| 1 | //===- MipsISelLowering.cpp - Mips 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 Mips uses to lower LLVM code into a |
| 10 | // selection DAG. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "MipsISelLowering.h" |
| 15 | #include "MCTargetDesc/MipsBaseInfo.h" |
| 16 | #include "MCTargetDesc/MipsInstPrinter.h" |
| 17 | #include "MCTargetDesc/MipsMCTargetDesc.h" |
| 18 | #include "MipsCCState.h" |
| 19 | #include "MipsInstrInfo.h" |
| 20 | #include "MipsMachineFunction.h" |
| 21 | #include "MipsRegisterInfo.h" |
| 22 | #include "MipsSubtarget.h" |
| 23 | #include "MipsTargetMachine.h" |
| 24 | #include "MipsTargetObjectFile.h" |
| 25 | #include "llvm/ADT/APFloat.h" |
| 26 | #include "llvm/ADT/ArrayRef.h" |
| 27 | #include "llvm/ADT/SmallVector.h" |
| 28 | #include "llvm/ADT/Statistic.h" |
| 29 | #include "llvm/ADT/StringRef.h" |
| 30 | #include "llvm/ADT/StringSwitch.h" |
| 31 | #include "llvm/CodeGen/CallingConvLower.h" |
| 32 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
| 33 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 34 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 35 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 36 | #include "llvm/CodeGen/MachineFunction.h" |
| 37 | #include "llvm/CodeGen/MachineInstr.h" |
| 38 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 39 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
| 40 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 41 | #include "llvm/CodeGen/MachineOperand.h" |
| 42 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 43 | #include "llvm/CodeGen/SelectionDAG.h" |
| 44 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 45 | #include "llvm/CodeGen/TargetFrameLowering.h" |
| 46 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 47 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 48 | #include "llvm/CodeGen/ValueTypes.h" |
| 49 | #include "llvm/CodeGenTypes/MachineValueType.h" |
| 50 | #include "llvm/IR/CallingConv.h" |
| 51 | #include "llvm/IR/Constants.h" |
| 52 | #include "llvm/IR/DataLayout.h" |
| 53 | #include "llvm/IR/DebugLoc.h" |
| 54 | #include "llvm/IR/DerivedTypes.h" |
| 55 | #include "llvm/IR/Function.h" |
| 56 | #include "llvm/IR/GlobalValue.h" |
| 57 | #include "llvm/IR/Module.h" |
| 58 | #include "llvm/IR/Type.h" |
| 59 | #include "llvm/IR/Value.h" |
| 60 | #include "llvm/MC/MCContext.h" |
| 61 | #include "llvm/Support/Casting.h" |
| 62 | #include "llvm/Support/CodeGen.h" |
| 63 | #include "llvm/Support/CommandLine.h" |
| 64 | #include "llvm/Support/Compiler.h" |
| 65 | #include "llvm/Support/ErrorHandling.h" |
| 66 | #include "llvm/Support/MathExtras.h" |
| 67 | #include "llvm/Target/TargetMachine.h" |
| 68 | #include "llvm/Target/TargetOptions.h" |
| 69 | #include <algorithm> |
| 70 | #include <cassert> |
| 71 | #include <cctype> |
| 72 | #include <cstdint> |
| 73 | #include <deque> |
| 74 | #include <iterator> |
| 75 | #include <regex> |
| 76 | #include <string> |
| 77 | #include <utility> |
| 78 | #include <vector> |
| 79 | |
| 80 | using namespace llvm; |
| 81 | |
| 82 | #define DEBUG_TYPE "mips-lower" |
| 83 | |
| 84 | STATISTIC(NumTailCalls, "Number of tail calls" ); |
| 85 | |
| 86 | extern cl::opt<bool> EmitJalrReloc; |
| 87 | extern cl::opt<bool> NoZeroDivCheck; |
| 88 | |
| 89 | static cl::opt<bool> UseMipsTailCalls("mips-tail-calls" , cl::Hidden, |
| 90 | cl::desc("MIPS: permit tail calls." ), |
| 91 | cl::init(Val: false)); |
| 92 | |
| 93 | static const MCPhysReg Mips64DPRegs[8] = { |
| 94 | Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, |
| 95 | Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 |
| 96 | }; |
| 97 | |
| 98 | // The MIPS MSA ABI passes vector arguments in the integer register set. |
| 99 | // The number of integer registers used is dependant on the ABI used. |
| 100 | MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, |
| 101 | CallingConv::ID CC, |
| 102 | EVT VT) const { |
| 103 | if (!VT.isVector()) |
| 104 | return getRegisterType(Context, VT); |
| 105 | |
| 106 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) |
| 107 | return Subtarget.isABI_O32() || VT.getSizeInBits() == 32 ? MVT::i32 |
| 108 | : MVT::i64; |
| 109 | return getRegisterType(Context, VT: VT.getVectorElementType()); |
| 110 | } |
| 111 | |
| 112 | unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, |
| 113 | CallingConv::ID CC, |
| 114 | EVT VT) const { |
| 115 | if (VT.isVector()) { |
| 116 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) |
| 117 | return divideCeil(Numerator: VT.getSizeInBits(), Denominator: Subtarget.isABI_O32() ? 32 : 64); |
| 118 | return VT.getVectorNumElements() * |
| 119 | getNumRegisters(Context, VT: VT.getVectorElementType()); |
| 120 | } |
| 121 | return MipsTargetLowering::getNumRegisters(Context, VT); |
| 122 | } |
| 123 | |
| 124 | unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv( |
| 125 | LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT, |
| 126 | unsigned &NumIntermediates, MVT &RegisterVT) const { |
| 127 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) { |
| 128 | IntermediateVT = getRegisterTypeForCallingConv(Context, CC, VT); |
| 129 | RegisterVT = IntermediateVT.getSimpleVT(); |
| 130 | NumIntermediates = getNumRegistersForCallingConv(Context, CC, VT); |
| 131 | return NumIntermediates; |
| 132 | } |
| 133 | IntermediateVT = VT.getVectorElementType(); |
| 134 | NumIntermediates = VT.getVectorNumElements(); |
| 135 | RegisterVT = getRegisterType(Context, VT: IntermediateVT); |
| 136 | return NumIntermediates * getNumRegisters(Context, VT: IntermediateVT); |
| 137 | } |
| 138 | |
| 139 | SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { |
| 140 | MachineFunction &MF = DAG.getMachineFunction(); |
| 141 | MipsFunctionInfo *FI = MF.getInfo<MipsFunctionInfo>(); |
| 142 | return DAG.getRegister(Reg: FI->getGlobalBaseReg(MF), VT: Ty); |
| 143 | } |
| 144 | |
| 145 | SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, |
| 146 | SelectionDAG &DAG, |
| 147 | unsigned Flag) const { |
| 148 | return DAG.getTargetGlobalAddress(GV: N->getGlobal(), DL: SDLoc(N), VT: Ty, offset: 0, TargetFlags: Flag); |
| 149 | } |
| 150 | |
| 151 | SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, |
| 152 | SelectionDAG &DAG, |
| 153 | unsigned Flag) const { |
| 154 | return DAG.getTargetExternalSymbol(Sym: N->getSymbol(), VT: Ty, TargetFlags: Flag); |
| 155 | } |
| 156 | |
| 157 | SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, |
| 158 | SelectionDAG &DAG, |
| 159 | unsigned Flag) const { |
| 160 | return DAG.getTargetBlockAddress(BA: N->getBlockAddress(), VT: Ty, Offset: 0, TargetFlags: Flag); |
| 161 | } |
| 162 | |
| 163 | SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, |
| 164 | SelectionDAG &DAG, |
| 165 | unsigned Flag) const { |
| 166 | return DAG.getTargetJumpTable(JTI: N->getIndex(), VT: Ty, TargetFlags: Flag); |
| 167 | } |
| 168 | |
| 169 | SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, |
| 170 | SelectionDAG &DAG, |
| 171 | unsigned Flag) const { |
| 172 | return DAG.getTargetConstantPool(C: N->getConstVal(), VT: Ty, Align: N->getAlign(), |
| 173 | Offset: N->getOffset(), TargetFlags: Flag); |
| 174 | } |
| 175 | |
| 176 | MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, |
| 177 | const MipsSubtarget &STI) |
| 178 | : TargetLowering(TM, STI), Subtarget(STI), ABI(TM.getABI()) { |
| 179 | // Mips does not have i1 type, so use i32 for |
| 180 | // setcc operations results (slt, sgt, ...). |
| 181 | setBooleanContents(ZeroOrOneBooleanContent); |
| 182 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
| 183 | // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA |
| 184 | // does. Integer booleans still use 0 and 1. |
| 185 | if (Subtarget.hasMips32r6()) |
| 186 | setBooleanContents(IntTy: ZeroOrOneBooleanContent, |
| 187 | FloatTy: ZeroOrNegativeOneBooleanContent); |
| 188 | |
| 189 | // Load extented operations for i1 types must be promoted |
| 190 | for (MVT VT : MVT::integer_valuetypes()) { |
| 191 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
| 192 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
| 193 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
| 194 | } |
| 195 | |
| 196 | // MIPS doesn't have extending float->double load/store. Set LoadExtAction |
| 197 | // for f32, f16 |
| 198 | for (MVT VT : MVT::fp_valuetypes()) { |
| 199 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f32, Action: Expand); |
| 200 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f16, Action: Expand); |
| 201 | } |
| 202 | |
| 203 | // Set LoadExtAction for f16 vectors to Expand |
| 204 | for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) { |
| 205 | MVT F16VT = MVT::getVectorVT(VT: MVT::f16, NumElements: VT.getVectorNumElements()); |
| 206 | if (F16VT.isValid()) |
| 207 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: F16VT, Action: Expand); |
| 208 | } |
| 209 | |
| 210 | setTruncStoreAction(ValVT: MVT::f32, MemVT: MVT::f16, Action: Expand); |
| 211 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f16, Action: Expand); |
| 212 | |
| 213 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 214 | |
| 215 | // Used by legalize types to correctly generate the setcc result. |
| 216 | // Without this, every float setcc comes with a AND/OR with the result, |
| 217 | // we don't want this, since the fpcmp result goes to a flag register, |
| 218 | // which is used implicitly by brcond and select operations. |
| 219 | AddPromotedToType(Opc: ISD::SETCC, OrigVT: MVT::i1, DestVT: MVT::i32); |
| 220 | |
| 221 | // Mips Custom Operations |
| 222 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
| 223 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i32, Action: Custom); |
| 224 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i32, Action: Custom); |
| 225 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVT::i32, Action: Custom); |
| 226 | setOperationAction(Op: ISD::JumpTable, VT: MVT::i32, Action: Custom); |
| 227 | setOperationAction(Op: ISD::ConstantPool, VT: MVT::i32, Action: Custom); |
| 228 | setOperationAction(Op: ISD::SELECT, VT: MVT::f32, Action: Custom); |
| 229 | setOperationAction(Op: ISD::SELECT, VT: MVT::f64, Action: Custom); |
| 230 | setOperationAction(Op: ISD::SELECT, VT: MVT::i32, Action: Custom); |
| 231 | setOperationAction(Op: ISD::SETCC, VT: MVT::f32, Action: Custom); |
| 232 | setOperationAction(Op: ISD::SETCC, VT: MVT::f64, Action: Custom); |
| 233 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Custom); |
| 234 | setOperationAction(Op: ISD::FABS, VT: MVT::f32, Action: Custom); |
| 235 | setOperationAction(Op: ISD::FABS, VT: MVT::f64, Action: Custom); |
| 236 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f32, Action: Custom); |
| 237 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f64, Action: Custom); |
| 238 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i32, Action: Custom); |
| 239 | setOperationAction(Op: ISD::STRICT_FP_TO_SINT, VT: MVT::i32, Action: Custom); |
| 240 | setOperationAction(Op: ISD::STRICT_FP_TO_UINT, VT: MVT::i32, Action: Custom); |
| 241 | |
| 242 | setOperationAction(Op: ISD::STRICT_FSETCC, VT: MVT::f32, Action: Custom); |
| 243 | setOperationAction(Op: ISD::STRICT_FSETCCS, VT: MVT::f32, Action: Custom); |
| 244 | setOperationAction(Op: ISD::STRICT_FSETCC, VT: MVT::f64, Action: Custom); |
| 245 | setOperationAction(Op: ISD::STRICT_FSETCCS, VT: MVT::f64, Action: Custom); |
| 246 | |
| 247 | if (Subtarget.hasMips32r2() || |
| 248 | getTargetMachine().getTargetTriple().isOSLinux()) |
| 249 | setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Custom); |
| 250 | |
| 251 | // Lower fmin/fmax/fclass operations for MIPS R6. |
| 252 | if (Subtarget.hasMips32r6()) { |
| 253 | setOperationAction(Op: ISD::FMINNUM_IEEE, VT: MVT::f32, Action: Legal); |
| 254 | setOperationAction(Op: ISD::FMAXNUM_IEEE, VT: MVT::f32, Action: Legal); |
| 255 | setOperationAction(Op: ISD::FMINNUM, VT: MVT::f32, Action: Legal); |
| 256 | setOperationAction(Op: ISD::FMAXNUM, VT: MVT::f32, Action: Legal); |
| 257 | setOperationAction(Op: ISD::FMINNUM_IEEE, VT: MVT::f64, Action: Legal); |
| 258 | setOperationAction(Op: ISD::FMAXNUM_IEEE, VT: MVT::f64, Action: Legal); |
| 259 | setOperationAction(Op: ISD::FMINNUM, VT: MVT::f64, Action: Legal); |
| 260 | setOperationAction(Op: ISD::FMAXNUM, VT: MVT::f64, Action: Legal); |
| 261 | setOperationAction(Op: ISD::IS_FPCLASS, VT: MVT::f32, Action: Legal); |
| 262 | setOperationAction(Op: ISD::IS_FPCLASS, VT: MVT::f64, Action: Legal); |
| 263 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f32, Action: Legal); |
| 264 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f64, Action: Legal); |
| 265 | } else { |
| 266 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f32, Action: Custom); |
| 267 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f64, Action: Custom); |
| 268 | } |
| 269 | |
| 270 | if (Subtarget.isGP64bit()) { |
| 271 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i64, Action: Custom); |
| 272 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i64, Action: Custom); |
| 273 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVT::i64, Action: Custom); |
| 274 | setOperationAction(Op: ISD::JumpTable, VT: MVT::i64, Action: Custom); |
| 275 | setOperationAction(Op: ISD::ConstantPool, VT: MVT::i64, Action: Custom); |
| 276 | setOperationAction(Op: ISD::SELECT, VT: MVT::i64, Action: Custom); |
| 277 | if (Subtarget.hasMips64r6()) { |
| 278 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Legal); |
| 279 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Legal); |
| 280 | } else { |
| 281 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Custom); |
| 282 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Custom); |
| 283 | } |
| 284 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i64, Action: Custom); |
| 285 | setOperationAction(Op: ISD::STRICT_FP_TO_UINT, VT: MVT::i64, Action: Custom); |
| 286 | setOperationAction(Op: ISD::STRICT_FP_TO_SINT, VT: MVT::i64, Action: Custom); |
| 287 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i64, Action: Custom); |
| 288 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i64, Action: Custom); |
| 289 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i64, Action: Custom); |
| 290 | } |
| 291 | |
| 292 | if (!Subtarget.isGP64bit()) { |
| 293 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i32, Action: Custom); |
| 294 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i32, Action: Custom); |
| 295 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i32, Action: Custom); |
| 296 | } |
| 297 | |
| 298 | setOperationAction(Op: ISD::EH_DWARF_CFA, VT: MVT::i32, Action: Custom); |
| 299 | if (Subtarget.isGP64bit()) |
| 300 | setOperationAction(Op: ISD::EH_DWARF_CFA, VT: MVT::i64, Action: Custom); |
| 301 | |
| 302 | setOperationAction(Op: ISD::SDIV, VT: MVT::i32, Action: Expand); |
| 303 | setOperationAction(Op: ISD::SREM, VT: MVT::i32, Action: Expand); |
| 304 | setOperationAction(Op: ISD::UDIV, VT: MVT::i32, Action: Expand); |
| 305 | setOperationAction(Op: ISD::UREM, VT: MVT::i32, Action: Expand); |
| 306 | setOperationAction(Op: ISD::SDIV, VT: MVT::i64, Action: Expand); |
| 307 | setOperationAction(Op: ISD::SREM, VT: MVT::i64, Action: Expand); |
| 308 | setOperationAction(Op: ISD::UDIV, VT: MVT::i64, Action: Expand); |
| 309 | setOperationAction(Op: ISD::UREM, VT: MVT::i64, Action: Expand); |
| 310 | |
| 311 | // Operations not directly supported by Mips. |
| 312 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f32, Action: Expand); |
| 313 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f64, Action: Expand); |
| 314 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Expand); |
| 315 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Expand); |
| 316 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Expand); |
| 317 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i64, Action: Expand); |
| 318 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f32, Action: Expand); |
| 319 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f64, Action: Expand); |
| 320 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i32, Action: Expand); |
| 321 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i64, Action: Expand); |
| 322 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i32, Action: Expand); |
| 323 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i64, Action: Expand); |
| 324 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand); |
| 325 | |
| 326 | if (Subtarget.hasCnMips()) { |
| 327 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, Action: Legal); |
| 328 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, Action: Legal); |
| 329 | } else { |
| 330 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, Action: Expand); |
| 331 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, Action: Expand); |
| 332 | } |
| 333 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
| 334 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
| 335 | setOperationAction(Op: ISD::ROTL, VT: MVT::i32, Action: Expand); |
| 336 | setOperationAction(Op: ISD::ROTL, VT: MVT::i64, Action: Expand); |
| 337 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i32, Action: Expand); |
| 338 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i64, Action: Expand); |
| 339 | |
| 340 | if (!Subtarget.hasMips32r2()) |
| 341 | setOperationAction(Op: ISD::ROTR, VT: MVT::i32, Action: Expand); |
| 342 | |
| 343 | if (!Subtarget.hasMips64r2()) |
| 344 | setOperationAction(Op: ISD::ROTR, VT: MVT::i64, Action: Expand); |
| 345 | |
| 346 | setOperationAction(Op: ISD::FSIN, VT: MVT::f32, Action: Expand); |
| 347 | setOperationAction(Op: ISD::FSIN, VT: MVT::f64, Action: Expand); |
| 348 | setOperationAction(Op: ISD::FCOS, VT: MVT::f32, Action: Expand); |
| 349 | setOperationAction(Op: ISD::FCOS, VT: MVT::f64, Action: Expand); |
| 350 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f32, Action: Expand); |
| 351 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f64, Action: Expand); |
| 352 | setOperationAction(Op: ISD::FPOW, VT: MVT::f32, Action: Expand); |
| 353 | setOperationAction(Op: ISD::FPOW, VT: MVT::f64, Action: Expand); |
| 354 | setOperationAction(Op: ISD::FLOG, VT: MVT::f32, Action: Expand); |
| 355 | setOperationAction(Op: ISD::FLOG2, VT: MVT::f32, Action: Expand); |
| 356 | setOperationAction(Op: ISD::FLOG10, VT: MVT::f32, Action: Expand); |
| 357 | setOperationAction(Op: ISD::FEXP, VT: MVT::f32, Action: Expand); |
| 358 | setOperationAction(Op: ISD::FMA, VT: MVT::f32, Action: Expand); |
| 359 | setOperationAction(Op: ISD::FMA, VT: MVT::f64, Action: Expand); |
| 360 | setOperationAction(Op: ISD::FREM, VT: MVT::f32, Action: LibCall); |
| 361 | setOperationAction(Op: ISD::FREM, VT: MVT::f64, Action: LibCall); |
| 362 | |
| 363 | // Lower f16 conversion operations into library calls |
| 364 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f32, Action: Expand); |
| 365 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f32, Action: Expand); |
| 366 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f64, Action: Expand); |
| 367 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f64, Action: Expand); |
| 368 | |
| 369 | setOperationAction(Op: ISD::EH_RETURN, VT: MVT::Other, Action: Custom); |
| 370 | |
| 371 | setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom); |
| 372 | setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Custom); |
| 373 | setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand); |
| 374 | setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand); |
| 375 | |
| 376 | // Use the default for now |
| 377 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
| 378 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
| 379 | |
| 380 | if (!Subtarget.isGP64bit()) { |
| 381 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i64, Action: Expand); |
| 382 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i64, Action: Expand); |
| 383 | } |
| 384 | |
| 385 | if (!Subtarget.hasMips32r2()) { |
| 386 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i8, Action: Expand); |
| 387 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i16, Action: Expand); |
| 388 | } |
| 389 | |
| 390 | // MIPS16 lacks MIPS32's clz and clo instructions. |
| 391 | if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) |
| 392 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
| 393 | if (!Subtarget.hasMips64()) |
| 394 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
| 395 | |
| 396 | if (!Subtarget.hasMips32r2()) |
| 397 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Expand); |
| 398 | if (!Subtarget.hasMips64r2()) |
| 399 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i64, Action: Expand); |
| 400 | |
| 401 | if (Subtarget.isGP64bit() && Subtarget.hasMips64r6()) { |
| 402 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
| 403 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
| 404 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
| 405 | setTruncStoreAction(ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
| 406 | } else if (Subtarget.isGP64bit()) { |
| 407 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
| 408 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
| 409 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
| 410 | setTruncStoreAction(ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
| 411 | } |
| 412 | |
| 413 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal); |
| 414 | |
| 415 | setTargetDAGCombine({ISD::SDIVREM, ISD::UDIVREM, ISD::SELECT, ISD::AND, |
| 416 | ISD::OR, ISD::ADD, ISD::SUB, ISD::AssertZext, ISD::SHL, |
| 417 | ISD::SIGN_EXTEND}); |
| 418 | |
| 419 | // R5900 has no LL/SC instructions for atomic operations |
| 420 | if (Subtarget.isR5900()) |
| 421 | setMaxAtomicSizeInBitsSupported(0); |
| 422 | else if (Subtarget.isGP64bit()) |
| 423 | setMaxAtomicSizeInBitsSupported(64); |
| 424 | else |
| 425 | setMaxAtomicSizeInBitsSupported(32); |
| 426 | |
| 427 | setMinFunctionAlignment(Subtarget.isGP64bit() ? Align(8) : Align(4)); |
| 428 | |
| 429 | // The arguments on the stack are defined in terms of 4-byte slots on O32 |
| 430 | // and 8-byte slots on N32/N64. |
| 431 | setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? Align(8) |
| 432 | : Align(4)); |
| 433 | |
| 434 | setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP); |
| 435 | |
| 436 | MaxStoresPerMemcpy = 16; |
| 437 | |
| 438 | isMicroMips = Subtarget.inMicroMipsMode(); |
| 439 | } |
| 440 | |
| 441 | const MipsTargetLowering * |
| 442 | MipsTargetLowering::create(const MipsTargetMachine &TM, |
| 443 | const MipsSubtarget &STI) { |
| 444 | if (STI.inMips16Mode()) |
| 445 | return createMips16TargetLowering(TM, STI); |
| 446 | |
| 447 | return createMipsSETargetLowering(TM, STI); |
| 448 | } |
| 449 | |
| 450 | // Create a fast isel object. |
| 451 | FastISel *MipsTargetLowering::createFastISel( |
| 452 | FunctionLoweringInfo &funcInfo, const TargetLibraryInfo *libInfo, |
| 453 | const LibcallLoweringInfo *libcallLowering) const { |
| 454 | const MipsTargetMachine &TM = |
| 455 | static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget()); |
| 456 | |
| 457 | // We support only the standard encoding [MIPS32,MIPS32R5] ISAs. |
| 458 | bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() && |
| 459 | !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() && |
| 460 | !Subtarget.inMicroMipsMode(); |
| 461 | |
| 462 | // Disable if either of the following is true: |
| 463 | // We do not generate PIC, the ABI is not O32, XGOT is being used. |
| 464 | if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || |
| 465 | Subtarget.useXGOT()) |
| 466 | UseFastISel = false; |
| 467 | |
| 468 | return UseFastISel ? Mips::createFastISel(funcInfo, libInfo, libcallLowering) |
| 469 | : nullptr; |
| 470 | } |
| 471 | |
| 472 | EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, |
| 473 | EVT VT) const { |
| 474 | if (!VT.isVector()) |
| 475 | return MVT::i32; |
| 476 | return VT.changeVectorElementTypeToInteger(); |
| 477 | } |
| 478 | |
| 479 | static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, |
| 480 | TargetLowering::DAGCombinerInfo &DCI, |
| 481 | const MipsSubtarget &Subtarget) { |
| 482 | if (DCI.isBeforeLegalizeOps()) |
| 483 | return SDValue(); |
| 484 | |
| 485 | EVT Ty = N->getValueType(ResNo: 0); |
| 486 | unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; |
| 487 | unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; |
| 488 | unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : |
| 489 | MipsISD::DivRemU16; |
| 490 | SDLoc DL(N); |
| 491 | |
| 492 | SDValue DivRem = DAG.getNode(Opcode: Opc, DL, VT: MVT::Glue, |
| 493 | N1: N->getOperand(Num: 0), N2: N->getOperand(Num: 1)); |
| 494 | SDValue InChain = DAG.getEntryNode(); |
| 495 | SDValue InGlue = DivRem; |
| 496 | |
| 497 | // insert MFLO |
| 498 | if (N->hasAnyUseOfValue(Value: 0)) { |
| 499 | SDValue CopyFromLo = DAG.getCopyFromReg(Chain: InChain, dl: DL, Reg: LO, VT: Ty, |
| 500 | Glue: InGlue); |
| 501 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(N, 0), To: CopyFromLo); |
| 502 | InChain = CopyFromLo.getValue(R: 1); |
| 503 | InGlue = CopyFromLo.getValue(R: 2); |
| 504 | } |
| 505 | |
| 506 | // insert MFHI |
| 507 | if (N->hasAnyUseOfValue(Value: 1)) { |
| 508 | SDValue CopyFromHi = DAG.getCopyFromReg(Chain: InChain, dl: DL, |
| 509 | Reg: HI, VT: Ty, Glue: InGlue); |
| 510 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(N, 1), To: CopyFromHi); |
| 511 | } |
| 512 | |
| 513 | return SDValue(); |
| 514 | } |
| 515 | |
| 516 | static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { |
| 517 | switch (CC) { |
| 518 | default: llvm_unreachable("Unknown fp condition code!" ); |
| 519 | case ISD::SETEQ: |
| 520 | case ISD::SETOEQ: return Mips::FCOND_OEQ; |
| 521 | case ISD::SETUNE: return Mips::FCOND_UNE; |
| 522 | case ISD::SETLT: |
| 523 | case ISD::SETOLT: return Mips::FCOND_OLT; |
| 524 | case ISD::SETGT: |
| 525 | case ISD::SETOGT: return Mips::FCOND_OGT; |
| 526 | case ISD::SETLE: |
| 527 | case ISD::SETOLE: return Mips::FCOND_OLE; |
| 528 | case ISD::SETGE: |
| 529 | case ISD::SETOGE: return Mips::FCOND_OGE; |
| 530 | case ISD::SETULT: return Mips::FCOND_ULT; |
| 531 | case ISD::SETULE: return Mips::FCOND_ULE; |
| 532 | case ISD::SETUGT: return Mips::FCOND_UGT; |
| 533 | case ISD::SETUGE: return Mips::FCOND_UGE; |
| 534 | case ISD::SETUO: return Mips::FCOND_UN; |
| 535 | case ISD::SETO: return Mips::FCOND_OR; |
| 536 | case ISD::SETNE: |
| 537 | case ISD::SETONE: return Mips::FCOND_ONE; |
| 538 | case ISD::SETUEQ: return Mips::FCOND_UEQ; |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | /// This function returns true if the floating point conditional branches and |
| 543 | /// conditional moves which use condition code CC should be inverted. |
| 544 | static bool invertFPCondCodeUser(Mips::CondCode CC) { |
| 545 | if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) |
| 546 | return false; |
| 547 | |
| 548 | assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && |
| 549 | "Illegal Condition Code" ); |
| 550 | |
| 551 | return true; |
| 552 | } |
| 553 | |
| 554 | // Creates and returns an FPCmp node from a setcc node. |
| 555 | // Returns Op if setcc is not a floating point comparison. |
| 556 | static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { |
| 557 | // must be a SETCC node |
| 558 | if (Op.getOpcode() != ISD::SETCC && Op.getOpcode() != ISD::STRICT_FSETCC && |
| 559 | Op.getOpcode() != ISD::STRICT_FSETCCS) |
| 560 | return Op; |
| 561 | |
| 562 | SDValue LHS = Op.getOperand(i: 0); |
| 563 | |
| 564 | if (!LHS.getValueType().isFloatingPoint()) |
| 565 | return Op; |
| 566 | |
| 567 | SDValue RHS = Op.getOperand(i: 1); |
| 568 | SDLoc DL(Op); |
| 569 | |
| 570 | // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of |
| 571 | // node if necessary. |
| 572 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 2))->get(); |
| 573 | |
| 574 | return DAG.getNode(Opcode: MipsISD::FPCmp, DL, VT: MVT::Glue, N1: LHS, N2: RHS, |
| 575 | N3: DAG.getConstant(Val: condCodeToFCC(CC), DL, VT: MVT::i32)); |
| 576 | } |
| 577 | |
| 578 | // Creates and returns a CMovFPT/F node. |
| 579 | static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, |
| 580 | SDValue False, const SDLoc &DL) { |
| 581 | ConstantSDNode *CC = cast<ConstantSDNode>(Val: Cond.getOperand(i: 2)); |
| 582 | bool invert = invertFPCondCodeUser(CC: (Mips::CondCode)CC->getSExtValue()); |
| 583 | SDValue FCC0 = DAG.getRegister(Reg: Mips::FCC0, VT: MVT::i32); |
| 584 | |
| 585 | return DAG.getNode(Opcode: (invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, |
| 586 | VT: True.getValueType(), N1: True, N2: FCC0, N3: False, N4: Cond); |
| 587 | } |
| 588 | |
| 589 | static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, |
| 590 | TargetLowering::DAGCombinerInfo &DCI, |
| 591 | const MipsSubtarget &Subtarget) { |
| 592 | if (DCI.isBeforeLegalizeOps()) |
| 593 | return SDValue(); |
| 594 | |
| 595 | SDValue SetCC = N->getOperand(Num: 0); |
| 596 | |
| 597 | if ((SetCC.getOpcode() != ISD::SETCC) || |
| 598 | !SetCC.getOperand(i: 0).getValueType().isInteger()) |
| 599 | return SDValue(); |
| 600 | |
| 601 | SDValue False = N->getOperand(Num: 2); |
| 602 | EVT FalseTy = False.getValueType(); |
| 603 | |
| 604 | if (!FalseTy.isInteger()) |
| 605 | return SDValue(); |
| 606 | |
| 607 | ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(Val&: False); |
| 608 | |
| 609 | // If the RHS (False) is 0, we swap the order of the operands |
| 610 | // of ISD::SELECT (obviously also inverting the condition) so that we can |
| 611 | // take advantage of conditional moves using the $0 register. |
| 612 | // Example: |
| 613 | // return (a != 0) ? x : 0; |
| 614 | // load $reg, x |
| 615 | // movz $reg, $0, a |
| 616 | if (!FalseC) |
| 617 | return SDValue(); |
| 618 | |
| 619 | const SDLoc DL(N); |
| 620 | |
| 621 | if (!FalseC->getZExtValue()) { |
| 622 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: SetCC.getOperand(i: 2))->get(); |
| 623 | SDValue True = N->getOperand(Num: 1); |
| 624 | |
| 625 | SetCC = DAG.getSetCC(DL, VT: SetCC.getValueType(), LHS: SetCC.getOperand(i: 0), |
| 626 | RHS: SetCC.getOperand(i: 1), |
| 627 | Cond: ISD::getSetCCInverse(Operation: CC, Type: SetCC.getValueType())); |
| 628 | |
| 629 | return DAG.getNode(Opcode: ISD::SELECT, DL, VT: FalseTy, N1: SetCC, N2: False, N3: True); |
| 630 | } |
| 631 | |
| 632 | // If both operands are integer constants there's a possibility that we |
| 633 | // can do some interesting optimizations. |
| 634 | SDValue True = N->getOperand(Num: 1); |
| 635 | ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(Val&: True); |
| 636 | |
| 637 | if (!TrueC || !True.getValueType().isInteger()) |
| 638 | return SDValue(); |
| 639 | |
| 640 | // We'll also ignore MVT::i64 operands as this optimizations proves |
| 641 | // to be ineffective because of the required sign extensions as the result |
| 642 | // of a SETCC operator is always MVT::i32 for non-vector types. |
| 643 | if (True.getValueType() == MVT::i64) |
| 644 | return SDValue(); |
| 645 | |
| 646 | int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); |
| 647 | |
| 648 | // 1) (a < x) ? y : y-1 |
| 649 | // slti $reg1, a, x |
| 650 | // addiu $reg2, $reg1, y-1 |
| 651 | if (Diff == 1) |
| 652 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: SetCC.getValueType(), N1: SetCC, N2: False); |
| 653 | |
| 654 | // 2) (a < x) ? y-1 : y |
| 655 | // slti $reg1, a, x |
| 656 | // xor $reg1, $reg1, 1 |
| 657 | // addiu $reg2, $reg1, y-1 |
| 658 | if (Diff == -1) { |
| 659 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: SetCC.getOperand(i: 2))->get(); |
| 660 | SetCC = DAG.getSetCC(DL, VT: SetCC.getValueType(), LHS: SetCC.getOperand(i: 0), |
| 661 | RHS: SetCC.getOperand(i: 1), |
| 662 | Cond: ISD::getSetCCInverse(Operation: CC, Type: SetCC.getValueType())); |
| 663 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: SetCC.getValueType(), N1: SetCC, N2: True); |
| 664 | } |
| 665 | |
| 666 | // Could not optimize. |
| 667 | return SDValue(); |
| 668 | } |
| 669 | |
| 670 | static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, |
| 671 | TargetLowering::DAGCombinerInfo &DCI, |
| 672 | const MipsSubtarget &Subtarget) { |
| 673 | if (DCI.isBeforeLegalizeOps()) |
| 674 | return SDValue(); |
| 675 | |
| 676 | SDValue ValueIfTrue = N->getOperand(Num: 0), ValueIfFalse = N->getOperand(Num: 2); |
| 677 | |
| 678 | ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(Val&: ValueIfFalse); |
| 679 | if (!FalseC || FalseC->getZExtValue()) |
| 680 | return SDValue(); |
| 681 | |
| 682 | // Since RHS (False) is 0, we swap the order of the True/False operands |
| 683 | // (obviously also inverting the condition) so that we can |
| 684 | // take advantage of conditional moves using the $0 register. |
| 685 | // Example: |
| 686 | // return (a != 0) ? x : 0; |
| 687 | // load $reg, x |
| 688 | // movz $reg, $0, a |
| 689 | unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : |
| 690 | MipsISD::CMovFP_T; |
| 691 | |
| 692 | SDValue FCC = N->getOperand(Num: 1), Glue = N->getOperand(Num: 3); |
| 693 | return DAG.getNode(Opcode: Opc, DL: SDLoc(N), VT: ValueIfFalse.getValueType(), |
| 694 | N1: ValueIfFalse, N2: FCC, N3: ValueIfTrue, N4: Glue); |
| 695 | } |
| 696 | |
| 697 | static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, |
| 698 | TargetLowering::DAGCombinerInfo &DCI, |
| 699 | const MipsSubtarget &Subtarget) { |
| 700 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) |
| 701 | return SDValue(); |
| 702 | |
| 703 | SDValue FirstOperand = N->getOperand(Num: 0); |
| 704 | unsigned FirstOperandOpc = FirstOperand.getOpcode(); |
| 705 | SDValue Mask = N->getOperand(Num: 1); |
| 706 | EVT ValTy = N->getValueType(ResNo: 0); |
| 707 | SDLoc DL(N); |
| 708 | |
| 709 | uint64_t Pos = 0; |
| 710 | unsigned SMPos, SMSize; |
| 711 | ConstantSDNode *CN; |
| 712 | SDValue NewOperand; |
| 713 | unsigned Opc; |
| 714 | |
| 715 | // Op's second operand must be a shifted mask. |
| 716 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: Mask)) || |
| 717 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos, MaskLen&: SMSize)) |
| 718 | return SDValue(); |
| 719 | |
| 720 | if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { |
| 721 | // Pattern match EXT. |
| 722 | // $dst = and ((sra or srl) $src , pos), (2**size - 1) |
| 723 | // => ext $dst, $src, pos, size |
| 724 | |
| 725 | // The second operand of the shift must be an immediate. |
| 726 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1)))) |
| 727 | return SDValue(); |
| 728 | |
| 729 | Pos = CN->getZExtValue(); |
| 730 | |
| 731 | // Return if the shifted mask does not start at bit 0 or the sum of its size |
| 732 | // and Pos exceeds the word's size. |
| 733 | if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) |
| 734 | return SDValue(); |
| 735 | |
| 736 | Opc = MipsISD::Ext; |
| 737 | NewOperand = FirstOperand.getOperand(i: 0); |
| 738 | } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) { |
| 739 | // Pattern match CINS. |
| 740 | // $dst = and (shl $src , pos), mask |
| 741 | // => cins $dst, $src, pos, size |
| 742 | // mask is a shifted mask with consecutive 1's, pos = shift amount, |
| 743 | // size = population count. |
| 744 | |
| 745 | // The second operand of the shift must be an immediate. |
| 746 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1)))) |
| 747 | return SDValue(); |
| 748 | |
| 749 | Pos = CN->getZExtValue(); |
| 750 | |
| 751 | if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 || |
| 752 | Pos + SMSize > ValTy.getSizeInBits()) |
| 753 | return SDValue(); |
| 754 | |
| 755 | NewOperand = FirstOperand.getOperand(i: 0); |
| 756 | // SMSize is 'location' (position) in this case, not size. |
| 757 | SMSize--; |
| 758 | Opc = MipsISD::CIns; |
| 759 | } else { |
| 760 | // Pattern match EXT. |
| 761 | // $dst = and $src, (2**size - 1) , if size > 16 |
| 762 | // => ext $dst, $src, pos, size , pos = 0 |
| 763 | |
| 764 | // If the mask is <= 0xffff, andi can be used instead. |
| 765 | if (CN->getZExtValue() <= 0xffff) |
| 766 | return SDValue(); |
| 767 | |
| 768 | // Return if the mask doesn't start at position 0. |
| 769 | if (SMPos) |
| 770 | return SDValue(); |
| 771 | |
| 772 | Opc = MipsISD::Ext; |
| 773 | NewOperand = FirstOperand; |
| 774 | } |
| 775 | return DAG.getNode(Opcode: Opc, DL, VT: ValTy, N1: NewOperand, |
| 776 | N2: DAG.getConstant(Val: Pos, DL, VT: MVT::i32), |
| 777 | N3: DAG.getConstant(Val: SMSize, DL, VT: MVT::i32)); |
| 778 | } |
| 779 | |
| 780 | static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, |
| 781 | TargetLowering::DAGCombinerInfo &DCI, |
| 782 | const MipsSubtarget &Subtarget) { |
| 783 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) |
| 784 | return SDValue(); |
| 785 | |
| 786 | SDValue FirstOperand = N->getOperand(Num: 0), SecondOperand = N->getOperand(Num: 1); |
| 787 | unsigned SMPos0, SMSize0, SMPos1, SMSize1; |
| 788 | ConstantSDNode *CN, *CN1; |
| 789 | |
| 790 | if ((FirstOperand.getOpcode() == ISD::AND && |
| 791 | SecondOperand.getOpcode() == ISD::SHL) || |
| 792 | (FirstOperand.getOpcode() == ISD::SHL && |
| 793 | SecondOperand.getOpcode() == ISD::AND)) { |
| 794 | // Pattern match INS. |
| 795 | // $dst = or (and $src1, (2**size0 - 1)), (shl $src2, size0) |
| 796 | // ==> ins $src1, $src2, pos, size, pos = size0, size = 32 - pos; |
| 797 | // Or: |
| 798 | // $dst = or (shl $src2, size0), (and $src1, (2**size0 - 1)) |
| 799 | // ==> ins $src1, $src2, pos, size, pos = size0, size = 32 - pos; |
| 800 | SDValue AndOperand0 = FirstOperand.getOpcode() == ISD::AND |
| 801 | ? FirstOperand.getOperand(i: 0) |
| 802 | : SecondOperand.getOperand(i: 0); |
| 803 | SDValue ShlOperand0 = FirstOperand.getOpcode() == ISD::AND |
| 804 | ? SecondOperand.getOperand(i: 0) |
| 805 | : FirstOperand.getOperand(i: 0); |
| 806 | SDValue AndMask = FirstOperand.getOpcode() == ISD::AND |
| 807 | ? FirstOperand.getOperand(i: 1) |
| 808 | : SecondOperand.getOperand(i: 1); |
| 809 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: AndMask)) || |
| 810 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos0, MaskLen&: SMSize0)) |
| 811 | return SDValue(); |
| 812 | |
| 813 | SDValue ShlShift = FirstOperand.getOpcode() == ISD::AND |
| 814 | ? SecondOperand.getOperand(i: 1) |
| 815 | : FirstOperand.getOperand(i: 1); |
| 816 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: ShlShift))) |
| 817 | return SDValue(); |
| 818 | uint64_t ShlShiftValue = CN->getZExtValue(); |
| 819 | |
| 820 | if (SMPos0 != 0 || SMSize0 != ShlShiftValue) |
| 821 | return SDValue(); |
| 822 | |
| 823 | SDLoc DL(N); |
| 824 | EVT ValTy = N->getValueType(ResNo: 0); |
| 825 | SMPos1 = ShlShiftValue; |
| 826 | assert(SMPos1 < ValTy.getSizeInBits()); |
| 827 | SMSize1 = (ValTy == MVT::i64 ? 64 : 32) - SMPos1; |
| 828 | return DAG.getNode(Opcode: MipsISD::Ins, DL, VT: ValTy, N1: ShlOperand0, |
| 829 | N2: DAG.getConstant(Val: SMPos1, DL, VT: MVT::i32), |
| 830 | N3: DAG.getConstant(Val: SMSize1, DL, VT: MVT::i32), N4: AndOperand0); |
| 831 | } |
| 832 | |
| 833 | // See if Op's first operand matches (and $src1 , mask0). |
| 834 | if (FirstOperand.getOpcode() != ISD::AND) |
| 835 | return SDValue(); |
| 836 | |
| 837 | // Pattern match INS. |
| 838 | // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), |
| 839 | // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 |
| 840 | // => ins $dst, $src, size, pos, $src1 |
| 841 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1))) || |
| 842 | !isShiftedMask_64(Value: ~CN->getSExtValue(), MaskIdx&: SMPos0, MaskLen&: SMSize0)) |
| 843 | return SDValue(); |
| 844 | |
| 845 | // See if Op's second operand matches (and (shl $src, pos), mask1). |
| 846 | if (SecondOperand.getOpcode() == ISD::AND && |
| 847 | SecondOperand.getOperand(i: 0).getOpcode() == ISD::SHL) { |
| 848 | |
| 849 | if (!(CN = dyn_cast<ConstantSDNode>(Val: SecondOperand.getOperand(i: 1))) || |
| 850 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos1, MaskLen&: SMSize1)) |
| 851 | return SDValue(); |
| 852 | |
| 853 | // The shift masks must have the same position and size. |
| 854 | if (SMPos0 != SMPos1 || SMSize0 != SMSize1) |
| 855 | return SDValue(); |
| 856 | |
| 857 | SDValue Shl = SecondOperand.getOperand(i: 0); |
| 858 | |
| 859 | if (!(CN = dyn_cast<ConstantSDNode>(Val: Shl.getOperand(i: 1)))) |
| 860 | return SDValue(); |
| 861 | |
| 862 | unsigned Shamt = CN->getZExtValue(); |
| 863 | |
| 864 | // Return if the shift amount and the first bit position of mask are not the |
| 865 | // same. |
| 866 | EVT ValTy = N->getValueType(ResNo: 0); |
| 867 | if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) |
| 868 | return SDValue(); |
| 869 | |
| 870 | SDLoc DL(N); |
| 871 | return DAG.getNode(Opcode: MipsISD::Ins, DL, VT: ValTy, N1: Shl.getOperand(i: 0), |
| 872 | N2: DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32), |
| 873 | N3: DAG.getConstant(Val: SMSize0, DL, VT: MVT::i32), |
| 874 | N4: FirstOperand.getOperand(i: 0)); |
| 875 | } else { |
| 876 | // Pattern match DINS. |
| 877 | // $dst = or (and $src, mask0), mask1 |
| 878 | // where mask0 = ((1 << SMSize0) -1) << SMPos0 |
| 879 | // => dins $dst, $src, pos, size |
| 880 | if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) && |
| 881 | ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) || |
| 882 | (SMSize0 + SMPos0 <= 32))) { |
| 883 | // Check if AND instruction has constant as argument |
| 884 | bool isConstCase = SecondOperand.getOpcode() != ISD::AND; |
| 885 | if (SecondOperand.getOpcode() == ISD::AND) { |
| 886 | if (!(CN1 = dyn_cast<ConstantSDNode>(Val: SecondOperand->getOperand(Num: 1)))) |
| 887 | return SDValue(); |
| 888 | } else { |
| 889 | if (!(CN1 = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1)))) |
| 890 | return SDValue(); |
| 891 | } |
| 892 | // Don't generate INS if constant OR operand doesn't fit into bits |
| 893 | // cleared by constant AND operand. |
| 894 | if (CN->getSExtValue() & CN1->getSExtValue()) |
| 895 | return SDValue(); |
| 896 | |
| 897 | SDLoc DL(N); |
| 898 | EVT ValTy = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
| 899 | SDValue Const1; |
| 900 | SDValue SrlX; |
| 901 | if (!isConstCase) { |
| 902 | Const1 = DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32); |
| 903 | SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: SecondOperand->getValueType(ResNo: 0), |
| 904 | N1: SecondOperand, N2: Const1); |
| 905 | } |
| 906 | return DAG.getNode( |
| 907 | Opcode: MipsISD::Ins, DL, VT: N->getValueType(ResNo: 0), |
| 908 | N1: isConstCase |
| 909 | ? DAG.getSignedConstant(Val: CN1->getSExtValue() >> SMPos0, DL, VT: ValTy) |
| 910 | : SrlX, |
| 911 | N2: DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32), |
| 912 | N3: DAG.getConstant(Val: ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31 |
| 913 | : SMSize0, |
| 914 | DL, VT: MVT::i32), |
| 915 | N4: FirstOperand->getOperand(Num: 0)); |
| 916 | } |
| 917 | return SDValue(); |
| 918 | } |
| 919 | } |
| 920 | |
| 921 | static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG, |
| 922 | const MipsSubtarget &Subtarget) { |
| 923 | // ROOTNode must have a multiplication as an operand for the match to be |
| 924 | // successful. |
| 925 | if (ROOTNode->getOperand(Num: 0).getOpcode() != ISD::MUL && |
| 926 | ROOTNode->getOperand(Num: 1).getOpcode() != ISD::MUL) |
| 927 | return SDValue(); |
| 928 | |
| 929 | // In the case where we have a multiplication as the left operand of |
| 930 | // of a subtraction, we can't combine into a MipsISD::MSub node as the |
| 931 | // the instruction definition of msub(u) places the multiplication on |
| 932 | // on the right. |
| 933 | if (ROOTNode->getOpcode() == ISD::SUB && |
| 934 | ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL) |
| 935 | return SDValue(); |
| 936 | |
| 937 | // We don't handle vector types here. |
| 938 | if (ROOTNode->getValueType(ResNo: 0).isVector()) |
| 939 | return SDValue(); |
| 940 | |
| 941 | // For MIPS64, madd / msub instructions are inefficent to use with 64 bit |
| 942 | // arithmetic. E.g. |
| 943 | // (add (mul a b) c) => |
| 944 | // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in |
| 945 | // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32) |
| 946 | // or |
| 947 | // MIPS64R2: (dins (mflo res) (mfhi res) 32 32) |
| 948 | // |
| 949 | // The overhead of setting up the Hi/Lo registers and reassembling the |
| 950 | // result makes this a dubious optimzation for MIPS64. The core of the |
| 951 | // problem is that Hi/Lo contain the upper and lower 32 bits of the |
| 952 | // operand and result. |
| 953 | // |
| 954 | // It requires a chain of 4 add/mul for MIPS64R2 to get better code |
| 955 | // density than doing it naively, 5 for MIPS64. Additionally, using |
| 956 | // madd/msub on MIPS64 requires the operands actually be 32 bit sign |
| 957 | // extended operands, not true 64 bit values. |
| 958 | // |
| 959 | // FIXME: For the moment, disable this completely for MIPS64. |
| 960 | if (Subtarget.hasMips64()) |
| 961 | return SDValue(); |
| 962 | |
| 963 | SDValue Mult = ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL |
| 964 | ? ROOTNode->getOperand(Num: 0) |
| 965 | : ROOTNode->getOperand(Num: 1); |
| 966 | |
| 967 | SDValue AddOperand = ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL |
| 968 | ? ROOTNode->getOperand(Num: 1) |
| 969 | : ROOTNode->getOperand(Num: 0); |
| 970 | |
| 971 | // Transform this to a MADD only if the user of this node is the add. |
| 972 | // If there are other users of the mul, this function returns here. |
| 973 | if (!Mult.hasOneUse()) |
| 974 | return SDValue(); |
| 975 | |
| 976 | // maddu and madd are unusual instructions in that on MIPS64 bits 63..31 |
| 977 | // must be in canonical form, i.e. sign extended. For MIPS32, the operands |
| 978 | // of the multiply must have 32 or more sign bits, otherwise we cannot |
| 979 | // perform this optimization. We have to check this here as we're performing |
| 980 | // this optimization pre-legalization. |
| 981 | SDValue MultLHS = Mult->getOperand(Num: 0); |
| 982 | SDValue MultRHS = Mult->getOperand(Num: 1); |
| 983 | |
| 984 | bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND && |
| 985 | MultRHS->getOpcode() == ISD::SIGN_EXTEND; |
| 986 | bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND && |
| 987 | MultRHS->getOpcode() == ISD::ZERO_EXTEND; |
| 988 | |
| 989 | if (!IsSigned && !IsUnsigned) |
| 990 | return SDValue(); |
| 991 | |
| 992 | // Initialize accumulator. |
| 993 | SDLoc DL(ROOTNode); |
| 994 | SDValue BottomHalf, TopHalf; |
| 995 | std::tie(args&: BottomHalf, args&: TopHalf) = |
| 996 | CurDAG.SplitScalar(N: AddOperand, DL, LoVT: MVT::i32, HiVT: MVT::i32); |
| 997 | SDValue ACCIn = |
| 998 | CurDAG.getNode(Opcode: MipsISD::MTLOHI, DL, VT: MVT::Untyped, N1: BottomHalf, N2: TopHalf); |
| 999 | |
| 1000 | // Create MipsMAdd(u) / MipsMSub(u) node. |
| 1001 | bool IsAdd = ROOTNode->getOpcode() == ISD::ADD; |
| 1002 | unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd) |
| 1003 | : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub); |
| 1004 | SDValue MAddOps[3] = { |
| 1005 | CurDAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: Mult->getOperand(Num: 0)), |
| 1006 | CurDAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: Mult->getOperand(Num: 1)), ACCIn}; |
| 1007 | SDValue MAdd = CurDAG.getNode(Opcode, DL, VT: MVT::Untyped, Ops: MAddOps); |
| 1008 | |
| 1009 | SDValue ResLo = CurDAG.getNode(Opcode: MipsISD::MFLO, DL, VT: MVT::i32, Operand: MAdd); |
| 1010 | SDValue ResHi = CurDAG.getNode(Opcode: MipsISD::MFHI, DL, VT: MVT::i32, Operand: MAdd); |
| 1011 | SDValue Combined = |
| 1012 | CurDAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VT: MVT::i64, N1: ResLo, N2: ResHi); |
| 1013 | return Combined; |
| 1014 | } |
| 1015 | |
| 1016 | static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG, |
| 1017 | TargetLowering::DAGCombinerInfo &DCI, |
| 1018 | const MipsSubtarget &Subtarget) { |
| 1019 | // (sub v0 (mul v1, v2)) => (msub v1, v2, v0) |
| 1020 | if (DCI.isBeforeLegalizeOps()) { |
| 1021 | if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && |
| 1022 | !Subtarget.inMips16Mode() && N->getValueType(ResNo: 0) == MVT::i64) |
| 1023 | return performMADD_MSUBCombine(ROOTNode: N, CurDAG&: DAG, Subtarget); |
| 1024 | |
| 1025 | return SDValue(); |
| 1026 | } |
| 1027 | |
| 1028 | return SDValue(); |
| 1029 | } |
| 1030 | |
| 1031 | static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, |
| 1032 | TargetLowering::DAGCombinerInfo &DCI, |
| 1033 | const MipsSubtarget &Subtarget) { |
| 1034 | // (add v0 (mul v1, v2)) => (madd v1, v2, v0) |
| 1035 | if (DCI.isBeforeLegalizeOps()) { |
| 1036 | if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && |
| 1037 | !Subtarget.inMips16Mode() && N->getValueType(ResNo: 0) == MVT::i64) |
| 1038 | return performMADD_MSUBCombine(ROOTNode: N, CurDAG&: DAG, Subtarget); |
| 1039 | |
| 1040 | return SDValue(); |
| 1041 | } |
| 1042 | |
| 1043 | // When loading from a jump table, push the Lo node to the position that |
| 1044 | // allows folding it into a load immediate. |
| 1045 | // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) |
| 1046 | // (add (add abs_lo(tjt), v1), v0) => (add (add v0, v1), abs_lo(tjt)) |
| 1047 | SDValue InnerAdd = N->getOperand(Num: 1); |
| 1048 | SDValue Index = N->getOperand(Num: 0); |
| 1049 | if (InnerAdd.getOpcode() != ISD::ADD) |
| 1050 | std::swap(a&: InnerAdd, b&: Index); |
| 1051 | if (InnerAdd.getOpcode() != ISD::ADD) |
| 1052 | return SDValue(); |
| 1053 | |
| 1054 | SDValue Lo = InnerAdd.getOperand(i: 0); |
| 1055 | SDValue Other = InnerAdd.getOperand(i: 1); |
| 1056 | if (Lo.getOpcode() != MipsISD::Lo) |
| 1057 | std::swap(a&: Lo, b&: Other); |
| 1058 | |
| 1059 | if ((Lo.getOpcode() != MipsISD::Lo) || |
| 1060 | (Lo.getOperand(i: 0).getOpcode() != ISD::TargetJumpTable)) |
| 1061 | return SDValue(); |
| 1062 | |
| 1063 | EVT ValTy = N->getValueType(ResNo: 0); |
| 1064 | SDLoc DL(N); |
| 1065 | |
| 1066 | SDValue Add1 = DAG.getNode(Opcode: ISD::ADD, DL, VT: ValTy, N1: Index, N2: Other); |
| 1067 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: ValTy, N1: Add1, N2: Lo); |
| 1068 | } |
| 1069 | |
| 1070 | static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, |
| 1071 | TargetLowering::DAGCombinerInfo &DCI, |
| 1072 | const MipsSubtarget &Subtarget) { |
| 1073 | // Pattern match CINS. |
| 1074 | // $dst = shl (and $src , imm), pos |
| 1075 | // => cins $dst, $src, pos, size |
| 1076 | |
| 1077 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips()) |
| 1078 | return SDValue(); |
| 1079 | |
| 1080 | SDValue FirstOperand = N->getOperand(Num: 0); |
| 1081 | unsigned FirstOperandOpc = FirstOperand.getOpcode(); |
| 1082 | SDValue SecondOperand = N->getOperand(Num: 1); |
| 1083 | EVT ValTy = N->getValueType(ResNo: 0); |
| 1084 | SDLoc DL(N); |
| 1085 | |
| 1086 | uint64_t Pos = 0; |
| 1087 | unsigned SMPos, SMSize; |
| 1088 | ConstantSDNode *CN; |
| 1089 | SDValue NewOperand; |
| 1090 | |
| 1091 | // The second operand of the shift must be an immediate. |
| 1092 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: SecondOperand))) |
| 1093 | return SDValue(); |
| 1094 | |
| 1095 | Pos = CN->getZExtValue(); |
| 1096 | |
| 1097 | if (Pos >= ValTy.getSizeInBits()) |
| 1098 | return SDValue(); |
| 1099 | |
| 1100 | if (FirstOperandOpc != ISD::AND) |
| 1101 | return SDValue(); |
| 1102 | |
| 1103 | // AND's second operand must be a shifted mask. |
| 1104 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1))) || |
| 1105 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos, MaskLen&: SMSize)) |
| 1106 | return SDValue(); |
| 1107 | |
| 1108 | // Return if the shifted mask does not start at bit 0 or the sum of its size |
| 1109 | // and Pos exceeds the word's size. |
| 1110 | if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits()) |
| 1111 | return SDValue(); |
| 1112 | |
| 1113 | NewOperand = FirstOperand.getOperand(i: 0); |
| 1114 | // SMSize is 'location' (position) in this case, not size. |
| 1115 | SMSize--; |
| 1116 | |
| 1117 | return DAG.getNode(Opcode: MipsISD::CIns, DL, VT: ValTy, N1: NewOperand, |
| 1118 | N2: DAG.getConstant(Val: Pos, DL, VT: MVT::i32), |
| 1119 | N3: DAG.getConstant(Val: SMSize, DL, VT: MVT::i32)); |
| 1120 | } |
| 1121 | |
| 1122 | static SDValue performSignExtendCombine(SDNode *N, SelectionDAG &DAG, |
| 1123 | TargetLowering::DAGCombinerInfo &DCI, |
| 1124 | const MipsSubtarget &Subtarget) { |
| 1125 | if (DCI.Level != AfterLegalizeDAG || !Subtarget.isGP64bit()) { |
| 1126 | return SDValue(); |
| 1127 | } |
| 1128 | |
| 1129 | SDValue N0 = N->getOperand(Num: 0); |
| 1130 | EVT VT = N->getValueType(ResNo: 0); |
| 1131 | |
| 1132 | // Pattern match XOR. |
| 1133 | // $dst = sign_extend (xor (trunc $src, i32), imm) |
| 1134 | // => $dst = xor (signext_inreg $src, i32), imm |
| 1135 | if (N0.getOpcode() == ISD::XOR && |
| 1136 | N0.getOperand(i: 0).getOpcode() == ISD::TRUNCATE && |
| 1137 | N0.getOperand(i: 1).getOpcode() == ISD::Constant) { |
| 1138 | SDValue TruncateSource = N0.getOperand(i: 0).getOperand(i: 0); |
| 1139 | auto *ConstantOperand = dyn_cast<ConstantSDNode>(Val: N0->getOperand(Num: 1)); |
| 1140 | |
| 1141 | SDValue FirstOperand = |
| 1142 | DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: SDLoc(N0), VT, N1: TruncateSource, |
| 1143 | N2: DAG.getValueType(N0.getOperand(i: 0).getValueType())); |
| 1144 | |
| 1145 | int64_t ConstImm = ConstantOperand->getSExtValue(); |
| 1146 | return DAG.getNode(Opcode: ISD::XOR, DL: SDLoc(N0), VT, N1: FirstOperand, |
| 1147 | N2: DAG.getConstant(Val: ConstImm, DL: SDLoc(N0), VT)); |
| 1148 | } |
| 1149 | |
| 1150 | return SDValue(); |
| 1151 | } |
| 1152 | |
| 1153 | SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) |
| 1154 | const { |
| 1155 | SelectionDAG &DAG = DCI.DAG; |
| 1156 | unsigned Opc = N->getOpcode(); |
| 1157 | |
| 1158 | switch (Opc) { |
| 1159 | default: break; |
| 1160 | case ISD::SDIVREM: |
| 1161 | case ISD::UDIVREM: |
| 1162 | return performDivRemCombine(N, DAG, DCI, Subtarget); |
| 1163 | case ISD::SELECT: |
| 1164 | return performSELECTCombine(N, DAG, DCI, Subtarget); |
| 1165 | case MipsISD::CMovFP_F: |
| 1166 | case MipsISD::CMovFP_T: |
| 1167 | return performCMovFPCombine(N, DAG, DCI, Subtarget); |
| 1168 | case ISD::AND: |
| 1169 | return performANDCombine(N, DAG, DCI, Subtarget); |
| 1170 | case ISD::OR: |
| 1171 | return performORCombine(N, DAG, DCI, Subtarget); |
| 1172 | case ISD::ADD: |
| 1173 | return performADDCombine(N, DAG, DCI, Subtarget); |
| 1174 | case ISD::SHL: |
| 1175 | return performSHLCombine(N, DAG, DCI, Subtarget); |
| 1176 | case ISD::SUB: |
| 1177 | return performSUBCombine(N, DAG, DCI, Subtarget); |
| 1178 | case ISD::SIGN_EXTEND: |
| 1179 | return performSignExtendCombine(N, DAG, DCI, Subtarget); |
| 1180 | } |
| 1181 | |
| 1182 | return SDValue(); |
| 1183 | } |
| 1184 | |
| 1185 | bool MipsTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
| 1186 | return Subtarget.hasMips32(); |
| 1187 | } |
| 1188 | |
| 1189 | bool MipsTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const { |
| 1190 | return Subtarget.hasMips32(); |
| 1191 | } |
| 1192 | |
| 1193 | bool MipsTargetLowering::hasBitTest(SDValue X, SDValue Y) const { |
| 1194 | // We can use ANDI+SLTIU as a bit test. Y contains the bit position. |
| 1195 | // For MIPSR2 or later, we may be able to use the `ext` instruction or its |
| 1196 | // double-word variants. |
| 1197 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: Y)) |
| 1198 | return C->getAPIntValue().ule(RHS: 15); |
| 1199 | |
| 1200 | return false; |
| 1201 | } |
| 1202 | |
| 1203 | bool MipsTargetLowering::shouldFoldConstantShiftPairToMask( |
| 1204 | const SDNode *N) const { |
| 1205 | assert(((N->getOpcode() == ISD::SHL && |
| 1206 | N->getOperand(0).getOpcode() == ISD::SRL) || |
| 1207 | (N->getOpcode() == ISD::SRL && |
| 1208 | N->getOperand(0).getOpcode() == ISD::SHL)) && |
| 1209 | "Expected shift-shift mask" ); |
| 1210 | |
| 1211 | if (N->getOperand(Num: 0).getValueType().isVector()) |
| 1212 | return false; |
| 1213 | return true; |
| 1214 | } |
| 1215 | |
| 1216 | void |
| 1217 | MipsTargetLowering::ReplaceNodeResults(SDNode *N, |
| 1218 | SmallVectorImpl<SDValue> &Results, |
| 1219 | SelectionDAG &DAG) const { |
| 1220 | return LowerOperationWrapper(N, Results, DAG); |
| 1221 | } |
| 1222 | |
| 1223 | SDValue MipsTargetLowering:: |
| 1224 | LowerOperation(SDValue Op, SelectionDAG &DAG) const |
| 1225 | { |
| 1226 | switch (Op.getOpcode()) |
| 1227 | { |
| 1228 | case ISD::BRCOND: return lowerBRCOND(Op, DAG); |
| 1229 | case ISD::ConstantPool: return lowerConstantPool(Op, DAG); |
| 1230 | case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); |
| 1231 | case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); |
| 1232 | case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); |
| 1233 | case ISD::JumpTable: return lowerJumpTable(Op, DAG); |
| 1234 | case ISD::SELECT: return lowerSELECT(Op, DAG); |
| 1235 | case ISD::SETCC: return lowerSETCC(Op, DAG); |
| 1236 | case ISD::STRICT_FSETCC: |
| 1237 | case ISD::STRICT_FSETCCS: |
| 1238 | return lowerFSETCC(Op, DAG); |
| 1239 | case ISD::VASTART: return lowerVASTART(Op, DAG); |
| 1240 | case ISD::VAARG: return lowerVAARG(Op, DAG); |
| 1241 | case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); |
| 1242 | case ISD::FABS: return lowerFABS(Op, DAG); |
| 1243 | case ISD::FCANONICALIZE: |
| 1244 | return lowerFCANONICALIZE(Op, DAG); |
| 1245 | case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); |
| 1246 | case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); |
| 1247 | case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); |
| 1248 | case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); |
| 1249 | case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); |
| 1250 | case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, IsSRA: true); |
| 1251 | case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, IsSRA: false); |
| 1252 | case ISD::LOAD: return lowerLOAD(Op, DAG); |
| 1253 | case ISD::STORE: return lowerSTORE(Op, DAG); |
| 1254 | case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG); |
| 1255 | case ISD::STRICT_FP_TO_SINT: |
| 1256 | case ISD::STRICT_FP_TO_UINT: |
| 1257 | return lowerSTRICT_FP_TO_INT(Op, DAG); |
| 1258 | case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); |
| 1259 | case ISD::READCYCLECOUNTER: |
| 1260 | return lowerREADCYCLECOUNTER(Op, DAG); |
| 1261 | } |
| 1262 | return SDValue(); |
| 1263 | } |
| 1264 | |
| 1265 | //===----------------------------------------------------------------------===// |
| 1266 | // Lower helper functions |
| 1267 | //===----------------------------------------------------------------------===// |
| 1268 | |
| 1269 | // addLiveIn - This helper function adds the specified physical register to the |
| 1270 | // MachineFunction as a live in value. It also creates a corresponding |
| 1271 | // virtual register for it. |
| 1272 | static unsigned |
| 1273 | addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) |
| 1274 | { |
| 1275 | Register VReg = MF.getRegInfo().createVirtualRegister(RegClass: RC); |
| 1276 | MF.getRegInfo().addLiveIn(Reg: PReg, vreg: VReg); |
| 1277 | return VReg; |
| 1278 | } |
| 1279 | |
| 1280 | static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, |
| 1281 | MachineBasicBlock &MBB, |
| 1282 | const TargetInstrInfo &TII, |
| 1283 | bool Is64Bit, bool IsMicroMips) { |
| 1284 | if (NoZeroDivCheck) |
| 1285 | return &MBB; |
| 1286 | |
| 1287 | // Insert instruction "teq $divisor_reg, $zero, 7". |
| 1288 | MachineBasicBlock::iterator I(MI); |
| 1289 | MachineInstrBuilder MIB; |
| 1290 | MachineOperand &Divisor = MI.getOperand(i: 2); |
| 1291 | MIB = BuildMI(BB&: MBB, I: std::next(x: I), MIMD: MI.getDebugLoc(), |
| 1292 | MCID: TII.get(Opcode: IsMicroMips ? Mips::TEQ_MM : Mips::TEQ)) |
| 1293 | .addReg(RegNo: Divisor.getReg(), Flags: getKillRegState(B: Divisor.isKill())) |
| 1294 | .addReg(RegNo: Mips::ZERO) |
| 1295 | .addImm(Val: 7); |
| 1296 | |
| 1297 | // Use the 32-bit sub-register if this is a 64-bit division. |
| 1298 | if (Is64Bit) |
| 1299 | MIB->getOperand(i: 0).setSubReg(Mips::sub_32); |
| 1300 | |
| 1301 | // Clear Divisor's kill flag. |
| 1302 | Divisor.setIsKill(false); |
| 1303 | |
| 1304 | // We would normally delete the original instruction here but in this case |
| 1305 | // we only needed to inject an additional instruction rather than replace it. |
| 1306 | |
| 1307 | return &MBB; |
| 1308 | } |
| 1309 | |
| 1310 | MachineBasicBlock * |
| 1311 | MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
| 1312 | MachineBasicBlock *BB) const { |
| 1313 | switch (MI.getOpcode()) { |
| 1314 | default: |
| 1315 | llvm_unreachable("Unexpected instr type to insert" ); |
| 1316 | case Mips::ATOMIC_LOAD_ADD_I8: |
| 1317 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1318 | case Mips::ATOMIC_LOAD_ADD_I16: |
| 1319 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1320 | case Mips::ATOMIC_LOAD_ADD_I32: |
| 1321 | return emitAtomicBinary(MI, BB); |
| 1322 | case Mips::ATOMIC_LOAD_ADD_I64: |
| 1323 | return emitAtomicBinary(MI, BB); |
| 1324 | |
| 1325 | case Mips::ATOMIC_LOAD_AND_I8: |
| 1326 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1327 | case Mips::ATOMIC_LOAD_AND_I16: |
| 1328 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1329 | case Mips::ATOMIC_LOAD_AND_I32: |
| 1330 | return emitAtomicBinary(MI, BB); |
| 1331 | case Mips::ATOMIC_LOAD_AND_I64: |
| 1332 | return emitAtomicBinary(MI, BB); |
| 1333 | |
| 1334 | case Mips::ATOMIC_LOAD_OR_I8: |
| 1335 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1336 | case Mips::ATOMIC_LOAD_OR_I16: |
| 1337 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1338 | case Mips::ATOMIC_LOAD_OR_I32: |
| 1339 | return emitAtomicBinary(MI, BB); |
| 1340 | case Mips::ATOMIC_LOAD_OR_I64: |
| 1341 | return emitAtomicBinary(MI, BB); |
| 1342 | |
| 1343 | case Mips::ATOMIC_LOAD_XOR_I8: |
| 1344 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1345 | case Mips::ATOMIC_LOAD_XOR_I16: |
| 1346 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1347 | case Mips::ATOMIC_LOAD_XOR_I32: |
| 1348 | return emitAtomicBinary(MI, BB); |
| 1349 | case Mips::ATOMIC_LOAD_XOR_I64: |
| 1350 | return emitAtomicBinary(MI, BB); |
| 1351 | |
| 1352 | case Mips::ATOMIC_LOAD_NAND_I8: |
| 1353 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1354 | case Mips::ATOMIC_LOAD_NAND_I16: |
| 1355 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1356 | case Mips::ATOMIC_LOAD_NAND_I32: |
| 1357 | return emitAtomicBinary(MI, BB); |
| 1358 | case Mips::ATOMIC_LOAD_NAND_I64: |
| 1359 | return emitAtomicBinary(MI, BB); |
| 1360 | |
| 1361 | case Mips::ATOMIC_LOAD_SUB_I8: |
| 1362 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1363 | case Mips::ATOMIC_LOAD_SUB_I16: |
| 1364 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1365 | case Mips::ATOMIC_LOAD_SUB_I32: |
| 1366 | return emitAtomicBinary(MI, BB); |
| 1367 | case Mips::ATOMIC_LOAD_SUB_I64: |
| 1368 | return emitAtomicBinary(MI, BB); |
| 1369 | |
| 1370 | case Mips::ATOMIC_SWAP_I8: |
| 1371 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1372 | case Mips::ATOMIC_SWAP_I16: |
| 1373 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1374 | case Mips::ATOMIC_SWAP_I32: |
| 1375 | return emitAtomicBinary(MI, BB); |
| 1376 | case Mips::ATOMIC_SWAP_I64: |
| 1377 | return emitAtomicBinary(MI, BB); |
| 1378 | |
| 1379 | case Mips::ATOMIC_CMP_SWAP_I8: |
| 1380 | return emitAtomicCmpSwapPartword(MI, BB, Size: 1); |
| 1381 | case Mips::ATOMIC_CMP_SWAP_I16: |
| 1382 | return emitAtomicCmpSwapPartword(MI, BB, Size: 2); |
| 1383 | case Mips::ATOMIC_CMP_SWAP_I32: |
| 1384 | return emitAtomicCmpSwap(MI, BB); |
| 1385 | case Mips::ATOMIC_CMP_SWAP_I64: |
| 1386 | return emitAtomicCmpSwap(MI, BB); |
| 1387 | |
| 1388 | case Mips::ATOMIC_LOAD_MIN_I8: |
| 1389 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1390 | case Mips::ATOMIC_LOAD_MIN_I16: |
| 1391 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1392 | case Mips::ATOMIC_LOAD_MIN_I32: |
| 1393 | return emitAtomicBinary(MI, BB); |
| 1394 | case Mips::ATOMIC_LOAD_MIN_I64: |
| 1395 | return emitAtomicBinary(MI, BB); |
| 1396 | |
| 1397 | case Mips::ATOMIC_LOAD_MAX_I8: |
| 1398 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1399 | case Mips::ATOMIC_LOAD_MAX_I16: |
| 1400 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1401 | case Mips::ATOMIC_LOAD_MAX_I32: |
| 1402 | return emitAtomicBinary(MI, BB); |
| 1403 | case Mips::ATOMIC_LOAD_MAX_I64: |
| 1404 | return emitAtomicBinary(MI, BB); |
| 1405 | |
| 1406 | case Mips::ATOMIC_LOAD_UMIN_I8: |
| 1407 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1408 | case Mips::ATOMIC_LOAD_UMIN_I16: |
| 1409 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1410 | case Mips::ATOMIC_LOAD_UMIN_I32: |
| 1411 | return emitAtomicBinary(MI, BB); |
| 1412 | case Mips::ATOMIC_LOAD_UMIN_I64: |
| 1413 | return emitAtomicBinary(MI, BB); |
| 1414 | |
| 1415 | case Mips::ATOMIC_LOAD_UMAX_I8: |
| 1416 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
| 1417 | case Mips::ATOMIC_LOAD_UMAX_I16: |
| 1418 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
| 1419 | case Mips::ATOMIC_LOAD_UMAX_I32: |
| 1420 | return emitAtomicBinary(MI, BB); |
| 1421 | case Mips::ATOMIC_LOAD_UMAX_I64: |
| 1422 | return emitAtomicBinary(MI, BB); |
| 1423 | |
| 1424 | case Mips::PseudoSDIV: |
| 1425 | case Mips::PseudoUDIV: |
| 1426 | case Mips::DIV: |
| 1427 | case Mips::DIVU: |
| 1428 | case Mips::MOD: |
| 1429 | case Mips::MODU: |
| 1430 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: false, |
| 1431 | IsMicroMips: false); |
| 1432 | case Mips::SDIV_MM_Pseudo: |
| 1433 | case Mips::UDIV_MM_Pseudo: |
| 1434 | case Mips::SDIV_MM: |
| 1435 | case Mips::UDIV_MM: |
| 1436 | case Mips::DIV_MMR6: |
| 1437 | case Mips::DIVU_MMR6: |
| 1438 | case Mips::MOD_MMR6: |
| 1439 | case Mips::MODU_MMR6: |
| 1440 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: false, IsMicroMips: true); |
| 1441 | case Mips::PseudoDSDIV: |
| 1442 | case Mips::PseudoDUDIV: |
| 1443 | case Mips::DDIV: |
| 1444 | case Mips::DDIVU: |
| 1445 | case Mips::DMOD: |
| 1446 | case Mips::DMODU: |
| 1447 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: true, IsMicroMips: false); |
| 1448 | |
| 1449 | case Mips::PseudoSELECT_I: |
| 1450 | case Mips::PseudoSELECT_I64: |
| 1451 | case Mips::PseudoSELECT_S: |
| 1452 | case Mips::PseudoSELECT_D32: |
| 1453 | case Mips::PseudoSELECT_D64: |
| 1454 | return emitPseudoSELECT(MI, BB, isFPCmp: false, Opc: Mips::BNE); |
| 1455 | case Mips::PseudoSELECTFP_F_I: |
| 1456 | case Mips::PseudoSELECTFP_F_I64: |
| 1457 | case Mips::PseudoSELECTFP_F_S: |
| 1458 | case Mips::PseudoSELECTFP_F_D32: |
| 1459 | case Mips::PseudoSELECTFP_F_D64: |
| 1460 | return emitPseudoSELECT(MI, BB, isFPCmp: true, Opc: Mips::BC1F); |
| 1461 | case Mips::PseudoSELECTFP_T_I: |
| 1462 | case Mips::PseudoSELECTFP_T_I64: |
| 1463 | case Mips::PseudoSELECTFP_T_S: |
| 1464 | case Mips::PseudoSELECTFP_T_D32: |
| 1465 | case Mips::PseudoSELECTFP_T_D64: |
| 1466 | return emitPseudoSELECT(MI, BB, isFPCmp: true, Opc: Mips::BC1T); |
| 1467 | case Mips::PseudoD_SELECT_I: |
| 1468 | case Mips::PseudoD_SELECT_I64: |
| 1469 | return emitPseudoD_SELECT(MI, BB); |
| 1470 | case Mips::LDR_W: |
| 1471 | return emitLDR_W(MI, BB); |
| 1472 | case Mips::LDR_D: |
| 1473 | return emitLDR_D(MI, BB); |
| 1474 | case Mips::STR_W: |
| 1475 | return emitSTR_W(MI, BB); |
| 1476 | case Mips::STR_D: |
| 1477 | return emitSTR_D(MI, BB); |
| 1478 | } |
| 1479 | } |
| 1480 | |
| 1481 | // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and |
| 1482 | // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) |
| 1483 | MachineBasicBlock * |
| 1484 | MipsTargetLowering::emitAtomicBinary(MachineInstr &MI, |
| 1485 | MachineBasicBlock *BB) const { |
| 1486 | |
| 1487 | MachineFunction *MF = BB->getParent(); |
| 1488 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
| 1489 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1490 | DebugLoc DL = MI.getDebugLoc(); |
| 1491 | |
| 1492 | unsigned AtomicOp; |
| 1493 | bool NeedsAdditionalReg = false; |
| 1494 | switch (MI.getOpcode()) { |
| 1495 | case Mips::ATOMIC_LOAD_ADD_I32: |
| 1496 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I32_POSTRA; |
| 1497 | break; |
| 1498 | case Mips::ATOMIC_LOAD_SUB_I32: |
| 1499 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I32_POSTRA; |
| 1500 | break; |
| 1501 | case Mips::ATOMIC_LOAD_AND_I32: |
| 1502 | AtomicOp = Mips::ATOMIC_LOAD_AND_I32_POSTRA; |
| 1503 | break; |
| 1504 | case Mips::ATOMIC_LOAD_OR_I32: |
| 1505 | AtomicOp = Mips::ATOMIC_LOAD_OR_I32_POSTRA; |
| 1506 | break; |
| 1507 | case Mips::ATOMIC_LOAD_XOR_I32: |
| 1508 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I32_POSTRA; |
| 1509 | break; |
| 1510 | case Mips::ATOMIC_LOAD_NAND_I32: |
| 1511 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I32_POSTRA; |
| 1512 | break; |
| 1513 | case Mips::ATOMIC_SWAP_I32: |
| 1514 | AtomicOp = Mips::ATOMIC_SWAP_I32_POSTRA; |
| 1515 | break; |
| 1516 | case Mips::ATOMIC_LOAD_ADD_I64: |
| 1517 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I64_POSTRA; |
| 1518 | break; |
| 1519 | case Mips::ATOMIC_LOAD_SUB_I64: |
| 1520 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I64_POSTRA; |
| 1521 | break; |
| 1522 | case Mips::ATOMIC_LOAD_AND_I64: |
| 1523 | AtomicOp = Mips::ATOMIC_LOAD_AND_I64_POSTRA; |
| 1524 | break; |
| 1525 | case Mips::ATOMIC_LOAD_OR_I64: |
| 1526 | AtomicOp = Mips::ATOMIC_LOAD_OR_I64_POSTRA; |
| 1527 | break; |
| 1528 | case Mips::ATOMIC_LOAD_XOR_I64: |
| 1529 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I64_POSTRA; |
| 1530 | break; |
| 1531 | case Mips::ATOMIC_LOAD_NAND_I64: |
| 1532 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I64_POSTRA; |
| 1533 | break; |
| 1534 | case Mips::ATOMIC_SWAP_I64: |
| 1535 | AtomicOp = Mips::ATOMIC_SWAP_I64_POSTRA; |
| 1536 | break; |
| 1537 | case Mips::ATOMIC_LOAD_MIN_I32: |
| 1538 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I32_POSTRA; |
| 1539 | NeedsAdditionalReg = true; |
| 1540 | break; |
| 1541 | case Mips::ATOMIC_LOAD_MAX_I32: |
| 1542 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I32_POSTRA; |
| 1543 | NeedsAdditionalReg = true; |
| 1544 | break; |
| 1545 | case Mips::ATOMIC_LOAD_UMIN_I32: |
| 1546 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I32_POSTRA; |
| 1547 | NeedsAdditionalReg = true; |
| 1548 | break; |
| 1549 | case Mips::ATOMIC_LOAD_UMAX_I32: |
| 1550 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I32_POSTRA; |
| 1551 | NeedsAdditionalReg = true; |
| 1552 | break; |
| 1553 | case Mips::ATOMIC_LOAD_MIN_I64: |
| 1554 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I64_POSTRA; |
| 1555 | NeedsAdditionalReg = true; |
| 1556 | break; |
| 1557 | case Mips::ATOMIC_LOAD_MAX_I64: |
| 1558 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I64_POSTRA; |
| 1559 | NeedsAdditionalReg = true; |
| 1560 | break; |
| 1561 | case Mips::ATOMIC_LOAD_UMIN_I64: |
| 1562 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I64_POSTRA; |
| 1563 | NeedsAdditionalReg = true; |
| 1564 | break; |
| 1565 | case Mips::ATOMIC_LOAD_UMAX_I64: |
| 1566 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I64_POSTRA; |
| 1567 | NeedsAdditionalReg = true; |
| 1568 | break; |
| 1569 | default: |
| 1570 | llvm_unreachable("Unknown pseudo atomic for replacement!" ); |
| 1571 | } |
| 1572 | |
| 1573 | Register OldVal = MI.getOperand(i: 0).getReg(); |
| 1574 | Register Ptr = MI.getOperand(i: 1).getReg(); |
| 1575 | Register Incr = MI.getOperand(i: 2).getReg(); |
| 1576 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: OldVal)); |
| 1577 | |
| 1578 | MachineBasicBlock::iterator II(MI); |
| 1579 | |
| 1580 | // The scratch registers here with the EarlyClobber | Define | Implicit |
| 1581 | // flags is used to persuade the register allocator and the machine |
| 1582 | // verifier to accept the usage of this register. This has to be a real |
| 1583 | // register which has an UNDEF value but is dead after the instruction which |
| 1584 | // is unique among the registers chosen for the instruction. |
| 1585 | |
| 1586 | // The EarlyClobber flag has the semantic properties that the operand it is |
| 1587 | // attached to is clobbered before the rest of the inputs are read. Hence it |
| 1588 | // must be unique among the operands to the instruction. |
| 1589 | // The Define flag is needed to coerce the machine verifier that an Undef |
| 1590 | // value isn't a problem. |
| 1591 | // The Dead flag is needed as the value in scratch isn't used by any other |
| 1592 | // instruction. Kill isn't used as Dead is more precise. |
| 1593 | // The implicit flag is here due to the interaction between the other flags |
| 1594 | // and the machine verifier. |
| 1595 | |
| 1596 | // For correctness purpose, a new pseudo is introduced here. We need this |
| 1597 | // new pseudo, so that FastRegisterAllocator does not see an ll/sc sequence |
| 1598 | // that is spread over >1 basic blocks. A register allocator which |
| 1599 | // introduces (or any codegen infact) a store, can violate the expectations |
| 1600 | // of the hardware. |
| 1601 | // |
| 1602 | // An atomic read-modify-write sequence starts with a linked load |
| 1603 | // instruction and ends with a store conditional instruction. The atomic |
| 1604 | // read-modify-write sequence fails if any of the following conditions |
| 1605 | // occur between the execution of ll and sc: |
| 1606 | // * A coherent store is completed by another process or coherent I/O |
| 1607 | // module into the block of synchronizable physical memory containing |
| 1608 | // the word. The size and alignment of the block is |
| 1609 | // implementation-dependent. |
| 1610 | // * A coherent store is executed between an LL and SC sequence on the |
| 1611 | // same processor to the block of synchornizable physical memory |
| 1612 | // containing the word. |
| 1613 | // |
| 1614 | |
| 1615 | Register PtrCopy = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: Ptr)); |
| 1616 | Register IncrCopy = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: Incr)); |
| 1617 | |
| 1618 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: IncrCopy).addReg(RegNo: Incr); |
| 1619 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: PtrCopy).addReg(RegNo: Ptr); |
| 1620 | |
| 1621 | MachineInstrBuilder MIB = |
| 1622 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
| 1623 | .addReg(RegNo: OldVal, Flags: RegState::Define | RegState::EarlyClobber) |
| 1624 | .addReg(RegNo: PtrCopy) |
| 1625 | .addReg(RegNo: IncrCopy) |
| 1626 | .addReg(RegNo: Scratch, Flags: RegState::Define | RegState::EarlyClobber | |
| 1627 | RegState::Implicit | RegState::Dead); |
| 1628 | if (NeedsAdditionalReg) { |
| 1629 | Register Scratch2 = |
| 1630 | RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: OldVal)); |
| 1631 | MIB.addReg(RegNo: Scratch2, Flags: RegState::Define | RegState::EarlyClobber | |
| 1632 | RegState::Implicit | RegState::Dead); |
| 1633 | } |
| 1634 | |
| 1635 | MI.eraseFromParent(); |
| 1636 | |
| 1637 | return BB; |
| 1638 | } |
| 1639 | |
| 1640 | MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( |
| 1641 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, |
| 1642 | unsigned SrcReg) const { |
| 1643 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1644 | const DebugLoc &DL = MI.getDebugLoc(); |
| 1645 | |
| 1646 | if (Subtarget.hasMips32r2() && Size == 1) { |
| 1647 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SEB), DestReg: DstReg).addReg(RegNo: SrcReg); |
| 1648 | return BB; |
| 1649 | } |
| 1650 | |
| 1651 | if (Subtarget.hasMips32r2() && Size == 2) { |
| 1652 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SEH), DestReg: DstReg).addReg(RegNo: SrcReg); |
| 1653 | return BB; |
| 1654 | } |
| 1655 | |
| 1656 | MachineFunction *MF = BB->getParent(); |
| 1657 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
| 1658 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
| 1659 | Register ScrReg = RegInfo.createVirtualRegister(RegClass: RC); |
| 1660 | |
| 1661 | assert(Size < 32); |
| 1662 | int64_t ShiftImm = 32 - (Size * 8); |
| 1663 | |
| 1664 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ScrReg).addReg(RegNo: SrcReg).addImm(Val: ShiftImm); |
| 1665 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SRA), DestReg: DstReg).addReg(RegNo: ScrReg).addImm(Val: ShiftImm); |
| 1666 | |
| 1667 | return BB; |
| 1668 | } |
| 1669 | |
| 1670 | MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( |
| 1671 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { |
| 1672 | assert((Size == 1 || Size == 2) && |
| 1673 | "Unsupported size for EmitAtomicBinaryPartial." ); |
| 1674 | |
| 1675 | MachineFunction *MF = BB->getParent(); |
| 1676 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
| 1677 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
| 1678 | const bool ArePtrs64bit = ABI.ArePtrs64bit(); |
| 1679 | const TargetRegisterClass *RCp = |
| 1680 | getRegClassFor(VT: ArePtrs64bit ? MVT::i64 : MVT::i32); |
| 1681 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1682 | DebugLoc DL = MI.getDebugLoc(); |
| 1683 | |
| 1684 | Register Dest = MI.getOperand(i: 0).getReg(); |
| 1685 | Register Ptr = MI.getOperand(i: 1).getReg(); |
| 1686 | Register Incr = MI.getOperand(i: 2).getReg(); |
| 1687 | |
| 1688 | Register AlignedAddr = RegInfo.createVirtualRegister(RegClass: RCp); |
| 1689 | Register ShiftAmt = RegInfo.createVirtualRegister(RegClass: RC); |
| 1690 | Register Mask = RegInfo.createVirtualRegister(RegClass: RC); |
| 1691 | Register Mask2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1692 | Register Incr2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1693 | Register MaskLSB2 = RegInfo.createVirtualRegister(RegClass: RCp); |
| 1694 | Register PtrLSB2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1695 | Register MaskUpper = RegInfo.createVirtualRegister(RegClass: RC); |
| 1696 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RC); |
| 1697 | Register Scratch2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1698 | Register Scratch3 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1699 | |
| 1700 | unsigned AtomicOp = 0; |
| 1701 | bool NeedsAdditionalReg = false; |
| 1702 | switch (MI.getOpcode()) { |
| 1703 | case Mips::ATOMIC_LOAD_NAND_I8: |
| 1704 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I8_POSTRA; |
| 1705 | break; |
| 1706 | case Mips::ATOMIC_LOAD_NAND_I16: |
| 1707 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I16_POSTRA; |
| 1708 | break; |
| 1709 | case Mips::ATOMIC_SWAP_I8: |
| 1710 | AtomicOp = Mips::ATOMIC_SWAP_I8_POSTRA; |
| 1711 | break; |
| 1712 | case Mips::ATOMIC_SWAP_I16: |
| 1713 | AtomicOp = Mips::ATOMIC_SWAP_I16_POSTRA; |
| 1714 | break; |
| 1715 | case Mips::ATOMIC_LOAD_ADD_I8: |
| 1716 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I8_POSTRA; |
| 1717 | break; |
| 1718 | case Mips::ATOMIC_LOAD_ADD_I16: |
| 1719 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I16_POSTRA; |
| 1720 | break; |
| 1721 | case Mips::ATOMIC_LOAD_SUB_I8: |
| 1722 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I8_POSTRA; |
| 1723 | break; |
| 1724 | case Mips::ATOMIC_LOAD_SUB_I16: |
| 1725 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I16_POSTRA; |
| 1726 | break; |
| 1727 | case Mips::ATOMIC_LOAD_AND_I8: |
| 1728 | AtomicOp = Mips::ATOMIC_LOAD_AND_I8_POSTRA; |
| 1729 | break; |
| 1730 | case Mips::ATOMIC_LOAD_AND_I16: |
| 1731 | AtomicOp = Mips::ATOMIC_LOAD_AND_I16_POSTRA; |
| 1732 | break; |
| 1733 | case Mips::ATOMIC_LOAD_OR_I8: |
| 1734 | AtomicOp = Mips::ATOMIC_LOAD_OR_I8_POSTRA; |
| 1735 | break; |
| 1736 | case Mips::ATOMIC_LOAD_OR_I16: |
| 1737 | AtomicOp = Mips::ATOMIC_LOAD_OR_I16_POSTRA; |
| 1738 | break; |
| 1739 | case Mips::ATOMIC_LOAD_XOR_I8: |
| 1740 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I8_POSTRA; |
| 1741 | break; |
| 1742 | case Mips::ATOMIC_LOAD_XOR_I16: |
| 1743 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I16_POSTRA; |
| 1744 | break; |
| 1745 | case Mips::ATOMIC_LOAD_MIN_I8: |
| 1746 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I8_POSTRA; |
| 1747 | NeedsAdditionalReg = true; |
| 1748 | break; |
| 1749 | case Mips::ATOMIC_LOAD_MIN_I16: |
| 1750 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I16_POSTRA; |
| 1751 | NeedsAdditionalReg = true; |
| 1752 | break; |
| 1753 | case Mips::ATOMIC_LOAD_MAX_I8: |
| 1754 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I8_POSTRA; |
| 1755 | NeedsAdditionalReg = true; |
| 1756 | break; |
| 1757 | case Mips::ATOMIC_LOAD_MAX_I16: |
| 1758 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I16_POSTRA; |
| 1759 | NeedsAdditionalReg = true; |
| 1760 | break; |
| 1761 | case Mips::ATOMIC_LOAD_UMIN_I8: |
| 1762 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I8_POSTRA; |
| 1763 | NeedsAdditionalReg = true; |
| 1764 | break; |
| 1765 | case Mips::ATOMIC_LOAD_UMIN_I16: |
| 1766 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I16_POSTRA; |
| 1767 | NeedsAdditionalReg = true; |
| 1768 | break; |
| 1769 | case Mips::ATOMIC_LOAD_UMAX_I8: |
| 1770 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I8_POSTRA; |
| 1771 | NeedsAdditionalReg = true; |
| 1772 | break; |
| 1773 | case Mips::ATOMIC_LOAD_UMAX_I16: |
| 1774 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I16_POSTRA; |
| 1775 | NeedsAdditionalReg = true; |
| 1776 | break; |
| 1777 | default: |
| 1778 | llvm_unreachable("Unknown subword atomic pseudo for expansion!" ); |
| 1779 | } |
| 1780 | |
| 1781 | // insert new blocks after the current block |
| 1782 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 1783 | MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
| 1784 | MachineFunction::iterator It = ++BB->getIterator(); |
| 1785 | MF->insert(MBBI: It, MBB: exitMBB); |
| 1786 | |
| 1787 | // Transfer the remainder of BB and its successor edges to exitMBB. |
| 1788 | exitMBB->splice(Where: exitMBB->begin(), Other: BB, |
| 1789 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 1790 | exitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 1791 | |
| 1792 | BB->addSuccessor(Succ: exitMBB, Prob: BranchProbability::getOne()); |
| 1793 | |
| 1794 | // thisMBB: |
| 1795 | // addiu masklsb2,$0,-4 # 0xfffffffc |
| 1796 | // and alignedaddr,ptr,masklsb2 |
| 1797 | // andi ptrlsb2,ptr,3 |
| 1798 | // sll shiftamt,ptrlsb2,3 |
| 1799 | // ori maskupper,$0,255 # 0xff |
| 1800 | // sll mask,maskupper,shiftamt |
| 1801 | // nor mask2,$0,mask |
| 1802 | // sll incr2,incr,shiftamt |
| 1803 | |
| 1804 | int64_t MaskImm = (Size == 1) ? 255 : 65535; |
| 1805 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ABI.GetPtrAddiuOp()), DestReg: MaskLSB2) |
| 1806 | .addReg(RegNo: ABI.GetNullPtr()).addImm(Val: -4); |
| 1807 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ABI.GetPtrAndOp()), DestReg: AlignedAddr) |
| 1808 | .addReg(RegNo: Ptr).addReg(RegNo: MaskLSB2); |
| 1809 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: PtrLSB2) |
| 1810 | .addReg(RegNo: Ptr, Flags: {}, SubReg: ArePtrs64bit ? Mips::sub_32 : 0) |
| 1811 | .addImm(Val: 3); |
| 1812 | if (Subtarget.isLittle()) { |
| 1813 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: PtrLSB2).addImm(Val: 3); |
| 1814 | } else { |
| 1815 | Register Off = RegInfo.createVirtualRegister(RegClass: RC); |
| 1816 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::XORi), DestReg: Off) |
| 1817 | .addReg(RegNo: PtrLSB2).addImm(Val: (Size == 1) ? 3 : 2); |
| 1818 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: Off).addImm(Val: 3); |
| 1819 | } |
| 1820 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ORi), DestReg: MaskUpper) |
| 1821 | .addReg(RegNo: Mips::ZERO).addImm(Val: MaskImm); |
| 1822 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Mask) |
| 1823 | .addReg(RegNo: MaskUpper).addReg(RegNo: ShiftAmt); |
| 1824 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::NOR), DestReg: Mask2).addReg(RegNo: Mips::ZERO).addReg(RegNo: Mask); |
| 1825 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Incr2).addReg(RegNo: Incr).addReg(RegNo: ShiftAmt); |
| 1826 | |
| 1827 | |
| 1828 | // The purposes of the flags on the scratch registers is explained in |
| 1829 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
| 1830 | // be undef, that is unique among registers chosen for the instruction. |
| 1831 | |
| 1832 | MachineInstrBuilder MIB = |
| 1833 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
| 1834 | .addReg(RegNo: Dest, Flags: RegState::Define | RegState::EarlyClobber) |
| 1835 | .addReg(RegNo: AlignedAddr) |
| 1836 | .addReg(RegNo: Incr2) |
| 1837 | .addReg(RegNo: Mask) |
| 1838 | .addReg(RegNo: Mask2) |
| 1839 | .addReg(RegNo: ShiftAmt) |
| 1840 | .addReg(RegNo: Scratch, Flags: RegState::EarlyClobber | RegState::Define | |
| 1841 | RegState::Dead | RegState::Implicit) |
| 1842 | .addReg(RegNo: Scratch2, Flags: RegState::EarlyClobber | RegState::Define | |
| 1843 | RegState::Dead | RegState::Implicit) |
| 1844 | .addReg(RegNo: Scratch3, Flags: RegState::EarlyClobber | RegState::Define | |
| 1845 | RegState::Dead | RegState::Implicit); |
| 1846 | if (NeedsAdditionalReg) { |
| 1847 | Register Scratch4 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1848 | MIB.addReg(RegNo: Scratch4, Flags: RegState::EarlyClobber | RegState::Define | |
| 1849 | RegState::Dead | RegState::Implicit); |
| 1850 | } |
| 1851 | |
| 1852 | MI.eraseFromParent(); // The instruction is gone now. |
| 1853 | |
| 1854 | return exitMBB; |
| 1855 | } |
| 1856 | |
| 1857 | // Lower atomic compare and swap to a pseudo instruction, taking care to |
| 1858 | // define a scratch register for the pseudo instruction's expansion. The |
| 1859 | // instruction is expanded after the register allocator as to prevent |
| 1860 | // the insertion of stores between the linked load and the store conditional. |
| 1861 | |
| 1862 | MachineBasicBlock * |
| 1863 | MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI, |
| 1864 | MachineBasicBlock *BB) const { |
| 1865 | |
| 1866 | assert((MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 || |
| 1867 | MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I64) && |
| 1868 | "Unsupported atomic pseudo for EmitAtomicCmpSwap." ); |
| 1869 | |
| 1870 | const unsigned Size = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ? 4 : 8; |
| 1871 | |
| 1872 | MachineFunction *MF = BB->getParent(); |
| 1873 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 1874 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::getIntegerVT(BitWidth: Size * 8)); |
| 1875 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1876 | DebugLoc DL = MI.getDebugLoc(); |
| 1877 | |
| 1878 | unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 |
| 1879 | ? Mips::ATOMIC_CMP_SWAP_I32_POSTRA |
| 1880 | : Mips::ATOMIC_CMP_SWAP_I64_POSTRA; |
| 1881 | Register Dest = MI.getOperand(i: 0).getReg(); |
| 1882 | Register Ptr = MI.getOperand(i: 1).getReg(); |
| 1883 | Register OldVal = MI.getOperand(i: 2).getReg(); |
| 1884 | Register NewVal = MI.getOperand(i: 3).getReg(); |
| 1885 | |
| 1886 | Register Scratch = MRI.createVirtualRegister(RegClass: RC); |
| 1887 | MachineBasicBlock::iterator II(MI); |
| 1888 | |
| 1889 | // We need to create copies of the various registers and kill them at the |
| 1890 | // atomic pseudo. If the copies are not made, when the atomic is expanded |
| 1891 | // after fast register allocation, the spills will end up outside of the |
| 1892 | // blocks that their values are defined in, causing livein errors. |
| 1893 | |
| 1894 | Register PtrCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: Ptr)); |
| 1895 | Register OldValCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OldVal)); |
| 1896 | Register NewValCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: NewVal)); |
| 1897 | |
| 1898 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: PtrCopy).addReg(RegNo: Ptr); |
| 1899 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: OldValCopy).addReg(RegNo: OldVal); |
| 1900 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: NewValCopy).addReg(RegNo: NewVal); |
| 1901 | |
| 1902 | // The purposes of the flags on the scratch registers is explained in |
| 1903 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
| 1904 | // be undef, that is unique among registers chosen for the instruction. |
| 1905 | |
| 1906 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
| 1907 | .addReg(RegNo: Dest, Flags: RegState::Define | RegState::EarlyClobber) |
| 1908 | .addReg(RegNo: PtrCopy, Flags: RegState::Kill) |
| 1909 | .addReg(RegNo: OldValCopy, Flags: RegState::Kill) |
| 1910 | .addReg(RegNo: NewValCopy, Flags: RegState::Kill) |
| 1911 | .addReg(RegNo: Scratch, Flags: RegState::EarlyClobber | RegState::Define | |
| 1912 | RegState::Dead | RegState::Implicit); |
| 1913 | |
| 1914 | MI.eraseFromParent(); // The instruction is gone now. |
| 1915 | |
| 1916 | return BB; |
| 1917 | } |
| 1918 | |
| 1919 | MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword( |
| 1920 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { |
| 1921 | assert((Size == 1 || Size == 2) && |
| 1922 | "Unsupported size for EmitAtomicCmpSwapPartial." ); |
| 1923 | |
| 1924 | MachineFunction *MF = BB->getParent(); |
| 1925 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
| 1926 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
| 1927 | const bool ArePtrs64bit = ABI.ArePtrs64bit(); |
| 1928 | const TargetRegisterClass *RCp = |
| 1929 | getRegClassFor(VT: ArePtrs64bit ? MVT::i64 : MVT::i32); |
| 1930 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 1931 | DebugLoc DL = MI.getDebugLoc(); |
| 1932 | |
| 1933 | Register Dest = MI.getOperand(i: 0).getReg(); |
| 1934 | Register Ptr = MI.getOperand(i: 1).getReg(); |
| 1935 | Register CmpVal = MI.getOperand(i: 2).getReg(); |
| 1936 | Register NewVal = MI.getOperand(i: 3).getReg(); |
| 1937 | |
| 1938 | Register AlignedAddr = RegInfo.createVirtualRegister(RegClass: RCp); |
| 1939 | Register ShiftAmt = RegInfo.createVirtualRegister(RegClass: RC); |
| 1940 | Register Mask = RegInfo.createVirtualRegister(RegClass: RC); |
| 1941 | Register Mask2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1942 | Register ShiftedCmpVal = RegInfo.createVirtualRegister(RegClass: RC); |
| 1943 | Register ShiftedNewVal = RegInfo.createVirtualRegister(RegClass: RC); |
| 1944 | Register MaskLSB2 = RegInfo.createVirtualRegister(RegClass: RCp); |
| 1945 | Register PtrLSB2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1946 | Register MaskUpper = RegInfo.createVirtualRegister(RegClass: RC); |
| 1947 | Register MaskedCmpVal = RegInfo.createVirtualRegister(RegClass: RC); |
| 1948 | Register MaskedNewVal = RegInfo.createVirtualRegister(RegClass: RC); |
| 1949 | unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I8 |
| 1950 | ? Mips::ATOMIC_CMP_SWAP_I8_POSTRA |
| 1951 | : Mips::ATOMIC_CMP_SWAP_I16_POSTRA; |
| 1952 | |
| 1953 | // The scratch registers here with the EarlyClobber | Define | Dead | Implicit |
| 1954 | // flags are used to coerce the register allocator and the machine verifier to |
| 1955 | // accept the usage of these registers. |
| 1956 | // The EarlyClobber flag has the semantic properties that the operand it is |
| 1957 | // attached to is clobbered before the rest of the inputs are read. Hence it |
| 1958 | // must be unique among the operands to the instruction. |
| 1959 | // The Define flag is needed to coerce the machine verifier that an Undef |
| 1960 | // value isn't a problem. |
| 1961 | // The Dead flag is needed as the value in scratch isn't used by any other |
| 1962 | // instruction. Kill isn't used as Dead is more precise. |
| 1963 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RC); |
| 1964 | Register Scratch2 = RegInfo.createVirtualRegister(RegClass: RC); |
| 1965 | |
| 1966 | // insert new blocks after the current block |
| 1967 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 1968 | MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
| 1969 | MachineFunction::iterator It = ++BB->getIterator(); |
| 1970 | MF->insert(MBBI: It, MBB: exitMBB); |
| 1971 | |
| 1972 | // Transfer the remainder of BB and its successor edges to exitMBB. |
| 1973 | exitMBB->splice(Where: exitMBB->begin(), Other: BB, |
| 1974 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 1975 | exitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 1976 | |
| 1977 | BB->addSuccessor(Succ: exitMBB, Prob: BranchProbability::getOne()); |
| 1978 | |
| 1979 | // thisMBB: |
| 1980 | // addiu masklsb2,$0,-4 # 0xfffffffc |
| 1981 | // and alignedaddr,ptr,masklsb2 |
| 1982 | // andi ptrlsb2,ptr,3 |
| 1983 | // xori ptrlsb2,ptrlsb2,3 # Only for BE |
| 1984 | // sll shiftamt,ptrlsb2,3 |
| 1985 | // ori maskupper,$0,255 # 0xff |
| 1986 | // sll mask,maskupper,shiftamt |
| 1987 | // nor mask2,$0,mask |
| 1988 | // andi maskedcmpval,cmpval,255 |
| 1989 | // sll shiftedcmpval,maskedcmpval,shiftamt |
| 1990 | // andi maskednewval,newval,255 |
| 1991 | // sll shiftednewval,maskednewval,shiftamt |
| 1992 | int64_t MaskImm = (Size == 1) ? 255 : 65535; |
| 1993 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), DestReg: MaskLSB2) |
| 1994 | .addReg(RegNo: ABI.GetNullPtr()).addImm(Val: -4); |
| 1995 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ArePtrs64bit ? Mips::AND64 : Mips::AND), DestReg: AlignedAddr) |
| 1996 | .addReg(RegNo: Ptr).addReg(RegNo: MaskLSB2); |
| 1997 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: PtrLSB2) |
| 1998 | .addReg(RegNo: Ptr, Flags: {}, SubReg: ArePtrs64bit ? Mips::sub_32 : 0) |
| 1999 | .addImm(Val: 3); |
| 2000 | if (Subtarget.isLittle()) { |
| 2001 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: PtrLSB2).addImm(Val: 3); |
| 2002 | } else { |
| 2003 | Register Off = RegInfo.createVirtualRegister(RegClass: RC); |
| 2004 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::XORi), DestReg: Off) |
| 2005 | .addReg(RegNo: PtrLSB2).addImm(Val: (Size == 1) ? 3 : 2); |
| 2006 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: Off).addImm(Val: 3); |
| 2007 | } |
| 2008 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ORi), DestReg: MaskUpper) |
| 2009 | .addReg(RegNo: Mips::ZERO).addImm(Val: MaskImm); |
| 2010 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Mask) |
| 2011 | .addReg(RegNo: MaskUpper).addReg(RegNo: ShiftAmt); |
| 2012 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::NOR), DestReg: Mask2).addReg(RegNo: Mips::ZERO).addReg(RegNo: Mask); |
| 2013 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: MaskedCmpVal) |
| 2014 | .addReg(RegNo: CmpVal).addImm(Val: MaskImm); |
| 2015 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: ShiftedCmpVal) |
| 2016 | .addReg(RegNo: MaskedCmpVal).addReg(RegNo: ShiftAmt); |
| 2017 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: MaskedNewVal) |
| 2018 | .addReg(RegNo: NewVal).addImm(Val: MaskImm); |
| 2019 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: ShiftedNewVal) |
| 2020 | .addReg(RegNo: MaskedNewVal).addReg(RegNo: ShiftAmt); |
| 2021 | |
| 2022 | // The purposes of the flags on the scratch registers are explained in |
| 2023 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
| 2024 | // be undef, that is unique among the register chosen for the instruction. |
| 2025 | |
| 2026 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
| 2027 | .addReg(RegNo: Dest, Flags: RegState::Define | RegState::EarlyClobber) |
| 2028 | .addReg(RegNo: AlignedAddr) |
| 2029 | .addReg(RegNo: Mask) |
| 2030 | .addReg(RegNo: ShiftedCmpVal) |
| 2031 | .addReg(RegNo: Mask2) |
| 2032 | .addReg(RegNo: ShiftedNewVal) |
| 2033 | .addReg(RegNo: ShiftAmt) |
| 2034 | .addReg(RegNo: Scratch, Flags: RegState::EarlyClobber | RegState::Define | |
| 2035 | RegState::Dead | RegState::Implicit) |
| 2036 | .addReg(RegNo: Scratch2, Flags: RegState::EarlyClobber | RegState::Define | |
| 2037 | RegState::Dead | RegState::Implicit); |
| 2038 | |
| 2039 | MI.eraseFromParent(); // The instruction is gone now. |
| 2040 | |
| 2041 | return exitMBB; |
| 2042 | } |
| 2043 | |
| 2044 | SDValue MipsTargetLowering::lowerREADCYCLECOUNTER(SDValue Op, |
| 2045 | SelectionDAG &DAG) const { |
| 2046 | SmallVector<SDValue, 3> Results; |
| 2047 | SDLoc DL(Op); |
| 2048 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2049 | unsigned RdhwrOpc, DestReg; |
| 2050 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2051 | |
| 2052 | if (PtrVT == MVT::i64) { |
| 2053 | RdhwrOpc = Mips::RDHWR64; |
| 2054 | DestReg = MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: MVT::i64)); |
| 2055 | SDNode *Rdhwr = DAG.getMachineNode(Opcode: RdhwrOpc, dl: DL, VT1: MVT::i64, VT2: MVT::Glue, |
| 2056 | Op1: DAG.getRegister(Reg: Mips::HWR2, VT: MVT::i32), |
| 2057 | Op2: DAG.getTargetConstant(Val: 0, DL, VT: MVT::i32)); |
| 2058 | SDValue Chain = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg: DestReg, |
| 2059 | N: SDValue(Rdhwr, 0), Glue: SDValue(Rdhwr, 1)); |
| 2060 | SDValue ResNode = |
| 2061 | DAG.getCopyFromReg(Chain, dl: DL, Reg: DestReg, VT: MVT::i64, Glue: Chain.getValue(R: 1)); |
| 2062 | Results.push_back(Elt: ResNode); |
| 2063 | Results.push_back(Elt: ResNode.getValue(R: 1)); |
| 2064 | } else { |
| 2065 | RdhwrOpc = Mips::RDHWR; |
| 2066 | DestReg = MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: MVT::i32)); |
| 2067 | SDNode *Rdhwr = DAG.getMachineNode(Opcode: RdhwrOpc, dl: DL, VT1: MVT::i32, VT2: MVT::Glue, |
| 2068 | Op1: DAG.getRegister(Reg: Mips::HWR2, VT: MVT::i32), |
| 2069 | Op2: DAG.getTargetConstant(Val: 0, DL, VT: MVT::i32)); |
| 2070 | SDValue Chain = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg: DestReg, |
| 2071 | N: SDValue(Rdhwr, 0), Glue: SDValue(Rdhwr, 1)); |
| 2072 | SDValue ResNode = |
| 2073 | DAG.getCopyFromReg(Chain, dl: DL, Reg: DestReg, VT: MVT::i32, Glue: Chain.getValue(R: 1)); |
| 2074 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VT: MVT::i64, N1: ResNode, |
| 2075 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32))); |
| 2076 | Results.push_back(Elt: ResNode.getValue(R: 1)); |
| 2077 | } |
| 2078 | |
| 2079 | return DAG.getMergeValues(Ops: Results, dl: DL); |
| 2080 | } |
| 2081 | |
| 2082 | SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { |
| 2083 | // The first operand is the chain, the second is the condition, the third is |
| 2084 | // the block to branch to if the condition is true. |
| 2085 | SDValue Chain = Op.getOperand(i: 0); |
| 2086 | SDValue Dest = Op.getOperand(i: 2); |
| 2087 | SDLoc DL(Op); |
| 2088 | |
| 2089 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
| 2090 | SDValue CondRes = createFPCmp(DAG, Op: Op.getOperand(i: 1)); |
| 2091 | |
| 2092 | // Return if flag is not set by a floating point comparison. |
| 2093 | if (CondRes.getOpcode() != MipsISD::FPCmp) |
| 2094 | return Op; |
| 2095 | |
| 2096 | SDValue CCNode = CondRes.getOperand(i: 2); |
| 2097 | Mips::CondCode CC = (Mips::CondCode)CCNode->getAsZExtVal(); |
| 2098 | unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; |
| 2099 | SDValue BrCode = DAG.getConstant(Val: Opc, DL, VT: MVT::i32); |
| 2100 | SDValue FCC0 = DAG.getRegister(Reg: Mips::FCC0, VT: MVT::i32); |
| 2101 | return DAG.getNode(Opcode: MipsISD::FPBrcond, DL, VT: Op.getValueType(), N1: Chain, N2: BrCode, |
| 2102 | N3: FCC0, N4: Dest, N5: CondRes); |
| 2103 | } |
| 2104 | |
| 2105 | SDValue MipsTargetLowering:: |
| 2106 | lowerSELECT(SDValue Op, SelectionDAG &DAG) const |
| 2107 | { |
| 2108 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
| 2109 | SDValue Cond = createFPCmp(DAG, Op: Op.getOperand(i: 0)); |
| 2110 | |
| 2111 | // Return if flag is not set by a floating point comparison. |
| 2112 | if (Cond.getOpcode() != MipsISD::FPCmp) |
| 2113 | return Op; |
| 2114 | |
| 2115 | return createCMovFP(DAG, Cond, True: Op.getOperand(i: 1), False: Op.getOperand(i: 2), |
| 2116 | DL: SDLoc(Op)); |
| 2117 | } |
| 2118 | |
| 2119 | SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
| 2120 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
| 2121 | SDValue Cond = createFPCmp(DAG, Op); |
| 2122 | |
| 2123 | assert(Cond.getOpcode() == MipsISD::FPCmp && |
| 2124 | "Floating point operand expected." ); |
| 2125 | |
| 2126 | SDLoc DL(Op); |
| 2127 | SDValue True = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2128 | SDValue False = DAG.getConstant(Val: 0, DL, VT: MVT::i32); |
| 2129 | |
| 2130 | return createCMovFP(DAG, Cond, True, False, DL); |
| 2131 | } |
| 2132 | |
| 2133 | SDValue MipsTargetLowering::lowerFSETCC(SDValue Op, SelectionDAG &DAG) const { |
| 2134 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
| 2135 | |
| 2136 | SDLoc DL(Op); |
| 2137 | SDValue Chain = Op.getOperand(i: 0); |
| 2138 | SDValue LHS = Op.getOperand(i: 1); |
| 2139 | SDValue RHS = Op.getOperand(i: 2); |
| 2140 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 3))->get(); |
| 2141 | |
| 2142 | SDValue Cond = DAG.getNode(Opcode: MipsISD::FPCmp, DL, VT: MVT::Glue, N1: LHS, N2: RHS, |
| 2143 | N3: DAG.getConstant(Val: condCodeToFCC(CC), DL, VT: MVT::i32)); |
| 2144 | SDValue True = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2145 | SDValue False = DAG.getConstant(Val: 0, DL, VT: MVT::i32); |
| 2146 | SDValue CMovFP = createCMovFP(DAG, Cond, True, False, DL); |
| 2147 | |
| 2148 | return DAG.getMergeValues(Ops: {CMovFP, Chain}, dl: DL); |
| 2149 | } |
| 2150 | |
| 2151 | SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, |
| 2152 | SelectionDAG &DAG) const { |
| 2153 | EVT Ty = Op.getValueType(); |
| 2154 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Val&: Op); |
| 2155 | const GlobalValue *GV = N->getGlobal(); |
| 2156 | |
| 2157 | if (GV->hasDLLImportStorageClass()) { |
| 2158 | assert(Subtarget.isTargetWindows() && |
| 2159 | "Windows is the only supported COFF target" ); |
| 2160 | return getDllimportVariable( |
| 2161 | N, DL: SDLoc(N), Ty, DAG, Chain: DAG.getEntryNode(), |
| 2162 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
| 2163 | } |
| 2164 | |
| 2165 | if (!isPositionIndependent()) { |
| 2166 | const MipsTargetObjectFile *TLOF = |
| 2167 | static_cast<const MipsTargetObjectFile *>( |
| 2168 | getTargetMachine().getObjFileLowering()); |
| 2169 | const GlobalObject *GO = GV->getAliaseeObject(); |
| 2170 | if (GO && TLOF->IsGlobalInSmallSection(GO, TM: getTargetMachine())) |
| 2171 | // %gp_rel relocation |
| 2172 | return getAddrGPRel(N, DL: SDLoc(N), Ty, DAG, IsN64: ABI.IsN64()); |
| 2173 | |
| 2174 | // %hi/%lo relocation |
| 2175 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 2176 | // %highest/%higher/%hi/%lo relocation |
| 2177 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 2178 | } |
| 2179 | |
| 2180 | // Every other architecture would use shouldAssumeDSOLocal in here, but |
| 2181 | // mips is special. |
| 2182 | // * In PIC code mips requires got loads even for local statics! |
| 2183 | // * To save on got entries, for local statics the got entry contains the |
| 2184 | // page and an additional add instruction takes care of the low bits. |
| 2185 | // * It is legal to access a hidden symbol with a non hidden undefined, |
| 2186 | // so one cannot guarantee that all access to a hidden symbol will know |
| 2187 | // it is hidden. |
| 2188 | // * Mips linkers don't support creating a page and a full got entry for |
| 2189 | // the same symbol. |
| 2190 | // * Given all that, we have to use a full got entry for hidden symbols :-( |
| 2191 | if (GV->hasLocalLinkage()) |
| 2192 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
| 2193 | |
| 2194 | if (Subtarget.useXGOT()) |
| 2195 | return getAddrGlobalLargeGOT( |
| 2196 | N, DL: SDLoc(N), Ty, DAG, HiFlag: MipsII::MO_GOT_HI16, LoFlag: MipsII::MO_GOT_LO16, |
| 2197 | Chain: DAG.getEntryNode(), |
| 2198 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
| 2199 | |
| 2200 | return getAddrGlobal( |
| 2201 | N, DL: SDLoc(N), Ty, DAG, |
| 2202 | Flag: (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT, |
| 2203 | Chain: DAG.getEntryNode(), PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
| 2204 | } |
| 2205 | |
| 2206 | SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, |
| 2207 | SelectionDAG &DAG) const { |
| 2208 | BlockAddressSDNode *N = cast<BlockAddressSDNode>(Val&: Op); |
| 2209 | EVT Ty = Op.getValueType(); |
| 2210 | |
| 2211 | if (!isPositionIndependent()) |
| 2212 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 2213 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 2214 | |
| 2215 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
| 2216 | } |
| 2217 | |
| 2218 | SDValue MipsTargetLowering:: |
| 2219 | lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const |
| 2220 | { |
| 2221 | // If the relocation model is PIC, use the General Dynamic TLS Model or |
| 2222 | // Local Dynamic TLS model, otherwise use the Initial Exec or |
| 2223 | // Local Exec TLS Model. |
| 2224 | |
| 2225 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2226 | if (DAG.getTarget().useEmulatedTLS()) |
| 2227 | return LowerToTLSEmulatedModel(GA, DAG); |
| 2228 | |
| 2229 | SDLoc DL(GA); |
| 2230 | const GlobalValue *GV = GA->getGlobal(); |
| 2231 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2232 | |
| 2233 | TLSModel::Model model = getTargetMachine().getTLSModel(GV); |
| 2234 | |
| 2235 | if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { |
| 2236 | // General Dynamic and Local Dynamic TLS Model. |
| 2237 | unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM |
| 2238 | : MipsII::MO_TLSGD; |
| 2239 | |
| 2240 | SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, TargetFlags: Flag); |
| 2241 | SDValue Argument = DAG.getNode(Opcode: MipsISD::Wrapper, DL, VT: PtrVT, |
| 2242 | N1: getGlobalReg(DAG, Ty: PtrVT), N2: TGA); |
| 2243 | unsigned PtrSize = PtrVT.getSizeInBits(); |
| 2244 | IntegerType *PtrTy = Type::getIntNTy(C&: *DAG.getContext(), N: PtrSize); |
| 2245 | |
| 2246 | SDValue TlsGetAddr = DAG.getExternalSymbol(Sym: "__tls_get_addr" , VT: PtrVT); |
| 2247 | |
| 2248 | ArgListTy Args; |
| 2249 | Args.emplace_back(args&: Argument, args&: PtrTy); |
| 2250 | |
| 2251 | TargetLowering::CallLoweringInfo CLI(DAG); |
| 2252 | CLI.setDebugLoc(DL) |
| 2253 | .setChain(DAG.getEntryNode()) |
| 2254 | .setLibCallee(CC: CallingConv::C, ResultType: PtrTy, Target: TlsGetAddr, ArgsList: std::move(Args)); |
| 2255 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
| 2256 | |
| 2257 | SDValue Ret = CallResult.first; |
| 2258 | |
| 2259 | if (model != TLSModel::LocalDynamic) |
| 2260 | return Ret; |
| 2261 | |
| 2262 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
| 2263 | TargetFlags: MipsII::MO_DTPREL_HI); |
| 2264 | SDValue Hi = DAG.getNode(Opcode: MipsISD::TlsHi, DL, VT: PtrVT, Operand: TGAHi); |
| 2265 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
| 2266 | TargetFlags: MipsII::MO_DTPREL_LO); |
| 2267 | SDValue Lo = DAG.getNode(Opcode: MipsISD::Lo, DL, VT: PtrVT, Operand: TGALo); |
| 2268 | SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Hi, N2: Ret); |
| 2269 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Add, N2: Lo); |
| 2270 | } |
| 2271 | |
| 2272 | SDValue Offset; |
| 2273 | if (model == TLSModel::InitialExec) { |
| 2274 | // Initial Exec TLS Model |
| 2275 | SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
| 2276 | TargetFlags: MipsII::MO_GOTTPREL); |
| 2277 | TGA = DAG.getNode(Opcode: MipsISD::Wrapper, DL, VT: PtrVT, N1: getGlobalReg(DAG, Ty: PtrVT), |
| 2278 | N2: TGA); |
| 2279 | Offset = |
| 2280 | DAG.getLoad(VT: PtrVT, dl: DL, Chain: DAG.getEntryNode(), Ptr: TGA, PtrInfo: MachinePointerInfo()); |
| 2281 | } else { |
| 2282 | // Local Exec TLS Model |
| 2283 | assert(model == TLSModel::LocalExec); |
| 2284 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
| 2285 | TargetFlags: MipsII::MO_TPREL_HI); |
| 2286 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
| 2287 | TargetFlags: MipsII::MO_TPREL_LO); |
| 2288 | SDValue Hi = DAG.getNode(Opcode: MipsISD::TlsHi, DL, VT: PtrVT, Operand: TGAHi); |
| 2289 | SDValue Lo = DAG.getNode(Opcode: MipsISD::Lo, DL, VT: PtrVT, Operand: TGALo); |
| 2290 | Offset = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Hi, N2: Lo); |
| 2291 | } |
| 2292 | |
| 2293 | SDValue ThreadPointer = DAG.getNode(Opcode: MipsISD::ThreadPointer, DL, VT: PtrVT); |
| 2294 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: ThreadPointer, N2: Offset); |
| 2295 | } |
| 2296 | |
| 2297 | SDValue MipsTargetLowering:: |
| 2298 | lowerJumpTable(SDValue Op, SelectionDAG &DAG) const |
| 2299 | { |
| 2300 | JumpTableSDNode *N = cast<JumpTableSDNode>(Val&: Op); |
| 2301 | EVT Ty = Op.getValueType(); |
| 2302 | |
| 2303 | if (!isPositionIndependent()) |
| 2304 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 2305 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 2306 | |
| 2307 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
| 2308 | } |
| 2309 | |
| 2310 | SDValue MipsTargetLowering:: |
| 2311 | lowerConstantPool(SDValue Op, SelectionDAG &DAG) const |
| 2312 | { |
| 2313 | ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Val&: Op); |
| 2314 | EVT Ty = Op.getValueType(); |
| 2315 | |
| 2316 | if (!isPositionIndependent()) { |
| 2317 | const MipsTargetObjectFile *TLOF = |
| 2318 | static_cast<const MipsTargetObjectFile *>( |
| 2319 | getTargetMachine().getObjFileLowering()); |
| 2320 | |
| 2321 | if (TLOF->IsConstantInSmallSection(DL: DAG.getDataLayout(), CN: N->getConstVal(), |
| 2322 | TM: getTargetMachine())) |
| 2323 | // %gp_rel relocation |
| 2324 | return getAddrGPRel(N, DL: SDLoc(N), Ty, DAG, IsN64: ABI.IsN64()); |
| 2325 | |
| 2326 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 2327 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 2328 | } |
| 2329 | |
| 2330 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
| 2331 | } |
| 2332 | |
| 2333 | SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
| 2334 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2335 | MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); |
| 2336 | |
| 2337 | SDLoc DL(Op); |
| 2338 | SDValue FI = DAG.getFrameIndex(FI: FuncInfo->getVarArgsFrameIndex(), |
| 2339 | VT: getPointerTy(DL: MF.getDataLayout())); |
| 2340 | |
| 2341 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
| 2342 | // memory location argument. |
| 2343 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
| 2344 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: FI, Ptr: Op.getOperand(i: 1), |
| 2345 | PtrInfo: MachinePointerInfo(SV)); |
| 2346 | } |
| 2347 | |
| 2348 | SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { |
| 2349 | SDNode *Node = Op.getNode(); |
| 2350 | EVT VT = Node->getValueType(ResNo: 0); |
| 2351 | SDValue Chain = Node->getOperand(Num: 0); |
| 2352 | SDValue VAListPtr = Node->getOperand(Num: 1); |
| 2353 | const Align Align = |
| 2354 | llvm::MaybeAlign(Node->getConstantOperandVal(Num: 3)).valueOrOne(); |
| 2355 | const Value *SV = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue(); |
| 2356 | SDLoc DL(Node); |
| 2357 | unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4; |
| 2358 | |
| 2359 | SDValue VAListLoad = DAG.getLoad(VT: getPointerTy(DL: DAG.getDataLayout()), dl: DL, Chain, |
| 2360 | Ptr: VAListPtr, PtrInfo: MachinePointerInfo(SV)); |
| 2361 | SDValue VAList = VAListLoad; |
| 2362 | |
| 2363 | // Re-align the pointer if necessary. |
| 2364 | // It should only ever be necessary for 64-bit types on O32 since the minimum |
| 2365 | // argument alignment is the same as the maximum type alignment for N32/N64. |
| 2366 | // |
| 2367 | // FIXME: We currently align too often. The code generator doesn't notice |
| 2368 | // when the pointer is still aligned from the last va_arg (or pair of |
| 2369 | // va_args for the i64 on O32 case). |
| 2370 | if (Align > getMinStackArgumentAlignment()) { |
| 2371 | VAList = DAG.getNode( |
| 2372 | Opcode: ISD::ADD, DL, VT: VAList.getValueType(), N1: VAList, |
| 2373 | N2: DAG.getConstant(Val: Align.value() - 1, DL, VT: VAList.getValueType())); |
| 2374 | |
| 2375 | VAList = DAG.getNode(Opcode: ISD::AND, DL, VT: VAList.getValueType(), N1: VAList, |
| 2376 | N2: DAG.getSignedConstant(Val: -(int64_t)Align.value(), DL, |
| 2377 | VT: VAList.getValueType())); |
| 2378 | } |
| 2379 | |
| 2380 | // Increment the pointer, VAList, to the next vaarg. |
| 2381 | auto &TD = DAG.getDataLayout(); |
| 2382 | unsigned ArgSizeInBytes = |
| 2383 | TD.getTypeAllocSize(Ty: VT.getTypeForEVT(Context&: *DAG.getContext())); |
| 2384 | SDValue Tmp3 = |
| 2385 | DAG.getNode(Opcode: ISD::ADD, DL, VT: VAList.getValueType(), N1: VAList, |
| 2386 | N2: DAG.getConstant(Val: alignTo(Value: ArgSizeInBytes, Align: ArgSlotSizeInBytes), |
| 2387 | DL, VT: VAList.getValueType())); |
| 2388 | // Store the incremented VAList to the legalized pointer |
| 2389 | Chain = DAG.getStore(Chain: VAListLoad.getValue(R: 1), dl: DL, Val: Tmp3, Ptr: VAListPtr, |
| 2390 | PtrInfo: MachinePointerInfo(SV)); |
| 2391 | |
| 2392 | // In big-endian mode we must adjust the pointer when the load size is smaller |
| 2393 | // than the argument slot size. We must also reduce the known alignment to |
| 2394 | // match. For example in the N64 ABI, we must add 4 bytes to the offset to get |
| 2395 | // the correct half of the slot, and reduce the alignment from 8 (slot |
| 2396 | // alignment) down to 4 (type alignment). |
| 2397 | if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { |
| 2398 | unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; |
| 2399 | VAList = DAG.getNode(Opcode: ISD::ADD, DL, VT: VAListPtr.getValueType(), N1: VAList, |
| 2400 | N2: DAG.getIntPtrConstant(Val: Adjustment, DL)); |
| 2401 | } |
| 2402 | // Load the actual argument out of the pointer VAList |
| 2403 | return DAG.getLoad(VT, dl: DL, Chain, Ptr: VAList, PtrInfo: MachinePointerInfo()); |
| 2404 | } |
| 2405 | |
| 2406 | static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, |
| 2407 | bool ) { |
| 2408 | EVT TyX = Op.getOperand(i: 0).getValueType(); |
| 2409 | EVT TyY = Op.getOperand(i: 1).getValueType(); |
| 2410 | SDLoc DL(Op); |
| 2411 | SDValue Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2412 | SDValue Const31 = DAG.getConstant(Val: 31, DL, VT: MVT::i32); |
| 2413 | SDValue Res; |
| 2414 | |
| 2415 | // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it |
| 2416 | // to i32. |
| 2417 | SDValue X = (TyX == MVT::f32) ? |
| 2418 | DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 0)) : |
| 2419 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 0), |
| 2420 | N2: Const1); |
| 2421 | SDValue Y = (TyY == MVT::f32) ? |
| 2422 | DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 1)) : |
| 2423 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 1), |
| 2424 | N2: Const1); |
| 2425 | |
| 2426 | if (HasExtractInsert) { |
| 2427 | // ext E, Y, 31, 1 ; extract bit31 of Y |
| 2428 | // ins X, E, 31, 1 ; insert extracted bit at bit31 of X |
| 2429 | SDValue E = DAG.getNode(Opcode: MipsISD::Ext, DL, VT: MVT::i32, N1: Y, N2: Const31, N3: Const1); |
| 2430 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i32, N1: E, N2: Const31, N3: Const1, N4: X); |
| 2431 | } else { |
| 2432 | // sll SllX, X, 1 |
| 2433 | // srl SrlX, SllX, 1 |
| 2434 | // srl SrlY, Y, 31 |
| 2435 | // sll SllY, SrlX, 31 |
| 2436 | // or Or, SrlX, SllY |
| 2437 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: X, N2: Const1); |
| 2438 | SDValue SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: SllX, N2: Const1); |
| 2439 | SDValue SrlY = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: Y, N2: Const31); |
| 2440 | SDValue SllY = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: SrlY, N2: Const31); |
| 2441 | Res = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i32, N1: SrlX, N2: SllY); |
| 2442 | } |
| 2443 | |
| 2444 | if (TyX == MVT::f32) |
| 2445 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: Res); |
| 2446 | |
| 2447 | SDValue LowX = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
| 2448 | N1: Op.getOperand(i: 0), |
| 2449 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
| 2450 | return DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, N1: LowX, N2: Res); |
| 2451 | } |
| 2452 | |
| 2453 | static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, |
| 2454 | bool ) { |
| 2455 | unsigned WidthX = Op.getOperand(i: 0).getValueSizeInBits(); |
| 2456 | unsigned WidthY = Op.getOperand(i: 1).getValueSizeInBits(); |
| 2457 | EVT TyX = MVT::getIntegerVT(BitWidth: WidthX), TyY = MVT::getIntegerVT(BitWidth: WidthY); |
| 2458 | SDLoc DL(Op); |
| 2459 | SDValue Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2460 | |
| 2461 | // Bitcast to integer nodes. |
| 2462 | SDValue X = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: TyX, Operand: Op.getOperand(i: 0)); |
| 2463 | SDValue Y = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: TyY, Operand: Op.getOperand(i: 1)); |
| 2464 | |
| 2465 | if (HasExtractInsert) { |
| 2466 | // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y |
| 2467 | // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X |
| 2468 | SDValue E = DAG.getNode(Opcode: MipsISD::Ext, DL, VT: TyY, N1: Y, |
| 2469 | N2: DAG.getConstant(Val: WidthY - 1, DL, VT: MVT::i32), N3: Const1); |
| 2470 | |
| 2471 | if (WidthX > WidthY) |
| 2472 | E = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: TyX, Operand: E); |
| 2473 | else if (WidthY > WidthX) |
| 2474 | E = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: TyX, Operand: E); |
| 2475 | |
| 2476 | SDValue I = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: TyX, N1: E, |
| 2477 | N2: DAG.getConstant(Val: WidthX - 1, DL, VT: MVT::i32), N3: Const1, |
| 2478 | N4: X); |
| 2479 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: I); |
| 2480 | } |
| 2481 | |
| 2482 | // (d)sll SllX, X, 1 |
| 2483 | // (d)srl SrlX, SllX, 1 |
| 2484 | // (d)srl SrlY, Y, width(Y)-1 |
| 2485 | // (d)sll SllY, SrlX, width(Y)-1 |
| 2486 | // or Or, SrlX, SllY |
| 2487 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: TyX, N1: X, N2: Const1); |
| 2488 | SDValue SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: TyX, N1: SllX, N2: Const1); |
| 2489 | SDValue SrlY = DAG.getNode(Opcode: ISD::SRL, DL, VT: TyY, N1: Y, |
| 2490 | N2: DAG.getConstant(Val: WidthY - 1, DL, VT: MVT::i32)); |
| 2491 | |
| 2492 | if (WidthX > WidthY) |
| 2493 | SrlY = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: TyX, Operand: SrlY); |
| 2494 | else if (WidthY > WidthX) |
| 2495 | SrlY = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: TyX, Operand: SrlY); |
| 2496 | |
| 2497 | SDValue SllY = DAG.getNode(Opcode: ISD::SHL, DL, VT: TyX, N1: SrlY, |
| 2498 | N2: DAG.getConstant(Val: WidthX - 1, DL, VT: MVT::i32)); |
| 2499 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT: TyX, N1: SrlX, N2: SllY); |
| 2500 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: Or); |
| 2501 | } |
| 2502 | |
| 2503 | SDValue |
| 2504 | MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { |
| 2505 | if (Subtarget.isGP64bit()) |
| 2506 | return lowerFCOPYSIGN64(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
| 2507 | |
| 2508 | return lowerFCOPYSIGN32(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
| 2509 | } |
| 2510 | |
| 2511 | SDValue MipsTargetLowering::lowerFABS32(SDValue Op, SelectionDAG &DAG, |
| 2512 | bool ) const { |
| 2513 | SDLoc DL(Op); |
| 2514 | SDValue Res, Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2515 | |
| 2516 | if (DAG.getTarget().Options.NoNaNsFPMath || Subtarget.inAbs2008Mode()) |
| 2517 | return DAG.getNode(Opcode: MipsISD::FAbs, DL, VT: Op.getValueType(), Operand: Op.getOperand(i: 0)); |
| 2518 | |
| 2519 | // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it |
| 2520 | // to i32. |
| 2521 | SDValue X = (Op.getValueType() == MVT::f32) |
| 2522 | ? DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 0)) |
| 2523 | : DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
| 2524 | N1: Op.getOperand(i: 0), N2: Const1); |
| 2525 | |
| 2526 | // Clear MSB. |
| 2527 | if (HasExtractInsert) |
| 2528 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i32, |
| 2529 | N1: DAG.getRegister(Reg: Mips::ZERO, VT: MVT::i32), |
| 2530 | N2: DAG.getConstant(Val: 31, DL, VT: MVT::i32), N3: Const1, N4: X); |
| 2531 | else { |
| 2532 | // TODO: Provide DAG patterns which transform (and x, cst) |
| 2533 | // back to a (shl (srl x (clz cst)) (clz cst)) sequence. |
| 2534 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: X, N2: Const1); |
| 2535 | Res = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: SllX, N2: Const1); |
| 2536 | } |
| 2537 | |
| 2538 | if (Op.getValueType() == MVT::f32) |
| 2539 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::f32, Operand: Res); |
| 2540 | |
| 2541 | // FIXME: For mips32r2, the sequence of (BuildPairF64 (ins (ExtractElementF64 |
| 2542 | // Op 1), $zero, 31 1) (ExtractElementF64 Op 0)) and the Op has one use, we |
| 2543 | // should be able to drop the usage of mfc1/mtc1 and rewrite the register in |
| 2544 | // place. |
| 2545 | SDValue LowX = |
| 2546 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 0), |
| 2547 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
| 2548 | return DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, N1: LowX, N2: Res); |
| 2549 | } |
| 2550 | |
| 2551 | SDValue MipsTargetLowering::lowerFABS64(SDValue Op, SelectionDAG &DAG, |
| 2552 | bool ) const { |
| 2553 | SDLoc DL(Op); |
| 2554 | SDValue Res, Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
| 2555 | |
| 2556 | if (DAG.getTarget().Options.NoNaNsFPMath || Subtarget.inAbs2008Mode()) |
| 2557 | return DAG.getNode(Opcode: MipsISD::FAbs, DL, VT: Op.getValueType(), Operand: Op.getOperand(i: 0)); |
| 2558 | |
| 2559 | // Bitcast to integer node. |
| 2560 | SDValue X = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i64, Operand: Op.getOperand(i: 0)); |
| 2561 | |
| 2562 | // Clear MSB. |
| 2563 | if (HasExtractInsert) |
| 2564 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i64, |
| 2565 | N1: DAG.getRegister(Reg: Mips::ZERO_64, VT: MVT::i64), |
| 2566 | N2: DAG.getConstant(Val: 63, DL, VT: MVT::i32), N3: Const1, N4: X); |
| 2567 | else { |
| 2568 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: X, N2: Const1); |
| 2569 | Res = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i64, N1: SllX, N2: Const1); |
| 2570 | } |
| 2571 | |
| 2572 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::f64, Operand: Res); |
| 2573 | } |
| 2574 | |
| 2575 | SDValue MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { |
| 2576 | if ((ABI.IsN32() || ABI.IsN64()) && (Op.getValueType() == MVT::f64)) |
| 2577 | return lowerFABS64(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
| 2578 | |
| 2579 | return lowerFABS32(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
| 2580 | } |
| 2581 | |
| 2582 | SDValue MipsTargetLowering::lowerFCANONICALIZE(SDValue Op, |
| 2583 | SelectionDAG &DAG) const { |
| 2584 | SDLoc DL(Op); |
| 2585 | EVT VT = Op.getValueType(); |
| 2586 | SDValue Operand = Op.getOperand(i: 0); |
| 2587 | SDNodeFlags Flags = Op->getFlags(); |
| 2588 | |
| 2589 | if (Flags.hasNoNaNs() || DAG.isKnownNeverNaN(Op: Operand)) |
| 2590 | return Operand; |
| 2591 | |
| 2592 | SDValue Quiet = DAG.getNode(Opcode: ISD::FADD, DL, VT, N1: Operand, N2: Operand); |
| 2593 | return DAG.getSelectCC(DL, LHS: Operand, RHS: Operand, True: Quiet, False: Operand, Cond: ISD::SETUO); |
| 2594 | } |
| 2595 | |
| 2596 | SDValue MipsTargetLowering:: |
| 2597 | lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { |
| 2598 | // check the depth |
| 2599 | if (Op.getConstantOperandVal(i: 0) != 0) { |
| 2600 | DAG.getContext()->emitError( |
| 2601 | ErrorStr: "return address can be determined only for current frame" ); |
| 2602 | return SDValue(); |
| 2603 | } |
| 2604 | |
| 2605 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2606 | MFI.setFrameAddressIsTaken(true); |
| 2607 | EVT VT = Op.getValueType(); |
| 2608 | SDLoc DL(Op); |
| 2609 | SDValue FrameAddr = DAG.getCopyFromReg( |
| 2610 | Chain: DAG.getEntryNode(), dl: DL, Reg: ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT); |
| 2611 | return FrameAddr; |
| 2612 | } |
| 2613 | |
| 2614 | SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, |
| 2615 | SelectionDAG &DAG) const { |
| 2616 | // check the depth |
| 2617 | if (Op.getConstantOperandVal(i: 0) != 0) { |
| 2618 | DAG.getContext()->emitError( |
| 2619 | ErrorStr: "return address can be determined only for current frame" ); |
| 2620 | return SDValue(); |
| 2621 | } |
| 2622 | |
| 2623 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2624 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2625 | MVT VT = Op.getSimpleValueType(); |
| 2626 | unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA; |
| 2627 | MFI.setReturnAddressIsTaken(true); |
| 2628 | |
| 2629 | // Return RA, which contains the return address. Mark it an implicit live-in. |
| 2630 | Register Reg = MF.addLiveIn(PReg: RA, RC: getRegClassFor(VT)); |
| 2631 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg, VT); |
| 2632 | } |
| 2633 | |
| 2634 | // An EH_RETURN is the result of lowering llvm.eh.return which in turn is |
| 2635 | // generated from __builtin_eh_return (offset, handler) |
| 2636 | // The effect of this is to adjust the stack pointer by "offset" |
| 2637 | // and then branch to "handler". |
| 2638 | SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) |
| 2639 | const { |
| 2640 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2641 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
| 2642 | |
| 2643 | MipsFI->setCallsEhReturn(); |
| 2644 | SDValue Chain = Op.getOperand(i: 0); |
| 2645 | SDValue Offset = Op.getOperand(i: 1); |
| 2646 | SDValue Handler = Op.getOperand(i: 2); |
| 2647 | SDLoc DL(Op); |
| 2648 | EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; |
| 2649 | |
| 2650 | // Store stack offset in V1, store jump target in V0. Glue CopyToReg and |
| 2651 | // EH_RETURN nodes, so that instructions are emitted back-to-back. |
| 2652 | unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1; |
| 2653 | unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0; |
| 2654 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: OffsetReg, N: Offset, Glue: SDValue()); |
| 2655 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: AddrReg, N: Handler, Glue: Chain.getValue(R: 1)); |
| 2656 | return DAG.getNode(Opcode: MipsISD::EH_RETURN, DL, VT: MVT::Other, N1: Chain, |
| 2657 | N2: DAG.getRegister(Reg: OffsetReg, VT: Ty), |
| 2658 | N3: DAG.getRegister(Reg: AddrReg, VT: getPointerTy(DL: MF.getDataLayout())), |
| 2659 | N4: Chain.getValue(R: 1)); |
| 2660 | } |
| 2661 | |
| 2662 | SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, |
| 2663 | SelectionDAG &DAG) const { |
| 2664 | // FIXME: Need pseudo-fence for 'singlethread' fences |
| 2665 | // FIXME: Set SType for weaker fences where supported/appropriate. |
| 2666 | unsigned SType = 0; |
| 2667 | SDLoc DL(Op); |
| 2668 | return DAG.getNode(Opcode: MipsISD::Sync, DL, VT: MVT::Other, N1: Op.getOperand(i: 0), |
| 2669 | N2: DAG.getConstant(Val: SType, DL, VT: MVT::i32)); |
| 2670 | } |
| 2671 | |
| 2672 | SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, |
| 2673 | SelectionDAG &DAG) const { |
| 2674 | SDLoc DL(Op); |
| 2675 | MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; |
| 2676 | |
| 2677 | SDValue Lo = Op.getOperand(i: 0), Hi = Op.getOperand(i: 1); |
| 2678 | SDValue Shamt = Op.getOperand(i: 2); |
| 2679 | // if shamt < (VT.bits): |
| 2680 | // lo = (shl lo, shamt) |
| 2681 | // hi = (or (shl hi, shamt) (srl (srl lo, 1), (xor shamt, (VT.bits-1)))) |
| 2682 | // else: |
| 2683 | // lo = 0 |
| 2684 | // hi = (shl lo, shamt[4:0]) |
| 2685 | SDValue Not = |
| 2686 | DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Shamt, |
| 2687 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT: MVT::i32)); |
| 2688 | SDValue ShiftRight1Lo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: Lo, |
| 2689 | N2: DAG.getConstant(Val: 1, DL, VT)); |
| 2690 | SDValue ShiftRightLo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: ShiftRight1Lo, N2: Not); |
| 2691 | SDValue ShiftLeftHi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, N2: Shamt); |
| 2692 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT, N1: ShiftLeftHi, N2: ShiftRightLo); |
| 2693 | SDValue ShiftLeftLo = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Lo, N2: Shamt); |
| 2694 | SDValue Cond = DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: Shamt, |
| 2695 | N2: DAG.getConstant(Val: VT.getSizeInBits(), DL, VT: MVT::i32)); |
| 2696 | Lo = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, |
| 2697 | N2: DAG.getConstant(Val: 0, DL, VT), N3: ShiftLeftLo); |
| 2698 | Hi = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, N2: ShiftLeftLo, N3: Or); |
| 2699 | |
| 2700 | SDValue Ops[2] = {Lo, Hi}; |
| 2701 | return DAG.getMergeValues(Ops, dl: DL); |
| 2702 | } |
| 2703 | |
| 2704 | SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, |
| 2705 | bool IsSRA) const { |
| 2706 | SDLoc DL(Op); |
| 2707 | SDValue Lo = Op.getOperand(i: 0), Hi = Op.getOperand(i: 1); |
| 2708 | SDValue Shamt = Op.getOperand(i: 2); |
| 2709 | MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; |
| 2710 | |
| 2711 | // if shamt < (VT.bits): |
| 2712 | // lo = (or (shl (shl hi, 1), (xor shamt, (VT.bits-1))) (srl lo, shamt)) |
| 2713 | // if isSRA: |
| 2714 | // hi = (sra hi, shamt) |
| 2715 | // else: |
| 2716 | // hi = (srl hi, shamt) |
| 2717 | // else: |
| 2718 | // if isSRA: |
| 2719 | // lo = (sra hi, shamt[4:0]) |
| 2720 | // hi = (sra hi, 31) |
| 2721 | // else: |
| 2722 | // lo = (srl hi, shamt[4:0]) |
| 2723 | // hi = 0 |
| 2724 | SDValue Not = |
| 2725 | DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Shamt, |
| 2726 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT: MVT::i32)); |
| 2727 | SDValue ShiftLeft1Hi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, |
| 2728 | N2: DAG.getConstant(Val: 1, DL, VT)); |
| 2729 | SDValue ShiftLeftHi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: ShiftLeft1Hi, N2: Not); |
| 2730 | SDValue ShiftRightLo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: Lo, N2: Shamt); |
| 2731 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT, N1: ShiftLeftHi, N2: ShiftRightLo); |
| 2732 | SDValue ShiftRightHi = DAG.getNode(Opcode: IsSRA ? ISD::SRA : ISD::SRL, |
| 2733 | DL, VT, N1: Hi, N2: Shamt); |
| 2734 | SDValue Cond = DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: Shamt, |
| 2735 | N2: DAG.getConstant(Val: VT.getSizeInBits(), DL, VT: MVT::i32)); |
| 2736 | SDValue Ext = DAG.getNode(Opcode: ISD::SRA, DL, VT, N1: Hi, |
| 2737 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT)); |
| 2738 | |
| 2739 | if (!(Subtarget.hasMips4() || Subtarget.hasMips32())) { |
| 2740 | SDVTList VTList = DAG.getVTList(VT1: VT, VT2: VT); |
| 2741 | return DAG.getNode(Opcode: Subtarget.isGP64bit() ? MipsISD::DOUBLE_SELECT_I64 |
| 2742 | : MipsISD::DOUBLE_SELECT_I, |
| 2743 | DL, VTList, N1: Cond, N2: ShiftRightHi, |
| 2744 | N3: IsSRA ? Ext : DAG.getConstant(Val: 0, DL, VT), N4: Or, |
| 2745 | N5: ShiftRightHi); |
| 2746 | } |
| 2747 | |
| 2748 | Lo = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, N2: ShiftRightHi, N3: Or); |
| 2749 | Hi = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, |
| 2750 | N2: IsSRA ? Ext : DAG.getConstant(Val: 0, DL, VT), N3: ShiftRightHi); |
| 2751 | |
| 2752 | SDValue Ops[2] = {Lo, Hi}; |
| 2753 | return DAG.getMergeValues(Ops, dl: DL); |
| 2754 | } |
| 2755 | |
| 2756 | static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, |
| 2757 | SDValue Chain, SDValue Src, unsigned Offset) { |
| 2758 | SDValue Ptr = LD->getBasePtr(); |
| 2759 | EVT VT = LD->getValueType(ResNo: 0), MemVT = LD->getMemoryVT(); |
| 2760 | EVT BasePtrVT = Ptr.getValueType(); |
| 2761 | SDLoc DL(LD); |
| 2762 | SDVTList VTList = DAG.getVTList(VT1: VT, VT2: MVT::Other); |
| 2763 | |
| 2764 | if (Offset) |
| 2765 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: BasePtrVT, N1: Ptr, |
| 2766 | N2: DAG.getConstant(Val: Offset, DL, VT: BasePtrVT)); |
| 2767 | |
| 2768 | SDValue Ops[] = { Chain, Ptr, Src }; |
| 2769 | return DAG.getMemIntrinsicNode(Opcode: Opc, dl: DL, VTList, Ops, MemVT, |
| 2770 | MMO: LD->getMemOperand()); |
| 2771 | } |
| 2772 | |
| 2773 | // Expand an unaligned 32 or 64-bit integer load node. |
| 2774 | SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
| 2775 | LoadSDNode *LD = cast<LoadSDNode>(Val&: Op); |
| 2776 | EVT MemVT = LD->getMemoryVT(); |
| 2777 | |
| 2778 | if (Subtarget.systemSupportsUnalignedAccess()) |
| 2779 | return Op; |
| 2780 | |
| 2781 | // Return if load is aligned or if MemVT is neither i32 nor i64. |
| 2782 | if ((LD->getAlign().value() >= (MemVT.getSizeInBits() / 8)) || |
| 2783 | ((MemVT != MVT::i32) && (MemVT != MVT::i64))) |
| 2784 | return SDValue(); |
| 2785 | |
| 2786 | bool IsLittle = Subtarget.isLittle(); |
| 2787 | EVT VT = Op.getValueType(); |
| 2788 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
| 2789 | SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); |
| 2790 | |
| 2791 | assert((VT == MVT::i32) || (VT == MVT::i64)); |
| 2792 | |
| 2793 | // Expand |
| 2794 | // (set dst, (i64 (load baseptr))) |
| 2795 | // to |
| 2796 | // (set tmp, (ldl (add baseptr, 7), undef)) |
| 2797 | // (set dst, (ldr baseptr, tmp)) |
| 2798 | if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { |
| 2799 | SDValue LDL = createLoadLR(Opc: MipsISD::LDL, DAG, LD, Chain, Src: Undef, |
| 2800 | Offset: IsLittle ? 7 : 0); |
| 2801 | return createLoadLR(Opc: MipsISD::LDR, DAG, LD, Chain: LDL.getValue(R: 1), Src: LDL, |
| 2802 | Offset: IsLittle ? 0 : 7); |
| 2803 | } |
| 2804 | |
| 2805 | SDValue LWL = createLoadLR(Opc: MipsISD::LWL, DAG, LD, Chain, Src: Undef, |
| 2806 | Offset: IsLittle ? 3 : 0); |
| 2807 | SDValue LWR = createLoadLR(Opc: MipsISD::LWR, DAG, LD, Chain: LWL.getValue(R: 1), Src: LWL, |
| 2808 | Offset: IsLittle ? 0 : 3); |
| 2809 | |
| 2810 | // Expand |
| 2811 | // (set dst, (i32 (load baseptr))) or |
| 2812 | // (set dst, (i64 (sextload baseptr))) or |
| 2813 | // (set dst, (i64 (extload baseptr))) |
| 2814 | // to |
| 2815 | // (set tmp, (lwl (add baseptr, 3), undef)) |
| 2816 | // (set dst, (lwr baseptr, tmp)) |
| 2817 | if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || |
| 2818 | (ExtType == ISD::EXTLOAD)) |
| 2819 | return LWR; |
| 2820 | |
| 2821 | assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); |
| 2822 | |
| 2823 | // Expand |
| 2824 | // (set dst, (i64 (zextload baseptr))) |
| 2825 | // to |
| 2826 | // (set tmp0, (lwl (add baseptr, 3), undef)) |
| 2827 | // (set tmp1, (lwr baseptr, tmp0)) |
| 2828 | // (set tmp2, (shl tmp1, 32)) |
| 2829 | // (set dst, (srl tmp2, 32)) |
| 2830 | SDLoc DL(LD); |
| 2831 | SDValue Const32 = DAG.getConstant(Val: 32, DL, VT: MVT::i32); |
| 2832 | SDValue SLL = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: LWR, N2: Const32); |
| 2833 | SDValue SRL = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i64, N1: SLL, N2: Const32); |
| 2834 | SDValue Ops[] = { SRL, LWR.getValue(R: 1) }; |
| 2835 | return DAG.getMergeValues(Ops, dl: DL); |
| 2836 | } |
| 2837 | |
| 2838 | static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, |
| 2839 | SDValue Chain, unsigned Offset) { |
| 2840 | SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); |
| 2841 | EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); |
| 2842 | SDLoc DL(SD); |
| 2843 | SDVTList VTList = DAG.getVTList(VT: MVT::Other); |
| 2844 | |
| 2845 | if (Offset) |
| 2846 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: BasePtrVT, N1: Ptr, |
| 2847 | N2: DAG.getConstant(Val: Offset, DL, VT: BasePtrVT)); |
| 2848 | |
| 2849 | SDValue Ops[] = { Chain, Value, Ptr }; |
| 2850 | return DAG.getMemIntrinsicNode(Opcode: Opc, dl: DL, VTList, Ops, MemVT, |
| 2851 | MMO: SD->getMemOperand()); |
| 2852 | } |
| 2853 | |
| 2854 | // Expand an unaligned 32 or 64-bit integer store node. |
| 2855 | static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, |
| 2856 | bool IsLittle) { |
| 2857 | SDValue Value = SD->getValue(), Chain = SD->getChain(); |
| 2858 | EVT VT = Value.getValueType(); |
| 2859 | |
| 2860 | // Expand |
| 2861 | // (store val, baseptr) or |
| 2862 | // (truncstore val, baseptr) |
| 2863 | // to |
| 2864 | // (swl val, (add baseptr, 3)) |
| 2865 | // (swr val, baseptr) |
| 2866 | if ((VT == MVT::i32) || SD->isTruncatingStore()) { |
| 2867 | SDValue SWL = createStoreLR(Opc: MipsISD::SWL, DAG, SD, Chain, |
| 2868 | Offset: IsLittle ? 3 : 0); |
| 2869 | return createStoreLR(Opc: MipsISD::SWR, DAG, SD, Chain: SWL, Offset: IsLittle ? 0 : 3); |
| 2870 | } |
| 2871 | |
| 2872 | assert(VT == MVT::i64); |
| 2873 | |
| 2874 | // Expand |
| 2875 | // (store val, baseptr) |
| 2876 | // to |
| 2877 | // (sdl val, (add baseptr, 7)) |
| 2878 | // (sdr val, baseptr) |
| 2879 | SDValue SDL = createStoreLR(Opc: MipsISD::SDL, DAG, SD, Chain, Offset: IsLittle ? 7 : 0); |
| 2880 | return createStoreLR(Opc: MipsISD::SDR, DAG, SD, Chain: SDL, Offset: IsLittle ? 0 : 7); |
| 2881 | } |
| 2882 | |
| 2883 | // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). |
| 2884 | static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG, |
| 2885 | bool SingleFloat) { |
| 2886 | SDValue Val = SD->getValue(); |
| 2887 | |
| 2888 | if (Val.getOpcode() != ISD::FP_TO_SINT || |
| 2889 | (Val.getValueSizeInBits() > 32 && SingleFloat)) |
| 2890 | return SDValue(); |
| 2891 | |
| 2892 | EVT FPTy = EVT::getFloatingPointVT(BitWidth: Val.getValueSizeInBits()); |
| 2893 | SDValue Tr = DAG.getNode(Opcode: MipsISD::TruncIntFP, DL: SDLoc(Val), VT: FPTy, |
| 2894 | Operand: Val.getOperand(i: 0)); |
| 2895 | return DAG.getStore(Chain: SD->getChain(), dl: SDLoc(SD), Val: Tr, Ptr: SD->getBasePtr(), |
| 2896 | PtrInfo: SD->getPointerInfo(), Alignment: SD->getAlign(), |
| 2897 | MMOFlags: SD->getMemOperand()->getFlags()); |
| 2898 | } |
| 2899 | |
| 2900 | SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
| 2901 | StoreSDNode *SD = cast<StoreSDNode>(Val&: Op); |
| 2902 | EVT MemVT = SD->getMemoryVT(); |
| 2903 | |
| 2904 | // Lower unaligned integer stores. |
| 2905 | if (!Subtarget.systemSupportsUnalignedAccess() && |
| 2906 | (SD->getAlign().value() < (MemVT.getSizeInBits() / 8)) && |
| 2907 | ((MemVT == MVT::i32) || (MemVT == MVT::i64))) |
| 2908 | return lowerUnalignedIntStore(SD, DAG, IsLittle: Subtarget.isLittle()); |
| 2909 | |
| 2910 | return lowerFP_TO_SINT_STORE(SD, DAG, SingleFloat: Subtarget.isSingleFloat()); |
| 2911 | } |
| 2912 | |
| 2913 | SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op, |
| 2914 | SelectionDAG &DAG) const { |
| 2915 | |
| 2916 | // Return a fixed StackObject with offset 0 which points to the old stack |
| 2917 | // pointer. |
| 2918 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2919 | EVT ValTy = Op->getValueType(ResNo: 0); |
| 2920 | int FI = MFI.CreateFixedObject(Size: Op.getValueSizeInBits() / 8, SPOffset: 0, IsImmutable: false); |
| 2921 | return DAG.getFrameIndex(FI, VT: ValTy); |
| 2922 | } |
| 2923 | |
| 2924 | SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, |
| 2925 | SelectionDAG &DAG) const { |
| 2926 | if (Op.getValueSizeInBits() > 32 && Subtarget.isSingleFloat()) |
| 2927 | return SDValue(); |
| 2928 | |
| 2929 | EVT FPTy = EVT::getFloatingPointVT(BitWidth: Op.getValueSizeInBits()); |
| 2930 | SDValue Trunc = DAG.getNode(Opcode: MipsISD::TruncIntFP, DL: SDLoc(Op), VT: FPTy, |
| 2931 | Operand: Op.getOperand(i: 0)); |
| 2932 | return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(Op), VT: Op.getValueType(), Operand: Trunc); |
| 2933 | } |
| 2934 | |
| 2935 | SDValue MipsTargetLowering::lowerSTRICT_FP_TO_INT(SDValue Op, |
| 2936 | SelectionDAG &DAG) const { |
| 2937 | assert(Op->isStrictFPOpcode()); |
| 2938 | SDValue SrcVal = Op.getOperand(i: 1); |
| 2939 | SDLoc Loc(Op); |
| 2940 | |
| 2941 | SDValue Result = |
| 2942 | DAG.getNode(Opcode: Op.getOpcode() == ISD::STRICT_FP_TO_SINT ? ISD::FP_TO_SINT |
| 2943 | : ISD::FP_TO_UINT, |
| 2944 | DL: Loc, VT: Op.getValueType(), Operand: SrcVal); |
| 2945 | |
| 2946 | return DAG.getMergeValues(Ops: {Result, Op.getOperand(i: 0)}, dl: Loc); |
| 2947 | } |
| 2948 | |
| 2949 | ArrayRef<MCPhysReg> MipsTargetLowering::getRoundingControlRegisters() const { |
| 2950 | static const MCPhysReg RCRegs[] = {Mips::FCR31}; |
| 2951 | return RCRegs; |
| 2952 | } |
| 2953 | |
| 2954 | //===----------------------------------------------------------------------===// |
| 2955 | // Calling Convention Implementation |
| 2956 | //===----------------------------------------------------------------------===// |
| 2957 | |
| 2958 | //===----------------------------------------------------------------------===// |
| 2959 | // TODO: Implement a generic logic using tblgen that can support this. |
| 2960 | // Mips O32 ABI rules: |
| 2961 | // --- |
| 2962 | // i32 - Passed in A0, A1, A2, A3 and stack |
| 2963 | // f32 - Only passed in f32 registers if no int reg has been used yet to hold |
| 2964 | // an argument. Otherwise, passed in A1, A2, A3 and stack. |
| 2965 | // f64 - Only passed in two aliased f32 registers if no int reg has been used |
| 2966 | // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is |
| 2967 | // not used, it must be shadowed. If only A3 is available, shadow it and |
| 2968 | // go to stack. |
| 2969 | // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack. |
| 2970 | // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3} |
| 2971 | // with the remainder spilled to the stack. |
| 2972 | // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases |
| 2973 | // spilling the remainder to the stack. |
| 2974 | // |
| 2975 | // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. |
| 2976 | //===----------------------------------------------------------------------===// |
| 2977 | |
| 2978 | static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, |
| 2979 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
| 2980 | Type *OrigTy, CCState &State, |
| 2981 | ArrayRef<MCPhysReg> F64Regs) { |
| 2982 | const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>( |
| 2983 | State.getMachineFunction().getSubtarget()); |
| 2984 | |
| 2985 | static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; |
| 2986 | |
| 2987 | static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; |
| 2988 | |
| 2989 | static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 }; |
| 2990 | |
| 2991 | // Do not process byval args here. |
| 2992 | if (ArgFlags.isByVal()) |
| 2993 | return true; |
| 2994 | |
| 2995 | // Promote i8 and i16 |
| 2996 | if (ArgFlags.isInReg() && !Subtarget.isLittle()) { |
| 2997 | if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { |
| 2998 | LocVT = MVT::i32; |
| 2999 | if (ArgFlags.isSExt()) |
| 3000 | LocInfo = CCValAssign::SExtUpper; |
| 3001 | else if (ArgFlags.isZExt()) |
| 3002 | LocInfo = CCValAssign::ZExtUpper; |
| 3003 | else |
| 3004 | LocInfo = CCValAssign::AExtUpper; |
| 3005 | } |
| 3006 | } |
| 3007 | |
| 3008 | // Promote i8 and i16 |
| 3009 | if (LocVT == MVT::i8 || LocVT == MVT::i16) { |
| 3010 | LocVT = MVT::i32; |
| 3011 | if (ArgFlags.isSExt()) |
| 3012 | LocInfo = CCValAssign::SExt; |
| 3013 | else if (ArgFlags.isZExt()) |
| 3014 | LocInfo = CCValAssign::ZExt; |
| 3015 | else |
| 3016 | LocInfo = CCValAssign::AExt; |
| 3017 | } |
| 3018 | |
| 3019 | unsigned Reg; |
| 3020 | |
| 3021 | // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following |
| 3022 | // is true: function is vararg, argument is 3rd or higher, there is previous |
| 3023 | // argument which is not f32 or f64. |
| 3024 | bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 || |
| 3025 | State.getFirstUnallocated(Regs: F32Regs) != ValNo; |
| 3026 | Align OrigAlign = ArgFlags.getNonZeroOrigAlign(); |
| 3027 | bool isI64 = (ValVT == MVT::i32 && OrigAlign == Align(8)); |
| 3028 | bool isVectorFloat = OrigTy->isVectorTy() && OrigTy->isFPOrFPVectorTy(); |
| 3029 | |
| 3030 | // The MIPS vector ABI for floats passes them in a pair of registers |
| 3031 | if (ValVT == MVT::i32 && isVectorFloat) { |
| 3032 | // This is the start of an vector that was scalarized into an unknown number |
| 3033 | // of components. It doesn't matter how many there are. Allocate one of the |
| 3034 | // notional 8 byte aligned registers which map onto the argument stack, and |
| 3035 | // shadow the register lost to alignment requirements. |
| 3036 | if (ArgFlags.isSplit()) { |
| 3037 | Reg = State.AllocateReg(Regs: FloatVectorIntRegs); |
| 3038 | if (Reg == Mips::A2) |
| 3039 | State.AllocateReg(Reg: Mips::A1); |
| 3040 | else if (Reg == 0) |
| 3041 | State.AllocateReg(Reg: Mips::A3); |
| 3042 | } else { |
| 3043 | // If we're an intermediate component of the split, we can just attempt to |
| 3044 | // allocate a register directly. |
| 3045 | Reg = State.AllocateReg(Regs: IntRegs); |
| 3046 | } |
| 3047 | } else if (ValVT == MVT::i32 || |
| 3048 | (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { |
| 3049 | Reg = State.AllocateReg(Regs: IntRegs); |
| 3050 | // If this is the first part of an i64 arg, |
| 3051 | // the allocated register must be either A0 or A2. |
| 3052 | if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) |
| 3053 | Reg = State.AllocateReg(Regs: IntRegs); |
| 3054 | LocVT = MVT::i32; |
| 3055 | } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { |
| 3056 | // Allocate int register and shadow next int register. If first |
| 3057 | // available register is Mips::A1 or Mips::A3, shadow it too. |
| 3058 | Reg = State.AllocateReg(Regs: IntRegs); |
| 3059 | if (Reg == Mips::A1 || Reg == Mips::A3) |
| 3060 | Reg = State.AllocateReg(Regs: IntRegs); |
| 3061 | |
| 3062 | if (Reg) { |
| 3063 | LocVT = MVT::i32; |
| 3064 | |
| 3065 | State.addLoc( |
| 3066 | V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 3067 | MCRegister HiReg = State.AllocateReg(Regs: IntRegs); |
| 3068 | assert(HiReg); |
| 3069 | State.addLoc( |
| 3070 | V: CCValAssign::getCustomReg(ValNo, ValVT, Reg: HiReg, LocVT, HTP: LocInfo)); |
| 3071 | return false; |
| 3072 | } |
| 3073 | } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { |
| 3074 | // we are guaranteed to find an available float register |
| 3075 | if (ValVT == MVT::f32) { |
| 3076 | Reg = State.AllocateReg(Regs: F32Regs); |
| 3077 | // Shadow int register |
| 3078 | State.AllocateReg(Regs: IntRegs); |
| 3079 | } else { |
| 3080 | Reg = State.AllocateReg(Regs: F64Regs); |
| 3081 | // Shadow int registers |
| 3082 | MCRegister Reg2 = State.AllocateReg(Regs: IntRegs); |
| 3083 | if (Reg2 == Mips::A1 || Reg2 == Mips::A3) |
| 3084 | State.AllocateReg(Regs: IntRegs); |
| 3085 | State.AllocateReg(Regs: IntRegs); |
| 3086 | } |
| 3087 | } else |
| 3088 | llvm_unreachable("Cannot handle this ValVT." ); |
| 3089 | |
| 3090 | if (!Reg) { |
| 3091 | unsigned Offset = State.AllocateStack(Size: ValVT.getStoreSize(), Alignment: OrigAlign); |
| 3092 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
| 3093 | } else |
| 3094 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 3095 | |
| 3096 | return false; |
| 3097 | } |
| 3098 | |
| 3099 | static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, MVT LocVT, |
| 3100 | CCValAssign::LocInfo LocInfo, |
| 3101 | ISD::ArgFlagsTy ArgFlags, Type *OrigTy, |
| 3102 | CCState &State) { |
| 3103 | static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; |
| 3104 | |
| 3105 | return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, OrigTy, State, |
| 3106 | F64Regs); |
| 3107 | } |
| 3108 | |
| 3109 | static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, MVT LocVT, |
| 3110 | CCValAssign::LocInfo LocInfo, |
| 3111 | ISD::ArgFlagsTy ArgFlags, Type *OrigTy, |
| 3112 | CCState &State) { |
| 3113 | static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; |
| 3114 | |
| 3115 | return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, OrigTy, State, |
| 3116 | F64Regs); |
| 3117 | } |
| 3118 | |
| 3119 | [[maybe_unused]] static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, |
| 3120 | CCValAssign::LocInfo LocInfo, |
| 3121 | ISD::ArgFlagsTy ArgFlags, Type *OrigTy, |
| 3122 | CCState &State); |
| 3123 | |
| 3124 | #include "MipsGenCallingConv.inc" |
| 3125 | |
| 3126 | CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{ |
| 3127 | return CC_Mips_FixedArg; |
| 3128 | } |
| 3129 | |
| 3130 | CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{ |
| 3131 | return RetCC_Mips; |
| 3132 | } |
| 3133 | //===----------------------------------------------------------------------===// |
| 3134 | // Call Calling Convention Implementation |
| 3135 | //===----------------------------------------------------------------------===// |
| 3136 | |
| 3137 | SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, |
| 3138 | SDValue Chain, SDValue Arg, |
| 3139 | const SDLoc &DL, bool IsTailCall, |
| 3140 | SelectionDAG &DAG) const { |
| 3141 | if (!IsTailCall) { |
| 3142 | SDValue PtrOff = |
| 3143 | DAG.getNode(Opcode: ISD::ADD, DL, VT: getPointerTy(DL: DAG.getDataLayout()), N1: StackPtr, |
| 3144 | N2: DAG.getIntPtrConstant(Val: Offset, DL)); |
| 3145 | return DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo()); |
| 3146 | } |
| 3147 | |
| 3148 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 3149 | int FI = MFI.CreateFixedObject(Size: Arg.getValueSizeInBits() / 8, SPOffset: Offset, IsImmutable: false); |
| 3150 | SDValue FIN = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 3151 | return DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: FIN, PtrInfo: MachinePointerInfo(), Alignment: MaybeAlign(), |
| 3152 | MMOFlags: MachineMemOperand::MOVolatile); |
| 3153 | } |
| 3154 | |
| 3155 | void MipsTargetLowering:: |
| 3156 | getOpndList(SmallVectorImpl<SDValue> &Ops, |
| 3157 | std::deque<std::pair<unsigned, SDValue>> &RegsToPass, |
| 3158 | bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, |
| 3159 | bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, |
| 3160 | SDValue Chain) const { |
| 3161 | // Insert node "GP copy globalreg" before call to function. |
| 3162 | // |
| 3163 | // R_MIPS_CALL* operators (emitted when non-internal functions are called |
| 3164 | // in PIC mode) allow symbols to be resolved via lazy binding. |
| 3165 | // The lazy binding stub requires GP to point to the GOT. |
| 3166 | // Note that we don't need GP to point to the GOT for indirect calls |
| 3167 | // (when R_MIPS_CALL* is not used for the call) because Mips linker generates |
| 3168 | // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs |
| 3169 | // used for the function (that is, Mips linker doesn't generate lazy binding |
| 3170 | // stub for a function whose address is taken in the program). |
| 3171 | if (IsPICCall && !InternalLinkage && IsCallReloc) { |
| 3172 | unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP; |
| 3173 | EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; |
| 3174 | RegsToPass.push_back(x: std::make_pair(x&: GPReg, y: getGlobalReg(DAG&: CLI.DAG, Ty))); |
| 3175 | } |
| 3176 | |
| 3177 | // Build a sequence of copy-to-reg nodes chained together with token |
| 3178 | // chain and flag operands which copy the outgoing args into registers. |
| 3179 | // The InGlue in necessary since all emitted instructions must be |
| 3180 | // stuck together. |
| 3181 | SDValue InGlue; |
| 3182 | |
| 3183 | for (auto &R : RegsToPass) { |
| 3184 | Chain = CLI.DAG.getCopyToReg(Chain, dl: CLI.DL, Reg: R.first, N: R.second, Glue: InGlue); |
| 3185 | InGlue = Chain.getValue(R: 1); |
| 3186 | } |
| 3187 | |
| 3188 | // Add argument registers to the end of the list so that they are |
| 3189 | // known live into the call. |
| 3190 | for (auto &R : RegsToPass) |
| 3191 | Ops.push_back(Elt: CLI.DAG.getRegister(Reg: R.first, VT: R.second.getValueType())); |
| 3192 | |
| 3193 | // Add a register mask operand representing the call-preserved registers. |
| 3194 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
| 3195 | const uint32_t *Mask = |
| 3196 | TRI->getCallPreservedMask(MF: CLI.DAG.getMachineFunction(), CLI.CallConv); |
| 3197 | assert(Mask && "Missing call preserved mask for calling convention" ); |
| 3198 | if (Subtarget.inMips16HardFloat()) { |
| 3199 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: CLI.Callee)) { |
| 3200 | StringRef Sym = G->getGlobal()->getName(); |
| 3201 | Function *F = G->getGlobal()->getParent()->getFunction(Name: Sym); |
| 3202 | if (F && F->hasFnAttribute(Kind: "__Mips16RetHelper" )) { |
| 3203 | Mask = MipsRegisterInfo::getMips16RetHelperMask(); |
| 3204 | } |
| 3205 | } |
| 3206 | } |
| 3207 | Ops.push_back(Elt: CLI.DAG.getRegisterMask(RegMask: Mask)); |
| 3208 | |
| 3209 | if (InGlue.getNode()) |
| 3210 | Ops.push_back(Elt: InGlue); |
| 3211 | } |
| 3212 | |
| 3213 | void MipsTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, |
| 3214 | SDNode *Node) const { |
| 3215 | switch (MI.getOpcode()) { |
| 3216 | default: |
| 3217 | return; |
| 3218 | case Mips::JALR: |
| 3219 | case Mips::JALRPseudo: |
| 3220 | case Mips::JALR64: |
| 3221 | case Mips::JALR64Pseudo: |
| 3222 | case Mips::JALR16_MM: |
| 3223 | case Mips::JALRC16_MMR6: |
| 3224 | case Mips::TAILCALLREG: |
| 3225 | case Mips::TAILCALLREG64: |
| 3226 | case Mips::TAILCALLR6REG: |
| 3227 | case Mips::TAILCALL64R6REG: |
| 3228 | case Mips::TAILCALLREG_MM: |
| 3229 | case Mips::TAILCALLREG_MMR6: { |
| 3230 | if (!EmitJalrReloc || |
| 3231 | Subtarget.inMips16Mode() || |
| 3232 | !isPositionIndependent() || |
| 3233 | Node->getNumOperands() < 1 || |
| 3234 | Node->getOperand(Num: 0).getNumOperands() < 2) { |
| 3235 | return; |
| 3236 | } |
| 3237 | // We are after the callee address, set by LowerCall(). |
| 3238 | // If added to MI, asm printer will emit .reloc R_MIPS_JALR for the |
| 3239 | // symbol. |
| 3240 | const SDValue TargetAddr = Node->getOperand(Num: 0).getOperand(i: 1); |
| 3241 | StringRef Sym; |
| 3242 | if (const GlobalAddressSDNode *G = |
| 3243 | dyn_cast_or_null<const GlobalAddressSDNode>(Val: TargetAddr)) { |
| 3244 | // We must not emit the R_MIPS_JALR relocation against data symbols |
| 3245 | // since this will cause run-time crashes if the linker replaces the |
| 3246 | // call instruction with a relative branch to the data symbol. |
| 3247 | if (!isa<Function>(Val: G->getGlobal())) { |
| 3248 | LLVM_DEBUG(dbgs() << "Not adding R_MIPS_JALR against data symbol " |
| 3249 | << G->getGlobal()->getName() << "\n" ); |
| 3250 | return; |
| 3251 | } |
| 3252 | Sym = G->getGlobal()->getName(); |
| 3253 | } |
| 3254 | else if (const ExternalSymbolSDNode *ES = |
| 3255 | dyn_cast_or_null<const ExternalSymbolSDNode>(Val: TargetAddr)) { |
| 3256 | Sym = ES->getSymbol(); |
| 3257 | } |
| 3258 | |
| 3259 | if (Sym.empty()) |
| 3260 | return; |
| 3261 | |
| 3262 | MachineFunction *MF = MI.getParent()->getParent(); |
| 3263 | MCSymbol *S = MF->getContext().getOrCreateSymbol(Name: Sym); |
| 3264 | LLVM_DEBUG(dbgs() << "Adding R_MIPS_JALR against " << Sym << "\n" ); |
| 3265 | MI.addOperand(Op: MachineOperand::CreateMCSymbol(Sym: S, TargetFlags: MipsII::MO_JALR)); |
| 3266 | } |
| 3267 | } |
| 3268 | } |
| 3269 | |
| 3270 | /// LowerCall - functions arguments are copied from virtual regs to |
| 3271 | /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. |
| 3272 | SDValue |
| 3273 | MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
| 3274 | SmallVectorImpl<SDValue> &InVals) const { |
| 3275 | SelectionDAG &DAG = CLI.DAG; |
| 3276 | SDLoc DL = CLI.DL; |
| 3277 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 3278 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 3279 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
| 3280 | SDValue Chain = CLI.Chain; |
| 3281 | SDValue Callee = CLI.Callee; |
| 3282 | bool &IsTailCall = CLI.IsTailCall; |
| 3283 | CallingConv::ID CallConv = CLI.CallConv; |
| 3284 | bool IsVarArg = CLI.IsVarArg; |
| 3285 | const CallBase *CB = CLI.CB; |
| 3286 | |
| 3287 | MachineFunction &MF = DAG.getMachineFunction(); |
| 3288 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 3289 | const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); |
| 3290 | MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); |
| 3291 | bool IsPIC = isPositionIndependent(); |
| 3292 | |
| 3293 | // Analyze operands of the call, assigning locations to each operand. |
| 3294 | SmallVector<CCValAssign, 16> ArgLocs; |
| 3295 | MipsCCState CCInfo( |
| 3296 | CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), |
| 3297 | MipsCCState::getSpecialCallingConvForCallee(Callee: Callee.getNode(), Subtarget)); |
| 3298 | |
| 3299 | const ExternalSymbolSDNode *ES = |
| 3300 | dyn_cast_or_null<const ExternalSymbolSDNode>(Val: Callee.getNode()); |
| 3301 | |
| 3302 | // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which |
| 3303 | // is during the lowering of a call with a byval argument which produces |
| 3304 | // a call to memcpy. For the O32 case, this causes the caller to allocate |
| 3305 | // stack space for the reserved argument area for the callee, then recursively |
| 3306 | // again for the memcpy call. In the NEWABI case, this doesn't occur as those |
| 3307 | // ABIs mandate that the callee allocates the reserved argument area. We do |
| 3308 | // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though. |
| 3309 | // |
| 3310 | // If the callee has a byval argument and memcpy is used, we are mandated |
| 3311 | // to already have produced a reserved argument area for the callee for O32. |
| 3312 | // Therefore, the reserved argument area can be reused for both calls. |
| 3313 | // |
| 3314 | // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START |
| 3315 | // present, as we have yet to hook that node onto the chain. |
| 3316 | // |
| 3317 | // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this |
| 3318 | // case. GCC does a similar trick, in that wherever possible, it calculates |
| 3319 | // the maximum out going argument area (including the reserved area), and |
| 3320 | // preallocates the stack space on entrance to the caller. |
| 3321 | // |
| 3322 | // FIXME: We should do the same for efficiency and space. |
| 3323 | |
| 3324 | // Note: The check on the calling convention below must match |
| 3325 | // MipsABIInfo::GetCalleeAllocdArgSizeInBytes(). |
| 3326 | bool MemcpyInByVal = ES && StringRef(ES->getSymbol()) == "memcpy" && |
| 3327 | CallConv != CallingConv::Fast && |
| 3328 | Chain.getOpcode() == ISD::CALLSEQ_START; |
| 3329 | |
| 3330 | // Allocate the reserved argument area. It seems strange to do this from the |
| 3331 | // caller side but removing it breaks the frame size calculation. |
| 3332 | unsigned ReservedArgArea = |
| 3333 | MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CC: CallConv); |
| 3334 | CCInfo.AllocateStack(Size: ReservedArgArea, Alignment: Align(1)); |
| 3335 | |
| 3336 | CCInfo.AnalyzeCallOperands(Outs, Fn: CC_Mips); |
| 3337 | |
| 3338 | // Get a count of how many bytes are to be pushed on the stack. |
| 3339 | unsigned StackSize = CCInfo.getStackSize(); |
| 3340 | |
| 3341 | // Call site info for function parameters tracking and call base type info. |
| 3342 | MachineFunction::CallSiteInfo CSInfo; |
| 3343 | // Set type id for call site info. |
| 3344 | if (MF.getTarget().Options.EmitCallGraphSection && CB && CB->isIndirectCall()) |
| 3345 | CSInfo = MachineFunction::CallSiteInfo(*CB); |
| 3346 | |
| 3347 | // Check if it's really possible to do a tail call. |
| 3348 | // For non-musttail calls, restrict to functions that won't require $gp |
| 3349 | // restoration. In PIC mode, calling external functions via tail call can |
| 3350 | // cause issues with $gp register handling (see D24763). |
| 3351 | bool IsMustTail = CLI.CB && CLI.CB->isMustTailCall(); |
| 3352 | bool CalleeIsLocal = true; |
| 3353 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
| 3354 | const GlobalValue *GV = G->getGlobal(); |
| 3355 | bool HasLocalLinkage = GV->hasLocalLinkage() || GV->hasPrivateLinkage(); |
| 3356 | bool HasHiddenVisibility = |
| 3357 | GV->hasHiddenVisibility() || GV->hasProtectedVisibility(); |
| 3358 | if (GV->isDeclarationForLinker()) |
| 3359 | CalleeIsLocal = HasLocalLinkage || HasHiddenVisibility; |
| 3360 | else |
| 3361 | CalleeIsLocal = GV->isDSOLocal(); |
| 3362 | } |
| 3363 | |
| 3364 | if (IsTailCall) { |
| 3365 | if (!UseMipsTailCalls) { |
| 3366 | IsTailCall = false; |
| 3367 | } else { |
| 3368 | bool Eligible = isEligibleForTailCallOptimization( |
| 3369 | CCInfo, NextStackOffset: StackSize, FI: *MF.getInfo<MipsFunctionInfo>()); |
| 3370 | if (!Eligible || !CalleeIsLocal) { |
| 3371 | IsTailCall = false; |
| 3372 | if (IsMustTail) |
| 3373 | report_fatal_error( |
| 3374 | reason: "failed to perform tail call elimination on a call " |
| 3375 | "site marked musttail" ); |
| 3376 | } |
| 3377 | } |
| 3378 | } |
| 3379 | |
| 3380 | if (IsTailCall) |
| 3381 | ++NumTailCalls; |
| 3382 | |
| 3383 | // Chain is the output chain of the last Load/Store or CopyToReg node. |
| 3384 | // ByValChain is the output chain of the last Memcpy node created for copying |
| 3385 | // byval arguments to the stack. |
| 3386 | unsigned StackAlignment = TFL->getStackAlignment(); |
| 3387 | StackSize = alignTo(Value: StackSize, Align: StackAlignment); |
| 3388 | |
| 3389 | if (!(IsTailCall || MemcpyInByVal)) |
| 3390 | Chain = DAG.getCALLSEQ_START(Chain, InSize: StackSize, OutSize: 0, DL); |
| 3391 | |
| 3392 | SDValue StackPtr = |
| 3393 | DAG.getCopyFromReg(Chain, dl: DL, Reg: ABI.IsN64() ? Mips::SP_64 : Mips::SP, |
| 3394 | VT: getPointerTy(DL: DAG.getDataLayout())); |
| 3395 | std::deque<std::pair<unsigned, SDValue>> RegsToPass; |
| 3396 | SmallVector<SDValue, 8> MemOpChains; |
| 3397 | |
| 3398 | CCInfo.rewindByValRegsInfo(); |
| 3399 | |
| 3400 | // Walk the register/memloc assignments, inserting copies/loads. |
| 3401 | for (unsigned i = 0, e = ArgLocs.size(), OutIdx = 0; i != e; ++i, ++OutIdx) { |
| 3402 | SDValue Arg = OutVals[OutIdx]; |
| 3403 | CCValAssign &VA = ArgLocs[i]; |
| 3404 | MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); |
| 3405 | ISD::ArgFlagsTy Flags = Outs[OutIdx].Flags; |
| 3406 | bool UseUpperBits = false; |
| 3407 | |
| 3408 | // ByVal Arg. |
| 3409 | if (Flags.isByVal()) { |
| 3410 | unsigned FirstByValReg, LastByValReg; |
| 3411 | unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); |
| 3412 | CCInfo.getInRegsParamInfo(InRegsParamRecordIndex: ByValIdx, BeginReg&: FirstByValReg, EndReg&: LastByValReg); |
| 3413 | |
| 3414 | assert(Flags.getByValSize() && |
| 3415 | "ByVal args of size 0 should have been ignored by front-end." ); |
| 3416 | assert(ByValIdx < CCInfo.getInRegsParamsCount()); |
| 3417 | assert(!IsTailCall && |
| 3418 | "Do not tail-call optimize if there is a byval argument." ); |
| 3419 | passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, |
| 3420 | FirstReg: FirstByValReg, LastReg: LastByValReg, Flags, isLittle: Subtarget.isLittle(), |
| 3421 | VA); |
| 3422 | CCInfo.nextInRegsParam(); |
| 3423 | continue; |
| 3424 | } |
| 3425 | |
| 3426 | // Promote the value if needed. |
| 3427 | switch (VA.getLocInfo()) { |
| 3428 | default: |
| 3429 | llvm_unreachable("Unknown loc info!" ); |
| 3430 | case CCValAssign::Full: |
| 3431 | if (VA.isRegLoc()) { |
| 3432 | if ((ValVT == MVT::f32 && LocVT == MVT::i32) || |
| 3433 | (ValVT == MVT::f64 && LocVT == MVT::i64) || |
| 3434 | (ValVT == MVT::i64 && LocVT == MVT::f64)) |
| 3435 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: LocVT, Operand: Arg); |
| 3436 | else if (ValVT == MVT::f64 && LocVT == MVT::i32) { |
| 3437 | SDValue Lo = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
| 3438 | N1: Arg, N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
| 3439 | SDValue Hi = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
| 3440 | N1: Arg, N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32)); |
| 3441 | if (!Subtarget.isLittle()) |
| 3442 | std::swap(a&: Lo, b&: Hi); |
| 3443 | |
| 3444 | assert(VA.needsCustom()); |
| 3445 | |
| 3446 | Register LocRegLo = VA.getLocReg(); |
| 3447 | Register LocRegHigh = ArgLocs[++i].getLocReg(); |
| 3448 | RegsToPass.push_back(x: std::make_pair(x&: LocRegLo, y&: Lo)); |
| 3449 | RegsToPass.push_back(x: std::make_pair(x&: LocRegHigh, y&: Hi)); |
| 3450 | continue; |
| 3451 | } |
| 3452 | } |
| 3453 | break; |
| 3454 | case CCValAssign::BCvt: |
| 3455 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: LocVT, Operand: Arg); |
| 3456 | break; |
| 3457 | case CCValAssign::SExtUpper: |
| 3458 | UseUpperBits = true; |
| 3459 | [[fallthrough]]; |
| 3460 | case CCValAssign::SExt: |
| 3461 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: LocVT, Operand: Arg); |
| 3462 | break; |
| 3463 | case CCValAssign::ZExtUpper: |
| 3464 | UseUpperBits = true; |
| 3465 | [[fallthrough]]; |
| 3466 | case CCValAssign::ZExt: |
| 3467 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: LocVT, Operand: Arg); |
| 3468 | break; |
| 3469 | case CCValAssign::AExtUpper: |
| 3470 | UseUpperBits = true; |
| 3471 | [[fallthrough]]; |
| 3472 | case CCValAssign::AExt: |
| 3473 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: LocVT, Operand: Arg); |
| 3474 | break; |
| 3475 | } |
| 3476 | |
| 3477 | if (UseUpperBits) { |
| 3478 | unsigned ValSizeInBits = Outs[OutIdx].ArgVT.getSizeInBits(); |
| 3479 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
| 3480 | Arg = DAG.getNode( |
| 3481 | Opcode: ISD::SHL, DL, VT: VA.getLocVT(), N1: Arg, |
| 3482 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
| 3483 | } |
| 3484 | |
| 3485 | // Arguments that can be passed on register must be kept at |
| 3486 | // RegsToPass vector |
| 3487 | if (VA.isRegLoc()) { |
| 3488 | RegsToPass.push_back(x: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
| 3489 | |
| 3490 | // If the parameter is passed through reg $D, which splits into |
| 3491 | // two physical registers, avoid creating call site info. |
| 3492 | if (Mips::AFGR64RegClass.contains(Reg: VA.getLocReg())) |
| 3493 | continue; |
| 3494 | |
| 3495 | // Collect CSInfo about which register passes which parameter. |
| 3496 | const TargetOptions &Options = DAG.getTarget().Options; |
| 3497 | if (Options.EmitCallSiteInfo) |
| 3498 | CSInfo.ArgRegPairs.emplace_back(Args: VA.getLocReg(), Args&: i); |
| 3499 | |
| 3500 | continue; |
| 3501 | } |
| 3502 | |
| 3503 | // Register can't get to this point... |
| 3504 | assert(VA.isMemLoc()); |
| 3505 | |
| 3506 | // emit ISD::STORE whichs stores the |
| 3507 | // parameter value to a stack Location |
| 3508 | MemOpChains.push_back(Elt: passArgOnStack(StackPtr, Offset: VA.getLocMemOffset(), |
| 3509 | Chain, Arg, DL, IsTailCall, DAG)); |
| 3510 | } |
| 3511 | |
| 3512 | // Transform all store nodes into one single node because all store |
| 3513 | // nodes are independent of each other. |
| 3514 | if (!MemOpChains.empty()) |
| 3515 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: MemOpChains); |
| 3516 | |
| 3517 | // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every |
| 3518 | // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol |
| 3519 | // node so that legalize doesn't hack it. |
| 3520 | |
| 3521 | EVT Ty = Callee.getValueType(); |
| 3522 | bool GlobalOrExternal = false, IsCallReloc = false; |
| 3523 | |
| 3524 | // The long-calls feature is ignored in case of PIC. |
| 3525 | // While we do not support -mshared / -mno-shared properly, |
| 3526 | // ignore long-calls in case of -mabicalls too. |
| 3527 | if (!Subtarget.isABICalls() && !IsPIC) { |
| 3528 | // If the function should be called using "long call", |
| 3529 | // get its address into a register to prevent using |
| 3530 | // of the `jal` instruction for the direct call. |
| 3531 | if (auto *N = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
| 3532 | if (Subtarget.useLongCalls()) |
| 3533 | Callee = Subtarget.hasSym32() |
| 3534 | ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 3535 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 3536 | } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
| 3537 | bool UseLongCalls = Subtarget.useLongCalls(); |
| 3538 | // If the function has long-call/far/near attribute |
| 3539 | // it overrides command line switch pased to the backend. |
| 3540 | if (auto *F = dyn_cast<Function>(Val: N->getGlobal())) { |
| 3541 | if (F->hasFnAttribute(Kind: "long-call" )) |
| 3542 | UseLongCalls = true; |
| 3543 | else if (F->hasFnAttribute(Kind: "short-call" )) |
| 3544 | UseLongCalls = false; |
| 3545 | } |
| 3546 | if (UseLongCalls) |
| 3547 | Callee = Subtarget.hasSym32() |
| 3548 | ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
| 3549 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
| 3550 | } |
| 3551 | } |
| 3552 | |
| 3553 | bool InternalLinkage = false; |
| 3554 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
| 3555 | if (Subtarget.isTargetCOFF() && |
| 3556 | G->getGlobal()->hasDLLImportStorageClass()) { |
| 3557 | assert(Subtarget.isTargetWindows() && |
| 3558 | "Windows is the only supported COFF target" ); |
| 3559 | auto PtrInfo = MachinePointerInfo(); |
| 3560 | Callee = DAG.getLoad(VT: Ty, dl: DL, Chain, |
| 3561 | Ptr: getDllimportSymbol(N: G, DL: SDLoc(G), Ty, DAG), PtrInfo); |
| 3562 | } else if (IsPIC) { |
| 3563 | const GlobalValue *Val = G->getGlobal(); |
| 3564 | InternalLinkage = Val->hasInternalLinkage(); |
| 3565 | |
| 3566 | if (InternalLinkage) |
| 3567 | Callee = getAddrLocal(N: G, DL, Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
| 3568 | else if (Subtarget.useXGOT()) { |
| 3569 | Callee = getAddrGlobalLargeGOT(N: G, DL, Ty, DAG, HiFlag: MipsII::MO_CALL_HI16, |
| 3570 | LoFlag: MipsII::MO_CALL_LO16, Chain, |
| 3571 | PtrInfo: FuncInfo->callPtrInfo(MF, GV: Val)); |
| 3572 | IsCallReloc = true; |
| 3573 | } else { |
| 3574 | Callee = getAddrGlobal(N: G, DL, Ty, DAG, Flag: MipsII::MO_GOT_CALL, Chain, |
| 3575 | PtrInfo: FuncInfo->callPtrInfo(MF, GV: Val)); |
| 3576 | IsCallReloc = true; |
| 3577 | } |
| 3578 | } else |
| 3579 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL, |
| 3580 | VT: getPointerTy(DL: DAG.getDataLayout()), offset: 0, |
| 3581 | TargetFlags: MipsII::MO_NO_FLAG); |
| 3582 | GlobalOrExternal = true; |
| 3583 | } |
| 3584 | else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
| 3585 | const char *Sym = S->getSymbol(); |
| 3586 | |
| 3587 | if (!IsPIC) // static |
| 3588 | Callee = DAG.getTargetExternalSymbol( |
| 3589 | Sym, VT: getPointerTy(DL: DAG.getDataLayout()), TargetFlags: MipsII::MO_NO_FLAG); |
| 3590 | else if (Subtarget.useXGOT()) { |
| 3591 | Callee = getAddrGlobalLargeGOT(N: S, DL, Ty, DAG, HiFlag: MipsII::MO_CALL_HI16, |
| 3592 | LoFlag: MipsII::MO_CALL_LO16, Chain, |
| 3593 | PtrInfo: FuncInfo->callPtrInfo(MF, ES: Sym)); |
| 3594 | IsCallReloc = true; |
| 3595 | } else { // PIC |
| 3596 | Callee = getAddrGlobal(N: S, DL, Ty, DAG, Flag: MipsII::MO_GOT_CALL, Chain, |
| 3597 | PtrInfo: FuncInfo->callPtrInfo(MF, ES: Sym)); |
| 3598 | IsCallReloc = true; |
| 3599 | } |
| 3600 | |
| 3601 | GlobalOrExternal = true; |
| 3602 | } |
| 3603 | |
| 3604 | SmallVector<SDValue, 8> Ops(1, Chain); |
| 3605 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 3606 | |
| 3607 | getOpndList(Ops, RegsToPass, IsPICCall: IsPIC, GlobalOrExternal, InternalLinkage, |
| 3608 | IsCallReloc, CLI, Callee, Chain); |
| 3609 | |
| 3610 | if (IsTailCall) { |
| 3611 | MF.getFrameInfo().setHasTailCall(); |
| 3612 | SDValue Ret = DAG.getNode(Opcode: MipsISD::TailCall, DL, VT: MVT::Other, Ops); |
| 3613 | DAG.addCallSiteInfo(Node: Ret.getNode(), CallInfo: std::move(CSInfo)); |
| 3614 | return Ret; |
| 3615 | } |
| 3616 | |
| 3617 | Chain = DAG.getNode(Opcode: MipsISD::JmpLink, DL, VTList: NodeTys, Ops); |
| 3618 | SDValue InGlue = Chain.getValue(R: 1); |
| 3619 | |
| 3620 | DAG.addCallSiteInfo(Node: Chain.getNode(), CallInfo: std::move(CSInfo)); |
| 3621 | |
| 3622 | // Create the CALLSEQ_END node in the case of where it is not a call to |
| 3623 | // memcpy. |
| 3624 | if (!(MemcpyInByVal)) { |
| 3625 | Chain = DAG.getCALLSEQ_END(Chain, Size1: StackSize, Size2: 0, Glue: InGlue, DL); |
| 3626 | InGlue = Chain.getValue(R: 1); |
| 3627 | } |
| 3628 | |
| 3629 | // Handle result values, copying them out of physregs into vregs that we |
| 3630 | // return. |
| 3631 | return LowerCallResult(Chain, InGlue, CallConv, isVarArg: IsVarArg, Ins, dl: DL, DAG, |
| 3632 | InVals, CLI); |
| 3633 | } |
| 3634 | |
| 3635 | /// LowerCallResult - Lower the result values of a call into the |
| 3636 | /// appropriate copies out of appropriate physical registers. |
| 3637 | SDValue MipsTargetLowering::LowerCallResult( |
| 3638 | SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool IsVarArg, |
| 3639 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 3640 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
| 3641 | TargetLowering::CallLoweringInfo &CLI) const { |
| 3642 | // Assign locations to each value returned by this call. |
| 3643 | SmallVector<CCValAssign, 16> RVLocs; |
| 3644 | MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
| 3645 | *DAG.getContext()); |
| 3646 | |
| 3647 | CCInfo.AnalyzeCallResult(Ins, Fn: RetCC_Mips); |
| 3648 | |
| 3649 | // Copy all of the result registers out of their specified physreg. |
| 3650 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
| 3651 | CCValAssign &VA = RVLocs[i]; |
| 3652 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
| 3653 | |
| 3654 | SDValue Val = DAG.getCopyFromReg(Chain, dl: DL, Reg: RVLocs[i].getLocReg(), |
| 3655 | VT: RVLocs[i].getLocVT(), Glue: InGlue); |
| 3656 | Chain = Val.getValue(R: 1); |
| 3657 | InGlue = Val.getValue(R: 2); |
| 3658 | |
| 3659 | if (VA.isUpperBitsInLoc()) { |
| 3660 | unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); |
| 3661 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
| 3662 | unsigned Shift = |
| 3663 | VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; |
| 3664 | Val = DAG.getNode( |
| 3665 | Opcode: Shift, DL, VT: VA.getLocVT(), N1: Val, |
| 3666 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
| 3667 | } |
| 3668 | |
| 3669 | switch (VA.getLocInfo()) { |
| 3670 | default: |
| 3671 | llvm_unreachable("Unknown loc info!" ); |
| 3672 | case CCValAssign::Full: |
| 3673 | break; |
| 3674 | case CCValAssign::BCvt: |
| 3675 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getValVT(), Operand: Val); |
| 3676 | break; |
| 3677 | case CCValAssign::AExt: |
| 3678 | case CCValAssign::AExtUpper: |
| 3679 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
| 3680 | break; |
| 3681 | case CCValAssign::ZExt: |
| 3682 | case CCValAssign::ZExtUpper: |
| 3683 | Val = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: Val, |
| 3684 | N2: DAG.getValueType(VA.getValVT())); |
| 3685 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
| 3686 | break; |
| 3687 | case CCValAssign::SExt: |
| 3688 | case CCValAssign::SExtUpper: |
| 3689 | Val = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: Val, |
| 3690 | N2: DAG.getValueType(VA.getValVT())); |
| 3691 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
| 3692 | break; |
| 3693 | } |
| 3694 | |
| 3695 | InVals.push_back(Elt: Val); |
| 3696 | } |
| 3697 | |
| 3698 | return Chain; |
| 3699 | } |
| 3700 | |
| 3701 | static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, |
| 3702 | EVT ArgVT, const SDLoc &DL, |
| 3703 | SelectionDAG &DAG) { |
| 3704 | MVT LocVT = VA.getLocVT(); |
| 3705 | EVT ValVT = VA.getValVT(); |
| 3706 | |
| 3707 | // Shift into the upper bits if necessary. |
| 3708 | switch (VA.getLocInfo()) { |
| 3709 | default: |
| 3710 | break; |
| 3711 | case CCValAssign::AExtUpper: |
| 3712 | case CCValAssign::SExtUpper: |
| 3713 | case CCValAssign::ZExtUpper: { |
| 3714 | unsigned ValSizeInBits = ArgVT.getSizeInBits(); |
| 3715 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
| 3716 | unsigned Opcode = |
| 3717 | VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; |
| 3718 | Val = DAG.getNode( |
| 3719 | Opcode, DL, VT: VA.getLocVT(), N1: Val, |
| 3720 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
| 3721 | break; |
| 3722 | } |
| 3723 | } |
| 3724 | |
| 3725 | // If this is an value smaller than the argument slot size (32-bit for O32, |
| 3726 | // 64-bit for N32/N64), it has been promoted in some way to the argument slot |
| 3727 | // size. Extract the value and insert any appropriate assertions regarding |
| 3728 | // sign/zero extension. |
| 3729 | switch (VA.getLocInfo()) { |
| 3730 | default: |
| 3731 | llvm_unreachable("Unknown loc info!" ); |
| 3732 | case CCValAssign::Full: |
| 3733 | break; |
| 3734 | case CCValAssign::AExtUpper: |
| 3735 | case CCValAssign::AExt: |
| 3736 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
| 3737 | break; |
| 3738 | case CCValAssign::SExtUpper: |
| 3739 | case CCValAssign::SExt: |
| 3740 | Val = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: LocVT, N1: Val, N2: DAG.getValueType(ValVT)); |
| 3741 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
| 3742 | break; |
| 3743 | case CCValAssign::ZExtUpper: |
| 3744 | case CCValAssign::ZExt: |
| 3745 | Val = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: LocVT, N1: Val, N2: DAG.getValueType(ValVT)); |
| 3746 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
| 3747 | break; |
| 3748 | case CCValAssign::BCvt: |
| 3749 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: ValVT, Operand: Val); |
| 3750 | break; |
| 3751 | } |
| 3752 | |
| 3753 | return Val; |
| 3754 | } |
| 3755 | |
| 3756 | //===----------------------------------------------------------------------===// |
| 3757 | // Formal Arguments Calling Convention Implementation |
| 3758 | //===----------------------------------------------------------------------===// |
| 3759 | /// LowerFormalArguments - transform physical registers into virtual registers |
| 3760 | /// and generate load operations for arguments places on the stack. |
| 3761 | SDValue MipsTargetLowering::LowerFormalArguments( |
| 3762 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 3763 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 3764 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 3765 | MachineFunction &MF = DAG.getMachineFunction(); |
| 3766 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 3767 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
| 3768 | |
| 3769 | MipsFI->setVarArgsFrameIndex(0); |
| 3770 | |
| 3771 | // Used with vargs to acumulate store chains. |
| 3772 | std::vector<SDValue> OutChains; |
| 3773 | |
| 3774 | // Assign locations to all of the incoming arguments. |
| 3775 | SmallVector<CCValAssign, 16> ArgLocs; |
| 3776 | MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, |
| 3777 | *DAG.getContext()); |
| 3778 | CCInfo.AllocateStack(Size: ABI.GetCalleeAllocdArgSizeInBytes(CC: CallConv), Alignment: Align(1)); |
| 3779 | const Function &Func = DAG.getMachineFunction().getFunction(); |
| 3780 | Function::const_arg_iterator FuncArg = Func.arg_begin(); |
| 3781 | |
| 3782 | if (Func.hasFnAttribute(Kind: "interrupt" ) && !Func.arg_empty()) |
| 3783 | report_fatal_error( |
| 3784 | reason: "Functions with the interrupt attribute cannot have arguments!" ); |
| 3785 | |
| 3786 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Mips_FixedArg); |
| 3787 | MipsFI->setFormalArgInfo(Size: CCInfo.getStackSize(), |
| 3788 | HasByval: CCInfo.getInRegsParamsCount() > 0); |
| 3789 | |
| 3790 | unsigned CurArgIdx = 0; |
| 3791 | CCInfo.rewindByValRegsInfo(); |
| 3792 | |
| 3793 | for (unsigned i = 0, e = ArgLocs.size(), InsIdx = 0; i != e; ++i, ++InsIdx) { |
| 3794 | CCValAssign &VA = ArgLocs[i]; |
| 3795 | if (Ins[InsIdx].isOrigArg()) { |
| 3796 | std::advance(i&: FuncArg, n: Ins[InsIdx].getOrigArgIndex() - CurArgIdx); |
| 3797 | CurArgIdx = Ins[InsIdx].getOrigArgIndex(); |
| 3798 | } |
| 3799 | EVT ValVT = VA.getValVT(); |
| 3800 | ISD::ArgFlagsTy Flags = Ins[InsIdx].Flags; |
| 3801 | bool IsRegLoc = VA.isRegLoc(); |
| 3802 | |
| 3803 | if (Flags.isByVal()) { |
| 3804 | assert(Ins[InsIdx].isOrigArg() && "Byval arguments cannot be implicit" ); |
| 3805 | unsigned FirstByValReg, LastByValReg; |
| 3806 | unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); |
| 3807 | CCInfo.getInRegsParamInfo(InRegsParamRecordIndex: ByValIdx, BeginReg&: FirstByValReg, EndReg&: LastByValReg); |
| 3808 | |
| 3809 | assert(Flags.getByValSize() && |
| 3810 | "ByVal args of size 0 should have been ignored by front-end." ); |
| 3811 | assert(ByValIdx < CCInfo.getInRegsParamsCount()); |
| 3812 | copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, FuncArg: &*FuncArg, |
| 3813 | FirstReg: FirstByValReg, LastReg: LastByValReg, VA, State&: CCInfo); |
| 3814 | CCInfo.nextInRegsParam(); |
| 3815 | continue; |
| 3816 | } |
| 3817 | |
| 3818 | // Arguments stored on registers |
| 3819 | if (IsRegLoc) { |
| 3820 | MVT RegVT = VA.getLocVT(); |
| 3821 | Register ArgReg = VA.getLocReg(); |
| 3822 | const TargetRegisterClass *RC = getRegClassFor(VT: RegVT); |
| 3823 | |
| 3824 | // Transform the arguments stored on |
| 3825 | // physical registers into virtual ones |
| 3826 | unsigned Reg = addLiveIn(MF&: DAG.getMachineFunction(), PReg: ArgReg, RC); |
| 3827 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RegVT); |
| 3828 | |
| 3829 | ArgValue = |
| 3830 | UnpackFromArgumentSlot(Val: ArgValue, VA, ArgVT: Ins[InsIdx].ArgVT, DL, DAG); |
| 3831 | |
| 3832 | // Handle floating point arguments passed in integer registers and |
| 3833 | // long double arguments passed in floating point registers. |
| 3834 | if ((RegVT == MVT::i32 && ValVT == MVT::f32) || |
| 3835 | (RegVT == MVT::i64 && ValVT == MVT::f64) || |
| 3836 | (RegVT == MVT::f64 && ValVT == MVT::i64)) |
| 3837 | ArgValue = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: ValVT, Operand: ArgValue); |
| 3838 | else if (ABI.IsO32() && RegVT == MVT::i32 && |
| 3839 | ValVT == MVT::f64) { |
| 3840 | assert(VA.needsCustom() && "Expected custom argument for f64 split" ); |
| 3841 | CCValAssign &NextVA = ArgLocs[++i]; |
| 3842 | unsigned Reg2 = |
| 3843 | addLiveIn(MF&: DAG.getMachineFunction(), PReg: NextVA.getLocReg(), RC); |
| 3844 | SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl: DL, Reg: Reg2, VT: RegVT); |
| 3845 | if (!Subtarget.isLittle()) |
| 3846 | std::swap(a&: ArgValue, b&: ArgValue2); |
| 3847 | ArgValue = DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, |
| 3848 | N1: ArgValue, N2: ArgValue2); |
| 3849 | } |
| 3850 | |
| 3851 | InVals.push_back(Elt: ArgValue); |
| 3852 | } else { // VA.isRegLoc() |
| 3853 | MVT LocVT = VA.getLocVT(); |
| 3854 | |
| 3855 | assert(!VA.needsCustom() && "unexpected custom memory argument" ); |
| 3856 | |
| 3857 | // Only arguments pased on the stack should make it here. |
| 3858 | assert(VA.isMemLoc()); |
| 3859 | |
| 3860 | // The stack pointer offset is relative to the caller stack frame. |
| 3861 | int FI = MFI.CreateFixedObject(Size: LocVT.getSizeInBits() / 8, |
| 3862 | SPOffset: VA.getLocMemOffset(), IsImmutable: true); |
| 3863 | |
| 3864 | // Create load nodes to retrieve arguments from the stack |
| 3865 | SDValue FIN = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 3866 | SDValue ArgValue = DAG.getLoad( |
| 3867 | VT: LocVT, dl: DL, Chain, Ptr: FIN, |
| 3868 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI)); |
| 3869 | OutChains.push_back(x: ArgValue.getValue(R: 1)); |
| 3870 | |
| 3871 | ArgValue = |
| 3872 | UnpackFromArgumentSlot(Val: ArgValue, VA, ArgVT: Ins[InsIdx].ArgVT, DL, DAG); |
| 3873 | |
| 3874 | InVals.push_back(Elt: ArgValue); |
| 3875 | } |
| 3876 | } |
| 3877 | |
| 3878 | for (unsigned i = 0, e = ArgLocs.size(), InsIdx = 0; i != e; ++i, ++InsIdx) { |
| 3879 | |
| 3880 | if (ArgLocs[i].needsCustom()) { |
| 3881 | ++i; |
| 3882 | continue; |
| 3883 | } |
| 3884 | |
| 3885 | // The mips ABIs for returning structs by value requires that we copy |
| 3886 | // the sret argument into $v0 for the return. Save the argument into |
| 3887 | // a virtual register so that we can access it from the return points. |
| 3888 | if (Ins[InsIdx].Flags.isSRet()) { |
| 3889 | unsigned Reg = MipsFI->getSRetReturnReg(); |
| 3890 | if (!Reg) { |
| 3891 | Reg = MF.getRegInfo().createVirtualRegister( |
| 3892 | RegClass: getRegClassFor(VT: ABI.IsN64() ? MVT::i64 : MVT::i32)); |
| 3893 | MipsFI->setSRetReturnReg(Reg); |
| 3894 | } |
| 3895 | SDValue Copy = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg, N: InVals[i]); |
| 3896 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Copy, N2: Chain); |
| 3897 | break; |
| 3898 | } |
| 3899 | } |
| 3900 | |
| 3901 | if (IsVarArg) |
| 3902 | writeVarArgRegs(OutChains, Chain, DL, DAG, State&: CCInfo); |
| 3903 | |
| 3904 | // All stores are grouped in one node to allow the matching between |
| 3905 | // the size of Ins and InVals. This only happens when on varg functions |
| 3906 | if (!OutChains.empty()) { |
| 3907 | OutChains.push_back(x: Chain); |
| 3908 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: OutChains); |
| 3909 | } |
| 3910 | |
| 3911 | return Chain; |
| 3912 | } |
| 3913 | |
| 3914 | //===----------------------------------------------------------------------===// |
| 3915 | // Return Value Calling Convention Implementation |
| 3916 | //===----------------------------------------------------------------------===// |
| 3917 | |
| 3918 | bool |
| 3919 | MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, |
| 3920 | MachineFunction &MF, bool IsVarArg, |
| 3921 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 3922 | LLVMContext &Context, const Type *RetTy) const { |
| 3923 | SmallVector<CCValAssign, 16> RVLocs; |
| 3924 | MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); |
| 3925 | return CCInfo.CheckReturn(Outs, Fn: RetCC_Mips); |
| 3926 | } |
| 3927 | |
| 3928 | bool MipsTargetLowering::shouldSignExtendTypeInLibCall(Type *Ty, |
| 3929 | bool IsSigned) const { |
| 3930 | if ((ABI.IsN32() || ABI.IsN64()) && Ty->isIntegerTy(Bitwidth: 32)) |
| 3931 | return true; |
| 3932 | |
| 3933 | return IsSigned; |
| 3934 | } |
| 3935 | |
| 3936 | SDValue |
| 3937 | MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps, |
| 3938 | const SDLoc &DL, |
| 3939 | SelectionDAG &DAG) const { |
| 3940 | MachineFunction &MF = DAG.getMachineFunction(); |
| 3941 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
| 3942 | |
| 3943 | MipsFI->setISR(); |
| 3944 | |
| 3945 | return DAG.getNode(Opcode: MipsISD::ERet, DL, VT: MVT::Other, Ops: RetOps); |
| 3946 | } |
| 3947 | |
| 3948 | SDValue |
| 3949 | MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
| 3950 | bool IsVarArg, |
| 3951 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 3952 | const SmallVectorImpl<SDValue> &OutVals, |
| 3953 | const SDLoc &DL, SelectionDAG &DAG) const { |
| 3954 | // CCValAssign - represent the assignment of |
| 3955 | // the return value to a location |
| 3956 | SmallVector<CCValAssign, 16> RVLocs; |
| 3957 | MachineFunction &MF = DAG.getMachineFunction(); |
| 3958 | |
| 3959 | // CCState - Info about the registers and stack slot. |
| 3960 | MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); |
| 3961 | |
| 3962 | // Analyze return values. |
| 3963 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Mips); |
| 3964 | |
| 3965 | SDValue Glue; |
| 3966 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 3967 | |
| 3968 | // Copy the result values into the output registers. |
| 3969 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
| 3970 | SDValue Val = OutVals[i]; |
| 3971 | CCValAssign &VA = RVLocs[i]; |
| 3972 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
| 3973 | bool UseUpperBits = false; |
| 3974 | |
| 3975 | switch (VA.getLocInfo()) { |
| 3976 | default: |
| 3977 | llvm_unreachable("Unknown loc info!" ); |
| 3978 | case CCValAssign::Full: |
| 3979 | break; |
| 3980 | case CCValAssign::BCvt: |
| 3981 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getLocVT(), Operand: Val); |
| 3982 | break; |
| 3983 | case CCValAssign::AExtUpper: |
| 3984 | UseUpperBits = true; |
| 3985 | [[fallthrough]]; |
| 3986 | case CCValAssign::AExt: |
| 3987 | Val = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
| 3988 | break; |
| 3989 | case CCValAssign::ZExtUpper: |
| 3990 | UseUpperBits = true; |
| 3991 | [[fallthrough]]; |
| 3992 | case CCValAssign::ZExt: |
| 3993 | Val = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
| 3994 | break; |
| 3995 | case CCValAssign::SExtUpper: |
| 3996 | UseUpperBits = true; |
| 3997 | [[fallthrough]]; |
| 3998 | case CCValAssign::SExt: |
| 3999 | Val = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
| 4000 | break; |
| 4001 | } |
| 4002 | |
| 4003 | if (UseUpperBits) { |
| 4004 | unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); |
| 4005 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
| 4006 | Val = DAG.getNode( |
| 4007 | Opcode: ISD::SHL, DL, VT: VA.getLocVT(), N1: Val, |
| 4008 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
| 4009 | } |
| 4010 | |
| 4011 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Val, Glue); |
| 4012 | |
| 4013 | // Guarantee that all emitted copies are stuck together with flags. |
| 4014 | Glue = Chain.getValue(R: 1); |
| 4015 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
| 4016 | } |
| 4017 | |
| 4018 | // The mips ABIs for returning structs by value requires that we copy |
| 4019 | // the sret argument into $v0 for the return. We saved the argument into |
| 4020 | // a virtual register in the entry block, so now we copy the value out |
| 4021 | // and into $v0. |
| 4022 | if (MF.getFunction().hasStructRetAttr()) { |
| 4023 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
| 4024 | unsigned Reg = MipsFI->getSRetReturnReg(); |
| 4025 | |
| 4026 | if (!Reg) |
| 4027 | llvm_unreachable("sret virtual register not created in the entry block" ); |
| 4028 | SDValue Val = |
| 4029 | DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 4030 | unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0; |
| 4031 | |
| 4032 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: V0, N: Val, Glue); |
| 4033 | Glue = Chain.getValue(R: 1); |
| 4034 | RetOps.push_back(Elt: DAG.getRegister(Reg: V0, VT: getPointerTy(DL: DAG.getDataLayout()))); |
| 4035 | } |
| 4036 | |
| 4037 | RetOps[0] = Chain; // Update chain. |
| 4038 | |
| 4039 | // Add the glue if we have it. |
| 4040 | if (Glue.getNode()) |
| 4041 | RetOps.push_back(Elt: Glue); |
| 4042 | |
| 4043 | // ISRs must use "eret". |
| 4044 | if (DAG.getMachineFunction().getFunction().hasFnAttribute(Kind: "interrupt" )) |
| 4045 | return LowerInterruptReturn(RetOps, DL, DAG); |
| 4046 | |
| 4047 | // Standard return on Mips is a "jr $ra" |
| 4048 | return DAG.getNode(Opcode: MipsISD::Ret, DL, VT: MVT::Other, Ops: RetOps); |
| 4049 | } |
| 4050 | |
| 4051 | //===----------------------------------------------------------------------===// |
| 4052 | // Mips Inline Assembly Support |
| 4053 | //===----------------------------------------------------------------------===// |
| 4054 | |
| 4055 | /// getConstraintType - Given a constraint letter, return the type of |
| 4056 | /// constraint it is for this target. |
| 4057 | MipsTargetLowering::ConstraintType |
| 4058 | MipsTargetLowering::getConstraintType(StringRef Constraint) const { |
| 4059 | // Mips specific constraints |
| 4060 | // GCC config/mips/constraints.md |
| 4061 | // |
| 4062 | // 'd' : An address register. Equivalent to r |
| 4063 | // unless generating MIPS16 code. |
| 4064 | // 'y' : Equivalent to r; retained for |
| 4065 | // backwards compatibility. |
| 4066 | // 'c' : A register suitable for use in an indirect |
| 4067 | // jump. This will always be $25 for -mabicalls. |
| 4068 | // 'l' : The lo register. 1 word storage. |
| 4069 | // 'x' : The hilo register pair. Double word storage. |
| 4070 | if (Constraint.size() == 1) { |
| 4071 | switch (Constraint[0]) { |
| 4072 | default : break; |
| 4073 | case 'd': |
| 4074 | case 'y': |
| 4075 | case 'f': |
| 4076 | case 'c': |
| 4077 | case 'l': |
| 4078 | case 'x': |
| 4079 | return C_RegisterClass; |
| 4080 | case 'R': |
| 4081 | return C_Memory; |
| 4082 | } |
| 4083 | } |
| 4084 | |
| 4085 | if (Constraint == "ZC" ) |
| 4086 | return C_Memory; |
| 4087 | |
| 4088 | return TargetLowering::getConstraintType(Constraint); |
| 4089 | } |
| 4090 | |
| 4091 | /// Examine constraint type and operand type and determine a weight value. |
| 4092 | /// This object must already have been set up with the operand type |
| 4093 | /// and the current alternative constraint selected. |
| 4094 | TargetLowering::ConstraintWeight |
| 4095 | MipsTargetLowering::getSingleConstraintMatchWeight( |
| 4096 | AsmOperandInfo &info, const char *constraint) const { |
| 4097 | ConstraintWeight weight = CW_Invalid; |
| 4098 | Value *CallOperandVal = info.CallOperandVal; |
| 4099 | // If we don't have a value, we can't do a match, |
| 4100 | // but allow it at the lowest weight. |
| 4101 | if (!CallOperandVal) |
| 4102 | return CW_Default; |
| 4103 | Type *type = CallOperandVal->getType(); |
| 4104 | // Look at the constraint type. |
| 4105 | switch (*constraint) { |
| 4106 | default: |
| 4107 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
| 4108 | break; |
| 4109 | case 'd': |
| 4110 | case 'y': |
| 4111 | if (type->isIntegerTy()) |
| 4112 | weight = CW_Register; |
| 4113 | break; |
| 4114 | case 'f': // FPU or MSA register |
| 4115 | if (Subtarget.hasMSA() && type->isVectorTy() && |
| 4116 | type->getPrimitiveSizeInBits().getFixedValue() == 128) |
| 4117 | weight = CW_Register; |
| 4118 | else if (type->isFloatTy()) |
| 4119 | weight = CW_Register; |
| 4120 | break; |
| 4121 | case 'c': // $25 for indirect jumps |
| 4122 | case 'l': // lo register |
| 4123 | case 'x': // hilo register pair |
| 4124 | if (type->isIntegerTy()) |
| 4125 | weight = CW_SpecificReg; |
| 4126 | break; |
| 4127 | case 'I': // signed 16 bit immediate |
| 4128 | case 'J': // integer zero |
| 4129 | case 'K': // unsigned 16 bit immediate |
| 4130 | case 'L': // signed 32 bit immediate where lower 16 bits are 0 |
| 4131 | case 'N': // immediate in the range of -65535 to -1 (inclusive) |
| 4132 | case 'O': // signed 15 bit immediate (+- 16383) |
| 4133 | case 'P': // immediate in the range of 65535 to 1 (inclusive) |
| 4134 | if (isa<ConstantInt>(Val: CallOperandVal)) |
| 4135 | weight = CW_Constant; |
| 4136 | break; |
| 4137 | case 'R': |
| 4138 | weight = CW_Memory; |
| 4139 | break; |
| 4140 | } |
| 4141 | return weight; |
| 4142 | } |
| 4143 | |
| 4144 | /// This is a helper function to parse a physical register string and split it |
| 4145 | /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag |
| 4146 | /// that is returned indicates whether parsing was successful. The second flag |
| 4147 | /// is true if the numeric part exists. |
| 4148 | static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix, |
| 4149 | unsigned long long &Reg) { |
| 4150 | if (C.front() != '{' || C.back() != '}') |
| 4151 | return std::make_pair(x: false, y: false); |
| 4152 | |
| 4153 | // Search for the first numeric character. |
| 4154 | StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; |
| 4155 | I = std::find_if(first: B, last: E, pred: isdigit); |
| 4156 | |
| 4157 | Prefix = StringRef(B, I - B); |
| 4158 | |
| 4159 | // The second flag is set to false if no numeric characters were found. |
| 4160 | if (I == E) |
| 4161 | return std::make_pair(x: true, y: false); |
| 4162 | |
| 4163 | // Parse the numeric characters. |
| 4164 | return std::make_pair(x: !getAsUnsignedInteger(Str: StringRef(I, E - I), Radix: 10, Result&: Reg), |
| 4165 | y: true); |
| 4166 | } |
| 4167 | |
| 4168 | EVT MipsTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, |
| 4169 | ISD::NodeType) const { |
| 4170 | bool Cond = !Subtarget.isABI_O32() && VT.getSizeInBits() == 32; |
| 4171 | EVT MinVT = getRegisterType(VT: Cond ? MVT::i64 : MVT::i32); |
| 4172 | return VT.bitsLT(VT: MinVT) ? MinVT : VT; |
| 4173 | } |
| 4174 | |
| 4175 | std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: |
| 4176 | parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { |
| 4177 | const TargetRegisterInfo *TRI = |
| 4178 | Subtarget.getRegisterInfo(); |
| 4179 | const TargetRegisterClass *RC; |
| 4180 | StringRef Prefix; |
| 4181 | unsigned long long Reg; |
| 4182 | |
| 4183 | std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); |
| 4184 | |
| 4185 | if (!R.first) |
| 4186 | return std::make_pair(x: 0U, y: nullptr); |
| 4187 | |
| 4188 | if ((Prefix == "hi" || Prefix == "lo" )) { // Parse hi/lo. |
| 4189 | // No numeric characters follow "hi" or "lo". |
| 4190 | if (R.second) |
| 4191 | return std::make_pair(x: 0U, y: nullptr); |
| 4192 | |
| 4193 | RC = TRI->getRegClass(i: Prefix == "hi" ? |
| 4194 | Mips::HI32RegClassID : Mips::LO32RegClassID); |
| 4195 | return std::make_pair(x: *(RC->begin()), y&: RC); |
| 4196 | } else if (Prefix.starts_with(Prefix: "$msa" )) { |
| 4197 | // Parse $msa(ir|csr|access|save|modify|request|map|unmap) |
| 4198 | |
| 4199 | // No numeric characters follow the name. |
| 4200 | if (R.second) |
| 4201 | return std::make_pair(x: 0U, y: nullptr); |
| 4202 | |
| 4203 | Reg = StringSwitch<unsigned long long>(Prefix) |
| 4204 | .Case(S: "$msair" , Value: Mips::MSAIR) |
| 4205 | .Case(S: "$msacsr" , Value: Mips::MSACSR) |
| 4206 | .Case(S: "$msaaccess" , Value: Mips::MSAAccess) |
| 4207 | .Case(S: "$msasave" , Value: Mips::MSASave) |
| 4208 | .Case(S: "$msamodify" , Value: Mips::MSAModify) |
| 4209 | .Case(S: "$msarequest" , Value: Mips::MSARequest) |
| 4210 | .Case(S: "$msamap" , Value: Mips::MSAMap) |
| 4211 | .Case(S: "$msaunmap" , Value: Mips::MSAUnmap) |
| 4212 | .Default(Value: 0); |
| 4213 | |
| 4214 | if (!Reg) |
| 4215 | return std::make_pair(x: 0U, y: nullptr); |
| 4216 | |
| 4217 | RC = TRI->getRegClass(i: Mips::MSACtrlRegClassID); |
| 4218 | return std::make_pair(x&: Reg, y&: RC); |
| 4219 | } |
| 4220 | |
| 4221 | if (!R.second) |
| 4222 | return std::make_pair(x: 0U, y: nullptr); |
| 4223 | |
| 4224 | if (Prefix == "$f" ) { // Parse $f0-$f31. |
| 4225 | // If the targets is single float only, always select 32-bit registers, |
| 4226 | // otherwise if the size of FP registers is 64-bit or Reg is an even number, |
| 4227 | // select the 64-bit register class. Otherwise, select the 32-bit register |
| 4228 | // class. |
| 4229 | if (VT == MVT::Other) { |
| 4230 | if (Subtarget.isSingleFloat()) |
| 4231 | VT = MVT::f32; |
| 4232 | else |
| 4233 | VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; |
| 4234 | } |
| 4235 | |
| 4236 | RC = getRegClassFor(VT); |
| 4237 | |
| 4238 | if (RC == &Mips::AFGR64RegClass) { |
| 4239 | assert(Reg % 2 == 0); |
| 4240 | Reg >>= 1; |
| 4241 | } |
| 4242 | } else if (Prefix == "$fcc" ) // Parse $fcc0-$fcc7. |
| 4243 | RC = TRI->getRegClass(i: Mips::FCCRegClassID); |
| 4244 | else if (Prefix == "$w" ) { // Parse $w0-$w31. |
| 4245 | RC = getRegClassFor(VT: (VT == MVT::Other) ? MVT::v16i8 : VT); |
| 4246 | } else { // Parse $0-$31. |
| 4247 | assert(Prefix == "$" ); |
| 4248 | RC = getRegClassFor(VT: (VT == MVT::Other) ? MVT::i32 : VT); |
| 4249 | } |
| 4250 | |
| 4251 | assert(Reg < RC->getNumRegs()); |
| 4252 | return std::make_pair(x: *(RC->begin() + Reg), y&: RC); |
| 4253 | } |
| 4254 | |
| 4255 | /// Given a register class constraint, like 'r', if this corresponds directly |
| 4256 | /// to an LLVM register class, return a register of 0 and the register class |
| 4257 | /// pointer. |
| 4258 | std::pair<unsigned, const TargetRegisterClass *> |
| 4259 | MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
| 4260 | StringRef Constraint, |
| 4261 | MVT VT) const { |
| 4262 | if (Constraint.size() == 1) { |
| 4263 | switch (Constraint[0]) { |
| 4264 | case 'd': // Address register. Same as 'r' unless generating MIPS16 code. |
| 4265 | case 'y': // Same as 'r'. Exists for compatibility. |
| 4266 | case 'r': |
| 4267 | if ((VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8 || |
| 4268 | VT == MVT::i1) || |
| 4269 | (VT == MVT::f32 && Subtarget.useSoftFloat())) { |
| 4270 | if (Subtarget.inMips16Mode()) |
| 4271 | return std::make_pair(x: 0U, y: &Mips::CPU16RegsRegClass); |
| 4272 | return std::make_pair(x: 0U, y: &Mips::GPR32RegClass); |
| 4273 | } |
| 4274 | if ((VT == MVT::i64 || (VT == MVT::f64 && Subtarget.useSoftFloat()) || |
| 4275 | (VT == MVT::f64 && Subtarget.isSingleFloat())) && |
| 4276 | !Subtarget.isGP64bit()) |
| 4277 | return std::make_pair(x: 0U, y: &Mips::GPR32RegClass); |
| 4278 | if ((VT == MVT::i64 || (VT == MVT::f64 && Subtarget.useSoftFloat()) || |
| 4279 | (VT == MVT::f64 && Subtarget.isSingleFloat())) && |
| 4280 | Subtarget.isGP64bit()) |
| 4281 | return std::make_pair(x: 0U, y: &Mips::GPR64RegClass); |
| 4282 | // This will generate an error message |
| 4283 | return std::make_pair(x: 0U, y: nullptr); |
| 4284 | case 'f': // FPU or MSA register |
| 4285 | if (VT == MVT::v16i8) |
| 4286 | return std::make_pair(x: 0U, y: &Mips::MSA128BRegClass); |
| 4287 | else if (VT == MVT::v8i16 || VT == MVT::v8f16) |
| 4288 | return std::make_pair(x: 0U, y: &Mips::MSA128HRegClass); |
| 4289 | else if (VT == MVT::v4i32 || VT == MVT::v4f32) |
| 4290 | return std::make_pair(x: 0U, y: &Mips::MSA128WRegClass); |
| 4291 | else if (VT == MVT::v2i64 || VT == MVT::v2f64) |
| 4292 | return std::make_pair(x: 0U, y: &Mips::MSA128DRegClass); |
| 4293 | else if (VT == MVT::f32) |
| 4294 | return std::make_pair(x: 0U, y: &Mips::FGR32RegClass); |
| 4295 | else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { |
| 4296 | if (Subtarget.isFP64bit()) |
| 4297 | return std::make_pair(x: 0U, y: &Mips::FGR64RegClass); |
| 4298 | return std::make_pair(x: 0U, y: &Mips::AFGR64RegClass); |
| 4299 | } |
| 4300 | break; |
| 4301 | case 'c': // register suitable for indirect jump |
| 4302 | if (VT == MVT::i32) |
| 4303 | return std::make_pair(x: (unsigned)Mips::T9, y: &Mips::GPR32RegClass); |
| 4304 | if (VT == MVT::i64) |
| 4305 | return std::make_pair(x: (unsigned)Mips::T9_64, y: &Mips::GPR64RegClass); |
| 4306 | // This will generate an error message |
| 4307 | return std::make_pair(x: 0U, y: nullptr); |
| 4308 | case 'l': // use the `lo` register to store values |
| 4309 | // that are no bigger than a word |
| 4310 | if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) |
| 4311 | return std::make_pair(x: (unsigned)Mips::LO0, y: &Mips::LO32RegClass); |
| 4312 | return std::make_pair(x: (unsigned)Mips::LO0_64, y: &Mips::LO64RegClass); |
| 4313 | case 'x': // use the concatenated `hi` and `lo` registers |
| 4314 | // to store doubleword values |
| 4315 | // Fixme: Not triggering the use of both hi and low |
| 4316 | // This will generate an error message |
| 4317 | return std::make_pair(x: 0U, y: nullptr); |
| 4318 | } |
| 4319 | } |
| 4320 | |
| 4321 | if (!Constraint.empty()) { |
| 4322 | std::pair<unsigned, const TargetRegisterClass *> R; |
| 4323 | R = parseRegForInlineAsmConstraint(C: Constraint, VT); |
| 4324 | |
| 4325 | if (R.second) |
| 4326 | return R; |
| 4327 | } |
| 4328 | |
| 4329 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
| 4330 | } |
| 4331 | |
| 4332 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
| 4333 | /// vector. If it is invalid, don't add anything to Ops. |
| 4334 | void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
| 4335 | StringRef Constraint, |
| 4336 | std::vector<SDValue> &Ops, |
| 4337 | SelectionDAG &DAG) const { |
| 4338 | SDLoc DL(Op); |
| 4339 | SDValue Result; |
| 4340 | |
| 4341 | // Only support length 1 constraints for now. |
| 4342 | if (Constraint.size() > 1) |
| 4343 | return; |
| 4344 | |
| 4345 | char ConstraintLetter = Constraint[0]; |
| 4346 | switch (ConstraintLetter) { |
| 4347 | default: break; // This will fall through to the generic implementation |
| 4348 | case 'I': // Signed 16 bit constant |
| 4349 | // If this fails, the parent routine will give an error |
| 4350 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4351 | EVT Type = Op.getValueType(); |
| 4352 | int64_t Val = C->getSExtValue(); |
| 4353 | if (isInt<16>(x: Val)) { |
| 4354 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
| 4355 | break; |
| 4356 | } |
| 4357 | } |
| 4358 | return; |
| 4359 | case 'J': // integer zero |
| 4360 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4361 | EVT Type = Op.getValueType(); |
| 4362 | int64_t Val = C->getZExtValue(); |
| 4363 | if (Val == 0) { |
| 4364 | Result = DAG.getTargetConstant(Val: 0, DL, VT: Type); |
| 4365 | break; |
| 4366 | } |
| 4367 | } |
| 4368 | return; |
| 4369 | case 'K': // unsigned 16 bit immediate |
| 4370 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4371 | EVT Type = Op.getValueType(); |
| 4372 | uint64_t Val = C->getZExtValue(); |
| 4373 | if (isUInt<16>(x: Val)) { |
| 4374 | Result = DAG.getTargetConstant(Val, DL, VT: Type); |
| 4375 | break; |
| 4376 | } |
| 4377 | } |
| 4378 | return; |
| 4379 | case 'L': // signed 32 bit immediate where lower 16 bits are 0 |
| 4380 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4381 | EVT Type = Op.getValueType(); |
| 4382 | int64_t Val = C->getSExtValue(); |
| 4383 | if ((isInt<32>(x: Val)) && ((Val & 0xffff) == 0)){ |
| 4384 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
| 4385 | break; |
| 4386 | } |
| 4387 | } |
| 4388 | return; |
| 4389 | case 'N': // immediate in the range of -65535 to -1 (inclusive) |
| 4390 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4391 | EVT Type = Op.getValueType(); |
| 4392 | int64_t Val = C->getSExtValue(); |
| 4393 | if ((Val >= -65535) && (Val <= -1)) { |
| 4394 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
| 4395 | break; |
| 4396 | } |
| 4397 | } |
| 4398 | return; |
| 4399 | case 'O': // signed 15 bit immediate |
| 4400 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4401 | EVT Type = Op.getValueType(); |
| 4402 | int64_t Val = C->getSExtValue(); |
| 4403 | if ((isInt<15>(x: Val))) { |
| 4404 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
| 4405 | break; |
| 4406 | } |
| 4407 | } |
| 4408 | return; |
| 4409 | case 'P': // immediate in the range of 1 to 65535 (inclusive) |
| 4410 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 4411 | EVT Type = Op.getValueType(); |
| 4412 | int64_t Val = C->getSExtValue(); |
| 4413 | if ((Val <= 65535) && (Val >= 1)) { |
| 4414 | Result = DAG.getTargetConstant(Val, DL, VT: Type); |
| 4415 | break; |
| 4416 | } |
| 4417 | } |
| 4418 | return; |
| 4419 | } |
| 4420 | |
| 4421 | if (Result.getNode()) { |
| 4422 | Ops.push_back(x: Result); |
| 4423 | return; |
| 4424 | } |
| 4425 | |
| 4426 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
| 4427 | } |
| 4428 | |
| 4429 | bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
| 4430 | const AddrMode &AM, Type *Ty, |
| 4431 | unsigned AS, |
| 4432 | Instruction *I) const { |
| 4433 | // No global is ever allowed as a base. |
| 4434 | if (AM.BaseGV) |
| 4435 | return false; |
| 4436 | |
| 4437 | switch (AM.Scale) { |
| 4438 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
| 4439 | break; |
| 4440 | case 1: |
| 4441 | if (!AM.HasBaseReg) // allow "r+i". |
| 4442 | break; |
| 4443 | return false; // disallow "r+r" or "r+r+i". |
| 4444 | default: |
| 4445 | return false; |
| 4446 | } |
| 4447 | |
| 4448 | return true; |
| 4449 | } |
| 4450 | |
| 4451 | bool |
| 4452 | MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
| 4453 | // The Mips target isn't yet aware of offsets. |
| 4454 | return false; |
| 4455 | } |
| 4456 | |
| 4457 | EVT MipsTargetLowering::getOptimalMemOpType( |
| 4458 | LLVMContext &Context, const MemOp &Op, |
| 4459 | const AttributeList &FuncAttributes) const { |
| 4460 | if (Subtarget.hasMips64()) |
| 4461 | return MVT::i64; |
| 4462 | |
| 4463 | return MVT::i32; |
| 4464 | } |
| 4465 | |
| 4466 | bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
| 4467 | bool ForCodeSize) const { |
| 4468 | if (VT != MVT::f32 && VT != MVT::f64) |
| 4469 | return false; |
| 4470 | if (Imm.isNegZero()) |
| 4471 | return false; |
| 4472 | return Imm.isZero(); |
| 4473 | } |
| 4474 | |
| 4475 | bool MipsTargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
| 4476 | return isInt<16>(x: Imm); |
| 4477 | } |
| 4478 | |
| 4479 | bool MipsTargetLowering::isLegalAddImmediate(int64_t Imm) const { |
| 4480 | return isInt<16>(x: Imm); |
| 4481 | } |
| 4482 | |
| 4483 | unsigned MipsTargetLowering::getJumpTableEncoding() const { |
| 4484 | if (!isPositionIndependent()) |
| 4485 | return MachineJumpTableInfo::EK_BlockAddress; |
| 4486 | if (ABI.IsN64()) |
| 4487 | return MachineJumpTableInfo::EK_GPRel64BlockAddress; |
| 4488 | return MachineJumpTableInfo::EK_GPRel32BlockAddress; |
| 4489 | } |
| 4490 | |
| 4491 | SDValue MipsTargetLowering::getPICJumpTableRelocBase(SDValue Table, |
| 4492 | SelectionDAG &DAG) const { |
| 4493 | if (!isPositionIndependent()) |
| 4494 | return Table; |
| 4495 | return DAG.getGLOBAL_OFFSET_TABLE(VT: getPointerTy(DL: DAG.getDataLayout())); |
| 4496 | } |
| 4497 | |
| 4498 | bool MipsTargetLowering::useSoftFloat() const { |
| 4499 | return Subtarget.useSoftFloat(); |
| 4500 | } |
| 4501 | |
| 4502 | void MipsTargetLowering::copyByValRegs( |
| 4503 | SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains, |
| 4504 | SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, |
| 4505 | SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, |
| 4506 | unsigned FirstReg, unsigned LastReg, const CCValAssign &VA, |
| 4507 | MipsCCState &State) const { |
| 4508 | MachineFunction &MF = DAG.getMachineFunction(); |
| 4509 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 4510 | unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); |
| 4511 | unsigned NumRegs = LastReg - FirstReg; |
| 4512 | unsigned RegAreaSize = NumRegs * GPRSizeInBytes; |
| 4513 | unsigned FrameObjSize = std::max(a: Flags.getByValSize(), b: RegAreaSize); |
| 4514 | int FrameObjOffset; |
| 4515 | ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); |
| 4516 | |
| 4517 | if (RegAreaSize) |
| 4518 | FrameObjOffset = |
| 4519 | (int)ABI.GetCalleeAllocdArgSizeInBytes(CC: State.getCallingConv()) - |
| 4520 | (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); |
| 4521 | else |
| 4522 | FrameObjOffset = VA.getLocMemOffset(); |
| 4523 | |
| 4524 | // Create frame object. |
| 4525 | EVT PtrTy = getPointerTy(DL: DAG.getDataLayout()); |
| 4526 | // Make the fixed object stored to mutable so that the load instructions |
| 4527 | // referencing it have their memory dependencies added. |
| 4528 | // Set the frame object as isAliased which clears the underlying objects |
| 4529 | // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all |
| 4530 | // stores as dependencies for loads referencing this fixed object. |
| 4531 | int FI = MFI.CreateFixedObject(Size: FrameObjSize, SPOffset: FrameObjOffset, IsImmutable: false, isAliased: true); |
| 4532 | SDValue FIN = DAG.getFrameIndex(FI, VT: PtrTy); |
| 4533 | InVals.push_back(Elt: FIN); |
| 4534 | |
| 4535 | if (!NumRegs) |
| 4536 | return; |
| 4537 | |
| 4538 | // Copy arg registers. |
| 4539 | MVT RegTy = MVT::getIntegerVT(BitWidth: GPRSizeInBytes * 8); |
| 4540 | const TargetRegisterClass *RC = getRegClassFor(VT: RegTy); |
| 4541 | |
| 4542 | for (unsigned I = 0; I < NumRegs; ++I) { |
| 4543 | unsigned ArgReg = ByValArgRegs[FirstReg + I]; |
| 4544 | unsigned VReg = addLiveIn(MF, PReg: ArgReg, RC); |
| 4545 | unsigned Offset = I * GPRSizeInBytes; |
| 4546 | SDValue StorePtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: FIN, |
| 4547 | N2: DAG.getConstant(Val: Offset, DL, VT: PtrTy)); |
| 4548 | SDValue Store = DAG.getStore(Chain, dl: DL, Val: DAG.getRegister(Reg: VReg, VT: RegTy), |
| 4549 | Ptr: StorePtr, PtrInfo: MachinePointerInfo(FuncArg, Offset)); |
| 4550 | OutChains.push_back(x: Store); |
| 4551 | } |
| 4552 | } |
| 4553 | |
| 4554 | // Copy byVal arg to registers and stack. |
| 4555 | void MipsTargetLowering::passByValArg( |
| 4556 | SDValue Chain, const SDLoc &DL, |
| 4557 | std::deque<std::pair<unsigned, SDValue>> &RegsToPass, |
| 4558 | SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, |
| 4559 | MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, |
| 4560 | unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, |
| 4561 | const CCValAssign &VA) const { |
| 4562 | unsigned ByValSizeInBytes = Flags.getByValSize(); |
| 4563 | unsigned OffsetInBytes = 0; // From beginning of struct |
| 4564 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
| 4565 | Align Alignment = |
| 4566 | std::min(a: Flags.getNonZeroByValAlign(), b: Align(RegSizeInBytes)); |
| 4567 | EVT PtrTy = getPointerTy(DL: DAG.getDataLayout()), |
| 4568 | RegTy = MVT::getIntegerVT(BitWidth: RegSizeInBytes * 8); |
| 4569 | unsigned NumRegs = LastReg - FirstReg; |
| 4570 | |
| 4571 | if (NumRegs) { |
| 4572 | ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs(); |
| 4573 | bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); |
| 4574 | unsigned I = 0; |
| 4575 | |
| 4576 | // Copy words to registers. |
| 4577 | for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { |
| 4578 | SDValue LoadPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
| 4579 | N2: DAG.getConstant(Val: OffsetInBytes, DL, VT: PtrTy)); |
| 4580 | SDValue LoadVal = DAG.getLoad(VT: RegTy, dl: DL, Chain, Ptr: LoadPtr, |
| 4581 | PtrInfo: MachinePointerInfo(), Alignment); |
| 4582 | MemOpChains.push_back(Elt: LoadVal.getValue(R: 1)); |
| 4583 | unsigned ArgReg = ArgRegs[FirstReg + I]; |
| 4584 | RegsToPass.push_back(x: std::make_pair(x&: ArgReg, y&: LoadVal)); |
| 4585 | } |
| 4586 | |
| 4587 | // Return if the struct has been fully copied. |
| 4588 | if (ByValSizeInBytes == OffsetInBytes) |
| 4589 | return; |
| 4590 | |
| 4591 | // Copy the remainder of the byval argument with sub-word loads and shifts. |
| 4592 | if (LeftoverBytes) { |
| 4593 | SDValue Val; |
| 4594 | |
| 4595 | for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; |
| 4596 | OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { |
| 4597 | unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; |
| 4598 | |
| 4599 | if (RemainingSizeInBytes < LoadSizeInBytes) |
| 4600 | continue; |
| 4601 | |
| 4602 | // Load subword. |
| 4603 | SDValue LoadPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
| 4604 | N2: DAG.getConstant(Val: OffsetInBytes, DL, |
| 4605 | VT: PtrTy)); |
| 4606 | SDValue LoadVal = DAG.getExtLoad( |
| 4607 | ExtType: ISD::ZEXTLOAD, dl: DL, VT: RegTy, Chain, Ptr: LoadPtr, PtrInfo: MachinePointerInfo(), |
| 4608 | MemVT: MVT::getIntegerVT(BitWidth: LoadSizeInBytes * 8), Alignment); |
| 4609 | MemOpChains.push_back(Elt: LoadVal.getValue(R: 1)); |
| 4610 | |
| 4611 | // Shift the loaded value. |
| 4612 | unsigned Shamt; |
| 4613 | |
| 4614 | if (isLittle) |
| 4615 | Shamt = TotalBytesLoaded * 8; |
| 4616 | else |
| 4617 | Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; |
| 4618 | |
| 4619 | SDValue Shift = DAG.getNode(Opcode: ISD::SHL, DL, VT: RegTy, N1: LoadVal, |
| 4620 | N2: DAG.getConstant(Val: Shamt, DL, VT: MVT::i32)); |
| 4621 | |
| 4622 | if (Val.getNode()) |
| 4623 | Val = DAG.getNode(Opcode: ISD::OR, DL, VT: RegTy, N1: Val, N2: Shift); |
| 4624 | else |
| 4625 | Val = Shift; |
| 4626 | |
| 4627 | OffsetInBytes += LoadSizeInBytes; |
| 4628 | TotalBytesLoaded += LoadSizeInBytes; |
| 4629 | Alignment = std::min(a: Alignment, b: Align(LoadSizeInBytes)); |
| 4630 | } |
| 4631 | |
| 4632 | unsigned ArgReg = ArgRegs[FirstReg + I]; |
| 4633 | RegsToPass.push_back(x: std::make_pair(x&: ArgReg, y&: Val)); |
| 4634 | return; |
| 4635 | } |
| 4636 | } |
| 4637 | |
| 4638 | // Copy remainder of byval arg to it with memcpy. |
| 4639 | unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; |
| 4640 | SDValue Src = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
| 4641 | N2: DAG.getConstant(Val: OffsetInBytes, DL, VT: PtrTy)); |
| 4642 | SDValue Dst = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: StackPtr, |
| 4643 | N2: DAG.getIntPtrConstant(Val: VA.getLocMemOffset(), DL)); |
| 4644 | Chain = DAG.getMemcpy( |
| 4645 | Chain, dl: DL, Dst, Src, Size: DAG.getConstant(Val: MemCpySize, DL, VT: PtrTy), |
| 4646 | Alignment: Align(Alignment), /*isVolatile=*/isVol: false, /*AlwaysInline=*/false, |
| 4647 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), SrcPtrInfo: MachinePointerInfo()); |
| 4648 | MemOpChains.push_back(Elt: Chain); |
| 4649 | } |
| 4650 | |
| 4651 | void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, |
| 4652 | SDValue Chain, const SDLoc &DL, |
| 4653 | SelectionDAG &DAG, |
| 4654 | CCState &State) const { |
| 4655 | ArrayRef<MCPhysReg> ArgRegs = ABI.getVarArgRegs(isGP64bit: Subtarget.isGP64bit()); |
| 4656 | unsigned Idx = State.getFirstUnallocated(Regs: ArgRegs); |
| 4657 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
| 4658 | MVT RegTy = MVT::getIntegerVT(BitWidth: RegSizeInBytes * 8); |
| 4659 | const TargetRegisterClass *RC = getRegClassFor(VT: RegTy); |
| 4660 | MachineFunction &MF = DAG.getMachineFunction(); |
| 4661 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 4662 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
| 4663 | |
| 4664 | // Offset of the first variable argument from stack pointer. |
| 4665 | int VaArgOffset; |
| 4666 | |
| 4667 | if (ArgRegs.size() == Idx) |
| 4668 | VaArgOffset = alignTo(Value: State.getStackSize(), Align: RegSizeInBytes); |
| 4669 | else { |
| 4670 | VaArgOffset = |
| 4671 | (int)ABI.GetCalleeAllocdArgSizeInBytes(CC: State.getCallingConv()) - |
| 4672 | (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); |
| 4673 | } |
| 4674 | |
| 4675 | // Record the frame index of the first variable argument |
| 4676 | // which is a value necessary to VASTART. |
| 4677 | int FI = MFI.CreateFixedObject(Size: RegSizeInBytes, SPOffset: VaArgOffset, IsImmutable: true); |
| 4678 | MipsFI->setVarArgsFrameIndex(FI); |
| 4679 | |
| 4680 | // Copy the integer registers that have not been used for argument passing |
| 4681 | // to the argument register save area. For O32, the save area is allocated |
| 4682 | // in the caller's stack frame, while for N32/64, it is allocated in the |
| 4683 | // callee's stack frame. |
| 4684 | for (unsigned I = Idx; I < ArgRegs.size(); |
| 4685 | ++I, VaArgOffset += RegSizeInBytes) { |
| 4686 | unsigned Reg = addLiveIn(MF, PReg: ArgRegs[I], RC); |
| 4687 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RegTy); |
| 4688 | FI = MFI.CreateFixedObject(Size: RegSizeInBytes, SPOffset: VaArgOffset, IsImmutable: true); |
| 4689 | SDValue PtrOff = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 4690 | SDValue Store = |
| 4691 | DAG.getStore(Chain, dl: DL, Val: ArgValue, Ptr: PtrOff, PtrInfo: MachinePointerInfo()); |
| 4692 | cast<StoreSDNode>(Val: Store.getNode())->getMemOperand()->setValue( |
| 4693 | (Value *)nullptr); |
| 4694 | OutChains.push_back(x: Store); |
| 4695 | } |
| 4696 | } |
| 4697 | |
| 4698 | void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, |
| 4699 | Align Alignment) const { |
| 4700 | const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); |
| 4701 | |
| 4702 | assert(Size && "Byval argument's size shouldn't be 0." ); |
| 4703 | |
| 4704 | Alignment = std::min(a: Alignment, b: TFL->getStackAlign()); |
| 4705 | |
| 4706 | unsigned FirstReg = 0; |
| 4707 | unsigned NumRegs = 0; |
| 4708 | |
| 4709 | if (State->getCallingConv() != CallingConv::Fast) { |
| 4710 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
| 4711 | ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs(); |
| 4712 | // FIXME: The O32 case actually describes no shadow registers. |
| 4713 | const MCPhysReg *ShadowRegs = |
| 4714 | ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs; |
| 4715 | |
| 4716 | // We used to check the size as well but we can't do that anymore since |
| 4717 | // CCState::HandleByVal() rounds up the size after calling this function. |
| 4718 | assert( |
| 4719 | Alignment >= Align(RegSizeInBytes) && |
| 4720 | "Byval argument's alignment should be a multiple of RegSizeInBytes." ); |
| 4721 | |
| 4722 | FirstReg = State->getFirstUnallocated(Regs: IntArgRegs); |
| 4723 | |
| 4724 | // If Alignment > RegSizeInBytes, the first arg register must be even. |
| 4725 | // FIXME: This condition happens to do the right thing but it's not the |
| 4726 | // right way to test it. We want to check that the stack frame offset |
| 4727 | // of the register is aligned. |
| 4728 | if ((Alignment > RegSizeInBytes) && (FirstReg % 2)) { |
| 4729 | State->AllocateReg(Reg: IntArgRegs[FirstReg], ShadowReg: ShadowRegs[FirstReg]); |
| 4730 | ++FirstReg; |
| 4731 | } |
| 4732 | |
| 4733 | // Mark the registers allocated. |
| 4734 | Size = alignTo(Value: Size, Align: RegSizeInBytes); |
| 4735 | for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); |
| 4736 | Size -= RegSizeInBytes, ++I, ++NumRegs) |
| 4737 | State->AllocateReg(Reg: IntArgRegs[I], ShadowReg: ShadowRegs[I]); |
| 4738 | } |
| 4739 | |
| 4740 | State->addInRegsParamInfo(RegBegin: FirstReg, RegEnd: FirstReg + NumRegs); |
| 4741 | } |
| 4742 | |
| 4743 | MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI, |
| 4744 | MachineBasicBlock *BB, |
| 4745 | bool isFPCmp, |
| 4746 | unsigned Opc) const { |
| 4747 | assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && |
| 4748 | "Subtarget already supports SELECT nodes with the use of" |
| 4749 | "conditional-move instructions." ); |
| 4750 | |
| 4751 | const TargetInstrInfo *TII = |
| 4752 | Subtarget.getInstrInfo(); |
| 4753 | DebugLoc DL = MI.getDebugLoc(); |
| 4754 | |
| 4755 | // To "insert" a SELECT instruction, we actually have to insert the |
| 4756 | // diamond control-flow pattern. The incoming instruction knows the |
| 4757 | // destination vreg to set, the condition code register to branch on, the |
| 4758 | // true/false values to select between, and a branch opcode to use. |
| 4759 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 4760 | MachineFunction::iterator It = ++BB->getIterator(); |
| 4761 | |
| 4762 | // thisMBB: |
| 4763 | // ... |
| 4764 | // TrueVal = ... |
| 4765 | // setcc r1, r2, r3 |
| 4766 | // bNE r1, r0, copy1MBB |
| 4767 | // fallthrough --> copy0MBB |
| 4768 | MachineBasicBlock *thisMBB = BB; |
| 4769 | MachineFunction *F = BB->getParent(); |
| 4770 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 4771 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 4772 | F->insert(MBBI: It, MBB: copy0MBB); |
| 4773 | F->insert(MBBI: It, MBB: sinkMBB); |
| 4774 | |
| 4775 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
| 4776 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
| 4777 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 4778 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 4779 | |
| 4780 | // Next, add the true and fallthrough blocks as its successors. |
| 4781 | BB->addSuccessor(Succ: copy0MBB); |
| 4782 | BB->addSuccessor(Succ: sinkMBB); |
| 4783 | |
| 4784 | if (isFPCmp) { |
| 4785 | // bc1[tf] cc, sinkMBB |
| 4786 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc)) |
| 4787 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 4788 | .addMBB(MBB: sinkMBB); |
| 4789 | } else { |
| 4790 | // bne rs, $0, sinkMBB |
| 4791 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc)) |
| 4792 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 4793 | .addReg(RegNo: Mips::ZERO) |
| 4794 | .addMBB(MBB: sinkMBB); |
| 4795 | } |
| 4796 | |
| 4797 | // copy0MBB: |
| 4798 | // %FalseValue = ... |
| 4799 | // # fallthrough to sinkMBB |
| 4800 | BB = copy0MBB; |
| 4801 | |
| 4802 | // Update machine-CFG edges |
| 4803 | BB->addSuccessor(Succ: sinkMBB); |
| 4804 | |
| 4805 | // sinkMBB: |
| 4806 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
| 4807 | // ... |
| 4808 | BB = sinkMBB; |
| 4809 | |
| 4810 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
| 4811 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 4812 | .addMBB(MBB: thisMBB) |
| 4813 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
| 4814 | .addMBB(MBB: copy0MBB); |
| 4815 | |
| 4816 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 4817 | |
| 4818 | return BB; |
| 4819 | } |
| 4820 | |
| 4821 | MachineBasicBlock * |
| 4822 | MipsTargetLowering::emitPseudoD_SELECT(MachineInstr &MI, |
| 4823 | MachineBasicBlock *BB) const { |
| 4824 | assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && |
| 4825 | "Subtarget already supports SELECT nodes with the use of" |
| 4826 | "conditional-move instructions." ); |
| 4827 | |
| 4828 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 4829 | DebugLoc DL = MI.getDebugLoc(); |
| 4830 | |
| 4831 | // D_SELECT substitutes two SELECT nodes that goes one after another and |
| 4832 | // have the same condition operand. On machines which don't have |
| 4833 | // conditional-move instruction, it reduces unnecessary branch instructions |
| 4834 | // which are result of using two diamond patterns that are result of two |
| 4835 | // SELECT pseudo instructions. |
| 4836 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 4837 | MachineFunction::iterator It = ++BB->getIterator(); |
| 4838 | |
| 4839 | // thisMBB: |
| 4840 | // ... |
| 4841 | // TrueVal = ... |
| 4842 | // setcc r1, r2, r3 |
| 4843 | // bNE r1, r0, copy1MBB |
| 4844 | // fallthrough --> copy0MBB |
| 4845 | MachineBasicBlock *thisMBB = BB; |
| 4846 | MachineFunction *F = BB->getParent(); |
| 4847 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 4848 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 4849 | F->insert(MBBI: It, MBB: copy0MBB); |
| 4850 | F->insert(MBBI: It, MBB: sinkMBB); |
| 4851 | |
| 4852 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
| 4853 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
| 4854 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
| 4855 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
| 4856 | |
| 4857 | // Next, add the true and fallthrough blocks as its successors. |
| 4858 | BB->addSuccessor(Succ: copy0MBB); |
| 4859 | BB->addSuccessor(Succ: sinkMBB); |
| 4860 | |
| 4861 | // bne rs, $0, sinkMBB |
| 4862 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::BNE)) |
| 4863 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 4864 | .addReg(RegNo: Mips::ZERO) |
| 4865 | .addMBB(MBB: sinkMBB); |
| 4866 | |
| 4867 | // copy0MBB: |
| 4868 | // %FalseValue = ... |
| 4869 | // # fallthrough to sinkMBB |
| 4870 | BB = copy0MBB; |
| 4871 | |
| 4872 | // Update machine-CFG edges |
| 4873 | BB->addSuccessor(Succ: sinkMBB); |
| 4874 | |
| 4875 | // sinkMBB: |
| 4876 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
| 4877 | // ... |
| 4878 | BB = sinkMBB; |
| 4879 | |
| 4880 | // Use two PHI nodes to select two reults |
| 4881 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
| 4882 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
| 4883 | .addMBB(MBB: thisMBB) |
| 4884 | .addReg(RegNo: MI.getOperand(i: 5).getReg()) |
| 4885 | .addMBB(MBB: copy0MBB); |
| 4886 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 1).getReg()) |
| 4887 | .addReg(RegNo: MI.getOperand(i: 4).getReg()) |
| 4888 | .addMBB(MBB: thisMBB) |
| 4889 | .addReg(RegNo: MI.getOperand(i: 6).getReg()) |
| 4890 | .addMBB(MBB: copy0MBB); |
| 4891 | |
| 4892 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 4893 | |
| 4894 | return BB; |
| 4895 | } |
| 4896 | |
| 4897 | // Copies the function MipsAsmParser::matchCPURegisterName. |
| 4898 | int MipsTargetLowering::getCPURegisterIndex(StringRef Name) const { |
| 4899 | int CC; |
| 4900 | |
| 4901 | CC = StringSwitch<unsigned>(Name) |
| 4902 | .Case(S: "zero" , Value: 0) |
| 4903 | .Case(S: "at" , Value: 1) |
| 4904 | .Case(S: "AT" , Value: 1) |
| 4905 | .Case(S: "a0" , Value: 4) |
| 4906 | .Case(S: "a1" , Value: 5) |
| 4907 | .Case(S: "a2" , Value: 6) |
| 4908 | .Case(S: "a3" , Value: 7) |
| 4909 | .Case(S: "v0" , Value: 2) |
| 4910 | .Case(S: "v1" , Value: 3) |
| 4911 | .Case(S: "s0" , Value: 16) |
| 4912 | .Case(S: "s1" , Value: 17) |
| 4913 | .Case(S: "s2" , Value: 18) |
| 4914 | .Case(S: "s3" , Value: 19) |
| 4915 | .Case(S: "s4" , Value: 20) |
| 4916 | .Case(S: "s5" , Value: 21) |
| 4917 | .Case(S: "s6" , Value: 22) |
| 4918 | .Case(S: "s7" , Value: 23) |
| 4919 | .Case(S: "k0" , Value: 26) |
| 4920 | .Case(S: "k1" , Value: 27) |
| 4921 | .Case(S: "gp" , Value: 28) |
| 4922 | .Case(S: "sp" , Value: 29) |
| 4923 | .Case(S: "fp" , Value: 30) |
| 4924 | .Case(S: "s8" , Value: 30) |
| 4925 | .Case(S: "ra" , Value: 31) |
| 4926 | .Case(S: "t0" , Value: 8) |
| 4927 | .Case(S: "t1" , Value: 9) |
| 4928 | .Case(S: "t2" , Value: 10) |
| 4929 | .Case(S: "t3" , Value: 11) |
| 4930 | .Case(S: "t4" , Value: 12) |
| 4931 | .Case(S: "t5" , Value: 13) |
| 4932 | .Case(S: "t6" , Value: 14) |
| 4933 | .Case(S: "t7" , Value: 15) |
| 4934 | .Case(S: "t8" , Value: 24) |
| 4935 | .Case(S: "t9" , Value: 25) |
| 4936 | .Default(Value: -1); |
| 4937 | |
| 4938 | if (!(ABI.IsN32() || ABI.IsN64())) |
| 4939 | return CC; |
| 4940 | |
| 4941 | // Although SGI documentation just cuts out t0-t3 for n32/n64, |
| 4942 | // GNU pushes the values of t0-t3 to override the o32/o64 values for t4-t7 |
| 4943 | // We are supporting both cases, so for t0-t3 we'll just push them to t4-t7. |
| 4944 | if (8 <= CC && CC <= 11) |
| 4945 | CC += 4; |
| 4946 | |
| 4947 | if (CC == -1) |
| 4948 | CC = StringSwitch<unsigned>(Name) |
| 4949 | .Case(S: "a4" , Value: 8) |
| 4950 | .Case(S: "a5" , Value: 9) |
| 4951 | .Case(S: "a6" , Value: 10) |
| 4952 | .Case(S: "a7" , Value: 11) |
| 4953 | .Case(S: "kt0" , Value: 26) |
| 4954 | .Case(S: "kt1" , Value: 27) |
| 4955 | .Default(Value: -1); |
| 4956 | |
| 4957 | return CC; |
| 4958 | } |
| 4959 | |
| 4960 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
| 4961 | // this table could be generated automatically from RegInfo. |
| 4962 | Register |
| 4963 | MipsTargetLowering::getRegisterByName(const char *RegName, LLT VT, |
| 4964 | const MachineFunction &MF) const { |
| 4965 | // 1. Delete symbol '$'. |
| 4966 | std::string newRegName = RegName; |
| 4967 | if (StringRef(RegName).starts_with(Prefix: "$" )) |
| 4968 | newRegName = StringRef(RegName).substr(Start: 1); |
| 4969 | |
| 4970 | // 2. Get register index value. |
| 4971 | std::smatch matchResult; |
| 4972 | int regIdx; |
| 4973 | static const std::regex matchStr("^[0-9]*$" ); |
| 4974 | if (std::regex_match(s: newRegName, m&: matchResult, re: matchStr)) |
| 4975 | regIdx = std::stoi(str: newRegName); |
| 4976 | else { |
| 4977 | newRegName = StringRef(newRegName).lower(); |
| 4978 | regIdx = getCPURegisterIndex(Name: StringRef(newRegName)); |
| 4979 | } |
| 4980 | |
| 4981 | // 3. Get register. |
| 4982 | if (regIdx >= 0 && regIdx < 32) { |
| 4983 | const MCRegisterInfo *MRI = MF.getContext().getRegisterInfo(); |
| 4984 | const MCRegisterClass &RC = Subtarget.isGP64bit() |
| 4985 | ? MRI->getRegClass(i: Mips::GPR64RegClassID) |
| 4986 | : MRI->getRegClass(i: Mips::GPR32RegClassID); |
| 4987 | return RC.getRegister(i: regIdx); |
| 4988 | } |
| 4989 | |
| 4990 | report_fatal_error( |
| 4991 | reason: Twine("Invalid register name \"" + StringRef(RegName) + "\"." )); |
| 4992 | } |
| 4993 | |
| 4994 | MachineBasicBlock *MipsTargetLowering::emitLDR_W(MachineInstr &MI, |
| 4995 | MachineBasicBlock *BB) const { |
| 4996 | MachineFunction *MF = BB->getParent(); |
| 4997 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 4998 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 4999 | const bool IsLittle = Subtarget.isLittle(); |
| 5000 | DebugLoc DL = MI.getDebugLoc(); |
| 5001 | |
| 5002 | Register Dest = MI.getOperand(i: 0).getReg(); |
| 5003 | Register Address = MI.getOperand(i: 1).getReg(); |
| 5004 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
| 5005 | |
| 5006 | MachineBasicBlock::iterator I(MI); |
| 5007 | |
| 5008 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
| 5009 | // Mips release 6 can load from adress that is not naturally-aligned. |
| 5010 | Register Temp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5011 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
| 5012 | .addDef(RegNo: Temp) |
| 5013 | .addUse(RegNo: Address) |
| 5014 | .addImm(Val: Imm); |
| 5015 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Dest).addUse(RegNo: Temp); |
| 5016 | } else { |
| 5017 | // Mips release 5 needs to use instructions that can load from an unaligned |
| 5018 | // memory address. |
| 5019 | Register LoadHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5020 | Register LoadFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5021 | Register Undef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5022 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: Undef); |
| 5023 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
| 5024 | .addDef(RegNo: LoadHalf) |
| 5025 | .addUse(RegNo: Address) |
| 5026 | .addImm(Val: Imm + (IsLittle ? 0 : 3)) |
| 5027 | .addUse(RegNo: Undef); |
| 5028 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
| 5029 | .addDef(RegNo: LoadFull) |
| 5030 | .addUse(RegNo: Address) |
| 5031 | .addImm(Val: Imm + (IsLittle ? 3 : 0)) |
| 5032 | .addUse(RegNo: LoadHalf); |
| 5033 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Dest).addUse(RegNo: LoadFull); |
| 5034 | } |
| 5035 | |
| 5036 | MI.eraseFromParent(); |
| 5037 | return BB; |
| 5038 | } |
| 5039 | |
| 5040 | MachineBasicBlock *MipsTargetLowering::emitLDR_D(MachineInstr &MI, |
| 5041 | MachineBasicBlock *BB) const { |
| 5042 | MachineFunction *MF = BB->getParent(); |
| 5043 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 5044 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 5045 | const bool IsLittle = Subtarget.isLittle(); |
| 5046 | DebugLoc DL = MI.getDebugLoc(); |
| 5047 | |
| 5048 | Register Dest = MI.getOperand(i: 0).getReg(); |
| 5049 | Register Address = MI.getOperand(i: 1).getReg(); |
| 5050 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
| 5051 | |
| 5052 | MachineBasicBlock::iterator I(MI); |
| 5053 | |
| 5054 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
| 5055 | // Mips release 6 can load from adress that is not naturally-aligned. |
| 5056 | if (Subtarget.isGP64bit()) { |
| 5057 | Register Temp = MRI.createVirtualRegister(RegClass: &Mips::GPR64RegClass); |
| 5058 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LD)) |
| 5059 | .addDef(RegNo: Temp) |
| 5060 | .addUse(RegNo: Address) |
| 5061 | .addImm(Val: Imm); |
| 5062 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_D)).addDef(RegNo: Dest).addUse(RegNo: Temp); |
| 5063 | } else { |
| 5064 | Register Wtemp = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
| 5065 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5066 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5067 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
| 5068 | .addDef(RegNo: Lo) |
| 5069 | .addUse(RegNo: Address) |
| 5070 | .addImm(Val: Imm + (IsLittle ? 0 : 4)); |
| 5071 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
| 5072 | .addDef(RegNo: Hi) |
| 5073 | .addUse(RegNo: Address) |
| 5074 | .addImm(Val: Imm + (IsLittle ? 4 : 0)); |
| 5075 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Wtemp).addUse(RegNo: Lo); |
| 5076 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::INSERT_W), DestReg: Dest) |
| 5077 | .addUse(RegNo: Wtemp) |
| 5078 | .addUse(RegNo: Hi) |
| 5079 | .addImm(Val: 1); |
| 5080 | } |
| 5081 | } else { |
| 5082 | // Mips release 5 needs to use instructions that can load from an unaligned |
| 5083 | // memory address. |
| 5084 | Register LoHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5085 | Register LoFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5086 | Register LoUndef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5087 | Register HiHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5088 | Register HiFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5089 | Register HiUndef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5090 | Register Wtemp = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
| 5091 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: LoUndef); |
| 5092 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
| 5093 | .addDef(RegNo: LoHalf) |
| 5094 | .addUse(RegNo: Address) |
| 5095 | .addImm(Val: Imm + (IsLittle ? 0 : 7)) |
| 5096 | .addUse(RegNo: LoUndef); |
| 5097 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
| 5098 | .addDef(RegNo: LoFull) |
| 5099 | .addUse(RegNo: Address) |
| 5100 | .addImm(Val: Imm + (IsLittle ? 3 : 4)) |
| 5101 | .addUse(RegNo: LoHalf); |
| 5102 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: HiUndef); |
| 5103 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
| 5104 | .addDef(RegNo: HiHalf) |
| 5105 | .addUse(RegNo: Address) |
| 5106 | .addImm(Val: Imm + (IsLittle ? 4 : 3)) |
| 5107 | .addUse(RegNo: HiUndef); |
| 5108 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
| 5109 | .addDef(RegNo: HiFull) |
| 5110 | .addUse(RegNo: Address) |
| 5111 | .addImm(Val: Imm + (IsLittle ? 7 : 0)) |
| 5112 | .addUse(RegNo: HiHalf); |
| 5113 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Wtemp).addUse(RegNo: LoFull); |
| 5114 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::INSERT_W), DestReg: Dest) |
| 5115 | .addUse(RegNo: Wtemp) |
| 5116 | .addUse(RegNo: HiFull) |
| 5117 | .addImm(Val: 1); |
| 5118 | } |
| 5119 | |
| 5120 | MI.eraseFromParent(); |
| 5121 | return BB; |
| 5122 | } |
| 5123 | |
| 5124 | MachineBasicBlock *MipsTargetLowering::emitSTR_W(MachineInstr &MI, |
| 5125 | MachineBasicBlock *BB) const { |
| 5126 | MachineFunction *MF = BB->getParent(); |
| 5127 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 5128 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 5129 | const bool IsLittle = Subtarget.isLittle(); |
| 5130 | DebugLoc DL = MI.getDebugLoc(); |
| 5131 | |
| 5132 | Register StoreVal = MI.getOperand(i: 0).getReg(); |
| 5133 | Register Address = MI.getOperand(i: 1).getReg(); |
| 5134 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
| 5135 | |
| 5136 | MachineBasicBlock::iterator I(MI); |
| 5137 | |
| 5138 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
| 5139 | // Mips release 6 can store to adress that is not naturally-aligned. |
| 5140 | Register BitcastW = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
| 5141 | Register Tmp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5142 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)).addDef(RegNo: BitcastW).addUse(RegNo: StoreVal); |
| 5143 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5144 | .addDef(RegNo: Tmp) |
| 5145 | .addUse(RegNo: BitcastW) |
| 5146 | .addImm(Val: 0); |
| 5147 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
| 5148 | .addUse(RegNo: Tmp) |
| 5149 | .addUse(RegNo: Address) |
| 5150 | .addImm(Val: Imm); |
| 5151 | } else { |
| 5152 | // Mips release 5 needs to use instructions that can store to an unaligned |
| 5153 | // memory address. |
| 5154 | Register Tmp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5155 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5156 | .addDef(RegNo: Tmp) |
| 5157 | .addUse(RegNo: StoreVal) |
| 5158 | .addImm(Val: 0); |
| 5159 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
| 5160 | .addUse(RegNo: Tmp) |
| 5161 | .addUse(RegNo: Address) |
| 5162 | .addImm(Val: Imm + (IsLittle ? 0 : 3)); |
| 5163 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
| 5164 | .addUse(RegNo: Tmp) |
| 5165 | .addUse(RegNo: Address) |
| 5166 | .addImm(Val: Imm + (IsLittle ? 3 : 0)); |
| 5167 | } |
| 5168 | |
| 5169 | MI.eraseFromParent(); |
| 5170 | |
| 5171 | return BB; |
| 5172 | } |
| 5173 | |
| 5174 | MachineBasicBlock *MipsTargetLowering::emitSTR_D(MachineInstr &MI, |
| 5175 | MachineBasicBlock *BB) const { |
| 5176 | MachineFunction *MF = BB->getParent(); |
| 5177 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
| 5178 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
| 5179 | const bool IsLittle = Subtarget.isLittle(); |
| 5180 | DebugLoc DL = MI.getDebugLoc(); |
| 5181 | |
| 5182 | Register StoreVal = MI.getOperand(i: 0).getReg(); |
| 5183 | Register Address = MI.getOperand(i: 1).getReg(); |
| 5184 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
| 5185 | |
| 5186 | MachineBasicBlock::iterator I(MI); |
| 5187 | |
| 5188 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
| 5189 | // Mips release 6 can store to adress that is not naturally-aligned. |
| 5190 | if (Subtarget.isGP64bit()) { |
| 5191 | Register BitcastD = MRI.createVirtualRegister(RegClass: &Mips::MSA128DRegClass); |
| 5192 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR64RegClass); |
| 5193 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)) |
| 5194 | .addDef(RegNo: BitcastD) |
| 5195 | .addUse(RegNo: StoreVal); |
| 5196 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_D)) |
| 5197 | .addDef(RegNo: Lo) |
| 5198 | .addUse(RegNo: BitcastD) |
| 5199 | .addImm(Val: 0); |
| 5200 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SD)) |
| 5201 | .addUse(RegNo: Lo) |
| 5202 | .addUse(RegNo: Address) |
| 5203 | .addImm(Val: Imm); |
| 5204 | } else { |
| 5205 | Register BitcastW = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
| 5206 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5207 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5208 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)) |
| 5209 | .addDef(RegNo: BitcastW) |
| 5210 | .addUse(RegNo: StoreVal); |
| 5211 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5212 | .addDef(RegNo: Lo) |
| 5213 | .addUse(RegNo: BitcastW) |
| 5214 | .addImm(Val: 0); |
| 5215 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5216 | .addDef(RegNo: Hi) |
| 5217 | .addUse(RegNo: BitcastW) |
| 5218 | .addImm(Val: 1); |
| 5219 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
| 5220 | .addUse(RegNo: Lo) |
| 5221 | .addUse(RegNo: Address) |
| 5222 | .addImm(Val: Imm + (IsLittle ? 0 : 4)); |
| 5223 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
| 5224 | .addUse(RegNo: Hi) |
| 5225 | .addUse(RegNo: Address) |
| 5226 | .addImm(Val: Imm + (IsLittle ? 4 : 0)); |
| 5227 | } |
| 5228 | } else { |
| 5229 | // Mips release 5 needs to use instructions that can store to an unaligned |
| 5230 | // memory address. |
| 5231 | Register Bitcast = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
| 5232 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5233 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
| 5234 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)).addDef(RegNo: Bitcast).addUse(RegNo: StoreVal); |
| 5235 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5236 | .addDef(RegNo: Lo) |
| 5237 | .addUse(RegNo: Bitcast) |
| 5238 | .addImm(Val: 0); |
| 5239 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
| 5240 | .addDef(RegNo: Hi) |
| 5241 | .addUse(RegNo: Bitcast) |
| 5242 | .addImm(Val: 1); |
| 5243 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
| 5244 | .addUse(RegNo: Lo) |
| 5245 | .addUse(RegNo: Address) |
| 5246 | .addImm(Val: Imm + (IsLittle ? 0 : 3)); |
| 5247 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
| 5248 | .addUse(RegNo: Lo) |
| 5249 | .addUse(RegNo: Address) |
| 5250 | .addImm(Val: Imm + (IsLittle ? 3 : 0)); |
| 5251 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
| 5252 | .addUse(RegNo: Hi) |
| 5253 | .addUse(RegNo: Address) |
| 5254 | .addImm(Val: Imm + (IsLittle ? 4 : 7)); |
| 5255 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
| 5256 | .addUse(RegNo: Hi) |
| 5257 | .addUse(RegNo: Address) |
| 5258 | .addImm(Val: Imm + (IsLittle ? 7 : 4)); |
| 5259 | } |
| 5260 | |
| 5261 | MI.eraseFromParent(); |
| 5262 | return BB; |
| 5263 | } |
| 5264 | |