | {DAG.getNode(LoongArchISD::MOVFCSR2GR, DL, Op.getValueType(), |
| 724 | DAG.getConstant(Imm, DL, GRLenVT)), |
| 725 | Op.getOperand(0)}, |
| 726 | DL); |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | |
| 731 | // Helper function that emits error message for intrinsics with void return |
| 732 | // value. |
| 733 | static SDValue emitIntrinsicErrorMessage(SDValue Op, StringRef ErrorMsg, |
| 734 | SelectionDAG &DAG) { |
| 735 | |
| 736 | DAG.getContext()->emitError("argument to '" + Op->getOperationName(0) + "' " + |
| 737 | ErrorMsg); |
| 738 | return Op.getOperand(0); |
| 739 | } |
| 740 | |
| 741 | SDValue LoongArchTargetLowering::lowerINTRINSIC_VOID(SDValue Op, |
| 742 | SelectionDAG &DAG) const { |
| 743 | SDLoc DL(Op); |
| 744 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 745 | SDValue Op0 = Op.getOperand(0); |
| 746 | uint64_t IntrinsicEnum = Op.getConstantOperandVal(1); |
| 747 | SDValue Op2 = Op.getOperand(2); |
| 748 | const StringRef ErrorMsgOOR = "out of range"; |
| 749 | |
| 750 | switch (IntrinsicEnum) { |
| 751 | default: |
| 752 | // TODO: Add more Intrinsics. |
| 753 | return SDValue(); |
| 754 | case Intrinsic::loongarch_cacop_d: |
| 755 | case Intrinsic::loongarch_cacop_w: { |
| 756 | if (IntrinsicEnum == Intrinsic::loongarch_cacop_d && !Subtarget.is64Bit()) { |
| 757 | DAG.getContext()->emitError( |
| 758 | "llvm.loongarch.cacop.d requires target: loongarch64"); |
| 759 | return Op.getOperand(0); |
| 760 | } |
| 761 | if (IntrinsicEnum == Intrinsic::loongarch_cacop_w && Subtarget.is64Bit()) { |
| 762 | DAG.getContext()->emitError( |
| 763 | "llvm.loongarch.cacop.w requires target: loongarch32"); |
| 764 | return Op.getOperand(0); |
| 765 | } |
| 766 | // call void @llvm.loongarch.cacop.[d/w](uimm5, rj, simm12) |
| 767 | unsigned Imm1 = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 768 | if (!isUInt<5>(Imm1)) |
| 769 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 770 | SDValue Op4 = Op.getOperand(4); |
| 771 | int Imm2 = cast<ConstantSDNode>(Op4)->getSExtValue(); |
| 772 | if (!isInt<12>(Imm2)) |
| 773 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 774 | |
| 775 | return Op; |
| 776 | } |
| 777 | |
| 778 | case Intrinsic::loongarch_dbar: { |
| 779 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 780 | if (!isUInt<15>(Imm)) |
| 781 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 782 | |
| 783 | return DAG.getNode(LoongArchISD::DBAR, DL, MVT::Other, Op0, |
| 784 | DAG.getConstant(Imm, DL, GRLenVT)); |
| 785 | } |
| 786 | case Intrinsic::loongarch_ibar: { |
| 787 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 788 | if (!isUInt<15>(Imm)) |
| 789 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 790 | |
| 791 | return DAG.getNode(LoongArchISD::IBAR, DL, MVT::Other, Op0, |
| 792 | DAG.getConstant(Imm, DL, GRLenVT)); |
| 793 | } |
| 794 | case Intrinsic::loongarch_break: { |
| 795 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 796 | if (!isUInt<15>(Imm)) |
| 797 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 798 | |
| 799 | return DAG.getNode(LoongArchISD::BREAK, DL, MVT::Other, Op0, |
| 800 | DAG.getConstant(Imm, DL, GRLenVT)); |
| 801 | } |
| 802 | case Intrinsic::loongarch_movgr2fcsr: { |
| 803 | if (!Subtarget.hasBasicF()) { |
| 804 | DAG.getContext()->emitError( |
| 805 | "llvm.loongarch.movgr2fcsr expects basic f target feature"); |
| 806 | return Op0; |
| 807 | } |
| 808 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 809 | if (!isUInt<2>(Imm)) |
| 810 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 811 | |
| 812 | return DAG.getNode( |
| 813 | LoongArchISD::MOVGR2FCSR, DL, MVT::Other, Op0, |
| 814 | DAG.getConstant(Imm, DL, GRLenVT), |
| 815 | DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, Op.getOperand(3))); |
| 816 | } |
| 817 | case Intrinsic::loongarch_syscall: { |
| 818 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 819 | if (!isUInt<15>(Imm)) |
| 820 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 821 | |
| 822 | return DAG.getNode(LoongArchISD::SYSCALL, DL, MVT::Other, Op0, |
| 823 | DAG.getConstant(Imm, DL, GRLenVT)); |
| 824 | } |
| 825 | #define IOCSRWR_CASE(NAME, NODE) \ |
| 826 | case Intrinsic::loongarch_##NAME: { \ |
| 827 | SDValue Op3 = Op.getOperand(3); \ |
| 828 | if (Subtarget.is64Bit()) \ |
| 829 | return DAG.getNode(LoongArchISD::NODE, DL, MVT::Other, Op0, \ |
| 830 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2), \ |
| 831 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op3)); \ |
| 832 | else \ |
| 833 | return DAG.getNode(LoongArchISD::NODE, DL, MVT::Other, Op0, Op2, Op3); \ |
| 834 | } |
| 835 | IOCSRWR_CASE(iocsrwr_b, IOCSRWR_B); |
| 836 | IOCSRWR_CASE(iocsrwr_h, IOCSRWR_H); |
| 837 | IOCSRWR_CASE(iocsrwr_w, IOCSRWR_W); |
| 838 | #undef IOCSRWR_CASE |
| 839 | case Intrinsic::loongarch_iocsrwr_d: { |
| 840 | if (Subtarget.is64Bit()) |
| 841 | return DAG.getNode( |
| 842 | LoongArchISD::IOCSRWR_D, DL, MVT::Other, Op0, Op2, |
| 843 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op.getOperand(3))); |
| 844 | else { |
| 845 | DAG.getContext()->emitError( |
| 846 | "llvm.loongarch.iocsrwr.d requires target: loongarch64"); |
| 847 | return Op.getOperand(0); |
| 848 | } |
| 849 | } |
| 850 | #define ASRT_LE_GT_CASE(NAME) \ |
| 851 | case Intrinsic::loongarch_##NAME: { \ |
| 852 | if (!Subtarget.is64Bit()) { \ |
| 853 | DAG.getContext()->emitError(Op->getOperationName(0) + \ |
| 854 | " requires target: loongarch64"); \ |
| 855 | return Op.getOperand(0); \ |
| 856 | } \ |
| 857 | return Op; \ |
| 858 | } |
| 859 | ASRT_LE_GT_CASE(asrtle_d) |
| 860 | ASRT_LE_GT_CASE(asrtgt_d) |
| 861 | #undef ASRT_LE_GT_CASE |
| 862 | case Intrinsic::loongarch_ldpte_d: { |
| 863 | unsigned Imm = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue(); |
| 864 | if (!isUInt<8>(Imm)) |
| 865 | return emitIntrinsicErrorMessage(Op, ErrorMsgOOR, DAG); |
| 866 | if (!Subtarget.is64Bit()) { |
| 867 | DAG.getContext()->emitError(Op->getOperationName(0) + |
| 868 | " requires target: loongarch64"); |
| 869 | return Op.getOperand(0); |
| 870 | } |
| 871 | return Op; |
| 872 | } |
| 873 | } |
| 874 | } |
| 875 | |
| 876 | SDValue LoongArchTargetLowering::lowerShiftLeftParts(SDValue Op, |
| 877 | SelectionDAG &DAG) const { |
| 878 | SDLoc DL(Op); |
| 879 | SDValue Lo = Op.getOperand(0); |
| 880 | SDValue Hi = Op.getOperand(1); |
| 881 | SDValue Shamt = Op.getOperand(2); |
| 882 | EVT VT = Lo.getValueType(); |
| 883 | |
| 884 | // if Shamt-GRLen < 0: // Shamt < GRLen |
| 885 | // Lo = Lo << Shamt |
| 886 | // Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (GRLen-1 ^ Shamt)) |
| 887 | // else: |
| 888 | // Lo = 0 |
| 889 | // Hi = Lo << (Shamt-GRLen) |
| 890 | |
| 891 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 892 | SDValue One = DAG.getConstant(1, DL, VT); |
| 893 | SDValue MinusGRLen = DAG.getConstant(-(int)Subtarget.getGRLen(), DL, VT); |
| 894 | SDValue GRLenMinus1 = DAG.getConstant(Subtarget.getGRLen() - 1, DL, VT); |
| 895 | SDValue ShamtMinusGRLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusGRLen); |
| 896 | SDValue GRLenMinus1Shamt = DAG.getNode(ISD::XOR, DL, VT, Shamt, GRLenMinus1); |
| 897 | |
| 898 | SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); |
| 899 | SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One); |
| 900 | SDValue ShiftRightLo = |
| 901 | DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, GRLenMinus1Shamt); |
| 902 | SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); |
| 903 | SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); |
| 904 | SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusGRLen); |
| 905 | |
| 906 | SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusGRLen, Zero, ISD::SETLT); |
| 907 | |
| 908 | Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero); |
| 909 | Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse); |
| 910 | |
| 911 | SDValue Parts[2] = {Lo, Hi}; |
| 912 | return DAG.getMergeValues(Parts, DL); |
| 913 | } |
| 914 | |
| 915 | SDValue LoongArchTargetLowering::lowerShiftRightParts(SDValue Op, |
| 916 | SelectionDAG &DAG, |
| 917 | bool IsSRA) const { |
| 918 | SDLoc DL(Op); |
| 919 | SDValue Lo = Op.getOperand(0); |
| 920 | SDValue Hi = Op.getOperand(1); |
| 921 | SDValue Shamt = Op.getOperand(2); |
| 922 | EVT VT = Lo.getValueType(); |
| 923 | |
| 924 | // SRA expansion: |
| 925 | // if Shamt-GRLen < 0: // Shamt < GRLen |
| 926 | // Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ GRLen-1)) |
| 927 | // Hi = Hi >>s Shamt |
| 928 | // else: |
| 929 | // Lo = Hi >>s (Shamt-GRLen); |
| 930 | // Hi = Hi >>s (GRLen-1) |
| 931 | // |
| 932 | // SRL expansion: |
| 933 | // if Shamt-GRLen < 0: // Shamt < GRLen |
| 934 | // Lo = (Lo >>u Shamt) | ((Hi << 1) << (ShAmt ^ GRLen-1)) |
| 935 | // Hi = Hi >>u Shamt |
| 936 | // else: |
| 937 | // Lo = Hi >>u (Shamt-GRLen); |
| 938 | // Hi = 0; |
| 939 | |
| 940 | unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL; |
| 941 | |
| 942 | SDValue Zero = DAG.getConstant(0, DL, VT); |
| 943 | SDValue One = DAG.getConstant(1, DL, VT); |
| 944 | SDValue MinusGRLen = DAG.getConstant(-(int)Subtarget.getGRLen(), DL, VT); |
| 945 | SDValue GRLenMinus1 = DAG.getConstant(Subtarget.getGRLen() - 1, DL, VT); |
| 946 | SDValue ShamtMinusGRLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusGRLen); |
| 947 | SDValue GRLenMinus1Shamt = DAG.getNode(ISD::XOR, DL, VT, Shamt, GRLenMinus1); |
| 948 | |
| 949 | SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); |
| 950 | SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One); |
| 951 | SDValue ShiftLeftHi = |
| 952 | DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, GRLenMinus1Shamt); |
| 953 | SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi); |
| 954 | SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt); |
| 955 | SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusGRLen); |
| 956 | SDValue HiFalse = |
| 957 | IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, GRLenMinus1) : Zero; |
| 958 | |
| 959 | SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusGRLen, Zero, ISD::SETLT); |
| 960 | |
| 961 | Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse); |
| 962 | Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse); |
| 963 | |
| 964 | SDValue Parts[2] = {Lo, Hi}; |
| 965 | return DAG.getMergeValues(Parts, DL); |
| 966 | } |
| 967 | |
| 968 | // Returns the opcode of the target-specific SDNode that implements the 32-bit |
| 969 | // form of the given Opcode. |
| 970 | static LoongArchISD::NodeType getLoongArchWOpcode(unsigned Opcode) { |
| 971 | switch (Opcode) { |
| 972 | default: |
| 973 | llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 973); |
| 974 | case ISD::SHL: |
| 975 | return LoongArchISD::SLL_W; |
| 976 | case ISD::SRA: |
| 977 | return LoongArchISD::SRA_W; |
| 978 | case ISD::SRL: |
| 979 | return LoongArchISD::SRL_W; |
| 980 | case ISD::ROTR: |
| 981 | return LoongArchISD::ROTR_W; |
| 982 | case ISD::ROTL: |
| 983 | return LoongArchISD::ROTL_W; |
| 984 | case ISD::CTTZ: |
| 985 | return LoongArchISD::CTZ_W; |
| 986 | case ISD::CTLZ: |
| 987 | return LoongArchISD::CLZ_W; |
| 988 | } |
| 989 | } |
| 990 | |
| 991 | // Converts the given i8/i16/i32 operation to a target-specific SelectionDAG |
| 992 | // node. Because i8/i16/i32 isn't a legal type for LA64, these operations would |
| 993 | // otherwise be promoted to i64, making it difficult to select the |
| 994 | // SLL_W/.../*W later one because the fact the operation was originally of |
| 995 | // type i8/i16/i32 is lost. |
| 996 | static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG, int NumOp, |
| 997 | unsigned ExtOpc = ISD::ANY_EXTEND) { |
| 998 | SDLoc DL(N); |
| 999 | LoongArchISD::NodeType WOpcode = getLoongArchWOpcode(N->getOpcode()); |
| 1000 | SDValue NewOp0, NewRes; |
| 1001 | |
| 1002 | switch (NumOp) { |
| 1003 | default: |
| 1004 | llvm_unreachable("Unexpected NumOp")::llvm::llvm_unreachable_internal("Unexpected NumOp", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 1004); |
| 1005 | case 1: { |
| 1006 | NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0)); |
| 1007 | NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0); |
| 1008 | break; |
| 1009 | } |
| 1010 | case 2: { |
| 1011 | NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0)); |
| 1012 | SDValue NewOp1 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(1)); |
| 1013 | NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1); |
| 1014 | break; |
| 1015 | } |
| 1016 | // TODO:Handle more NumOp. |
| 1017 | } |
| 1018 | |
| 1019 | // ReplaceNodeResults requires we maintain the same type for the return |
| 1020 | // value. |
| 1021 | return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes); |
| 1022 | } |
| 1023 | |
| 1024 | void LoongArchTargetLowering::ReplaceNodeResults( |
| 1025 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
| 1026 | SDLoc DL(N); |
| 1027 | EVT VT = N->getValueType(0); |
| 1028 | switch (N->getOpcode()) { |
| 1029 | default: |
| 1030 | llvm_unreachable("Don't know how to legalize this operation")::llvm::llvm_unreachable_internal("Don't know how to legalize this operation" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1030 ); |
| 1031 | case ISD::SHL: |
| 1032 | case ISD::SRA: |
| 1033 | case ISD::SRL: |
| 1034 | case ISD::ROTR: |
| 1035 | assert(VT == MVT::i32 && Subtarget.is64Bit() &&(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1036 , __extension__ __PRETTY_FUNCTION__)) |
| 1036 | "Unexpected custom legalisation")(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1036 , __extension__ __PRETTY_FUNCTION__)); |
| 1037 | if (N->getOperand(1).getOpcode() != ISD::Constant) { |
| 1038 | Results.push_back(customLegalizeToWOp(N, DAG, 2)); |
| 1039 | break; |
| 1040 | } |
| 1041 | break; |
| 1042 | case ISD::ROTL: |
| 1043 | ConstantSDNode *CN; |
| 1044 | if ((CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))) { |
| 1045 | Results.push_back(customLegalizeToWOp(N, DAG, 2)); |
| 1046 | break; |
| 1047 | } |
| 1048 | break; |
| 1049 | case ISD::FP_TO_SINT: { |
| 1050 | assert(VT == MVT::i32 && Subtarget.is64Bit() &&(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1051 , __extension__ __PRETTY_FUNCTION__)) |
| 1051 | "Unexpected custom legalisation")(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1051 , __extension__ __PRETTY_FUNCTION__)); |
| 1052 | SDValue Src = N->getOperand(0); |
| 1053 | EVT FVT = EVT::getFloatingPointVT(N->getValueSizeInBits(0)); |
| 1054 | if (getTypeAction(*DAG.getContext(), Src.getValueType()) != |
| 1055 | TargetLowering::TypeSoftenFloat) { |
| 1056 | SDValue Dst = DAG.getNode(LoongArchISD::FTINT, DL, FVT, Src); |
| 1057 | Results.push_back(DAG.getNode(ISD::BITCAST, DL, VT, Dst)); |
| 1058 | return; |
| 1059 | } |
| 1060 | // If the FP type needs to be softened, emit a library call using the 'si' |
| 1061 | // version. If we left it to default legalization we'd end up with 'di'. |
| 1062 | RTLIB::Libcall LC; |
| 1063 | LC = RTLIB::getFPTOSINT(Src.getValueType(), VT); |
| 1064 | MakeLibCallOptions CallOptions; |
| 1065 | EVT OpVT = Src.getValueType(); |
| 1066 | CallOptions.setTypeListBeforeSoften(OpVT, VT, true); |
| 1067 | SDValue Chain = SDValue(); |
| 1068 | SDValue Result; |
| 1069 | std::tie(Result, Chain) = |
| 1070 | makeLibCall(DAG, LC, VT, Src, CallOptions, DL, Chain); |
| 1071 | Results.push_back(Result); |
| 1072 | break; |
| 1073 | } |
| 1074 | case ISD::BITCAST: { |
| 1075 | SDValue Src = N->getOperand(0); |
| 1076 | EVT SrcVT = Src.getValueType(); |
| 1077 | if (VT == MVT::i32 && SrcVT == MVT::f32 && Subtarget.is64Bit() && |
| 1078 | Subtarget.hasBasicF()) { |
| 1079 | SDValue Dst = |
| 1080 | DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Src); |
| 1081 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Dst)); |
| 1082 | } |
| 1083 | break; |
| 1084 | } |
| 1085 | case ISD::FP_TO_UINT: { |
| 1086 | assert(VT == MVT::i32 && Subtarget.is64Bit() &&(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1087 , __extension__ __PRETTY_FUNCTION__)) |
| 1087 | "Unexpected custom legalisation")(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1087 , __extension__ __PRETTY_FUNCTION__)); |
| 1088 | auto &TLI = DAG.getTargetLoweringInfo(); |
| 1089 | SDValue Tmp1, Tmp2; |
| 1090 | TLI.expandFP_TO_UINT(N, Tmp1, Tmp2, DAG); |
| 1091 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Tmp1)); |
| 1092 | break; |
| 1093 | } |
| 1094 | case ISD::BSWAP: { |
| 1095 | SDValue Src = N->getOperand(0); |
| 1096 | assert((VT == MVT::i16 || VT == MVT::i32) &&(static_cast <bool> ((VT == MVT::i16 || VT == MVT::i32) && "Unexpected custom legalization") ? void (0) : __assert_fail ("(VT == MVT::i16 || VT == MVT::i32) && \"Unexpected custom legalization\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1097 , __extension__ __PRETTY_FUNCTION__)) |
| 1097 | "Unexpected custom legalization")(static_cast <bool> ((VT == MVT::i16 || VT == MVT::i32) && "Unexpected custom legalization") ? void (0) : __assert_fail ("(VT == MVT::i16 || VT == MVT::i32) && \"Unexpected custom legalization\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1097 , __extension__ __PRETTY_FUNCTION__)); |
| 1098 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1099 | SDValue NewSrc = DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, Src); |
| 1100 | SDValue Tmp; |
| 1101 | switch (VT.getSizeInBits()) { |
| 1102 | default: |
| 1103 | llvm_unreachable("Unexpected operand width")::llvm::llvm_unreachable_internal("Unexpected operand width", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1103); |
| 1104 | case 16: |
| 1105 | Tmp = DAG.getNode(LoongArchISD::REVB_2H, DL, GRLenVT, NewSrc); |
| 1106 | break; |
| 1107 | case 32: |
| 1108 | // Only LA64 will get to here due to the size mismatch between VT and |
| 1109 | // GRLenVT, LA32 lowering is directly defined in LoongArchInstrInfo. |
| 1110 | Tmp = DAG.getNode(LoongArchISD::REVB_2W, DL, GRLenVT, NewSrc); |
| 1111 | break; |
| 1112 | } |
| 1113 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Tmp)); |
| 1114 | break; |
| 1115 | } |
| 1116 | case ISD::BITREVERSE: { |
| 1117 | SDValue Src = N->getOperand(0); |
| 1118 | assert((VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) &&(static_cast <bool> ((VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) && "Unexpected custom legalization" ) ? void (0) : __assert_fail ("(VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) && \"Unexpected custom legalization\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1119 , __extension__ __PRETTY_FUNCTION__)) |
| 1119 | "Unexpected custom legalization")(static_cast <bool> ((VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) && "Unexpected custom legalization" ) ? void (0) : __assert_fail ("(VT == MVT::i8 || (VT == MVT::i32 && Subtarget.is64Bit())) && \"Unexpected custom legalization\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1119 , __extension__ __PRETTY_FUNCTION__)); |
| 1120 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1121 | SDValue NewSrc = DAG.getNode(ISD::ANY_EXTEND, DL, GRLenVT, Src); |
| 1122 | SDValue Tmp; |
| 1123 | switch (VT.getSizeInBits()) { |
| 1124 | default: |
| 1125 | llvm_unreachable("Unexpected operand width")::llvm::llvm_unreachable_internal("Unexpected operand width", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1125); |
| 1126 | case 8: |
| 1127 | Tmp = DAG.getNode(LoongArchISD::BITREV_4B, DL, GRLenVT, NewSrc); |
| 1128 | break; |
| 1129 | case 32: |
| 1130 | Tmp = DAG.getNode(LoongArchISD::BITREV_W, DL, GRLenVT, NewSrc); |
| 1131 | break; |
| 1132 | } |
| 1133 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Tmp)); |
| 1134 | break; |
| 1135 | } |
| 1136 | case ISD::CTLZ: |
| 1137 | case ISD::CTTZ: { |
| 1138 | assert(VT == MVT::i32 && Subtarget.is64Bit() &&(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1139 , __extension__ __PRETTY_FUNCTION__)) |
| 1139 | "Unexpected custom legalisation")(static_cast <bool> (VT == MVT::i32 && Subtarget .is64Bit() && "Unexpected custom legalisation") ? void (0) : __assert_fail ("VT == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1139 , __extension__ __PRETTY_FUNCTION__)); |
| 1140 | Results.push_back(customLegalizeToWOp(N, DAG, 1)); |
| 1141 | break; |
| 1142 | } |
| 1143 | case ISD::INTRINSIC_W_CHAIN: { |
| 1144 | SDValue Op0 = N->getOperand(0); |
| 1145 | EVT VT = N->getValueType(0); |
| 1146 | uint64_t Op1 = N->getConstantOperandVal(1); |
| 1147 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1148 | if (Op1 == Intrinsic::loongarch_movfcsr2gr) { |
| 1149 | if (!Subtarget.hasBasicF()) { |
| 1150 | DAG.getContext()->emitError( |
| 1151 | "llvm.loongarch.movfcsr2gr expects basic f target feature"); |
| 1152 | Results.push_back(DAG.getMergeValues( |
| 1153 | {DAG.getUNDEF(N->getValueType(0)), N->getOperand(0)}, SDLoc(N))); |
| 1154 | Results.push_back(N->getOperand(0)); |
| 1155 | return; |
| 1156 | } |
| 1157 | unsigned Imm = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(); |
| 1158 | if (!isUInt<2>(Imm)) { |
| 1159 | DAG.getContext()->emitError("argument to '" + N->getOperationName(0) + |
| 1160 | "' " + "out of range"); |
| 1161 | Results.push_back(DAG.getMergeValues( |
| 1162 | {DAG.getUNDEF(N->getValueType(0)), N->getOperand(0)}, SDLoc(N))); |
| 1163 | Results.push_back(N->getOperand(0)); |
| 1164 | return; |
| 1165 | } |
| 1166 | Results.push_back( |
| 1167 | DAG.getNode(ISD::TRUNCATE, DL, VT, |
| 1168 | DAG.getNode(LoongArchISD::MOVFCSR2GR, SDLoc(N), MVT::i64, |
| 1169 | DAG.getConstant(Imm, DL, GRLenVT)))); |
| 1170 | Results.push_back(N->getOperand(0)); |
| 1171 | return; |
| 1172 | } |
| 1173 | SDValue Op2 = N->getOperand(2); |
| 1174 | std::string Name = N->getOperationName(0); |
| 1175 | |
| 1176 | switch (Op1) { |
| 1177 | default: |
| 1178 | llvm_unreachable("Unexpected Intrinsic.")::llvm::llvm_unreachable_internal("Unexpected Intrinsic.", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 1178); |
| 1179 | #define CRC_CASE_EXT_BINARYOP(NAME, NODE) \ |
| 1180 | case Intrinsic::loongarch_##NAME: { \ |
| 1181 | Results.push_back(DAG.getNode( \ |
| 1182 | ISD::TRUNCATE, DL, VT, \ |
| 1183 | DAG.getNode( \ |
| 1184 | LoongArchISD::NODE, DL, MVT::i64, \ |
| 1185 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2), \ |
| 1186 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(3))))); \ |
| 1187 | Results.push_back(N->getOperand(0)); \ |
| 1188 | break; \ |
| 1189 | } |
| 1190 | CRC_CASE_EXT_BINARYOP(crc_w_b_w, CRC_W_B_W) |
| 1191 | CRC_CASE_EXT_BINARYOP(crc_w_h_w, CRC_W_H_W) |
| 1192 | CRC_CASE_EXT_BINARYOP(crc_w_w_w, CRC_W_W_W) |
| 1193 | CRC_CASE_EXT_BINARYOP(crcc_w_b_w, CRCC_W_B_W) |
| 1194 | CRC_CASE_EXT_BINARYOP(crcc_w_h_w, CRCC_W_H_W) |
| 1195 | CRC_CASE_EXT_BINARYOP(crcc_w_w_w, CRCC_W_W_W) |
| 1196 | #undef CRC_CASE_EXT_BINARYOP |
| 1197 | |
| 1198 | #define CRC_CASE_EXT_UNARYOP(NAME, NODE) \ |
| 1199 | case Intrinsic::loongarch_##NAME: { \ |
| 1200 | Results.push_back( \ |
| 1201 | DAG.getNode(ISD::TRUNCATE, DL, VT, \ |
| 1202 | DAG.getNode(LoongArchISD::NODE, DL, MVT::i64, Op2, \ |
| 1203 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, \ |
| 1204 | N->getOperand(3))))); \ |
| 1205 | Results.push_back(N->getOperand(0)); \ |
| 1206 | break; \ |
| 1207 | } |
| 1208 | CRC_CASE_EXT_UNARYOP(crc_w_d_w, CRC_W_D_W) |
| 1209 | CRC_CASE_EXT_UNARYOP(crcc_w_d_w, CRCC_W_D_W) |
| 1210 | #undef CRC_CASE_EXT_UNARYOP |
| 1211 | #define CSR_CASE(ID) \ |
| 1212 | case Intrinsic::loongarch_##ID: { \ |
| 1213 | if (!Subtarget.is64Bit()) { \ |
| 1214 | DAG.getContext()->emitError(Name + " requires target: loongarch64"); \ |
| 1215 | Results.push_back(DAG.getUNDEF(VT)); \ |
| 1216 | Results.push_back(N->getOperand(0)); \ |
| 1217 | } \ |
| 1218 | break; \ |
| 1219 | } |
| 1220 | CSR_CASE(csrrd_d); |
| 1221 | CSR_CASE(csrwr_d); |
| 1222 | CSR_CASE(csrxchg_d); |
| 1223 | CSR_CASE(iocsrrd_d); |
| 1224 | #undef CSR_CASE |
| 1225 | case Intrinsic::loongarch_csrrd_w: { |
| 1226 | unsigned Imm = cast<ConstantSDNode>(Op2)->getZExtValue(); |
| 1227 | if (!isUInt<14>(Imm)) { |
| 1228 | DAG.getContext()->emitError("argument to '" + Name + "' out of range"); |
| 1229 | Results.push_back(DAG.getUNDEF(VT)); |
| 1230 | Results.push_back(N->getOperand(0)); |
| 1231 | break; |
| 1232 | } |
| 1233 | |
| 1234 | Results.push_back( |
| 1235 | DAG.getNode(ISD::TRUNCATE, DL, VT, |
| 1236 | DAG.getNode(LoongArchISD::CSRRD, DL, GRLenVT, Op0, |
| 1237 | DAG.getConstant(Imm, DL, GRLenVT)))); |
| 1238 | Results.push_back(N->getOperand(0)); |
| 1239 | break; |
| 1240 | } |
| 1241 | case Intrinsic::loongarch_csrwr_w: { |
| 1242 | unsigned Imm = cast<ConstantSDNode>(N->getOperand(3))->getZExtValue(); |
| 1243 | if (!isUInt<14>(Imm)) { |
| 1244 | DAG.getContext()->emitError("argument to '" + Name + "' out of range"); |
| 1245 | Results.push_back(DAG.getUNDEF(VT)); |
| 1246 | Results.push_back(N->getOperand(0)); |
| 1247 | break; |
| 1248 | } |
| 1249 | |
| 1250 | Results.push_back(DAG.getNode( |
| 1251 | ISD::TRUNCATE, DL, VT, |
| 1252 | DAG.getNode(LoongArchISD::CSRWR, DL, GRLenVT, Op0, |
| 1253 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2), |
| 1254 | DAG.getConstant(Imm, DL, GRLenVT)))); |
| 1255 | Results.push_back(N->getOperand(0)); |
| 1256 | break; |
| 1257 | } |
| 1258 | case Intrinsic::loongarch_csrxchg_w: { |
| 1259 | unsigned Imm = cast<ConstantSDNode>(N->getOperand(4))->getZExtValue(); |
| 1260 | if (!isUInt<14>(Imm)) { |
| 1261 | DAG.getContext()->emitError("argument to '" + Name + "' out of range"); |
| 1262 | Results.push_back(DAG.getUNDEF(VT)); |
| 1263 | Results.push_back(N->getOperand(0)); |
| 1264 | break; |
| 1265 | } |
| 1266 | |
| 1267 | Results.push_back(DAG.getNode( |
| 1268 | ISD::TRUNCATE, DL, VT, |
| 1269 | DAG.getNode( |
| 1270 | LoongArchISD::CSRXCHG, DL, GRLenVT, Op0, |
| 1271 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2), |
| 1272 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(3)), |
| 1273 | DAG.getConstant(Imm, DL, GRLenVT)))); |
| 1274 | Results.push_back(N->getOperand(0)); |
| 1275 | break; |
| 1276 | } |
| 1277 | #define IOCSRRD_CASE(NAME, NODE) \ |
| 1278 | case Intrinsic::loongarch_##NAME: { \ |
| 1279 | Results.push_back(DAG.getNode( \ |
| 1280 | ISD::TRUNCATE, DL, N->getValueType(0), \ |
| 1281 | DAG.getNode(LoongArchISD::NODE, DL, MVT::i64, Op0, \ |
| 1282 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2)))); \ |
| 1283 | Results.push_back(N->getOperand(0)); \ |
| 1284 | break; \ |
| 1285 | } |
| 1286 | IOCSRRD_CASE(iocsrrd_b, IOCSRRD_B); |
| 1287 | IOCSRRD_CASE(iocsrrd_h, IOCSRRD_H); |
| 1288 | IOCSRRD_CASE(iocsrrd_w, IOCSRRD_W); |
| 1289 | #undef IOCSRRD_CASE |
| 1290 | case Intrinsic::loongarch_cpucfg: { |
| 1291 | Results.push_back(DAG.getNode( |
| 1292 | ISD::TRUNCATE, DL, VT, |
| 1293 | DAG.getNode(LoongArchISD::CPUCFG, DL, GRLenVT, Op0, |
| 1294 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op2)))); |
| 1295 | Results.push_back(Op0); |
| 1296 | break; |
| 1297 | } |
| 1298 | case Intrinsic::loongarch_lddir_d: { |
| 1299 | if (!Subtarget.is64Bit()) { |
| 1300 | DAG.getContext()->emitError(N->getOperationName(0) + |
| 1301 | " requires target: loongarch64"); |
| 1302 | Results.push_back(DAG.getUNDEF(VT)); |
| 1303 | Results.push_back(Op0); |
| 1304 | break; |
| 1305 | } |
| 1306 | break; |
| 1307 | } |
| 1308 | } |
| 1309 | break; |
| 1310 | } |
| 1311 | case ISD::READ_REGISTER: { |
| 1312 | if (Subtarget.is64Bit()) |
| 1313 | DAG.getContext()->emitError( |
| 1314 | "On LA64, only 64-bit registers can be read."); |
| 1315 | else |
| 1316 | DAG.getContext()->emitError( |
| 1317 | "On LA32, only 32-bit registers can be read."); |
| 1318 | Results.push_back(DAG.getUNDEF(VT)); |
| 1319 | Results.push_back(N->getOperand(0)); |
| 1320 | break; |
| 1321 | } |
| 1322 | } |
| 1323 | } |
| 1324 | |
| 1325 | static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, |
| 1326 | TargetLowering::DAGCombinerInfo &DCI, |
| 1327 | const LoongArchSubtarget &Subtarget) { |
| 1328 | if (DCI.isBeforeLegalizeOps()) |
| 1329 | return SDValue(); |
| 1330 | |
| 1331 | SDValue FirstOperand = N->getOperand(0); |
| 1332 | SDValue SecondOperand = N->getOperand(1); |
| 1333 | unsigned FirstOperandOpc = FirstOperand.getOpcode(); |
| 1334 | EVT ValTy = N->getValueType(0); |
| 1335 | SDLoc DL(N); |
| 1336 | uint64_t lsb, msb; |
| 1337 | unsigned SMIdx, SMLen; |
| 1338 | ConstantSDNode *CN; |
| 1339 | SDValue NewOperand; |
| 1340 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1341 | |
| 1342 | // Op's second operand must be a shifted mask. |
| 1343 | if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand)) || |
| 1344 | !isShiftedMask_64(CN->getZExtValue(), SMIdx, SMLen)) |
| 1345 | return SDValue(); |
| 1346 | |
| 1347 | if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { |
| 1348 | // Pattern match BSTRPICK. |
| 1349 | // $dst = and ((sra or srl) $src , lsb), (2**len - 1) |
| 1350 | // => BSTRPICK $dst, $src, msb, lsb |
| 1351 | // where msb = lsb + len - 1 |
| 1352 | |
| 1353 | // The second operand of the shift must be an immediate. |
| 1354 | if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1)))) |
| 1355 | return SDValue(); |
| 1356 | |
| 1357 | lsb = CN->getZExtValue(); |
| 1358 | |
| 1359 | // Return if the shifted mask does not start at bit 0 or the sum of its |
| 1360 | // length and lsb exceeds the word's size. |
| 1361 | if (SMIdx != 0 || lsb + SMLen > ValTy.getSizeInBits()) |
| 1362 | return SDValue(); |
| 1363 | |
| 1364 | NewOperand = FirstOperand.getOperand(0); |
| 1365 | } else { |
| 1366 | // Pattern match BSTRPICK. |
| 1367 | // $dst = and $src, (2**len- 1) , if len > 12 |
| 1368 | // => BSTRPICK $dst, $src, msb, lsb |
| 1369 | // where lsb = 0 and msb = len - 1 |
| 1370 | |
| 1371 | // If the mask is <= 0xfff, andi can be used instead. |
| 1372 | if (CN->getZExtValue() <= 0xfff) |
| 1373 | return SDValue(); |
| 1374 | |
| 1375 | // Return if the MSB exceeds. |
| 1376 | if (SMIdx + SMLen > ValTy.getSizeInBits()) |
| 1377 | return SDValue(); |
| 1378 | |
| 1379 | if (SMIdx > 0) { |
| 1380 | // Omit if the constant has more than 2 uses. This a conservative |
| 1381 | // decision. Whether it is a win depends on the HW microarchitecture. |
| 1382 | // However it should always be better for 1 and 2 uses. |
| 1383 | if (CN->use_size() > 2) |
| 1384 | return SDValue(); |
| 1385 | // Return if the constant can be composed by a single LU12I.W. |
| 1386 | if ((CN->getZExtValue() & 0xfff) == 0) |
| 1387 | return SDValue(); |
| 1388 | // Return if the constand can be composed by a single ADDI with |
| 1389 | // the zero register. |
| 1390 | if (CN->getSExtValue() >= -2048 && CN->getSExtValue() < 0) |
| 1391 | return SDValue(); |
| 1392 | } |
| 1393 | |
| 1394 | lsb = SMIdx; |
| 1395 | NewOperand = FirstOperand; |
| 1396 | } |
| 1397 | |
| 1398 | msb = lsb + SMLen - 1; |
| 1399 | SDValue NR0 = DAG.getNode(LoongArchISD::BSTRPICK, DL, ValTy, NewOperand, |
| 1400 | DAG.getConstant(msb, DL, GRLenVT), |
| 1401 | DAG.getConstant(lsb, DL, GRLenVT)); |
| 1402 | if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL || lsb == 0) |
| 1403 | return NR0; |
| 1404 | // Try to optimize to |
| 1405 | // bstrpick $Rd, $Rs, msb, lsb |
| 1406 | // slli $Rd, $Rd, lsb |
| 1407 | return DAG.getNode(ISD::SHL, DL, ValTy, NR0, |
| 1408 | DAG.getConstant(lsb, DL, GRLenVT)); |
| 1409 | } |
| 1410 | |
| 1411 | static SDValue performSRLCombine(SDNode *N, SelectionDAG &DAG, |
| 1412 | TargetLowering::DAGCombinerInfo &DCI, |
| 1413 | const LoongArchSubtarget &Subtarget) { |
| 1414 | if (DCI.isBeforeLegalizeOps()) |
| 1415 | return SDValue(); |
| 1416 | |
| 1417 | // $dst = srl (and $src, Mask), Shamt |
| 1418 | // => |
| 1419 | // BSTRPICK $dst, $src, MaskIdx+MaskLen-1, Shamt |
| 1420 | // when Mask is a shifted mask, and MaskIdx <= Shamt <= MaskIdx+MaskLen-1 |
| 1421 | // |
| 1422 | |
| 1423 | SDValue FirstOperand = N->getOperand(0); |
| 1424 | ConstantSDNode *CN; |
| 1425 | EVT ValTy = N->getValueType(0); |
| 1426 | SDLoc DL(N); |
| 1427 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1428 | unsigned MaskIdx, MaskLen; |
| 1429 | uint64_t Shamt; |
| 1430 | |
| 1431 | // The first operand must be an AND and the second operand of the AND must be |
| 1432 | // a shifted mask. |
| 1433 | if (FirstOperand.getOpcode() != ISD::AND || |
| 1434 | !(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) || |
| 1435 | !isShiftedMask_64(CN->getZExtValue(), MaskIdx, MaskLen)) |
| 1436 | return SDValue(); |
| 1437 | |
| 1438 | // The second operand (shift amount) must be an immediate. |
| 1439 | if (!(CN = dyn_cast<ConstantSDNode>(N->getOperand(1)))) |
| 1440 | return SDValue(); |
| 1441 | |
| 1442 | Shamt = CN->getZExtValue(); |
| 1443 | if (MaskIdx <= Shamt && Shamt <= MaskIdx + MaskLen - 1) |
| 1444 | return DAG.getNode(LoongArchISD::BSTRPICK, DL, ValTy, |
| 1445 | FirstOperand->getOperand(0), |
| 1446 | DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT), |
| 1447 | DAG.getConstant(Shamt, DL, GRLenVT)); |
| 1448 | |
| 1449 | return SDValue(); |
| 1450 | } |
| 1451 | |
| 1452 | static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, |
| 1453 | TargetLowering::DAGCombinerInfo &DCI, |
| 1454 | const LoongArchSubtarget &Subtarget) { |
| 1455 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 1456 | EVT ValTy = N->getValueType(0); |
| 1457 | SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); |
| 1458 | ConstantSDNode *CN0, *CN1; |
| 1459 | SDLoc DL(N); |
| 1460 | unsigned ValBits = ValTy.getSizeInBits(); |
| 1461 | unsigned MaskIdx0, MaskLen0, MaskIdx1, MaskLen1; |
| 1462 | unsigned Shamt; |
| 1463 | bool SwapAndRetried = false; |
| 1464 | |
| 1465 | if (DCI.isBeforeLegalizeOps()) |
| 1466 | return SDValue(); |
| 1467 | |
| 1468 | if (ValBits != 32 && ValBits != 64) |
| 1469 | return SDValue(); |
| 1470 | |
| 1471 | Retry: |
| 1472 | // 1st pattern to match BSTRINS: |
| 1473 | // R = or (and X, mask0), (and (shl Y, lsb), mask1) |
| 1474 | // where mask1 = (2**size - 1) << lsb, mask0 = ~mask1 |
| 1475 | // => |
| 1476 | // R = BSTRINS X, Y, msb, lsb (where msb = lsb + size - 1) |
| 1477 | if (N0.getOpcode() == ISD::AND && |
| 1478 | (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) && |
| 1479 | isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) && |
| 1480 | N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL && |
| 1481 | (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1482 | isShiftedMask_64(CN1->getZExtValue(), MaskIdx1, MaskLen1) && |
| 1483 | MaskIdx0 == MaskIdx1 && MaskLen0 == MaskLen1 && |
| 1484 | (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) && |
| 1485 | (Shamt = CN1->getZExtValue()) == MaskIdx0 && |
| 1486 | (MaskIdx0 + MaskLen0 <= ValBits)) { |
| 1487 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 1\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 1\n" ; } } while (false); |
| 1488 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0), |
| 1489 | N1.getOperand(0).getOperand(0), |
| 1490 | DAG.getConstant((MaskIdx0 + MaskLen0 - 1), DL, GRLenVT), |
| 1491 | DAG.getConstant(MaskIdx0, DL, GRLenVT)); |
| 1492 | } |
| 1493 | |
| 1494 | // 2nd pattern to match BSTRINS: |
| 1495 | // R = or (and X, mask0), (shl (and Y, mask1), lsb) |
| 1496 | // where mask1 = (2**size - 1), mask0 = ~(mask1 << lsb) |
| 1497 | // => |
| 1498 | // R = BSTRINS X, Y, msb, lsb (where msb = lsb + size - 1) |
| 1499 | if (N0.getOpcode() == ISD::AND && |
| 1500 | (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) && |
| 1501 | isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) && |
| 1502 | N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::AND && |
| 1503 | (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1504 | (Shamt = CN1->getZExtValue()) == MaskIdx0 && |
Although the value stored to 'Shamt' is used in the enclosing expression, the value is never actually read from 'Shamt' | |
| 1505 | (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) && |
| 1506 | isShiftedMask_64(CN1->getZExtValue(), MaskIdx1, MaskLen1) && |
| 1507 | MaskLen0 == MaskLen1 && MaskIdx1 == 0 && |
| 1508 | (MaskIdx0 + MaskLen0 <= ValBits)) { |
| 1509 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 2\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 2\n" ; } } while (false); |
| 1510 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0), |
| 1511 | N1.getOperand(0).getOperand(0), |
| 1512 | DAG.getConstant((MaskIdx0 + MaskLen0 - 1), DL, GRLenVT), |
| 1513 | DAG.getConstant(MaskIdx0, DL, GRLenVT)); |
| 1514 | } |
| 1515 | |
| 1516 | // 3rd pattern to match BSTRINS: |
| 1517 | // R = or (and X, mask0), (and Y, mask1) |
| 1518 | // where ~mask0 = (2**size - 1) << lsb, mask0 & mask1 = 0 |
| 1519 | // => |
| 1520 | // R = BSTRINS X, (shr (and Y, mask1), lsb), msb, lsb |
| 1521 | // where msb = lsb + size - 1 |
| 1522 | if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND && |
| 1523 | (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) && |
| 1524 | isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) && |
| 1525 | (MaskIdx0 + MaskLen0 <= 64) && |
| 1526 | (CN1 = dyn_cast<ConstantSDNode>(N1->getOperand(1))) && |
| 1527 | (CN1->getSExtValue() & CN0->getSExtValue()) == 0) { |
| 1528 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 3\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 3\n" ; } } while (false); |
| 1529 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0), |
| 1530 | DAG.getNode(ISD::SRL, DL, N1->getValueType(0), N1, |
| 1531 | DAG.getConstant(MaskIdx0, DL, GRLenVT)), |
| 1532 | DAG.getConstant(ValBits == 32 |
| 1533 | ? (MaskIdx0 + (MaskLen0 & 31) - 1) |
| 1534 | : (MaskIdx0 + MaskLen0 - 1), |
| 1535 | DL, GRLenVT), |
| 1536 | DAG.getConstant(MaskIdx0, DL, GRLenVT)); |
| 1537 | } |
| 1538 | |
| 1539 | // 4th pattern to match BSTRINS: |
| 1540 | // R = or (and X, mask), (shl Y, shamt) |
| 1541 | // where mask = (2**shamt - 1) |
| 1542 | // => |
| 1543 | // R = BSTRINS X, Y, ValBits - 1, shamt |
| 1544 | // where ValBits = 32 or 64 |
| 1545 | if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::SHL && |
| 1546 | (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) && |
| 1547 | isShiftedMask_64(CN0->getZExtValue(), MaskIdx0, MaskLen0) && |
| 1548 | MaskIdx0 == 0 && (CN1 = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1549 | (Shamt = CN1->getZExtValue()) == MaskLen0 && |
| 1550 | (MaskIdx0 + MaskLen0 <= ValBits)) { |
| 1551 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 4\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 4\n" ; } } while (false); |
| 1552 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0), |
| 1553 | N1.getOperand(0), |
| 1554 | DAG.getConstant((ValBits - 1), DL, GRLenVT), |
| 1555 | DAG.getConstant(Shamt, DL, GRLenVT)); |
| 1556 | } |
| 1557 | |
| 1558 | // 5th pattern to match BSTRINS: |
| 1559 | // R = or (and X, mask), const |
| 1560 | // where ~mask = (2**size - 1) << lsb, mask & const = 0 |
| 1561 | // => |
| 1562 | // R = BSTRINS X, (const >> lsb), msb, lsb |
| 1563 | // where msb = lsb + size - 1 |
| 1564 | if (N0.getOpcode() == ISD::AND && |
| 1565 | (CN0 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) && |
| 1566 | isShiftedMask_64(~CN0->getSExtValue(), MaskIdx0, MaskLen0) && |
| 1567 | (CN1 = dyn_cast<ConstantSDNode>(N1)) && |
| 1568 | (CN1->getSExtValue() & CN0->getSExtValue()) == 0) { |
| 1569 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 5\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 5\n" ; } } while (false); |
| 1570 | return DAG.getNode( |
| 1571 | LoongArchISD::BSTRINS, DL, ValTy, N0.getOperand(0), |
| 1572 | DAG.getConstant(CN1->getSExtValue() >> MaskIdx0, DL, ValTy), |
| 1573 | DAG.getConstant((MaskIdx0 + MaskLen0 - 1), DL, GRLenVT), |
| 1574 | DAG.getConstant(MaskIdx0, DL, GRLenVT)); |
| 1575 | } |
| 1576 | |
| 1577 | // 6th pattern. |
| 1578 | // a = b | ((c & mask) << shamt), where all positions in b to be overwritten |
| 1579 | // by the incoming bits are known to be zero. |
| 1580 | // => |
| 1581 | // a = BSTRINS b, c, shamt + MaskLen - 1, shamt |
| 1582 | // |
| 1583 | // Note that the 1st pattern is a special situation of the 6th, i.e. the 6th |
| 1584 | // pattern is more common than the 1st. So we put the 1st before the 6th in |
| 1585 | // order to match as many nodes as possible. |
| 1586 | ConstantSDNode *CNMask, *CNShamt; |
| 1587 | unsigned MaskIdx, MaskLen; |
| 1588 | if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::AND && |
| 1589 | (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) && |
| 1590 | isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen) && |
| 1591 | MaskIdx == 0 && (CNShamt = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1592 | CNShamt->getZExtValue() + MaskLen <= ValBits) { |
| 1593 | Shamt = CNShamt->getZExtValue(); |
| 1594 | APInt ShMask(ValBits, CNMask->getZExtValue() << Shamt); |
| 1595 | if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) { |
| 1596 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 6\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 6\n" ; } } while (false); |
| 1597 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0, |
| 1598 | N1.getOperand(0).getOperand(0), |
| 1599 | DAG.getConstant(Shamt + MaskLen - 1, DL, GRLenVT), |
| 1600 | DAG.getConstant(Shamt, DL, GRLenVT)); |
| 1601 | } |
| 1602 | } |
| 1603 | |
| 1604 | // 7th pattern. |
| 1605 | // a = b | ((c << shamt) & shifted_mask), where all positions in b to be |
| 1606 | // overwritten by the incoming bits are known to be zero. |
| 1607 | // => |
| 1608 | // a = BSTRINS b, c, MaskIdx + MaskLen - 1, MaskIdx |
| 1609 | // |
| 1610 | // Similarly, the 7th pattern is more common than the 2nd. So we put the 2nd |
| 1611 | // before the 7th in order to match as many nodes as possible. |
| 1612 | if (N1.getOpcode() == ISD::AND && |
| 1613 | (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1614 | isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen) && |
| 1615 | N1.getOperand(0).getOpcode() == ISD::SHL && |
| 1616 | (CNShamt = dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(1))) && |
| 1617 | CNShamt->getZExtValue() == MaskIdx) { |
| 1618 | APInt ShMask(ValBits, CNMask->getZExtValue()); |
| 1619 | if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) { |
| 1620 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 7\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 7\n" ; } } while (false); |
| 1621 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0, |
| 1622 | N1.getOperand(0).getOperand(0), |
| 1623 | DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT), |
| 1624 | DAG.getConstant(MaskIdx, DL, GRLenVT)); |
| 1625 | } |
| 1626 | } |
| 1627 | |
| 1628 | // (or a, b) and (or b, a) are equivalent, so swap the operands and retry. |
| 1629 | if (!SwapAndRetried) { |
| 1630 | std::swap(N0, N1); |
| 1631 | SwapAndRetried = true; |
| 1632 | goto Retry; |
| 1633 | } |
| 1634 | |
| 1635 | SwapAndRetried = false; |
| 1636 | Retry2: |
| 1637 | // 8th pattern. |
| 1638 | // a = b | (c & shifted_mask), where all positions in b to be overwritten by |
| 1639 | // the incoming bits are known to be zero. |
| 1640 | // => |
| 1641 | // a = BSTRINS b, c >> MaskIdx, MaskIdx + MaskLen - 1, MaskIdx |
| 1642 | // |
| 1643 | // Similarly, the 8th pattern is more common than the 4th and 5th patterns. So |
| 1644 | // we put it here in order to match as many nodes as possible or generate less |
| 1645 | // instructions. |
| 1646 | if (N1.getOpcode() == ISD::AND && |
| 1647 | (CNMask = dyn_cast<ConstantSDNode>(N1.getOperand(1))) && |
| 1648 | isShiftedMask_64(CNMask->getZExtValue(), MaskIdx, MaskLen)) { |
| 1649 | APInt ShMask(ValBits, CNMask->getZExtValue()); |
| 1650 | if (ShMask.isSubsetOf(DAG.computeKnownBits(N0).Zero)) { |
| 1651 | LLVM_DEBUG(dbgs() << "Perform OR combine: match pattern 8\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "Perform OR combine: match pattern 8\n" ; } } while (false); |
| 1652 | return DAG.getNode(LoongArchISD::BSTRINS, DL, ValTy, N0, |
| 1653 | DAG.getNode(ISD::SRL, DL, N1->getValueType(0), |
| 1654 | N1->getOperand(0), |
| 1655 | DAG.getConstant(MaskIdx, DL, GRLenVT)), |
| 1656 | DAG.getConstant(MaskIdx + MaskLen - 1, DL, GRLenVT), |
| 1657 | DAG.getConstant(MaskIdx, DL, GRLenVT)); |
| 1658 | } |
| 1659 | } |
| 1660 | // Swap N0/N1 and retry. |
| 1661 | if (!SwapAndRetried) { |
| 1662 | std::swap(N0, N1); |
| 1663 | SwapAndRetried = true; |
| 1664 | goto Retry2; |
| 1665 | } |
| 1666 | |
| 1667 | return SDValue(); |
| 1668 | } |
| 1669 | |
| 1670 | // Combine (loongarch_bitrev_w (loongarch_revb_2w X)) to loongarch_bitrev_4b. |
| 1671 | static SDValue performBITREV_WCombine(SDNode *N, SelectionDAG &DAG, |
| 1672 | TargetLowering::DAGCombinerInfo &DCI, |
| 1673 | const LoongArchSubtarget &Subtarget) { |
| 1674 | if (DCI.isBeforeLegalizeOps()) |
| 1675 | return SDValue(); |
| 1676 | |
| 1677 | SDValue Src = N->getOperand(0); |
| 1678 | if (Src.getOpcode() != LoongArchISD::REVB_2W) |
| 1679 | return SDValue(); |
| 1680 | |
| 1681 | return DAG.getNode(LoongArchISD::BITREV_4B, SDLoc(N), N->getValueType(0), |
| 1682 | Src.getOperand(0)); |
| 1683 | } |
| 1684 | |
| 1685 | SDValue LoongArchTargetLowering::PerformDAGCombine(SDNode *N, |
| 1686 | DAGCombinerInfo &DCI) const { |
| 1687 | SelectionDAG &DAG = DCI.DAG; |
| 1688 | switch (N->getOpcode()) { |
| 1689 | default: |
| 1690 | break; |
| 1691 | case ISD::AND: |
| 1692 | return performANDCombine(N, DAG, DCI, Subtarget); |
| 1693 | case ISD::OR: |
| 1694 | return performORCombine(N, DAG, DCI, Subtarget); |
| 1695 | case ISD::SRL: |
| 1696 | return performSRLCombine(N, DAG, DCI, Subtarget); |
| 1697 | case LoongArchISD::BITREV_W: |
| 1698 | return performBITREV_WCombine(N, DAG, DCI, Subtarget); |
| 1699 | } |
| 1700 | return SDValue(); |
| 1701 | } |
| 1702 | |
| 1703 | static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, |
| 1704 | MachineBasicBlock *MBB) { |
| 1705 | if (!ZeroDivCheck) |
| 1706 | return MBB; |
| 1707 | |
| 1708 | // Build instructions: |
| 1709 | // MBB: |
| 1710 | // div(or mod) $dst, $dividend, $divisor |
| 1711 | // bnez $divisor, SinkMBB |
| 1712 | // BreakMBB: |
| 1713 | // break 7 // BRK_DIVZERO |
| 1714 | // SinkMBB: |
| 1715 | // fallthrough |
| 1716 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); |
| 1717 | MachineFunction::iterator It = ++MBB->getIterator(); |
| 1718 | MachineFunction *MF = MBB->getParent(); |
| 1719 | auto BreakMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
| 1720 | auto SinkMBB = MF->CreateMachineBasicBlock(LLVM_BB); |
| 1721 | MF->insert(It, BreakMBB); |
| 1722 | MF->insert(It, SinkMBB); |
| 1723 | |
| 1724 | // Transfer the remainder of MBB and its successor edges to SinkMBB. |
| 1725 | SinkMBB->splice(SinkMBB->end(), MBB, std::next(MI.getIterator()), MBB->end()); |
| 1726 | SinkMBB->transferSuccessorsAndUpdatePHIs(MBB); |
| 1727 | |
| 1728 | const TargetInstrInfo &TII = *MF->getSubtarget().getInstrInfo(); |
| 1729 | DebugLoc DL = MI.getDebugLoc(); |
| 1730 | MachineOperand &Divisor = MI.getOperand(2); |
| 1731 | Register DivisorReg = Divisor.getReg(); |
| 1732 | |
| 1733 | // MBB: |
| 1734 | BuildMI(MBB, DL, TII.get(LoongArch::BNEZ)) |
| 1735 | .addReg(DivisorReg, getKillRegState(Divisor.isKill())) |
| 1736 | .addMBB(SinkMBB); |
| 1737 | MBB->addSuccessor(BreakMBB); |
| 1738 | MBB->addSuccessor(SinkMBB); |
| 1739 | |
| 1740 | // BreakMBB: |
| 1741 | // See linux header file arch/loongarch/include/uapi/asm/break.h for the |
| 1742 | // definition of BRK_DIVZERO. |
| 1743 | BuildMI(BreakMBB, DL, TII.get(LoongArch::BREAK)).addImm(7 /*BRK_DIVZERO*/); |
| 1744 | BreakMBB->addSuccessor(SinkMBB); |
| 1745 | |
| 1746 | // Clear Divisor's kill flag. |
| 1747 | Divisor.setIsKill(false); |
| 1748 | |
| 1749 | return SinkMBB; |
| 1750 | } |
| 1751 | |
| 1752 | MachineBasicBlock *LoongArchTargetLowering::EmitInstrWithCustomInserter( |
| 1753 | MachineInstr &MI, MachineBasicBlock *BB) const { |
| 1754 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1755 | DebugLoc DL = MI.getDebugLoc(); |
| 1756 | |
| 1757 | switch (MI.getOpcode()) { |
| 1758 | default: |
| 1759 | llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1759 ); |
| 1760 | case LoongArch::DIV_W: |
| 1761 | case LoongArch::DIV_WU: |
| 1762 | case LoongArch::MOD_W: |
| 1763 | case LoongArch::MOD_WU: |
| 1764 | case LoongArch::DIV_D: |
| 1765 | case LoongArch::DIV_DU: |
| 1766 | case LoongArch::MOD_D: |
| 1767 | case LoongArch::MOD_DU: |
| 1768 | return insertDivByZeroTrap(MI, BB); |
| 1769 | break; |
| 1770 | case LoongArch::WRFCSR: { |
| 1771 | BuildMI(*BB, MI, DL, TII->get(LoongArch::MOVGR2FCSR), |
| 1772 | LoongArch::FCSR0 + MI.getOperand(0).getImm()) |
| 1773 | .addReg(MI.getOperand(1).getReg()); |
| 1774 | MI.eraseFromParent(); |
| 1775 | return BB; |
| 1776 | } |
| 1777 | case LoongArch::RDFCSR: { |
| 1778 | MachineInstr *ReadFCSR = |
| 1779 | BuildMI(*BB, MI, DL, TII->get(LoongArch::MOVFCSR2GR), |
| 1780 | MI.getOperand(0).getReg()) |
| 1781 | .addReg(LoongArch::FCSR0 + MI.getOperand(1).getImm()); |
| 1782 | ReadFCSR->getOperand(1).setIsUndef(); |
| 1783 | MI.eraseFromParent(); |
| 1784 | return BB; |
| 1785 | } |
| 1786 | } |
| 1787 | } |
| 1788 | |
| 1789 | const char *LoongArchTargetLowering::getTargetNodeName(unsigned Opcode) const { |
| 1790 | switch ((LoongArchISD::NodeType)Opcode) { |
| 1791 | case LoongArchISD::FIRST_NUMBER: |
| 1792 | break; |
| 1793 | |
| 1794 | #define NODE_NAME_CASE(node) \ |
| 1795 | case LoongArchISD::node: \ |
| 1796 | return "LoongArchISD::" #node; |
| 1797 | |
| 1798 | // TODO: Add more target-dependent nodes later. |
| 1799 | NODE_NAME_CASE(CALL) |
| 1800 | NODE_NAME_CASE(RET) |
| 1801 | NODE_NAME_CASE(TAIL) |
| 1802 | NODE_NAME_CASE(SLL_W) |
| 1803 | NODE_NAME_CASE(SRA_W) |
| 1804 | NODE_NAME_CASE(SRL_W) |
| 1805 | NODE_NAME_CASE(BSTRINS) |
| 1806 | NODE_NAME_CASE(BSTRPICK) |
| 1807 | NODE_NAME_CASE(MOVGR2FR_W_LA64) |
| 1808 | NODE_NAME_CASE(MOVFR2GR_S_LA64) |
| 1809 | NODE_NAME_CASE(FTINT) |
| 1810 | NODE_NAME_CASE(REVB_2H) |
| 1811 | NODE_NAME_CASE(REVB_2W) |
| 1812 | NODE_NAME_CASE(BITREV_4B) |
| 1813 | NODE_NAME_CASE(BITREV_W) |
| 1814 | NODE_NAME_CASE(ROTR_W) |
| 1815 | NODE_NAME_CASE(ROTL_W) |
| 1816 | NODE_NAME_CASE(CLZ_W) |
| 1817 | NODE_NAME_CASE(CTZ_W) |
| 1818 | NODE_NAME_CASE(DBAR) |
| 1819 | NODE_NAME_CASE(IBAR) |
| 1820 | NODE_NAME_CASE(BREAK) |
| 1821 | NODE_NAME_CASE(SYSCALL) |
| 1822 | NODE_NAME_CASE(CRC_W_B_W) |
| 1823 | NODE_NAME_CASE(CRC_W_H_W) |
| 1824 | NODE_NAME_CASE(CRC_W_W_W) |
| 1825 | NODE_NAME_CASE(CRC_W_D_W) |
| 1826 | NODE_NAME_CASE(CRCC_W_B_W) |
| 1827 | NODE_NAME_CASE(CRCC_W_H_W) |
| 1828 | NODE_NAME_CASE(CRCC_W_W_W) |
| 1829 | NODE_NAME_CASE(CRCC_W_D_W) |
| 1830 | NODE_NAME_CASE(CSRRD) |
| 1831 | NODE_NAME_CASE(CSRWR) |
| 1832 | NODE_NAME_CASE(CSRXCHG) |
| 1833 | NODE_NAME_CASE(IOCSRRD_B) |
| 1834 | NODE_NAME_CASE(IOCSRRD_H) |
| 1835 | NODE_NAME_CASE(IOCSRRD_W) |
| 1836 | NODE_NAME_CASE(IOCSRRD_D) |
| 1837 | NODE_NAME_CASE(IOCSRWR_B) |
| 1838 | NODE_NAME_CASE(IOCSRWR_H) |
| 1839 | NODE_NAME_CASE(IOCSRWR_W) |
| 1840 | NODE_NAME_CASE(IOCSRWR_D) |
| 1841 | NODE_NAME_CASE(CPUCFG) |
| 1842 | NODE_NAME_CASE(MOVGR2FCSR) |
| 1843 | NODE_NAME_CASE(MOVFCSR2GR) |
| 1844 | NODE_NAME_CASE(CACOP_D) |
| 1845 | NODE_NAME_CASE(CACOP_W) |
| 1846 | } |
| 1847 | #undef NODE_NAME_CASE |
| 1848 | return nullptr; |
| 1849 | } |
| 1850 | |
| 1851 | //===----------------------------------------------------------------------===// |
| 1852 | // Calling Convention Implementation |
| 1853 | //===----------------------------------------------------------------------===// |
| 1854 | |
| 1855 | // Eight general-purpose registers a0-a7 used for passing integer arguments, |
| 1856 | // with a0-a1 reused to return values. Generally, the GPRs are used to pass |
| 1857 | // fixed-point arguments, and floating-point arguments when no FPR is available |
| 1858 | // or with soft float ABI. |
| 1859 | const MCPhysReg ArgGPRs[] = {LoongArch::R4, LoongArch::R5, LoongArch::R6, |
| 1860 | LoongArch::R7, LoongArch::R8, LoongArch::R9, |
| 1861 | LoongArch::R10, LoongArch::R11}; |
| 1862 | // Eight floating-point registers fa0-fa7 used for passing floating-point |
| 1863 | // arguments, and fa0-fa1 are also used to return values. |
| 1864 | const MCPhysReg ArgFPR32s[] = {LoongArch::F0, LoongArch::F1, LoongArch::F2, |
| 1865 | LoongArch::F3, LoongArch::F4, LoongArch::F5, |
| 1866 | LoongArch::F6, LoongArch::F7}; |
| 1867 | // FPR32 and FPR64 alias each other. |
| 1868 | const MCPhysReg ArgFPR64s[] = { |
| 1869 | LoongArch::F0_64, LoongArch::F1_64, LoongArch::F2_64, LoongArch::F3_64, |
| 1870 | LoongArch::F4_64, LoongArch::F5_64, LoongArch::F6_64, LoongArch::F7_64}; |
| 1871 | |
| 1872 | // Pass a 2*GRLen argument that has been split into two GRLen values through |
| 1873 | // registers or the stack as necessary. |
| 1874 | static bool CC_LoongArchAssign2GRLen(unsigned GRLen, CCState &State, |
| 1875 | CCValAssign VA1, ISD::ArgFlagsTy ArgFlags1, |
| 1876 | unsigned ValNo2, MVT ValVT2, MVT LocVT2, |
| 1877 | ISD::ArgFlagsTy ArgFlags2) { |
| 1878 | unsigned GRLenInBytes = GRLen / 8; |
| 1879 | if (Register Reg = State.AllocateReg(ArgGPRs)) { |
| 1880 | // At least one half can be passed via register. |
| 1881 | State.addLoc(CCValAssign::getReg(VA1.getValNo(), VA1.getValVT(), Reg, |
| 1882 | VA1.getLocVT(), CCValAssign::Full)); |
| 1883 | } else { |
| 1884 | // Both halves must be passed on the stack, with proper alignment. |
| 1885 | Align StackAlign = |
| 1886 | std::max(Align(GRLenInBytes), ArgFlags1.getNonZeroOrigAlign()); |
| 1887 | State.addLoc( |
| 1888 | CCValAssign::getMem(VA1.getValNo(), VA1.getValVT(), |
| 1889 | State.AllocateStack(GRLenInBytes, StackAlign), |
| 1890 | VA1.getLocVT(), CCValAssign::Full)); |
| 1891 | State.addLoc(CCValAssign::getMem( |
| 1892 | ValNo2, ValVT2, State.AllocateStack(GRLenInBytes, Align(GRLenInBytes)), |
| 1893 | LocVT2, CCValAssign::Full)); |
| 1894 | return false; |
| 1895 | } |
| 1896 | if (Register Reg = State.AllocateReg(ArgGPRs)) { |
| 1897 | // The second half can also be passed via register. |
| 1898 | State.addLoc( |
| 1899 | CCValAssign::getReg(ValNo2, ValVT2, Reg, LocVT2, CCValAssign::Full)); |
| 1900 | } else { |
| 1901 | // The second half is passed via the stack, without additional alignment. |
| 1902 | State.addLoc(CCValAssign::getMem( |
| 1903 | ValNo2, ValVT2, State.AllocateStack(GRLenInBytes, Align(GRLenInBytes)), |
| 1904 | LocVT2, CCValAssign::Full)); |
| 1905 | } |
| 1906 | return false; |
| 1907 | } |
| 1908 | |
| 1909 | // Implements the LoongArch calling convention. Returns true upon failure. |
| 1910 | static bool CC_LoongArch(const DataLayout &DL, LoongArchABI::ABI ABI, |
| 1911 | unsigned ValNo, MVT ValVT, |
| 1912 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
| 1913 | CCState &State, bool IsFixed, bool IsRet, |
| 1914 | Type *OrigTy) { |
| 1915 | unsigned GRLen = DL.getLargestLegalIntTypeSizeInBits(); |
| 1916 | assert((GRLen == 32 || GRLen == 64) && "Unspport GRLen")(static_cast <bool> ((GRLen == 32 || GRLen == 64) && "Unspport GRLen") ? void (0) : __assert_fail ("(GRLen == 32 || GRLen == 64) && \"Unspport GRLen\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1916 , __extension__ __PRETTY_FUNCTION__)); |
| 1917 | MVT GRLenVT = GRLen == 32 ? MVT::i32 : MVT::i64; |
| 1918 | MVT LocVT = ValVT; |
| 1919 | |
| 1920 | // Any return value split into more than two values can't be returned |
| 1921 | // directly. |
| 1922 | if (IsRet && ValNo > 1) |
| 1923 | return true; |
| 1924 | |
| 1925 | // If passing a variadic argument, or if no FPR is available. |
| 1926 | bool UseGPRForFloat = true; |
| 1927 | |
| 1928 | switch (ABI) { |
| 1929 | default: |
| 1930 | llvm_unreachable("Unexpected ABI")::llvm::llvm_unreachable_internal("Unexpected ABI", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 1930); |
| 1931 | case LoongArchABI::ABI_ILP32S: |
| 1932 | case LoongArchABI::ABI_LP64S: |
| 1933 | case LoongArchABI::ABI_ILP32F: |
| 1934 | case LoongArchABI::ABI_LP64F: |
| 1935 | report_fatal_error("Unimplemented ABI"); |
| 1936 | break; |
| 1937 | case LoongArchABI::ABI_ILP32D: |
| 1938 | case LoongArchABI::ABI_LP64D: |
| 1939 | UseGPRForFloat = !IsFixed; |
| 1940 | break; |
| 1941 | } |
| 1942 | |
| 1943 | // FPR32 and FPR64 alias each other. |
| 1944 | if (State.getFirstUnallocated(ArgFPR32s) == std::size(ArgFPR32s)) |
| 1945 | UseGPRForFloat = true; |
| 1946 | |
| 1947 | if (UseGPRForFloat && ValVT == MVT::f32) { |
| 1948 | LocVT = GRLenVT; |
| 1949 | LocInfo = CCValAssign::BCvt; |
| 1950 | } else if (UseGPRForFloat && GRLen == 64 && ValVT == MVT::f64) { |
| 1951 | LocVT = MVT::i64; |
| 1952 | LocInfo = CCValAssign::BCvt; |
| 1953 | } else if (UseGPRForFloat && GRLen == 32 && ValVT == MVT::f64) { |
| 1954 | // TODO: Handle passing f64 on LA32 with D feature. |
| 1955 | report_fatal_error("Passing f64 with GPR on LA32 is undefined"); |
| 1956 | } |
| 1957 | |
| 1958 | // If this is a variadic argument, the LoongArch calling convention requires |
| 1959 | // that it is assigned an 'even' or 'aligned' register if it has (2*GRLen)/8 |
| 1960 | // byte alignment. An aligned register should be used regardless of whether |
| 1961 | // the original argument was split during legalisation or not. The argument |
| 1962 | // will not be passed by registers if the original type is larger than |
| 1963 | // 2*GRLen, so the register alignment rule does not apply. |
| 1964 | unsigned TwoGRLenInBytes = (2 * GRLen) / 8; |
| 1965 | if (!IsFixed && ArgFlags.getNonZeroOrigAlign() == TwoGRLenInBytes && |
| 1966 | DL.getTypeAllocSize(OrigTy) == TwoGRLenInBytes) { |
| 1967 | unsigned RegIdx = State.getFirstUnallocated(ArgGPRs); |
| 1968 | // Skip 'odd' register if necessary. |
| 1969 | if (RegIdx != std::size(ArgGPRs) && RegIdx % 2 == 1) |
| 1970 | State.AllocateReg(ArgGPRs); |
| 1971 | } |
| 1972 | |
| 1973 | SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs(); |
| 1974 | SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags = |
| 1975 | State.getPendingArgFlags(); |
| 1976 | |
| 1977 | assert(PendingLocs.size() == PendingArgFlags.size() &&(static_cast <bool> (PendingLocs.size() == PendingArgFlags .size() && "PendingLocs and PendingArgFlags out of sync" ) ? void (0) : __assert_fail ("PendingLocs.size() == PendingArgFlags.size() && \"PendingLocs and PendingArgFlags out of sync\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1978 , __extension__ __PRETTY_FUNCTION__)) |
| 1978 | "PendingLocs and PendingArgFlags out of sync")(static_cast <bool> (PendingLocs.size() == PendingArgFlags .size() && "PendingLocs and PendingArgFlags out of sync" ) ? void (0) : __assert_fail ("PendingLocs.size() == PendingArgFlags.size() && \"PendingLocs and PendingArgFlags out of sync\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1978 , __extension__ __PRETTY_FUNCTION__)); |
| 1979 | |
| 1980 | // Split arguments might be passed indirectly, so keep track of the pending |
| 1981 | // values. |
| 1982 | if (ValVT.isScalarInteger() && (ArgFlags.isSplit() || !PendingLocs.empty())) { |
| 1983 | LocVT = GRLenVT; |
| 1984 | LocInfo = CCValAssign::Indirect; |
| 1985 | PendingLocs.push_back( |
| 1986 | CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo)); |
| 1987 | PendingArgFlags.push_back(ArgFlags); |
| 1988 | if (!ArgFlags.isSplitEnd()) { |
| 1989 | return false; |
| 1990 | } |
| 1991 | } |
| 1992 | |
| 1993 | // If the split argument only had two elements, it should be passed directly |
| 1994 | // in registers or on the stack. |
| 1995 | if (ValVT.isScalarInteger() && ArgFlags.isSplitEnd() && |
| 1996 | PendingLocs.size() <= 2) { |
| 1997 | assert(PendingLocs.size() == 2 && "Unexpected PendingLocs.size()")(static_cast <bool> (PendingLocs.size() == 2 && "Unexpected PendingLocs.size()") ? void (0) : __assert_fail ( "PendingLocs.size() == 2 && \"Unexpected PendingLocs.size()\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 1997 , __extension__ __PRETTY_FUNCTION__)); |
| 1998 | // Apply the normal calling convention rules to the first half of the |
| 1999 | // split argument. |
| 2000 | CCValAssign VA = PendingLocs[0]; |
| 2001 | ISD::ArgFlagsTy AF = PendingArgFlags[0]; |
| 2002 | PendingLocs.clear(); |
| 2003 | PendingArgFlags.clear(); |
| 2004 | return CC_LoongArchAssign2GRLen(GRLen, State, VA, AF, ValNo, ValVT, LocVT, |
| 2005 | ArgFlags); |
| 2006 | } |
| 2007 | |
| 2008 | // Allocate to a register if possible, or else a stack slot. |
| 2009 | Register Reg; |
| 2010 | unsigned StoreSizeBytes = GRLen / 8; |
| 2011 | Align StackAlign = Align(GRLen / 8); |
| 2012 | |
| 2013 | if (ValVT == MVT::f32 && !UseGPRForFloat) |
| 2014 | Reg = State.AllocateReg(ArgFPR32s); |
| 2015 | else if (ValVT == MVT::f64 && !UseGPRForFloat) |
| 2016 | Reg = State.AllocateReg(ArgFPR64s); |
| 2017 | else |
| 2018 | Reg = State.AllocateReg(ArgGPRs); |
| 2019 | |
| 2020 | unsigned StackOffset = |
| 2021 | Reg ? 0 : State.AllocateStack(StoreSizeBytes, StackAlign); |
| 2022 | |
| 2023 | // If we reach this point and PendingLocs is non-empty, we must be at the |
| 2024 | // end of a split argument that must be passed indirectly. |
| 2025 | if (!PendingLocs.empty()) { |
| 2026 | assert(ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()")(static_cast <bool> (ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()" ) ? void (0) : __assert_fail ("ArgFlags.isSplitEnd() && \"Expected ArgFlags.isSplitEnd()\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2026 , __extension__ __PRETTY_FUNCTION__)); |
| 2027 | assert(PendingLocs.size() > 2 && "Unexpected PendingLocs.size()")(static_cast <bool> (PendingLocs.size() > 2 && "Unexpected PendingLocs.size()") ? void (0) : __assert_fail ( "PendingLocs.size() > 2 && \"Unexpected PendingLocs.size()\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2027 , __extension__ __PRETTY_FUNCTION__)); |
| 2028 | for (auto &It : PendingLocs) { |
| 2029 | if (Reg) |
| 2030 | It.convertToReg(Reg); |
| 2031 | else |
| 2032 | It.convertToMem(StackOffset); |
| 2033 | State.addLoc(It); |
| 2034 | } |
| 2035 | PendingLocs.clear(); |
| 2036 | PendingArgFlags.clear(); |
| 2037 | return false; |
| 2038 | } |
| 2039 | assert((!UseGPRForFloat || LocVT == GRLenVT) &&(static_cast <bool> ((!UseGPRForFloat || LocVT == GRLenVT ) && "Expected an GRLenVT at this stage") ? void (0) : __assert_fail ("(!UseGPRForFloat || LocVT == GRLenVT) && \"Expected an GRLenVT at this stage\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2040 , __extension__ __PRETTY_FUNCTION__)) |
| 2040 | "Expected an GRLenVT at this stage")(static_cast <bool> ((!UseGPRForFloat || LocVT == GRLenVT ) && "Expected an GRLenVT at this stage") ? void (0) : __assert_fail ("(!UseGPRForFloat || LocVT == GRLenVT) && \"Expected an GRLenVT at this stage\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2040 , __extension__ __PRETTY_FUNCTION__)); |
| 2041 | |
| 2042 | if (Reg) { |
| 2043 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
| 2044 | return false; |
| 2045 | } |
| 2046 | |
| 2047 | // When a floating-point value is passed on the stack, no bit-cast is needed. |
| 2048 | if (ValVT.isFloatingPoint()) { |
| 2049 | LocVT = ValVT; |
| 2050 | LocInfo = CCValAssign::Full; |
| 2051 | } |
| 2052 | |
| 2053 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo)); |
| 2054 | return false; |
| 2055 | } |
| 2056 | |
| 2057 | void LoongArchTargetLowering::analyzeInputArgs( |
| 2058 | MachineFunction &MF, CCState &CCInfo, |
| 2059 | const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet, |
| 2060 | LoongArchCCAssignFn Fn) const { |
| 2061 | FunctionType *FType = MF.getFunction().getFunctionType(); |
| 2062 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { |
| 2063 | MVT ArgVT = Ins[i].VT; |
| 2064 | Type *ArgTy = nullptr; |
| 2065 | if (IsRet) |
| 2066 | ArgTy = FType->getReturnType(); |
| 2067 | else if (Ins[i].isOrigArg()) |
| 2068 | ArgTy = FType->getParamType(Ins[i].getOrigArgIndex()); |
| 2069 | LoongArchABI::ABI ABI = |
| 2070 | MF.getSubtarget<LoongArchSubtarget>().getTargetABI(); |
| 2071 | if (Fn(MF.getDataLayout(), ABI, i, ArgVT, CCValAssign::Full, Ins[i].Flags, |
| 2072 | CCInfo, /*IsFixed=*/true, IsRet, ArgTy)) { |
| 2073 | LLVM_DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "InputArg #" << i << " has unhandled type " << ArgVT << '\n' ; } } while (false) |
| 2074 | << ArgVT << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "InputArg #" << i << " has unhandled type " << ArgVT << '\n' ; } } while (false); |
| 2075 | llvm_unreachable("")::llvm::llvm_unreachable_internal("", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 2075); |
| 2076 | } |
| 2077 | } |
| 2078 | } |
| 2079 | |
| 2080 | void LoongArchTargetLowering::analyzeOutputArgs( |
| 2081 | MachineFunction &MF, CCState &CCInfo, |
| 2082 | const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet, |
| 2083 | CallLoweringInfo *CLI, LoongArchCCAssignFn Fn) const { |
| 2084 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
| 2085 | MVT ArgVT = Outs[i].VT; |
| 2086 | Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr; |
| 2087 | LoongArchABI::ABI ABI = |
| 2088 | MF.getSubtarget<LoongArchSubtarget>().getTargetABI(); |
| 2089 | if (Fn(MF.getDataLayout(), ABI, i, ArgVT, CCValAssign::Full, Outs[i].Flags, |
| 2090 | CCInfo, Outs[i].IsFixed, IsRet, OrigTy)) { |
| 2091 | LLVM_DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "OutputArg #" << i << " has unhandled type " << ArgVT << "\n" ; } } while (false) |
| 2092 | << ArgVT << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loongarch-isel-lowering")) { dbgs() << "OutputArg #" << i << " has unhandled type " << ArgVT << "\n" ; } } while (false); |
| 2093 | llvm_unreachable("")::llvm::llvm_unreachable_internal("", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 2093); |
| 2094 | } |
| 2095 | } |
| 2096 | } |
| 2097 | |
| 2098 | // Convert Val to a ValVT. Should not be called for CCValAssign::Indirect |
| 2099 | // values. |
| 2100 | static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val, |
| 2101 | const CCValAssign &VA, const SDLoc &DL) { |
| 2102 | switch (VA.getLocInfo()) { |
| 2103 | default: |
| 2104 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2104 ); |
| 2105 | case CCValAssign::Full: |
| 2106 | case CCValAssign::Indirect: |
| 2107 | break; |
| 2108 | case CCValAssign::BCvt: |
| 2109 | if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) |
| 2110 | Val = DAG.getNode(LoongArchISD::MOVGR2FR_W_LA64, DL, MVT::f32, Val); |
| 2111 | else |
| 2112 | Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); |
| 2113 | break; |
| 2114 | } |
| 2115 | return Val; |
| 2116 | } |
| 2117 | |
| 2118 | static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain, |
| 2119 | const CCValAssign &VA, const SDLoc &DL, |
| 2120 | const LoongArchTargetLowering &TLI) { |
| 2121 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2122 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
| 2123 | EVT LocVT = VA.getLocVT(); |
| 2124 | SDValue Val; |
| 2125 | const TargetRegisterClass *RC = TLI.getRegClassFor(LocVT.getSimpleVT()); |
| 2126 | Register VReg = RegInfo.createVirtualRegister(RC); |
| 2127 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
| 2128 | Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT); |
| 2129 | |
| 2130 | return convertLocVTToValVT(DAG, Val, VA, DL); |
| 2131 | } |
| 2132 | |
| 2133 | // The caller is responsible for loading the full value if the argument is |
| 2134 | // passed with CCValAssign::Indirect. |
| 2135 | static SDValue unpackFromMemLoc(SelectionDAG &DAG, SDValue Chain, |
| 2136 | const CCValAssign &VA, const SDLoc &DL) { |
| 2137 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2138 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2139 | EVT ValVT = VA.getValVT(); |
| 2140 | int FI = MFI.CreateFixedObject(ValVT.getStoreSize(), VA.getLocMemOffset(), |
| 2141 | /*IsImmutable=*/true); |
| 2142 | SDValue FIN = DAG.getFrameIndex( |
| 2143 | FI, MVT::getIntegerVT(DAG.getDataLayout().getPointerSizeInBits(0))); |
| 2144 | |
| 2145 | ISD::LoadExtType ExtType; |
| 2146 | switch (VA.getLocInfo()) { |
| 2147 | default: |
| 2148 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2148 ); |
| 2149 | case CCValAssign::Full: |
| 2150 | case CCValAssign::Indirect: |
| 2151 | case CCValAssign::BCvt: |
| 2152 | ExtType = ISD::NON_EXTLOAD; |
| 2153 | break; |
| 2154 | } |
| 2155 | return DAG.getExtLoad( |
| 2156 | ExtType, DL, VA.getLocVT(), Chain, FIN, |
| 2157 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), ValVT); |
| 2158 | } |
| 2159 | |
| 2160 | static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDValue Val, |
| 2161 | const CCValAssign &VA, const SDLoc &DL) { |
| 2162 | EVT LocVT = VA.getLocVT(); |
| 2163 | |
| 2164 | switch (VA.getLocInfo()) { |
| 2165 | default: |
| 2166 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2166 ); |
| 2167 | case CCValAssign::Full: |
| 2168 | break; |
| 2169 | case CCValAssign::BCvt: |
| 2170 | if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) |
| 2171 | Val = DAG.getNode(LoongArchISD::MOVFR2GR_S_LA64, DL, MVT::i64, Val); |
| 2172 | else |
| 2173 | Val = DAG.getNode(ISD::BITCAST, DL, LocVT, Val); |
| 2174 | break; |
| 2175 | } |
| 2176 | return Val; |
| 2177 | } |
| 2178 | |
| 2179 | static bool CC_LoongArch_GHC(unsigned ValNo, MVT ValVT, MVT LocVT, |
| 2180 | CCValAssign::LocInfo LocInfo, |
| 2181 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
| 2182 | if (LocVT == MVT::i32 || LocVT == MVT::i64) { |
| 2183 | // Pass in STG registers: Base, Sp, Hp, R1, R2, R3, R4, R5, SpLim |
| 2184 | // s0 s1 s2 s3 s4 s5 s6 s7 s8 |
| 2185 | static const MCPhysReg GPRList[] = { |
| 2186 | LoongArch::R23, LoongArch::R24, LoongArch::R25, LoongArch::R26, LoongArch::R27, |
| 2187 | LoongArch::R28, LoongArch::R29, LoongArch::R30, LoongArch::R31}; |
| 2188 | if (unsigned Reg = State.AllocateReg(GPRList)) { |
| 2189 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
| 2190 | return false; |
| 2191 | } |
| 2192 | } |
| 2193 | |
| 2194 | if (LocVT == MVT::f32) { |
| 2195 | // Pass in STG registers: F1, F2, F3, F4 |
| 2196 | // fs0,fs1,fs2,fs3 |
| 2197 | static const MCPhysReg FPR32List[] = {LoongArch::F24, LoongArch::F25, |
| 2198 | LoongArch::F26, LoongArch::F27}; |
| 2199 | if (unsigned Reg = State.AllocateReg(FPR32List)) { |
| 2200 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
| 2201 | return false; |
| 2202 | } |
| 2203 | } |
| 2204 | |
| 2205 | if (LocVT == MVT::f64) { |
| 2206 | // Pass in STG registers: D1, D2, D3, D4 |
| 2207 | // fs4,fs5,fs6,fs7 |
| 2208 | static const MCPhysReg FPR64List[] = {LoongArch::F28_64, LoongArch::F29_64, |
| 2209 | LoongArch::F30_64, LoongArch::F31_64}; |
| 2210 | if (unsigned Reg = State.AllocateReg(FPR64List)) { |
| 2211 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
| 2212 | return false; |
| 2213 | } |
| 2214 | } |
| 2215 | |
| 2216 | report_fatal_error("No registers left in GHC calling convention"); |
| 2217 | return true; |
| 2218 | } |
| 2219 | |
| 2220 | // Transform physical registers into virtual registers. |
| 2221 | SDValue LoongArchTargetLowering::LowerFormalArguments( |
| 2222 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 2223 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 2224 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 2225 | |
| 2226 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2227 | |
| 2228 | switch (CallConv) { |
| 2229 | default: |
| 2230 | llvm_unreachable("Unsupported calling convention")::llvm::llvm_unreachable_internal("Unsupported calling convention" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2230 ); |
| 2231 | case CallingConv::C: |
| 2232 | case CallingConv::Fast: |
| 2233 | break; |
| 2234 | case CallingConv::GHC: |
| 2235 | if (!MF.getSubtarget().hasFeature(LoongArch::FeatureBasicF) || |
| 2236 | !MF.getSubtarget().hasFeature(LoongArch::FeatureBasicD)) |
| 2237 | report_fatal_error( |
| 2238 | "GHC calling convention requires the F and D extensions"); |
| 2239 | } |
| 2240 | |
| 2241 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
| 2242 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 2243 | unsigned GRLenInBytes = Subtarget.getGRLen() / 8; |
| 2244 | // Used with varargs to acumulate store chains. |
| 2245 | std::vector<SDValue> OutChains; |
| 2246 | |
| 2247 | // Assign locations to all of the incoming arguments. |
| 2248 | SmallVector<CCValAssign> ArgLocs; |
| 2249 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
| 2250 | |
| 2251 | if (CallConv == CallingConv::GHC) |
| 2252 | CCInfo.AnalyzeFormalArguments(Ins, CC_LoongArch_GHC); |
| 2253 | else |
| 2254 | analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false, CC_LoongArch); |
| 2255 | |
| 2256 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
| 2257 | CCValAssign &VA = ArgLocs[i]; |
| 2258 | SDValue ArgValue; |
| 2259 | if (VA.isRegLoc()) |
| 2260 | ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, *this); |
| 2261 | else |
| 2262 | ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL); |
| 2263 | if (VA.getLocInfo() == CCValAssign::Indirect) { |
| 2264 | // If the original argument was split and passed by reference, we need to |
| 2265 | // load all parts of it here (using the same address). |
| 2266 | InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue, |
| 2267 | MachinePointerInfo())); |
| 2268 | unsigned ArgIndex = Ins[i].OrigArgIndex; |
| 2269 | unsigned ArgPartOffset = Ins[i].PartOffset; |
| 2270 | assert(ArgPartOffset == 0)(static_cast <bool> (ArgPartOffset == 0) ? void (0) : __assert_fail ("ArgPartOffset == 0", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 2270, __extension__ __PRETTY_FUNCTION__)); |
| 2271 | while (i + 1 != e && Ins[i + 1].OrigArgIndex == ArgIndex) { |
| 2272 | CCValAssign &PartVA = ArgLocs[i + 1]; |
| 2273 | unsigned PartOffset = Ins[i + 1].PartOffset - ArgPartOffset; |
| 2274 | SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL); |
| 2275 | SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue, Offset); |
| 2276 | InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address, |
| 2277 | MachinePointerInfo())); |
| 2278 | ++i; |
| 2279 | } |
| 2280 | continue; |
| 2281 | } |
| 2282 | InVals.push_back(ArgValue); |
| 2283 | } |
| 2284 | |
| 2285 | if (IsVarArg) { |
| 2286 | ArrayRef<MCPhysReg> ArgRegs = ArrayRef(ArgGPRs); |
| 2287 | unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs); |
| 2288 | const TargetRegisterClass *RC = &LoongArch::GPRRegClass; |
| 2289 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2290 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
| 2291 | auto *LoongArchFI = MF.getInfo<LoongArchMachineFunctionInfo>(); |
| 2292 | |
| 2293 | // Offset of the first variable argument from stack pointer, and size of |
| 2294 | // the vararg save area. For now, the varargs save area is either zero or |
| 2295 | // large enough to hold a0-a7. |
| 2296 | int VaArgOffset, VarArgsSaveSize; |
| 2297 | |
| 2298 | // If all registers are allocated, then all varargs must be passed on the |
| 2299 | // stack and we don't need to save any argregs. |
| 2300 | if (ArgRegs.size() == Idx) { |
| 2301 | VaArgOffset = CCInfo.getNextStackOffset(); |
| 2302 | VarArgsSaveSize = 0; |
| 2303 | } else { |
| 2304 | VarArgsSaveSize = GRLenInBytes * (ArgRegs.size() - Idx); |
| 2305 | VaArgOffset = -VarArgsSaveSize; |
| 2306 | } |
| 2307 | |
| 2308 | // Record the frame index of the first variable argument |
| 2309 | // which is a value necessary to VASTART. |
| 2310 | int FI = MFI.CreateFixedObject(GRLenInBytes, VaArgOffset, true); |
| 2311 | LoongArchFI->setVarArgsFrameIndex(FI); |
| 2312 | |
| 2313 | // If saving an odd number of registers then create an extra stack slot to |
| 2314 | // ensure that the frame pointer is 2*GRLen-aligned, which in turn ensures |
| 2315 | // offsets to even-numbered registered remain 2*GRLen-aligned. |
| 2316 | if (Idx % 2) { |
| 2317 | MFI.CreateFixedObject(GRLenInBytes, VaArgOffset - (int)GRLenInBytes, |
| 2318 | true); |
| 2319 | VarArgsSaveSize += GRLenInBytes; |
| 2320 | } |
| 2321 | |
| 2322 | // Copy the integer registers that may have been used for passing varargs |
| 2323 | // to the vararg save area. |
| 2324 | for (unsigned I = Idx; I < ArgRegs.size(); |
| 2325 | ++I, VaArgOffset += GRLenInBytes) { |
| 2326 | const Register Reg = RegInfo.createVirtualRegister(RC); |
| 2327 | RegInfo.addLiveIn(ArgRegs[I], Reg); |
| 2328 | SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, GRLenVT); |
| 2329 | FI = MFI.CreateFixedObject(GRLenInBytes, VaArgOffset, true); |
| 2330 | SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); |
| 2331 | SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, |
| 2332 | MachinePointerInfo::getFixedStack(MF, FI)); |
| 2333 | cast<StoreSDNode>(Store.getNode()) |
| 2334 | ->getMemOperand() |
| 2335 | ->setValue((Value *)nullptr); |
| 2336 | OutChains.push_back(Store); |
| 2337 | } |
| 2338 | LoongArchFI->setVarArgsSaveSize(VarArgsSaveSize); |
| 2339 | } |
| 2340 | |
| 2341 | // All stores are grouped in one node to allow the matching between |
| 2342 | // the size of Ins and InVals. This only happens for vararg functions. |
| 2343 | if (!OutChains.empty()) { |
| 2344 | OutChains.push_back(Chain); |
| 2345 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); |
| 2346 | } |
| 2347 | |
| 2348 | return Chain; |
| 2349 | } |
| 2350 | |
| 2351 | bool LoongArchTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { |
| 2352 | return CI->isTailCall(); |
| 2353 | } |
| 2354 | |
| 2355 | // Check if the return value is used as only a return value, as otherwise |
| 2356 | // we can't perform a tail-call. |
| 2357 | bool LoongArchTargetLowering::isUsedByReturnOnly(SDNode *N, |
| 2358 | SDValue &Chain) const { |
| 2359 | if (N->getNumValues() != 1) |
| 2360 | return false; |
| 2361 | if (!N->hasNUsesOfValue(1, 0)) |
| 2362 | return false; |
| 2363 | |
| 2364 | SDNode *Copy = *N->use_begin(); |
| 2365 | if (Copy->getOpcode() != ISD::CopyToReg) |
| 2366 | return false; |
| 2367 | |
| 2368 | // If the ISD::CopyToReg has a glue operand, we conservatively assume it |
| 2369 | // isn't safe to perform a tail call. |
| 2370 | if (Copy->getGluedNode()) |
| 2371 | return false; |
| 2372 | |
| 2373 | // The copy must be used by a LoongArchISD::RET, and nothing else. |
| 2374 | bool HasRet = false; |
| 2375 | for (SDNode *Node : Copy->uses()) { |
| 2376 | if (Node->getOpcode() != LoongArchISD::RET) |
| 2377 | return false; |
| 2378 | HasRet = true; |
| 2379 | } |
| 2380 | |
| 2381 | if (!HasRet) |
| 2382 | return false; |
| 2383 | |
| 2384 | Chain = Copy->getOperand(0); |
| 2385 | return true; |
| 2386 | } |
| 2387 | |
| 2388 | // Check whether the call is eligible for tail call optimization. |
| 2389 | bool LoongArchTargetLowering::isEligibleForTailCallOptimization( |
| 2390 | CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF, |
| 2391 | const SmallVectorImpl<CCValAssign> &ArgLocs) const { |
| 2392 | |
| 2393 | auto CalleeCC = CLI.CallConv; |
| 2394 | auto &Outs = CLI.Outs; |
| 2395 | auto &Caller = MF.getFunction(); |
| 2396 | auto CallerCC = Caller.getCallingConv(); |
| 2397 | |
| 2398 | // Do not tail call opt if the stack is used to pass parameters. |
| 2399 | if (CCInfo.getNextStackOffset() != 0) |
| 2400 | return false; |
| 2401 | |
| 2402 | // Do not tail call opt if any parameters need to be passed indirectly. |
| 2403 | for (auto &VA : ArgLocs) |
| 2404 | if (VA.getLocInfo() == CCValAssign::Indirect) |
| 2405 | return false; |
| 2406 | |
| 2407 | // Do not tail call opt if either caller or callee uses struct return |
| 2408 | // semantics. |
| 2409 | auto IsCallerStructRet = Caller.hasStructRetAttr(); |
| 2410 | auto IsCalleeStructRet = Outs.empty() ? false : Outs[0].Flags.isSRet(); |
| 2411 | if (IsCallerStructRet || IsCalleeStructRet) |
| 2412 | return false; |
| 2413 | |
| 2414 | // Do not tail call opt if either the callee or caller has a byval argument. |
| 2415 | for (auto &Arg : Outs) |
| 2416 | if (Arg.Flags.isByVal()) |
| 2417 | return false; |
| 2418 | |
| 2419 | // The callee has to preserve all registers the caller needs to preserve. |
| 2420 | const LoongArchRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
| 2421 | const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC); |
| 2422 | if (CalleeCC != CallerCC) { |
| 2423 | const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC); |
| 2424 | if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved)) |
| 2425 | return false; |
| 2426 | } |
| 2427 | return true; |
| 2428 | } |
| 2429 | |
| 2430 | static Align getPrefTypeAlign(EVT VT, SelectionDAG &DAG) { |
| 2431 | return DAG.getDataLayout().getPrefTypeAlign( |
| 2432 | VT.getTypeForEVT(*DAG.getContext())); |
| 2433 | } |
| 2434 | |
| 2435 | // Lower a call to a callseq_start + CALL + callseq_end chain, and add input |
| 2436 | // and output parameter nodes. |
| 2437 | SDValue |
| 2438 | LoongArchTargetLowering::LowerCall(CallLoweringInfo &CLI, |
| 2439 | SmallVectorImpl<SDValue> &InVals) const { |
| 2440 | SelectionDAG &DAG = CLI.DAG; |
| 2441 | SDLoc &DL = CLI.DL; |
| 2442 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 2443 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 2444 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
| 2445 | SDValue Chain = CLI.Chain; |
| 2446 | SDValue Callee = CLI.Callee; |
| 2447 | CallingConv::ID CallConv = CLI.CallConv; |
| 2448 | bool IsVarArg = CLI.IsVarArg; |
| 2449 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
| 2450 | MVT GRLenVT = Subtarget.getGRLenVT(); |
| 2451 | bool &IsTailCall = CLI.IsTailCall; |
| 2452 | |
| 2453 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2454 | |
| 2455 | // Analyze the operands of the call, assigning locations to each operand. |
| 2456 | SmallVector<CCValAssign> ArgLocs; |
| 2457 | CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
| 2458 | |
| 2459 | if (CallConv == CallingConv::GHC) |
| 2460 | ArgCCInfo.AnalyzeCallOperands(Outs, CC_LoongArch_GHC); |
| 2461 | else |
| 2462 | analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI, CC_LoongArch); |
| 2463 | |
| 2464 | // Check if it's really possible to do a tail call. |
| 2465 | if (IsTailCall) |
| 2466 | IsTailCall = isEligibleForTailCallOptimization(ArgCCInfo, CLI, MF, ArgLocs); |
| 2467 | |
| 2468 | if (IsTailCall) |
| 2469 | ++NumTailCalls; |
| 2470 | else if (CLI.CB && CLI.CB->isMustTailCall()) |
| 2471 | report_fatal_error("failed to perform tail call elimination on a call " |
| 2472 | "site marked musttail"); |
| 2473 | |
| 2474 | // Get a count of how many bytes are to be pushed on the stack. |
| 2475 | unsigned NumBytes = ArgCCInfo.getNextStackOffset(); |
| 2476 | |
| 2477 | // Create local copies for byval args. |
| 2478 | SmallVector<SDValue> ByValArgs; |
| 2479 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
| 2480 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
| 2481 | if (!Flags.isByVal()) |
| 2482 | continue; |
| 2483 | |
| 2484 | SDValue Arg = OutVals[i]; |
| 2485 | unsigned Size = Flags.getByValSize(); |
| 2486 | Align Alignment = Flags.getNonZeroByValAlign(); |
| 2487 | |
| 2488 | int FI = |
| 2489 | MF.getFrameInfo().CreateStackObject(Size, Alignment, /*isSS=*/false); |
| 2490 | SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); |
| 2491 | SDValue SizeNode = DAG.getConstant(Size, DL, GRLenVT); |
| 2492 | |
| 2493 | Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Alignment, |
| 2494 | /*IsVolatile=*/false, |
| 2495 | /*AlwaysInline=*/false, /*isTailCall=*/IsTailCall, |
| 2496 | MachinePointerInfo(), MachinePointerInfo()); |
| 2497 | ByValArgs.push_back(FIPtr); |
| 2498 | } |
| 2499 | |
| 2500 | if (!IsTailCall) |
| 2501 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL); |
| 2502 | |
| 2503 | // Copy argument values to their designated locations. |
| 2504 | SmallVector<std::pair<Register, SDValue>> RegsToPass; |
| 2505 | SmallVector<SDValue> MemOpChains; |
| 2506 | SDValue StackPtr; |
| 2507 | for (unsigned i = 0, j = 0, e = ArgLocs.size(); i != e; ++i) { |
| 2508 | CCValAssign &VA = ArgLocs[i]; |
| 2509 | SDValue ArgValue = OutVals[i]; |
| 2510 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
| 2511 | |
| 2512 | // Promote the value if needed. |
| 2513 | // For now, only handle fully promoted and indirect arguments. |
| 2514 | if (VA.getLocInfo() == CCValAssign::Indirect) { |
| 2515 | // Store the argument in a stack slot and pass its address. |
| 2516 | Align StackAlign = |
| 2517 | std::max(getPrefTypeAlign(Outs[i].ArgVT, DAG), |
| 2518 | getPrefTypeAlign(ArgValue.getValueType(), DAG)); |
| 2519 | TypeSize StoredSize = ArgValue.getValueType().getStoreSize(); |
| 2520 | // If the original argument was split and passed by reference, we need to |
| 2521 | // store the required parts of it here (and pass just one address). |
| 2522 | unsigned ArgIndex = Outs[i].OrigArgIndex; |
| 2523 | unsigned ArgPartOffset = Outs[i].PartOffset; |
| 2524 | assert(ArgPartOffset == 0)(static_cast <bool> (ArgPartOffset == 0) ? void (0) : __assert_fail ("ArgPartOffset == 0", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 2524, __extension__ __PRETTY_FUNCTION__)); |
| 2525 | // Calculate the total size to store. We don't have access to what we're |
| 2526 | // actually storing other than performing the loop and collecting the |
| 2527 | // info. |
| 2528 | SmallVector<std::pair<SDValue, SDValue>> Parts; |
| 2529 | while (i + 1 != e && Outs[i + 1].OrigArgIndex == ArgIndex) { |
| 2530 | SDValue PartValue = OutVals[i + 1]; |
| 2531 | unsigned PartOffset = Outs[i + 1].PartOffset - ArgPartOffset; |
| 2532 | SDValue Offset = DAG.getIntPtrConstant(PartOffset, DL); |
| 2533 | EVT PartVT = PartValue.getValueType(); |
| 2534 | |
| 2535 | StoredSize += PartVT.getStoreSize(); |
| 2536 | StackAlign = std::max(StackAlign, getPrefTypeAlign(PartVT, DAG)); |
| 2537 | Parts.push_back(std::make_pair(PartValue, Offset)); |
| 2538 | ++i; |
| 2539 | } |
| 2540 | SDValue SpillSlot = DAG.CreateStackTemporary(StoredSize, StackAlign); |
| 2541 | int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); |
| 2542 | MemOpChains.push_back( |
| 2543 | DAG.getStore(Chain, DL, ArgValue, SpillSlot, |
| 2544 | MachinePointerInfo::getFixedStack(MF, FI))); |
| 2545 | for (const auto &Part : Parts) { |
| 2546 | SDValue PartValue = Part.first; |
| 2547 | SDValue PartOffset = Part.second; |
| 2548 | SDValue Address = |
| 2549 | DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot, PartOffset); |
| 2550 | MemOpChains.push_back( |
| 2551 | DAG.getStore(Chain, DL, PartValue, Address, |
| 2552 | MachinePointerInfo::getFixedStack(MF, FI))); |
| 2553 | } |
| 2554 | ArgValue = SpillSlot; |
| 2555 | } else { |
| 2556 | ArgValue = convertValVTToLocVT(DAG, ArgValue, VA, DL); |
| 2557 | } |
| 2558 | |
| 2559 | // Use local copy if it is a byval arg. |
| 2560 | if (Flags.isByVal()) |
| 2561 | ArgValue = ByValArgs[j++]; |
| 2562 | |
| 2563 | if (VA.isRegLoc()) { |
| 2564 | // Queue up the argument copies and emit them at the end. |
| 2565 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue)); |
| 2566 | } else { |
| 2567 | assert(VA.isMemLoc() && "Argument not register or memory")(static_cast <bool> (VA.isMemLoc() && "Argument not register or memory" ) ? void (0) : __assert_fail ("VA.isMemLoc() && \"Argument not register or memory\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2567 , __extension__ __PRETTY_FUNCTION__)); |
| 2568 | assert(!IsTailCall && "Tail call not allowed if stack is used "(static_cast <bool> (!IsTailCall && "Tail call not allowed if stack is used " "for passing parameters") ? void (0) : __assert_fail ("!IsTailCall && \"Tail call not allowed if stack is used \" \"for passing parameters\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2569 , __extension__ __PRETTY_FUNCTION__)) |
| 2569 | "for passing parameters")(static_cast <bool> (!IsTailCall && "Tail call not allowed if stack is used " "for passing parameters") ? void (0) : __assert_fail ("!IsTailCall && \"Tail call not allowed if stack is used \" \"for passing parameters\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2569 , __extension__ __PRETTY_FUNCTION__)); |
| 2570 | |
| 2571 | // Work out the address of the stack slot. |
| 2572 | if (!StackPtr.getNode()) |
| 2573 | StackPtr = DAG.getCopyFromReg(Chain, DL, LoongArch::R3, PtrVT); |
| 2574 | SDValue Address = |
| 2575 | DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, |
| 2576 | DAG.getIntPtrConstant(VA.getLocMemOffset(), DL)); |
| 2577 | |
| 2578 | // Emit the store. |
| 2579 | MemOpChains.push_back( |
| 2580 | DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo())); |
| 2581 | } |
| 2582 | } |
| 2583 | |
| 2584 | // Join the stores, which are independent of one another. |
| 2585 | if (!MemOpChains.empty()) |
| 2586 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); |
| 2587 | |
| 2588 | SDValue Glue; |
| 2589 | |
| 2590 | // Build a sequence of copy-to-reg nodes, chained and glued together. |
| 2591 | for (auto &Reg : RegsToPass) { |
| 2592 | Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, Glue); |
| 2593 | Glue = Chain.getValue(1); |
| 2594 | } |
| 2595 | |
| 2596 | // If the callee is a GlobalAddress/ExternalSymbol node, turn it into a |
| 2597 | // TargetGlobalAddress/TargetExternalSymbol node so that legalize won't |
| 2598 | // split it and then direct call can be matched by PseudoCALL. |
| 2599 | if (GlobalAddressSDNode *S = dyn_cast<GlobalAddressSDNode>(Callee)) { |
| 2600 | const GlobalValue *GV = S->getGlobal(); |
| 2601 | unsigned OpFlags = |
| 2602 | getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV) |
| 2603 | ? LoongArchII::MO_CALL |
| 2604 | : LoongArchII::MO_CALL_PLT; |
| 2605 | Callee = DAG.getTargetGlobalAddress(S->getGlobal(), DL, PtrVT, 0, OpFlags); |
| 2606 | } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { |
| 2607 | unsigned OpFlags = getTargetMachine().shouldAssumeDSOLocal( |
| 2608 | *MF.getFunction().getParent(), nullptr) |
| 2609 | ? LoongArchII::MO_CALL |
| 2610 | : LoongArchII::MO_CALL_PLT; |
| 2611 | Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, OpFlags); |
| 2612 | } |
| 2613 | |
| 2614 | // The first call operand is the chain and the second is the target address. |
| 2615 | SmallVector<SDValue> Ops; |
| 2616 | Ops.push_back(Chain); |
| 2617 | Ops.push_back(Callee); |
| 2618 | |
| 2619 | // Add argument registers to the end of the list so that they are |
| 2620 | // known live into the call. |
| 2621 | for (auto &Reg : RegsToPass) |
| 2622 | Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType())); |
| 2623 | |
| 2624 | if (!IsTailCall) { |
| 2625 | // Add a register mask operand representing the call-preserved registers. |
| 2626 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
| 2627 | const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv); |
| 2628 | assert(Mask && "Missing call preserved mask for calling convention")(static_cast <bool> (Mask && "Missing call preserved mask for calling convention" ) ? void (0) : __assert_fail ("Mask && \"Missing call preserved mask for calling convention\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2628 , __extension__ __PRETTY_FUNCTION__)); |
| 2629 | Ops.push_back(DAG.getRegisterMask(Mask)); |
| 2630 | } |
| 2631 | |
| 2632 | // Glue the call to the argument copies, if any. |
| 2633 | if (Glue.getNode()) |
| 2634 | Ops.push_back(Glue); |
| 2635 | |
| 2636 | // Emit the call. |
| 2637 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
| 2638 | |
| 2639 | if (IsTailCall) { |
| 2640 | MF.getFrameInfo().setHasTailCall(); |
| 2641 | return DAG.getNode(LoongArchISD::TAIL, DL, NodeTys, Ops); |
| 2642 | } |
| 2643 | |
| 2644 | Chain = DAG.getNode(LoongArchISD::CALL, DL, NodeTys, Ops); |
| 2645 | DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge); |
| 2646 | Glue = Chain.getValue(1); |
| 2647 | |
| 2648 | // Mark the end of the call, which is glued to the call itself. |
| 2649 | Chain = DAG.getCALLSEQ_END(Chain, NumBytes, 0, Glue, DL); |
| 2650 | Glue = Chain.getValue(1); |
| 2651 | |
| 2652 | // Assign locations to each value returned by this call. |
| 2653 | SmallVector<CCValAssign> RVLocs; |
| 2654 | CCState RetCCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); |
| 2655 | analyzeInputArgs(MF, RetCCInfo, Ins, /*IsRet=*/true, CC_LoongArch); |
| 2656 | |
| 2657 | // Copy all of the result registers out of their specified physreg. |
| 2658 | for (auto &VA : RVLocs) { |
| 2659 | // Copy the value out. |
| 2660 | SDValue RetValue = |
| 2661 | DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), Glue); |
| 2662 | // Glue the RetValue to the end of the call sequence. |
| 2663 | Chain = RetValue.getValue(1); |
| 2664 | Glue = RetValue.getValue(2); |
| 2665 | |
| 2666 | RetValue = convertLocVTToValVT(DAG, RetValue, VA, DL); |
| 2667 | |
| 2668 | InVals.push_back(RetValue); |
| 2669 | } |
| 2670 | |
| 2671 | return Chain; |
| 2672 | } |
| 2673 | |
| 2674 | bool LoongArchTargetLowering::CanLowerReturn( |
| 2675 | CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg, |
| 2676 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const { |
| 2677 | SmallVector<CCValAssign> RVLocs; |
| 2678 | CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); |
| 2679 | |
| 2680 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
| 2681 | LoongArchABI::ABI ABI = |
| 2682 | MF.getSubtarget<LoongArchSubtarget>().getTargetABI(); |
| 2683 | if (CC_LoongArch(MF.getDataLayout(), ABI, i, Outs[i].VT, CCValAssign::Full, |
| 2684 | Outs[i].Flags, CCInfo, /*IsFixed=*/true, /*IsRet=*/true, |
| 2685 | nullptr)) |
| 2686 | return false; |
| 2687 | } |
| 2688 | return true; |
| 2689 | } |
| 2690 | |
| 2691 | SDValue LoongArchTargetLowering::LowerReturn( |
| 2692 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 2693 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 2694 | const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, |
| 2695 | SelectionDAG &DAG) const { |
| 2696 | // Stores the assignment of the return value to a location. |
| 2697 | SmallVector<CCValAssign> RVLocs; |
| 2698 | |
| 2699 | // Info about the registers and stack slot. |
| 2700 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
| 2701 | *DAG.getContext()); |
| 2702 | |
| 2703 | analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true, |
| 2704 | nullptr, CC_LoongArch); |
| 2705 | if (CallConv == CallingConv::GHC && !RVLocs.empty()) |
| 2706 | report_fatal_error("GHC functions return void only"); |
| 2707 | SDValue Glue; |
| 2708 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 2709 | |
| 2710 | // Copy the result values into the output registers. |
| 2711 | for (unsigned i = 0, e = RVLocs.size(); i < e; ++i) { |
| 2712 | CCValAssign &VA = RVLocs[i]; |
| 2713 | assert(VA.isRegLoc() && "Can only return in registers!")(static_cast <bool> (VA.isRegLoc() && "Can only return in registers!" ) ? void (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2713 , __extension__ __PRETTY_FUNCTION__)); |
| 2714 | |
| 2715 | // Handle a 'normal' return. |
| 2716 | SDValue Val = convertValVTToLocVT(DAG, OutVals[i], VA, DL); |
| 2717 | Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Glue); |
| 2718 | |
| 2719 | // Guarantee that all emitted copies are stuck together. |
| 2720 | Glue = Chain.getValue(1); |
| 2721 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
| 2722 | } |
| 2723 | |
| 2724 | RetOps[0] = Chain; // Update chain. |
| 2725 | |
| 2726 | // Add the glue node if we have it. |
| 2727 | if (Glue.getNode()) |
| 2728 | RetOps.push_back(Glue); |
| 2729 | |
| 2730 | return DAG.getNode(LoongArchISD::RET, DL, MVT::Other, RetOps); |
| 2731 | } |
| 2732 | |
| 2733 | bool LoongArchTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
| 2734 | bool ForCodeSize) const { |
| 2735 | // TODO: Maybe need more checks here after vector extension is supported. |
| 2736 | if (VT == MVT::f32 && !Subtarget.hasBasicF()) |
| 2737 | return false; |
| 2738 | if (VT == MVT::f64 && !Subtarget.hasBasicD()) |
| 2739 | return false; |
| 2740 | return (Imm.isZero() || Imm.isExactlyValue(+1.0)); |
| 2741 | } |
| 2742 | |
| 2743 | bool LoongArchTargetLowering::isCheapToSpeculateCttz(Type *) const { |
| 2744 | return true; |
| 2745 | } |
| 2746 | |
| 2747 | bool LoongArchTargetLowering::isCheapToSpeculateCtlz(Type *) const { |
| 2748 | return true; |
| 2749 | } |
| 2750 | |
| 2751 | bool LoongArchTargetLowering::shouldInsertFencesForAtomic( |
| 2752 | const Instruction *I) const { |
| 2753 | if (!Subtarget.is64Bit()) |
| 2754 | return isa<LoadInst>(I) || isa<StoreInst>(I); |
| 2755 | |
| 2756 | if (isa<LoadInst>(I)) |
| 2757 | return true; |
| 2758 | |
| 2759 | // On LA64, atomic store operations with IntegerBitWidth of 32 and 64 do not |
| 2760 | // require fences beacuse we can use amswap_db.[w/d]. |
| 2761 | if (isa<StoreInst>(I)) { |
| 2762 | unsigned Size = I->getOperand(0)->getType()->getIntegerBitWidth(); |
| 2763 | return (Size == 8 || Size == 16); |
| 2764 | } |
| 2765 | |
| 2766 | return false; |
| 2767 | } |
| 2768 | |
| 2769 | EVT LoongArchTargetLowering::getSetCCResultType(const DataLayout &DL, |
| 2770 | LLVMContext &Context, |
| 2771 | EVT VT) const { |
| 2772 | if (!VT.isVector()) |
| 2773 | return getPointerTy(DL); |
| 2774 | return VT.changeVectorElementTypeToInteger(); |
| 2775 | } |
| 2776 | |
| 2777 | bool LoongArchTargetLowering::hasAndNot(SDValue Y) const { |
| 2778 | // TODO: Support vectors. |
| 2779 | return Y.getValueType().isScalarInteger() && !isa<ConstantSDNode>(Y); |
| 2780 | } |
| 2781 | |
| 2782 | bool LoongArchTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
| 2783 | const CallInst &I, |
| 2784 | MachineFunction &MF, |
| 2785 | unsigned Intrinsic) const { |
| 2786 | switch (Intrinsic) { |
| 2787 | default: |
| 2788 | return false; |
| 2789 | case Intrinsic::loongarch_masked_atomicrmw_xchg_i32: |
| 2790 | case Intrinsic::loongarch_masked_atomicrmw_add_i32: |
| 2791 | case Intrinsic::loongarch_masked_atomicrmw_sub_i32: |
| 2792 | case Intrinsic::loongarch_masked_atomicrmw_nand_i32: |
| 2793 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
| 2794 | Info.memVT = MVT::i32; |
| 2795 | Info.ptrVal = I.getArgOperand(0); |
| 2796 | Info.offset = 0; |
| 2797 | Info.align = Align(4); |
| 2798 | Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore | |
| 2799 | MachineMemOperand::MOVolatile; |
| 2800 | return true; |
| 2801 | // TODO: Add more Intrinsics later. |
| 2802 | } |
| 2803 | } |
| 2804 | |
| 2805 | TargetLowering::AtomicExpansionKind |
| 2806 | LoongArchTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { |
| 2807 | // TODO: Add more AtomicRMWInst that needs to be extended. |
| 2808 | |
| 2809 | // Since floating-point operation requires a non-trivial set of data |
| 2810 | // operations, use CmpXChg to expand. |
| 2811 | if (AI->isFloatingPointOperation() || |
| 2812 | AI->getOperation() == AtomicRMWInst::UIncWrap || |
| 2813 | AI->getOperation() == AtomicRMWInst::UDecWrap) |
| 2814 | return AtomicExpansionKind::CmpXChg; |
| 2815 | |
| 2816 | unsigned Size = AI->getType()->getPrimitiveSizeInBits(); |
| 2817 | if (Size == 8 || Size == 16) |
| 2818 | return AtomicExpansionKind::MaskedIntrinsic; |
| 2819 | return AtomicExpansionKind::None; |
| 2820 | } |
| 2821 | |
| 2822 | static Intrinsic::ID |
| 2823 | getIntrinsicForMaskedAtomicRMWBinOp(unsigned GRLen, |
| 2824 | AtomicRMWInst::BinOp BinOp) { |
| 2825 | if (GRLen == 64) { |
| 2826 | switch (BinOp) { |
| 2827 | default: |
| 2828 | llvm_unreachable("Unexpected AtomicRMW BinOp")::llvm::llvm_unreachable_internal("Unexpected AtomicRMW BinOp" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2828 ); |
| 2829 | case AtomicRMWInst::Xchg: |
| 2830 | return Intrinsic::loongarch_masked_atomicrmw_xchg_i64; |
| 2831 | case AtomicRMWInst::Add: |
| 2832 | return Intrinsic::loongarch_masked_atomicrmw_add_i64; |
| 2833 | case AtomicRMWInst::Sub: |
| 2834 | return Intrinsic::loongarch_masked_atomicrmw_sub_i64; |
| 2835 | case AtomicRMWInst::Nand: |
| 2836 | return Intrinsic::loongarch_masked_atomicrmw_nand_i64; |
| 2837 | case AtomicRMWInst::UMax: |
| 2838 | return Intrinsic::loongarch_masked_atomicrmw_umax_i64; |
| 2839 | case AtomicRMWInst::UMin: |
| 2840 | return Intrinsic::loongarch_masked_atomicrmw_umin_i64; |
| 2841 | case AtomicRMWInst::Max: |
| 2842 | return Intrinsic::loongarch_masked_atomicrmw_max_i64; |
| 2843 | case AtomicRMWInst::Min: |
| 2844 | return Intrinsic::loongarch_masked_atomicrmw_min_i64; |
| 2845 | // TODO: support other AtomicRMWInst. |
| 2846 | } |
| 2847 | } |
| 2848 | |
| 2849 | if (GRLen == 32) { |
| 2850 | switch (BinOp) { |
| 2851 | default: |
| 2852 | llvm_unreachable("Unexpected AtomicRMW BinOp")::llvm::llvm_unreachable_internal("Unexpected AtomicRMW BinOp" , "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp", 2852 ); |
| 2853 | case AtomicRMWInst::Xchg: |
| 2854 | return Intrinsic::loongarch_masked_atomicrmw_xchg_i32; |
| 2855 | case AtomicRMWInst::Add: |
| 2856 | return Intrinsic::loongarch_masked_atomicrmw_add_i32; |
| 2857 | case AtomicRMWInst::Sub: |
| 2858 | return Intrinsic::loongarch_masked_atomicrmw_sub_i32; |
| 2859 | case AtomicRMWInst::Nand: |
| 2860 | return Intrinsic::loongarch_masked_atomicrmw_nand_i32; |
| 2861 | // TODO: support other AtomicRMWInst. |
| 2862 | } |
| 2863 | } |
| 2864 | |
| 2865 | llvm_unreachable("Unexpected GRLen\n")::llvm::llvm_unreachable_internal("Unexpected GRLen\n", "llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp" , 2865); |
| 2866 | } |
| 2867 | |
| 2868 | TargetLowering::AtomicExpansionKind |
| 2869 | LoongArchTargetLowering::shouldExpandAtomicCmpXchgInIR( |
| 2870 | AtomicCmpXchgInst *CI) const { |
| 2871 | unsigned Size = CI->getCompareOperand()->getType()->getPrimitiveSizeInBits(); |
| 2872 | if (Size == 8 || Size == 16) |
| 2873 | return AtomicExpansionKind::MaskedIntrinsic; |
| 2874 | return AtomicExpansionKind::None; |
| 2875 | } |
| 2876 | |
| 2877 | Value *LoongArchTargetLowering::emitMaskedAtomicCmpXchgIntrinsic( |
| 2878 | IRBuilderBase &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr, |
| 2879 | Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const { |
| 2880 | Value *Ordering = |
| 2881 | Builder.getIntN(Subtarget.getGRLen(), static_cast<uint64_t>(Ord)); |
| 2882 | |
| 2883 | // TODO: Support cmpxchg on LA32. |
| 2884 | Intrinsic::ID CmpXchgIntrID = Intrinsic::loongarch_masked_cmpxchg_i64; |
| 2885 | CmpVal = Builder.CreateSExt(CmpVal, Builder.getInt64Ty()); |
| 2886 | NewVal = Builder.CreateSExt(NewVal, Builder.getInt64Ty()); |
| 2887 | Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty()); |
| 2888 | Type *Tys[] = {AlignedAddr->getType()}; |
| 2889 | Function *MaskedCmpXchg = |
| 2890 | Intrinsic::getDeclaration(CI->getModule(), CmpXchgIntrID, Tys); |
| 2891 | Value *Result = Builder.CreateCall( |
| 2892 | MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering}); |
| 2893 | Result = Builder.CreateTrunc(Result, Builder.getInt32Ty()); |
| 2894 | return Result; |
| 2895 | } |
| 2896 | |
| 2897 | Value *LoongArchTargetLowering::emitMaskedAtomicRMWIntrinsic( |
| 2898 | IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr, |
| 2899 | Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const { |
| 2900 | unsigned GRLen = Subtarget.getGRLen(); |
| 2901 | Value *Ordering = |
| 2902 | Builder.getIntN(GRLen, static_cast<uint64_t>(AI->getOrdering())); |
| 2903 | Type *Tys[] = {AlignedAddr->getType()}; |
| 2904 | Function *LlwOpScwLoop = Intrinsic::getDeclaration( |
| 2905 | AI->getModule(), |
| 2906 | getIntrinsicForMaskedAtomicRMWBinOp(GRLen, AI->getOperation()), Tys); |
| 2907 | |
| 2908 | if (GRLen == 64) { |
| 2909 | Incr = Builder.CreateSExt(Incr, Builder.getInt64Ty()); |
| 2910 | Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty()); |
| 2911 | ShiftAmt = Builder.CreateSExt(ShiftAmt, Builder.getInt64Ty()); |
| 2912 | } |
| 2913 | |
| 2914 | Value *Result; |
| 2915 | |
| 2916 | // Must pass the shift amount needed to sign extend the loaded value prior |
| 2917 | // to performing a signed comparison for min/max. ShiftAmt is the number of |
| 2918 | // bits to shift the value into position. Pass GRLen-ShiftAmt-ValWidth, which |
| 2919 | // is the number of bits to left+right shift the value in order to |
| 2920 | // sign-extend. |
| 2921 | if (AI->getOperation() == AtomicRMWInst::Min || |
| 2922 | AI->getOperation() == AtomicRMWInst::Max) { |
| 2923 | const DataLayout &DL = AI->getModule()->getDataLayout(); |
| 2924 | unsigned ValWidth = |
| 2925 | DL.getTypeStoreSizeInBits(AI->getValOperand()->getType()); |
| 2926 | Value *SextShamt = |
| 2927 | Builder.CreateSub(Builder.getIntN(GRLen, GRLen - ValWidth), ShiftAmt); |
| 2928 | Result = Builder.CreateCall(LlwOpScwLoop, |
| 2929 | {AlignedAddr, Incr, Mask, SextShamt, Ordering}); |
| 2930 | } else { |
| 2931 | Result = |
| 2932 | Builder.CreateCall(LlwOpScwLoop, {AlignedAddr, Incr, Mask, Ordering}); |
| 2933 | } |
| 2934 | |
| 2935 | if (GRLen == 64) |
| 2936 | Result = Builder.CreateTrunc(Result, Builder.getInt32Ty()); |
| 2937 | return Result; |
| 2938 | } |
| 2939 | |
| 2940 | bool LoongArchTargetLowering::isFMAFasterThanFMulAndFAdd( |
| 2941 | const MachineFunction &MF, EVT VT) const { |
| 2942 | VT = VT.getScalarType(); |
| 2943 | |
| 2944 | if (!VT.isSimple()) |
| 2945 | return false; |
| 2946 | |
| 2947 | switch (VT.getSimpleVT().SimpleTy) { |
| 2948 | case MVT::f32: |
| 2949 | case MVT::f64: |
| 2950 | return true; |
| 2951 | default: |
| 2952 | break; |
| 2953 | } |
| 2954 | |
| 2955 | return false; |
| 2956 | } |
| 2957 | |
| 2958 | Register LoongArchTargetLowering::getExceptionPointerRegister( |
| 2959 | const Constant *PersonalityFn) const { |
| 2960 | return LoongArch::R4; |
| 2961 | } |
| 2962 | |
| 2963 | Register LoongArchTargetLowering::getExceptionSelectorRegister( |
| 2964 | const Constant *PersonalityFn) const { |
| 2965 | return LoongArch::R5; |
| 2966 | } |
| 2967 | |
| 2968 | //===----------------------------------------------------------------------===// |
| 2969 | // LoongArch Inline Assembly Support |
| 2970 | //===----------------------------------------------------------------------===// |
| 2971 | |
| 2972 | LoongArchTargetLowering::ConstraintType |
| 2973 | LoongArchTargetLowering::getConstraintType(StringRef Constraint) const { |
| 2974 | // LoongArch specific constraints in GCC: config/loongarch/constraints.md |
| 2975 | // |
| 2976 | // 'f': A floating-point register (if available). |
| 2977 | // 'k': A memory operand whose address is formed by a base register and |
| 2978 | // (optionally scaled) index register. |
| 2979 | // 'l': A signed 16-bit constant. |
| 2980 | // 'm': A memory operand whose address is formed by a base register and |
| 2981 | // offset that is suitable for use in instructions with the same |
| 2982 | // addressing mode as st.w and ld.w. |
| 2983 | // 'I': A signed 12-bit constant (for arithmetic instructions). |
| 2984 | // 'J': Integer zero. |
| 2985 | // 'K': An unsigned 12-bit constant (for logic instructions). |
| 2986 | // "ZB": An address that is held in a general-purpose register. The offset is |
| 2987 | // zero. |
| 2988 | // "ZC": A memory operand whose address is formed by a base register and |
| 2989 | // offset that is suitable for use in instructions with the same |
| 2990 | // addressing mode as ll.w and sc.w. |
| 2991 | if (Constraint.size() == 1) { |
| 2992 | switch (Constraint[0]) { |
| 2993 | default: |
| 2994 | break; |
| 2995 | case 'f': |
| 2996 | return C_RegisterClass; |
| 2997 | case 'l': |
| 2998 | case 'I': |
| 2999 | case 'J': |
| 3000 | case 'K': |
| 3001 | return C_Immediate; |
| 3002 | case 'k': |
| 3003 | return C_Memory; |
| 3004 | } |
| 3005 | } |
| 3006 | |
| 3007 | if (Constraint == "ZC" || Constraint == "ZB") |
| 3008 | return C_Memory; |
| 3009 | |
| 3010 | // 'm' is handled here. |
| 3011 | return TargetLowering::getConstraintType(Constraint); |
| 3012 | } |
| 3013 | |
| 3014 | unsigned LoongArchTargetLowering::getInlineAsmMemConstraint( |
| 3015 | StringRef ConstraintCode) const { |
| 3016 | return StringSwitch<unsigned>(ConstraintCode) |
| 3017 | .Case("k", InlineAsm::Constraint_k) |
| 3018 | .Case("ZB", InlineAsm::Constraint_ZB) |
| 3019 | .Case("ZC", InlineAsm::Constraint_ZC) |
| 3020 | .Default(TargetLowering::getInlineAsmMemConstraint(ConstraintCode)); |
| 3021 | } |
| 3022 | |
| 3023 | std::pair<unsigned, const TargetRegisterClass *> |
| 3024 | LoongArchTargetLowering::getRegForInlineAsmConstraint( |
| 3025 | const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { |
| 3026 | // First, see if this is a constraint that directly corresponds to a LoongArch |
| 3027 | // register class. |
| 3028 | if (Constraint.size() == 1) { |
| 3029 | switch (Constraint[0]) { |
| 3030 | case 'r': |
| 3031 | // TODO: Support fixed vectors up to GRLen? |
| 3032 | if (VT.isVector()) |
| 3033 | break; |
| 3034 | return std::make_pair(0U, &LoongArch::GPRRegClass); |
| 3035 | case 'f': |
| 3036 | if (Subtarget.hasBasicF() && VT == MVT::f32) |
| 3037 | return std::make_pair(0U, &LoongArch::FPR32RegClass); |
| 3038 | if (Subtarget.hasBasicD() && VT == MVT::f64) |
| 3039 | return std::make_pair(0U, &LoongArch::FPR64RegClass); |
| 3040 | break; |
| 3041 | default: |
| 3042 | break; |
| 3043 | } |
| 3044 | } |
| 3045 | |
| 3046 | // TargetLowering::getRegForInlineAsmConstraint uses the name of the TableGen |
| 3047 | // record (e.g. the "R0" in `def R0`) to choose registers for InlineAsm |
| 3048 | // constraints while the official register name is prefixed with a '$'. So we |
| 3049 | // clip the '$' from the original constraint string (e.g. {$r0} to {r0}.) |
| 3050 | // before it being parsed. And TargetLowering::getRegForInlineAsmConstraint is |
| 3051 | // case insensitive, so no need to convert the constraint to upper case here. |
| 3052 | // |
| 3053 | // For now, no need to support ABI names (e.g. `$a0`) as clang will correctly |
| 3054 | // decode the usage of register name aliases into their official names. And |
| 3055 | // AFAIK, the not yet upstreamed `rustc` for LoongArch will always use |
| 3056 | // official register names. |
| 3057 | if (Constraint.startswith("{$r") || Constraint.startswith("{$f")) { |
| 3058 | bool IsFP = Constraint[2] == 'f'; |
| 3059 | std::pair<StringRef, StringRef> Temp = Constraint.split('$'); |
| 3060 | std::pair<unsigned, const TargetRegisterClass *> R; |
| 3061 | R = TargetLowering::getRegForInlineAsmConstraint( |
| 3062 | TRI, join_items("", Temp.first, Temp.second), VT); |
| 3063 | // Match those names to the widest floating point register type available. |
| 3064 | if (IsFP) { |
| 3065 | unsigned RegNo = R.first; |
| 3066 | if (LoongArch::F0 <= RegNo && RegNo <= LoongArch::F31) { |
| 3067 | if (Subtarget.hasBasicD() && (VT == MVT::f64 || VT == MVT::Other)) { |
| 3068 | unsigned DReg = RegNo - LoongArch::F0 + LoongArch::F0_64; |
| 3069 | return std::make_pair(DReg, &LoongArch::FPR64RegClass); |
| 3070 | } |
| 3071 | } |
| 3072 | } |
| 3073 | return R; |
| 3074 | } |
| 3075 | |
| 3076 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
| 3077 | } |
| 3078 | |
| 3079 | void LoongArchTargetLowering::LowerAsmOperandForConstraint( |
| 3080 | SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops, |
| 3081 | SelectionDAG &DAG) const { |
| 3082 | // Currently only support length 1 constraints. |
| 3083 | if (Constraint.length() == 1) { |
| 3084 | switch (Constraint[0]) { |
| 3085 | case 'l': |
| 3086 | // Validate & create a 16-bit signed immediate operand. |
| 3087 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 3088 | uint64_t CVal = C->getSExtValue(); |
| 3089 | if (isInt<16>(CVal)) |
| 3090 | Ops.push_back( |
| 3091 | DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getGRLenVT())); |
| 3092 | } |
| 3093 | return; |
| 3094 | case 'I': |
| 3095 | // Validate & create a 12-bit signed immediate operand. |
| 3096 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 3097 | uint64_t CVal = C->getSExtValue(); |
| 3098 | if (isInt<12>(CVal)) |
| 3099 | Ops.push_back( |
| 3100 | DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getGRLenVT())); |
| 3101 | } |
| 3102 | return; |
| 3103 | case 'J': |
| 3104 | // Validate & create an integer zero operand. |
| 3105 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) |
| 3106 | if (C->getZExtValue() == 0) |
| 3107 | Ops.push_back( |
| 3108 | DAG.getTargetConstant(0, SDLoc(Op), Subtarget.getGRLenVT())); |
| 3109 | return; |
| 3110 | case 'K': |
| 3111 | // Validate & create a 12-bit unsigned immediate operand. |
| 3112 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { |
| 3113 | uint64_t CVal = C->getZExtValue(); |
| 3114 | if (isUInt<12>(CVal)) |
| 3115 | Ops.push_back( |
| 3116 | DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getGRLenVT())); |
| 3117 | } |
| 3118 | return; |
| 3119 | default: |
| 3120 | break; |
| 3121 | } |
| 3122 | } |
| 3123 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
| 3124 | } |
| 3125 | |
| 3126 | #define GET_REGISTER_MATCHER |
| 3127 | #include "LoongArchGenAsmMatcher.inc" |
| 3128 | |
| 3129 | Register |
| 3130 | LoongArchTargetLowering::getRegisterByName(const char *RegName, LLT VT, |
| 3131 | const MachineFunction &MF) const { |
| 3132 | std::pair<StringRef, StringRef> Name = StringRef(RegName).split('$'); |
| 3133 | std::string NewRegName = Name.second.str(); |
| 3134 | Register Reg = MatchRegisterAltName(NewRegName); |
| 3135 | if (Reg == LoongArch::NoRegister) |
| 3136 | Reg = MatchRegisterName(NewRegName); |
| 3137 | if (Reg == LoongArch::NoRegister) |
| 3138 | report_fatal_error( |
| 3139 | Twine("Invalid register name \"" + StringRef(RegName) + "\".")); |
| 3140 | BitVector ReservedRegs = Subtarget.getRegisterInfo()->getReservedRegs(MF); |
| 3141 | if (!ReservedRegs.test(Reg)) |
| 3142 | report_fatal_error(Twine("Trying to obtain non-reserved register \"" + |
| 3143 | StringRef(RegName) + "\".")); |
| 3144 | return Reg; |
| 3145 | } |
| 3146 | |
| 3147 | bool LoongArchTargetLowering::decomposeMulByConstant(LLVMContext &Context, |
| 3148 | EVT VT, SDValue C) const { |
| 3149 | // TODO: Support vectors. |
| 3150 | if (!VT.isScalarInteger()) |
| 3151 | return false; |
| 3152 | |
| 3153 | // Omit the optimization if the data size exceeds GRLen. |
| 3154 | if (VT.getSizeInBits() > Subtarget.getGRLen()) |
| 3155 | return false; |
| 3156 | |
| 3157 | if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) { |
| 3158 | const APInt &Imm = ConstNode->getAPIntValue(); |
| 3159 | // Break MUL into (SLLI + ADD/SUB) or ALSL. |
| 3160 | if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2() || |
| 3161 | (1 - Imm).isPowerOf2() || (-1 - Imm).isPowerOf2()) |
| 3162 | return true; |
| 3163 | // Break MUL into (ALSL x, (SLLI x, imm0), imm1). |
| 3164 | if (ConstNode->hasOneUse() && |
| 3165 | ((Imm - 2).isPowerOf2() || (Imm - 4).isPowerOf2() || |
| 3166 | (Imm - 8).isPowerOf2() || (Imm - 16).isPowerOf2())) |
| 3167 | return true; |
| 3168 | // Break (MUL x, imm) into (ADD (SLLI x, s0), (SLLI x, s1)), |
| 3169 | // in which the immediate has two set bits. Or Break (MUL x, imm) |
| 3170 | // into (SUB (SLLI x, s0), (SLLI x, s1)), in which the immediate |
| 3171 | // equals to (1 << s0) - (1 << s1). |
| 3172 | if (ConstNode->hasOneUse() && !(Imm.sge(-2048) && Imm.sle(4095))) { |
| 3173 | unsigned Shifts = Imm.countr_zero(); |
| 3174 | // Reject immediates which can be composed via a single LUI. |
| 3175 | if (Shifts >= 12) |
| 3176 | return false; |
| 3177 | // Reject multiplications can be optimized to |
| 3178 | // (SLLI (ALSL x, x, 1/2/3/4), s). |
| 3179 | APInt ImmPop = Imm.ashr(Shifts); |
| 3180 | if (ImmPop == 3 || ImmPop == 5 || ImmPop == 9 || ImmPop == 17) |
| 3181 | return false; |
| 3182 | // We do not consider the case `(-Imm - ImmSmall).isPowerOf2()`, |
| 3183 | // since it needs one more instruction than other 3 cases. |
| 3184 | APInt ImmSmall = APInt(Imm.getBitWidth(), 1ULL << Shifts, true); |
| 3185 | if ((Imm - ImmSmall).isPowerOf2() || (Imm + ImmSmall).isPowerOf2() || |
| 3186 | (ImmSmall - Imm).isPowerOf2()) |
| 3187 | return true; |
| 3188 | } |
| 3189 | } |
| 3190 | |
| 3191 | return false; |
| 3192 | } |
| 3193 | |
| 3194 | bool LoongArchTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
| 3195 | const AddrMode &AM, |
| 3196 | Type *Ty, unsigned AS, |
| 3197 | Instruction *I) const { |
| 3198 | // LoongArch has four basic addressing modes: |
| 3199 | // 1. reg |
| 3200 | // 2. reg + 12-bit signed offset |
| 3201 | // 3. reg + 14-bit signed offset left-shifted by 2 |
| 3202 | // 4. reg1 + reg2 |
| 3203 | // TODO: Add more checks after support vector extension. |
| 3204 | |
| 3205 | // No global is ever allowed as a base. |
| 3206 | if (AM.BaseGV) |
| 3207 | return false; |
| 3208 | |
| 3209 | // Require a 12 or 14 bit signed offset. |
| 3210 | if (!isInt<12>(AM.BaseOffs) || !isShiftedInt<14, 2>(AM.BaseOffs)) |
| 3211 | return false; |
| 3212 | |
| 3213 | switch (AM.Scale) { |
| 3214 | case 0: |
| 3215 | // "i" is not allowed. |
| 3216 | if (!AM.HasBaseReg) |
| 3217 | return false; |
| 3218 | // Otherwise we have "r+i". |
| 3219 | break; |
| 3220 | case 1: |
| 3221 | // "r+r+i" is not allowed. |
| 3222 | if (AM.HasBaseReg && AM.BaseOffs != 0) |
| 3223 | return false; |
| 3224 | // Otherwise we have "r+r" or "r+i". |
| 3225 | break; |
| 3226 | case 2: |
| 3227 | // "2*r+r" or "2*r+i" is not allowed. |
| 3228 | if (AM.HasBaseReg || AM.BaseOffs) |
| 3229 | return false; |
| 3230 | // Otherwise we have "r+r". |
| 3231 | break; |
| 3232 | default: |
| 3233 | return false; |
| 3234 | } |
| 3235 | |
| 3236 | return true; |
| 3237 | } |
| 3238 | |
| 3239 | bool LoongArchTargetLowering::hasAndNotCompare(SDValue Y) const { |
| 3240 | // TODO: Support vectors. |
| 3241 | if (Y.getValueType().isVector()) |
| 3242 | return false; |
| 3243 | |
| 3244 | return !isa<ConstantSDNode>(Y); |
| 3245 | } |