| 1 | //===-- SparcISelLowering.cpp - Sparc 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 implements the interfaces that Sparc uses to lower LLVM code into a |
| 10 | // selection DAG. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "SparcISelLowering.h" |
| 15 | #include "MCTargetDesc/SparcMCTargetDesc.h" |
| 16 | #include "SparcMachineFunctionInfo.h" |
| 17 | #include "SparcRegisterInfo.h" |
| 18 | #include "SparcSelectionDAGInfo.h" |
| 19 | #include "SparcTargetMachine.h" |
| 20 | #include "SparcTargetObjectFile.h" |
| 21 | #include "llvm/ADT/StringExtras.h" |
| 22 | #include "llvm/ADT/StringSwitch.h" |
| 23 | #include "llvm/BinaryFormat/ELF.h" |
| 24 | #include "llvm/CodeGen/CallingConvLower.h" |
| 25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 26 | #include "llvm/CodeGen/MachineFunction.h" |
| 27 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 29 | #include "llvm/CodeGen/SelectionDAG.h" |
| 30 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 31 | #include "llvm/CodeGen/TargetLowering.h" |
| 32 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
| 33 | #include "llvm/IR/DerivedTypes.h" |
| 34 | #include "llvm/IR/DiagnosticInfo.h" |
| 35 | #include "llvm/IR/Function.h" |
| 36 | #include "llvm/IR/IRBuilder.h" |
| 37 | #include "llvm/IR/Module.h" |
| 38 | #include "llvm/Support/ErrorHandling.h" |
| 39 | #include "llvm/Support/KnownBits.h" |
| 40 | using namespace llvm; |
| 41 | |
| 42 | |
| 43 | //===----------------------------------------------------------------------===// |
| 44 | // Calling Convention Implementation |
| 45 | //===----------------------------------------------------------------------===// |
| 46 | |
| 47 | static bool CC_Sparc_Assign_SRet(unsigned &ValNo, MVT &ValVT, |
| 48 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
| 49 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
| 50 | { |
| 51 | assert (ArgFlags.isSRet()); |
| 52 | |
| 53 | // Assign SRet argument. |
| 54 | State.addLoc(V: CCValAssign::getCustomMem(ValNo, ValVT, |
| 55 | Offset: 0, |
| 56 | LocVT, HTP: LocInfo)); |
| 57 | return true; |
| 58 | } |
| 59 | |
| 60 | static bool CC_Sparc_Assign_Split_64(unsigned &ValNo, MVT &ValVT, |
| 61 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
| 62 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
| 63 | { |
| 64 | static const MCPhysReg RegList[] = { |
| 65 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
| 66 | }; |
| 67 | // Try to get first reg. |
| 68 | if (Register Reg = State.AllocateReg(Regs: RegList)) { |
| 69 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 70 | } else { |
| 71 | // Assign whole thing in stack. |
| 72 | State.addLoc(V: CCValAssign::getCustomMem( |
| 73 | ValNo, ValVT, Offset: State.AllocateStack(Size: 8, Alignment: Align(4)), LocVT, HTP: LocInfo)); |
| 74 | return true; |
| 75 | } |
| 76 | |
| 77 | // Try to get second reg. |
| 78 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
| 79 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 80 | else |
| 81 | State.addLoc(V: CCValAssign::getCustomMem( |
| 82 | ValNo, ValVT, Offset: State.AllocateStack(Size: 4, Alignment: Align(4)), LocVT, HTP: LocInfo)); |
| 83 | return true; |
| 84 | } |
| 85 | |
| 86 | static bool CC_Sparc_Assign_Ret_Split_64(unsigned &ValNo, MVT &ValVT, |
| 87 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
| 88 | ISD::ArgFlagsTy &ArgFlags, CCState &State) |
| 89 | { |
| 90 | static const MCPhysReg RegList[] = { |
| 91 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
| 92 | }; |
| 93 | |
| 94 | // Try to get first reg. |
| 95 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
| 96 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 97 | else |
| 98 | return false; |
| 99 | |
| 100 | // Try to get second reg. |
| 101 | if (Register Reg = State.AllocateReg(Regs: RegList)) |
| 102 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 103 | else |
| 104 | return false; |
| 105 | |
| 106 | return true; |
| 107 | } |
| 108 | |
| 109 | // Allocate a full-sized argument for the 64-bit ABI. |
| 110 | static bool Analyze_CC_Sparc64_Full(bool IsReturn, unsigned &ValNo, MVT &ValVT, |
| 111 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
| 112 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 113 | assert((LocVT == MVT::f32 || LocVT == MVT::f128 |
| 114 | || LocVT.getSizeInBits() == 64) && |
| 115 | "Can't handle non-64 bits locations" ); |
| 116 | |
| 117 | // Stack space is allocated for all arguments starting from [%fp+BIAS+128]. |
| 118 | unsigned size = (LocVT == MVT::f128) ? 16 : 8; |
| 119 | Align alignment = |
| 120 | (LocVT == MVT::f128 || ArgFlags.isSplit()) ? Align(16) : Align(8); |
| 121 | unsigned Offset = State.AllocateStack(Size: size, Alignment: alignment); |
| 122 | unsigned Reg = 0; |
| 123 | |
| 124 | if (LocVT == MVT::i64 && Offset < 6*8) |
| 125 | // Promote integers to %i0-%i5. |
| 126 | Reg = SP::I0 + Offset/8; |
| 127 | else if (LocVT == MVT::f64 && Offset < 16*8) |
| 128 | // Promote doubles to %d0-%d30. (Which LLVM calls D0-D15). |
| 129 | Reg = SP::D0 + Offset/8; |
| 130 | else if (LocVT == MVT::f32 && Offset < 16*8) |
| 131 | // Promote floats to %f1, %f3, ... |
| 132 | Reg = SP::F1 + Offset/4; |
| 133 | else if (LocVT == MVT::f128 && Offset < 16*8) |
| 134 | // Promote long doubles to %q0-%q28. (Which LLVM calls Q0-Q7). |
| 135 | Reg = SP::Q0 + Offset/16; |
| 136 | |
| 137 | // Promote to register when possible, otherwise use the stack slot. |
| 138 | if (Reg) { |
| 139 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 140 | return true; |
| 141 | } |
| 142 | |
| 143 | // Bail out if this is a return CC and we run out of registers to place |
| 144 | // values into. |
| 145 | if (IsReturn) |
| 146 | return false; |
| 147 | |
| 148 | // This argument goes on the stack in an 8-byte slot. |
| 149 | // When passing floats, LocVT is smaller than 8 bytes. Adjust the offset to |
| 150 | // the right-aligned float. The first 4 bytes of the stack slot are undefined. |
| 151 | if (LocVT == MVT::f32) |
| 152 | Offset += 4; |
| 153 | |
| 154 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
| 155 | return true; |
| 156 | } |
| 157 | |
| 158 | // Allocate a half-sized argument for the 64-bit ABI. |
| 159 | // |
| 160 | // This is used when passing { float, int } structs by value in registers. |
| 161 | static bool Analyze_CC_Sparc64_Half(bool IsReturn, unsigned &ValNo, MVT &ValVT, |
| 162 | MVT &LocVT, CCValAssign::LocInfo &LocInfo, |
| 163 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 164 | assert(LocVT.getSizeInBits() == 32 && "Can't handle non-32 bits locations" ); |
| 165 | unsigned Offset = State.AllocateStack(Size: 4, Alignment: Align(4)); |
| 166 | |
| 167 | if (LocVT == MVT::f32 && Offset < 16*8) { |
| 168 | // Promote floats to %f0-%f31. |
| 169 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg: SP::F0 + Offset/4, |
| 170 | LocVT, HTP: LocInfo)); |
| 171 | return true; |
| 172 | } |
| 173 | |
| 174 | if (LocVT == MVT::i32 && Offset < 6*8) { |
| 175 | // Promote integers to %i0-%i5, using half the register. |
| 176 | unsigned Reg = SP::I0 + Offset/8; |
| 177 | LocVT = MVT::i64; |
| 178 | LocInfo = CCValAssign::AExt; |
| 179 | |
| 180 | // Set the Custom bit if this i32 goes in the high bits of a register. |
| 181 | if (Offset % 8 == 0) |
| 182 | State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, |
| 183 | LocVT, HTP: LocInfo)); |
| 184 | else |
| 185 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
| 186 | return true; |
| 187 | } |
| 188 | |
| 189 | // Bail out if this is a return CC and we run out of registers to place |
| 190 | // values into. |
| 191 | if (IsReturn) |
| 192 | return false; |
| 193 | |
| 194 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
| 195 | return true; |
| 196 | } |
| 197 | |
| 198 | static bool CC_Sparc64_Full(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
| 199 | CCValAssign::LocInfo &LocInfo, |
| 200 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 201 | return Analyze_CC_Sparc64_Full(IsReturn: false, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
| 202 | State); |
| 203 | } |
| 204 | |
| 205 | static bool CC_Sparc64_Half(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
| 206 | CCValAssign::LocInfo &LocInfo, |
| 207 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 208 | return Analyze_CC_Sparc64_Half(IsReturn: false, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
| 209 | State); |
| 210 | } |
| 211 | |
| 212 | static bool RetCC_Sparc64_Full(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
| 213 | CCValAssign::LocInfo &LocInfo, |
| 214 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 215 | return Analyze_CC_Sparc64_Full(IsReturn: true, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
| 216 | State); |
| 217 | } |
| 218 | |
| 219 | static bool RetCC_Sparc64_Half(unsigned &ValNo, MVT &ValVT, MVT &LocVT, |
| 220 | CCValAssign::LocInfo &LocInfo, |
| 221 | ISD::ArgFlagsTy &ArgFlags, CCState &State) { |
| 222 | return Analyze_CC_Sparc64_Half(IsReturn: true, ValNo, ValVT, LocVT, LocInfo, ArgFlags, |
| 223 | State); |
| 224 | } |
| 225 | |
| 226 | #define GET_CALLING_CONV_IMPL |
| 227 | #include "SparcGenCallingConv.inc" |
| 228 | |
| 229 | // The calling conventions in SparcCallingConv.td are described in terms of the |
| 230 | // callee's register window. This function translates registers to the |
| 231 | // corresponding caller window %o register. |
| 232 | static unsigned toCallerWindow(unsigned Reg) { |
| 233 | static_assert(SP::I0 + 7 == SP::I7 && SP::O0 + 7 == SP::O7, |
| 234 | "Unexpected enum" ); |
| 235 | if (Reg >= SP::I0 && Reg <= SP::I7) |
| 236 | return Reg - SP::I0 + SP::O0; |
| 237 | return Reg; |
| 238 | } |
| 239 | |
| 240 | bool SparcTargetLowering::CanLowerReturn( |
| 241 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, |
| 242 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context, |
| 243 | const Type *RetTy) const { |
| 244 | SmallVector<CCValAssign, 16> RVLocs; |
| 245 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
| 246 | return CCInfo.CheckReturn(Outs, Fn: Subtarget->is64Bit() ? RetCC_Sparc64 |
| 247 | : RetCC_Sparc32); |
| 248 | } |
| 249 | |
| 250 | SDValue |
| 251 | SparcTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
| 252 | bool IsVarArg, |
| 253 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 254 | const SmallVectorImpl<SDValue> &OutVals, |
| 255 | const SDLoc &DL, SelectionDAG &DAG) const { |
| 256 | if (Subtarget->is64Bit()) |
| 257 | return LowerReturn_64(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); |
| 258 | return LowerReturn_32(Chain, CallConv, IsVarArg, Outs, OutVals, DL, DAG); |
| 259 | } |
| 260 | |
| 261 | SDValue |
| 262 | SparcTargetLowering::LowerReturn_32(SDValue Chain, CallingConv::ID CallConv, |
| 263 | bool IsVarArg, |
| 264 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 265 | const SmallVectorImpl<SDValue> &OutVals, |
| 266 | const SDLoc &DL, SelectionDAG &DAG) const { |
| 267 | MachineFunction &MF = DAG.getMachineFunction(); |
| 268 | |
| 269 | // CCValAssign - represent the assignment of the return value to locations. |
| 270 | SmallVector<CCValAssign, 16> RVLocs; |
| 271 | |
| 272 | // CCState - Info about the registers and stack slot. |
| 273 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
| 274 | *DAG.getContext()); |
| 275 | |
| 276 | // Analyze return values. |
| 277 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Sparc32); |
| 278 | |
| 279 | SDValue Glue; |
| 280 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 281 | // Make room for the return address offset. |
| 282 | RetOps.push_back(Elt: SDValue()); |
| 283 | |
| 284 | // Copy the result values into the output registers. |
| 285 | for (unsigned i = 0, realRVLocIdx = 0; |
| 286 | i != RVLocs.size(); |
| 287 | ++i, ++realRVLocIdx) { |
| 288 | CCValAssign &VA = RVLocs[i]; |
| 289 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
| 290 | |
| 291 | SDValue Arg = OutVals[realRVLocIdx]; |
| 292 | |
| 293 | if (VA.needsCustom()) { |
| 294 | assert(VA.getLocVT() == MVT::v2i32); |
| 295 | // Legalize ret v2i32 -> ret 2 x i32 (Basically: do what would |
| 296 | // happen by default if this wasn't a legal type) |
| 297 | |
| 298 | SDValue Part0 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: MVT::i32, |
| 299 | N1: Arg, |
| 300 | N2: DAG.getConstant(Val: 0, DL, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 301 | SDValue Part1 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: MVT::i32, |
| 302 | N1: Arg, |
| 303 | N2: DAG.getConstant(Val: 1, DL, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 304 | |
| 305 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Part0, Glue); |
| 306 | Glue = Chain.getValue(R: 1); |
| 307 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
| 308 | VA = RVLocs[++i]; // skip ahead to next loc |
| 309 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Part1, |
| 310 | Glue); |
| 311 | } else |
| 312 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Arg, Glue); |
| 313 | |
| 314 | // Guarantee that all emitted copies are stuck together with flags. |
| 315 | Glue = Chain.getValue(R: 1); |
| 316 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
| 317 | } |
| 318 | |
| 319 | unsigned RetAddrOffset = 8; // Call Inst + Delay Slot |
| 320 | // If the function returns a struct, copy the SRetReturnReg to I0 |
| 321 | if (MF.getFunction().hasStructRetAttr()) { |
| 322 | SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); |
| 323 | Register Reg = SFI->getSRetReturnReg(); |
| 324 | if (!Reg) |
| 325 | llvm_unreachable("sret virtual register not created in the entry block" ); |
| 326 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 327 | SDValue Val = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: PtrVT); |
| 328 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: SP::I0, N: Val, Glue); |
| 329 | Glue = Chain.getValue(R: 1); |
| 330 | RetOps.push_back(Elt: DAG.getRegister(Reg: SP::I0, VT: PtrVT)); |
| 331 | RetAddrOffset = 12; // CallInst + Delay Slot + Unimp |
| 332 | } |
| 333 | |
| 334 | RetOps[0] = Chain; // Update chain. |
| 335 | RetOps[1] = DAG.getConstant(Val: RetAddrOffset, DL, VT: MVT::i32); |
| 336 | |
| 337 | // Add the glue if we have it. |
| 338 | if (Glue.getNode()) |
| 339 | RetOps.push_back(Elt: Glue); |
| 340 | |
| 341 | return DAG.getNode(Opcode: SPISD::RET_GLUE, DL, VT: MVT::Other, Ops: RetOps); |
| 342 | } |
| 343 | |
| 344 | // Lower return values for the 64-bit ABI. |
| 345 | // Return values are passed the exactly the same way as function arguments. |
| 346 | SDValue |
| 347 | SparcTargetLowering::LowerReturn_64(SDValue Chain, CallingConv::ID CallConv, |
| 348 | bool IsVarArg, |
| 349 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
| 350 | const SmallVectorImpl<SDValue> &OutVals, |
| 351 | const SDLoc &DL, SelectionDAG &DAG) const { |
| 352 | // CCValAssign - represent the assignment of the return value to locations. |
| 353 | SmallVector<CCValAssign, 16> RVLocs; |
| 354 | |
| 355 | // CCState - Info about the registers and stack slot. |
| 356 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
| 357 | *DAG.getContext()); |
| 358 | |
| 359 | // Analyze return values. |
| 360 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Sparc64); |
| 361 | |
| 362 | SDValue Glue; |
| 363 | SmallVector<SDValue, 4> RetOps(1, Chain); |
| 364 | |
| 365 | // The second operand on the return instruction is the return address offset. |
| 366 | // The return address is always %i7+8 with the 64-bit ABI. |
| 367 | RetOps.push_back(Elt: DAG.getConstant(Val: 8, DL, VT: MVT::i32)); |
| 368 | |
| 369 | // Copy the result values into the output registers. |
| 370 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
| 371 | CCValAssign &VA = RVLocs[i]; |
| 372 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
| 373 | SDValue OutVal = OutVals[i]; |
| 374 | |
| 375 | // Integer return values must be sign or zero extended by the callee. |
| 376 | switch (VA.getLocInfo()) { |
| 377 | case CCValAssign::Full: break; |
| 378 | case CCValAssign::SExt: |
| 379 | OutVal = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
| 380 | break; |
| 381 | case CCValAssign::ZExt: |
| 382 | OutVal = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
| 383 | break; |
| 384 | case CCValAssign::AExt: |
| 385 | OutVal = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: OutVal); |
| 386 | break; |
| 387 | default: |
| 388 | llvm_unreachable("Unknown loc info!" ); |
| 389 | } |
| 390 | |
| 391 | // The custom bit on an i32 return value indicates that it should be passed |
| 392 | // in the high bits of the register. |
| 393 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { |
| 394 | OutVal = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: OutVal, |
| 395 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
| 396 | |
| 397 | // The next value may go in the low bits of the same register. |
| 398 | // Handle both at once. |
| 399 | if (i+1 < RVLocs.size() && RVLocs[i+1].getLocReg() == VA.getLocReg()) { |
| 400 | SDValue NV = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: MVT::i64, Operand: OutVals[i+1]); |
| 401 | OutVal = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i64, N1: OutVal, N2: NV); |
| 402 | // Skip the next value, it's already done. |
| 403 | ++i; |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: OutVal, Glue); |
| 408 | |
| 409 | // Guarantee that all emitted copies are stuck together with flags. |
| 410 | Glue = Chain.getValue(R: 1); |
| 411 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
| 412 | } |
| 413 | |
| 414 | RetOps[0] = Chain; // Update chain. |
| 415 | |
| 416 | // Add the flag if we have it. |
| 417 | if (Glue.getNode()) |
| 418 | RetOps.push_back(Elt: Glue); |
| 419 | |
| 420 | return DAG.getNode(Opcode: SPISD::RET_GLUE, DL, VT: MVT::Other, Ops: RetOps); |
| 421 | } |
| 422 | |
| 423 | SDValue SparcTargetLowering::LowerFormalArguments( |
| 424 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 425 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 426 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 427 | if (Subtarget->is64Bit()) |
| 428 | return LowerFormalArguments_64(Chain, CallConv, isVarArg: IsVarArg, Ins, |
| 429 | dl: DL, DAG, InVals); |
| 430 | return LowerFormalArguments_32(Chain, CallConv, isVarArg: IsVarArg, Ins, |
| 431 | dl: DL, DAG, InVals); |
| 432 | } |
| 433 | |
| 434 | /// LowerFormalArguments32 - V8 uses a very simple ABI, where all values are |
| 435 | /// passed in either one or two GPRs, including FP values. TODO: we should |
| 436 | /// pass FP values in FP registers for fastcc functions. |
| 437 | SDValue SparcTargetLowering::LowerFormalArguments_32( |
| 438 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
| 439 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
| 440 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 441 | MachineFunction &MF = DAG.getMachineFunction(); |
| 442 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
| 443 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
| 444 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 445 | |
| 446 | // Assign locations to all of the incoming arguments. |
| 447 | SmallVector<CCValAssign, 16> ArgLocs; |
| 448 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
| 449 | *DAG.getContext()); |
| 450 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Sparc32); |
| 451 | |
| 452 | const unsigned StackOffset = 92; |
| 453 | bool IsLittleEndian = DAG.getDataLayout().isLittleEndian(); |
| 454 | |
| 455 | unsigned InIdx = 0; |
| 456 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i, ++InIdx) { |
| 457 | CCValAssign &VA = ArgLocs[i]; |
| 458 | EVT LocVT = VA.getLocVT(); |
| 459 | |
| 460 | if (Ins[InIdx].Flags.isSRet()) { |
| 461 | if (InIdx != 0) |
| 462 | report_fatal_error(reason: "sparc only supports sret on the first parameter" ); |
| 463 | // Get SRet from [%fp+64]. |
| 464 | int FrameIdx = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: 64, IsImmutable: true); |
| 465 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
| 466 | SDValue Arg = |
| 467 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
| 468 | InVals.push_back(Elt: Arg); |
| 469 | continue; |
| 470 | } |
| 471 | |
| 472 | SDValue Arg; |
| 473 | if (VA.isRegLoc()) { |
| 474 | if (VA.needsCustom()) { |
| 475 | assert(VA.getLocVT() == MVT::f64 || VA.getLocVT() == MVT::v2i32); |
| 476 | |
| 477 | Register VRegHi = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
| 478 | MF.getRegInfo().addLiveIn(Reg: VA.getLocReg(), vreg: VRegHi); |
| 479 | SDValue HiVal = DAG.getCopyFromReg(Chain, dl, Reg: VRegHi, VT: MVT::i32); |
| 480 | |
| 481 | assert(i+1 < e); |
| 482 | CCValAssign &NextVA = ArgLocs[++i]; |
| 483 | |
| 484 | SDValue LoVal; |
| 485 | if (NextVA.isMemLoc()) { |
| 486 | int FrameIdx = MF.getFrameInfo(). |
| 487 | CreateFixedObject(Size: 4, SPOffset: StackOffset+NextVA.getLocMemOffset(),IsImmutable: true); |
| 488 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
| 489 | LoVal = DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
| 490 | } else { |
| 491 | Register loReg = MF.addLiveIn(PReg: NextVA.getLocReg(), |
| 492 | RC: &SP::IntRegsRegClass); |
| 493 | LoVal = DAG.getCopyFromReg(Chain, dl, Reg: loReg, VT: MVT::i32); |
| 494 | } |
| 495 | |
| 496 | if (IsLittleEndian) |
| 497 | std::swap(a&: LoVal, b&: HiVal); |
| 498 | |
| 499 | SDValue WholeValue = |
| 500 | DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: LoVal, N2: HiVal); |
| 501 | WholeValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getLocVT(), Operand: WholeValue); |
| 502 | InVals.push_back(Elt: WholeValue); |
| 503 | continue; |
| 504 | } |
| 505 | Register VReg = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
| 506 | MF.getRegInfo().addLiveIn(Reg: VA.getLocReg(), vreg: VReg); |
| 507 | Arg = DAG.getCopyFromReg(Chain, dl, Reg: VReg, VT: MVT::i32); |
| 508 | if (VA.getLocInfo() != CCValAssign::Indirect) { |
| 509 | if (VA.getLocVT() == MVT::f32) |
| 510 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::f32, Operand: Arg); |
| 511 | else if (VA.getLocVT() != MVT::i32) { |
| 512 | Arg = DAG.getNode(Opcode: ISD::AssertSext, DL: dl, VT: MVT::i32, N1: Arg, |
| 513 | N2: DAG.getValueType(VA.getLocVT())); |
| 514 | Arg = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
| 515 | } |
| 516 | InVals.push_back(Elt: Arg); |
| 517 | continue; |
| 518 | } |
| 519 | } else { |
| 520 | assert(VA.isMemLoc()); |
| 521 | |
| 522 | unsigned Offset = VA.getLocMemOffset() + StackOffset; |
| 523 | |
| 524 | if (VA.needsCustom()) { |
| 525 | assert(VA.getValVT() == MVT::f64 || VA.getValVT() == MVT::v2i32); |
| 526 | // If it is double-word aligned, just load. |
| 527 | if (Offset % 8 == 0) { |
| 528 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 8, SPOffset: Offset, IsImmutable: true); |
| 529 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
| 530 | SDValue Load = DAG.getLoad(VT: VA.getValVT(), dl, Chain, Ptr: FIPtr, |
| 531 | PtrInfo: MachinePointerInfo()); |
| 532 | InVals.push_back(Elt: Load); |
| 533 | continue; |
| 534 | } |
| 535 | |
| 536 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: Offset, IsImmutable: true); |
| 537 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
| 538 | SDValue HiVal = |
| 539 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
| 540 | int FI2 = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: Offset + 4, IsImmutable: true); |
| 541 | SDValue FIPtr2 = DAG.getFrameIndex(FI: FI2, VT: PtrVT); |
| 542 | |
| 543 | SDValue LoVal = |
| 544 | DAG.getLoad(VT: MVT::i32, dl, Chain, Ptr: FIPtr2, PtrInfo: MachinePointerInfo()); |
| 545 | |
| 546 | if (IsLittleEndian) |
| 547 | std::swap(a&: LoVal, b&: HiVal); |
| 548 | |
| 549 | SDValue WholeValue = |
| 550 | DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: LoVal, N2: HiVal); |
| 551 | WholeValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getValVT(), Operand: WholeValue); |
| 552 | InVals.push_back(Elt: WholeValue); |
| 553 | continue; |
| 554 | } |
| 555 | |
| 556 | int FI = MF.getFrameInfo().CreateFixedObject(Size: LocVT.getSizeInBits() / 8, |
| 557 | SPOffset: Offset, IsImmutable: true); |
| 558 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: PtrVT); |
| 559 | SDValue Load = DAG.getLoad(VT: LocVT, dl, Chain, Ptr: FIPtr, |
| 560 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI)); |
| 561 | if (VA.getLocInfo() != CCValAssign::Indirect) { |
| 562 | InVals.push_back(Elt: Load); |
| 563 | continue; |
| 564 | } |
| 565 | Arg = Load; |
| 566 | } |
| 567 | |
| 568 | assert(VA.getLocInfo() == CCValAssign::Indirect); |
| 569 | |
| 570 | SDValue ArgValue = |
| 571 | DAG.getLoad(VT: VA.getValVT(), dl, Chain, Ptr: Arg, PtrInfo: MachinePointerInfo()); |
| 572 | InVals.push_back(Elt: ArgValue); |
| 573 | |
| 574 | unsigned ArgIndex = Ins[InIdx].OrigArgIndex; |
| 575 | assert(Ins[InIdx].PartOffset == 0); |
| 576 | while (i + 1 != e && Ins[InIdx + 1].OrigArgIndex == ArgIndex) { |
| 577 | CCValAssign &PartVA = ArgLocs[i + 1]; |
| 578 | unsigned PartOffset = Ins[InIdx + 1].PartOffset; |
| 579 | SDValue Address = DAG.getMemBasePlusOffset( |
| 580 | Base: ArgValue, Offset: TypeSize::getFixed(ExactSize: PartOffset), DL: dl); |
| 581 | InVals.push_back(Elt: DAG.getLoad(VT: PartVA.getValVT(), dl, Chain, Ptr: Address, |
| 582 | PtrInfo: MachinePointerInfo())); |
| 583 | ++i; |
| 584 | ++InIdx; |
| 585 | } |
| 586 | } |
| 587 | |
| 588 | if (MF.getFunction().hasStructRetAttr()) { |
| 589 | // Copy the SRet Argument to SRetReturnReg. |
| 590 | SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>(); |
| 591 | Register Reg = SFI->getSRetReturnReg(); |
| 592 | if (!Reg) { |
| 593 | Reg = MF.getRegInfo().createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
| 594 | SFI->setSRetReturnReg(Reg); |
| 595 | } |
| 596 | SDValue Copy = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl, Reg, N: InVals[0]); |
| 597 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Copy, N2: Chain); |
| 598 | } |
| 599 | |
| 600 | // Store remaining ArgRegs to the stack if this is a varargs function. |
| 601 | if (isVarArg) { |
| 602 | static const MCPhysReg ArgRegs[] = { |
| 603 | SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5 |
| 604 | }; |
| 605 | unsigned NumAllocated = CCInfo.getFirstUnallocated(Regs: ArgRegs); |
| 606 | const MCPhysReg *CurArgReg = ArgRegs+NumAllocated, *ArgRegEnd = ArgRegs+6; |
| 607 | unsigned ArgOffset = CCInfo.getStackSize(); |
| 608 | if (NumAllocated == 6) |
| 609 | ArgOffset += StackOffset; |
| 610 | else { |
| 611 | assert(!ArgOffset); |
| 612 | ArgOffset = 68+4*NumAllocated; |
| 613 | } |
| 614 | |
| 615 | // Remember the vararg offset for the va_start implementation. |
| 616 | FuncInfo->setVarArgsFrameOffset(ArgOffset); |
| 617 | |
| 618 | std::vector<SDValue> OutChains; |
| 619 | |
| 620 | for (; CurArgReg != ArgRegEnd; ++CurArgReg) { |
| 621 | Register VReg = RegInfo.createVirtualRegister(RegClass: &SP::IntRegsRegClass); |
| 622 | MF.getRegInfo().addLiveIn(Reg: *CurArgReg, vreg: VReg); |
| 623 | SDValue Arg = DAG.getCopyFromReg(Chain: DAG.getRoot(), dl, Reg: VReg, VT: MVT::i32); |
| 624 | |
| 625 | int FrameIdx = MF.getFrameInfo().CreateFixedObject(Size: 4, SPOffset: ArgOffset, |
| 626 | IsImmutable: true); |
| 627 | SDValue FIPtr = DAG.getFrameIndex(FI: FrameIdx, VT: MVT::i32); |
| 628 | |
| 629 | OutChains.push_back( |
| 630 | x: DAG.getStore(Chain: DAG.getRoot(), dl, Val: Arg, Ptr: FIPtr, PtrInfo: MachinePointerInfo())); |
| 631 | ArgOffset += 4; |
| 632 | } |
| 633 | |
| 634 | if (!OutChains.empty()) { |
| 635 | OutChains.push_back(x: Chain); |
| 636 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | return Chain; |
| 641 | } |
| 642 | |
| 643 | // Lower formal arguments for the 64 bit ABI. |
| 644 | SDValue SparcTargetLowering::LowerFormalArguments_64( |
| 645 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
| 646 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
| 647 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
| 648 | MachineFunction &MF = DAG.getMachineFunction(); |
| 649 | |
| 650 | // Analyze arguments according to CC_Sparc64. |
| 651 | SmallVector<CCValAssign, 16> ArgLocs; |
| 652 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, |
| 653 | *DAG.getContext()); |
| 654 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Sparc64); |
| 655 | |
| 656 | // The argument array begins at %fp+BIAS+128, after the register save area. |
| 657 | const unsigned ArgArea = 128; |
| 658 | |
| 659 | for (const CCValAssign &VA : ArgLocs) { |
| 660 | if (VA.isRegLoc()) { |
| 661 | // This argument is passed in a register. |
| 662 | // All integer register arguments are promoted by the caller to i64. |
| 663 | |
| 664 | // Create a virtual register for the promoted live-in value. |
| 665 | Register VReg = MF.addLiveIn(PReg: VA.getLocReg(), |
| 666 | RC: getRegClassFor(VT: VA.getLocVT())); |
| 667 | SDValue Arg = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: VA.getLocVT()); |
| 668 | |
| 669 | // Get the high bits for i32 struct elements. |
| 670 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) |
| 671 | Arg = DAG.getNode(Opcode: ISD::SRL, DL, VT: VA.getLocVT(), N1: Arg, |
| 672 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
| 673 | |
| 674 | // The caller promoted the argument, so insert an Assert?ext SDNode so we |
| 675 | // won't promote the value again in this function. |
| 676 | switch (VA.getLocInfo()) { |
| 677 | case CCValAssign::SExt: |
| 678 | Arg = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: Arg, |
| 679 | N2: DAG.getValueType(VA.getValVT())); |
| 680 | break; |
| 681 | case CCValAssign::ZExt: |
| 682 | Arg = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: Arg, |
| 683 | N2: DAG.getValueType(VA.getValVT())); |
| 684 | break; |
| 685 | default: |
| 686 | break; |
| 687 | } |
| 688 | |
| 689 | // Truncate the register down to the argument type. |
| 690 | if (VA.isExtInLoc()) |
| 691 | Arg = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Arg); |
| 692 | |
| 693 | InVals.push_back(Elt: Arg); |
| 694 | continue; |
| 695 | } |
| 696 | |
| 697 | // The registers are exhausted. This argument was passed on the stack. |
| 698 | assert(VA.isMemLoc()); |
| 699 | // The CC_Sparc64_Full/Half functions compute stack offsets relative to the |
| 700 | // beginning of the arguments area at %fp+BIAS+128. |
| 701 | unsigned Offset = VA.getLocMemOffset() + ArgArea; |
| 702 | unsigned ValSize = VA.getValVT().getSizeInBits() / 8; |
| 703 | // Adjust offset for extended arguments, SPARC is big-endian. |
| 704 | // The caller will have written the full slot with extended bytes, but we |
| 705 | // prefer our own extending loads. |
| 706 | if (VA.isExtInLoc()) |
| 707 | Offset += 8 - ValSize; |
| 708 | int FI = MF.getFrameInfo().CreateFixedObject(Size: ValSize, SPOffset: Offset, IsImmutable: true); |
| 709 | InVals.push_back( |
| 710 | Elt: DAG.getLoad(VT: VA.getValVT(), dl: DL, Chain, |
| 711 | Ptr: DAG.getFrameIndex(FI, VT: getPointerTy(DL: MF.getDataLayout())), |
| 712 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
| 713 | } |
| 714 | |
| 715 | if (!IsVarArg) |
| 716 | return Chain; |
| 717 | |
| 718 | // This function takes variable arguments, some of which may have been passed |
| 719 | // in registers %i0-%i5. Variable floating point arguments are never passed |
| 720 | // in floating point registers. They go on %i0-%i5 or on the stack like |
| 721 | // integer arguments. |
| 722 | // |
| 723 | // The va_start intrinsic needs to know the offset to the first variable |
| 724 | // argument. |
| 725 | unsigned ArgOffset = CCInfo.getStackSize(); |
| 726 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
| 727 | // Skip the 128 bytes of register save area. |
| 728 | FuncInfo->setVarArgsFrameOffset(ArgOffset + ArgArea + |
| 729 | Subtarget->getStackPointerBias()); |
| 730 | |
| 731 | // Save the variable arguments that were passed in registers. |
| 732 | // The caller is required to reserve stack space for 6 arguments regardless |
| 733 | // of how many arguments were actually passed. |
| 734 | SmallVector<SDValue, 8> OutChains; |
| 735 | for (; ArgOffset < 6*8; ArgOffset += 8) { |
| 736 | Register VReg = MF.addLiveIn(PReg: SP::I0 + ArgOffset/8, RC: &SP::I64RegsRegClass); |
| 737 | SDValue VArg = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: MVT::i64); |
| 738 | int FI = MF.getFrameInfo().CreateFixedObject(Size: 8, SPOffset: ArgOffset + ArgArea, IsImmutable: true); |
| 739 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 740 | OutChains.push_back( |
| 741 | Elt: DAG.getStore(Chain, dl: DL, Val: VArg, Ptr: DAG.getFrameIndex(FI, VT: PtrVT), |
| 742 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
| 743 | } |
| 744 | |
| 745 | if (!OutChains.empty()) |
| 746 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: OutChains); |
| 747 | |
| 748 | return Chain; |
| 749 | } |
| 750 | |
| 751 | // Check whether any of the argument registers are reserved |
| 752 | static bool isAnyArgRegReserved(const SparcRegisterInfo *TRI, |
| 753 | const MachineFunction &MF) { |
| 754 | // The register window design means that outgoing parameters at O* |
| 755 | // will appear in the callee as I*. |
| 756 | // Be conservative and check both sides of the register names. |
| 757 | bool Outgoing = |
| 758 | llvm::any_of(Range: SP::GPROutgoingArgRegClass, P: [TRI, &MF](MCPhysReg r) { |
| 759 | return TRI->isReservedReg(MF, Reg: r); |
| 760 | }); |
| 761 | bool Incoming = |
| 762 | llvm::any_of(Range: SP::GPRIncomingArgRegClass, P: [TRI, &MF](MCPhysReg r) { |
| 763 | return TRI->isReservedReg(MF, Reg: r); |
| 764 | }); |
| 765 | return Outgoing || Incoming; |
| 766 | } |
| 767 | |
| 768 | static void emitReservedArgRegCallError(const MachineFunction &MF) { |
| 769 | const Function &F = MF.getFunction(); |
| 770 | F.getContext().diagnose(DI: DiagnosticInfoUnsupported{ |
| 771 | F, ("SPARC doesn't support" |
| 772 | " function calls if any of the argument registers is reserved." )}); |
| 773 | } |
| 774 | |
| 775 | SDValue |
| 776 | SparcTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
| 777 | SmallVectorImpl<SDValue> &InVals) const { |
| 778 | if (Subtarget->is64Bit()) |
| 779 | return LowerCall_64(CLI, InVals); |
| 780 | return LowerCall_32(CLI, InVals); |
| 781 | } |
| 782 | |
| 783 | static bool hasReturnsTwiceAttr(SelectionDAG &DAG, SDValue Callee, |
| 784 | const CallBase *Call) { |
| 785 | if (Call) |
| 786 | return Call->hasFnAttr(Kind: Attribute::ReturnsTwice); |
| 787 | |
| 788 | const Function *CalleeFn = nullptr; |
| 789 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
| 790 | CalleeFn = dyn_cast<Function>(Val: G->getGlobal()); |
| 791 | } else if (ExternalSymbolSDNode *E = |
| 792 | dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
| 793 | const Function &Fn = DAG.getMachineFunction().getFunction(); |
| 794 | const Module *M = Fn.getParent(); |
| 795 | const char *CalleeName = E->getSymbol(); |
| 796 | CalleeFn = M->getFunction(Name: CalleeName); |
| 797 | } |
| 798 | |
| 799 | if (!CalleeFn) |
| 800 | return false; |
| 801 | return CalleeFn->hasFnAttribute(Kind: Attribute::ReturnsTwice); |
| 802 | } |
| 803 | |
| 804 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
| 805 | /// for tail call optimization. |
| 806 | bool SparcTargetLowering::IsEligibleForTailCallOptimization( |
| 807 | CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF) const { |
| 808 | |
| 809 | auto &Outs = CLI.Outs; |
| 810 | auto &Caller = MF.getFunction(); |
| 811 | |
| 812 | // Do not tail call opt functions with "disable-tail-calls" attribute. |
| 813 | if (Caller.getFnAttribute(Kind: "disable-tail-calls" ).getValueAsString() == "true" ) |
| 814 | return false; |
| 815 | |
| 816 | // Do not tail call opt if the stack is used to pass parameters. |
| 817 | // 64-bit targets have a slightly higher limit since the ABI requires |
| 818 | // to allocate some space even when all the parameters fit inside registers. |
| 819 | unsigned StackSizeLimit = Subtarget->is64Bit() ? 48 : 0; |
| 820 | if (CCInfo.getStackSize() > StackSizeLimit) |
| 821 | return false; |
| 822 | |
| 823 | // Do not tail call opt if either the callee or caller returns |
| 824 | // a struct and the other does not. |
| 825 | if (!Outs.empty() && Caller.hasStructRetAttr() != Outs[0].Flags.isSRet()) |
| 826 | return false; |
| 827 | |
| 828 | // Byval parameters hand the function a pointer directly into the stack area |
| 829 | // we want to reuse during a tail call. |
| 830 | for (auto &Arg : Outs) |
| 831 | if (Arg.Flags.isByVal()) |
| 832 | return false; |
| 833 | |
| 834 | return true; |
| 835 | } |
| 836 | |
| 837 | // Lower a call for the 32-bit ABI. |
| 838 | SDValue |
| 839 | SparcTargetLowering::LowerCall_32(TargetLowering::CallLoweringInfo &CLI, |
| 840 | SmallVectorImpl<SDValue> &InVals) const { |
| 841 | SelectionDAG &DAG = CLI.DAG; |
| 842 | SDLoc &dl = CLI.DL; |
| 843 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
| 844 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
| 845 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
| 846 | SDValue Chain = CLI.Chain; |
| 847 | SDValue Callee = CLI.Callee; |
| 848 | bool &isTailCall = CLI.IsTailCall; |
| 849 | CallingConv::ID CallConv = CLI.CallConv; |
| 850 | bool isVarArg = CLI.IsVarArg; |
| 851 | MachineFunction &MF = DAG.getMachineFunction(); |
| 852 | LLVMContext &Ctx = *DAG.getContext(); |
| 853 | EVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
| 854 | |
| 855 | // Analyze operands of the call, assigning locations to each operand. |
| 856 | SmallVector<CCValAssign, 16> ArgLocs; |
| 857 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
| 858 | *DAG.getContext()); |
| 859 | CCInfo.AnalyzeCallOperands(Outs, Fn: CC_Sparc32); |
| 860 | |
| 861 | isTailCall = isTailCall && IsEligibleForTailCallOptimization( |
| 862 | CCInfo, CLI, MF&: DAG.getMachineFunction()); |
| 863 | |
| 864 | // Get the size of the outgoing arguments stack space requirement. |
| 865 | unsigned ArgsSize = CCInfo.getStackSize(); |
| 866 | |
| 867 | // Keep stack frames 8-byte aligned. |
| 868 | ArgsSize = (ArgsSize+7) & ~7; |
| 869 | |
| 870 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 871 | |
| 872 | // Create local copies for byval args. |
| 873 | SmallVector<SDValue, 8> ByValArgs; |
| 874 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { |
| 875 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
| 876 | if (!Flags.isByVal()) |
| 877 | continue; |
| 878 | |
| 879 | SDValue Arg = OutVals[i]; |
| 880 | unsigned Size = Flags.getByValSize(); |
| 881 | Align Alignment = Flags.getNonZeroByValAlign(); |
| 882 | |
| 883 | if (Size > 0U) { |
| 884 | int FI = MFI.CreateStackObject(Size, Alignment, isSpillSlot: false); |
| 885 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 886 | SDValue SizeNode = DAG.getConstant(Val: Size, DL: dl, VT: MVT::i32); |
| 887 | |
| 888 | Chain = |
| 889 | DAG.getMemcpy(Chain, dl, Dst: FIPtr, Src: Arg, Size: SizeNode, DstAlign: Alignment, SrcAlign: Alignment, |
| 890 | isVol: false, // isVolatile, |
| 891 | AlwaysInline: (Size <= 32), // AlwaysInline if size <= 32, |
| 892 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), |
| 893 | SrcPtrInfo: MachinePointerInfo()); |
| 894 | ByValArgs.push_back(Elt: FIPtr); |
| 895 | } |
| 896 | else { |
| 897 | SDValue nullVal; |
| 898 | ByValArgs.push_back(Elt: nullVal); |
| 899 | } |
| 900 | } |
| 901 | |
| 902 | assert(!isTailCall || ArgsSize == 0); |
| 903 | |
| 904 | if (!isTailCall) |
| 905 | Chain = DAG.getCALLSEQ_START(Chain, InSize: ArgsSize, OutSize: 0, DL: dl); |
| 906 | |
| 907 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
| 908 | SmallVector<SDValue, 8> MemOpChains; |
| 909 | |
| 910 | const unsigned StackOffset = 92; |
| 911 | bool hasStructRetAttr = false; |
| 912 | unsigned SRetArgSize = 0; |
| 913 | // Walk the register/memloc assignments, inserting copies/loads. |
| 914 | for (unsigned i = 0, realArgIdx = 0, byvalArgIdx = 0, e = ArgLocs.size(); |
| 915 | i != e; |
| 916 | ++i, ++realArgIdx) { |
| 917 | CCValAssign &VA = ArgLocs[i]; |
| 918 | SDValue Arg = OutVals[realArgIdx]; |
| 919 | |
| 920 | ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; |
| 921 | |
| 922 | // Use local copy if it is a byval arg. |
| 923 | if (Flags.isByVal()) { |
| 924 | Arg = ByValArgs[byvalArgIdx++]; |
| 925 | if (!Arg) { |
| 926 | continue; |
| 927 | } |
| 928 | } |
| 929 | |
| 930 | // Promote the value if needed. |
| 931 | switch (VA.getLocInfo()) { |
| 932 | default: llvm_unreachable("Unknown loc info!" ); |
| 933 | case CCValAssign::Full: |
| 934 | case CCValAssign::Indirect: |
| 935 | break; |
| 936 | case CCValAssign::SExt: |
| 937 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
| 938 | break; |
| 939 | case CCValAssign::ZExt: |
| 940 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
| 941 | break; |
| 942 | case CCValAssign::AExt: |
| 943 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
| 944 | break; |
| 945 | case CCValAssign::BCvt: |
| 946 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
| 947 | break; |
| 948 | } |
| 949 | |
| 950 | if (Flags.isSRet()) { |
| 951 | assert(VA.needsCustom()); |
| 952 | |
| 953 | if (isTailCall) |
| 954 | continue; |
| 955 | |
| 956 | // store SRet argument in %sp+64 |
| 957 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
| 958 | SDValue PtrOff = DAG.getIntPtrConstant(Val: 64, DL: dl); |
| 959 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 960 | MemOpChains.push_back( |
| 961 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 962 | hasStructRetAttr = true; |
| 963 | // sret only allowed on first argument |
| 964 | assert(Outs[realArgIdx].OrigArgIndex == 0); |
| 965 | SRetArgSize = |
| 966 | DAG.getDataLayout().getTypeAllocSize(Ty: CLI.getArgs()[0].IndirectType); |
| 967 | continue; |
| 968 | } |
| 969 | |
| 970 | if (VA.needsCustom()) { |
| 971 | assert(VA.getLocVT() == MVT::f64 || VA.getLocVT() == MVT::v2i32); |
| 972 | |
| 973 | if (VA.isMemLoc()) { |
| 974 | unsigned Offset = VA.getLocMemOffset() + StackOffset; |
| 975 | // if it is double-word aligned, just store. |
| 976 | if (Offset % 8 == 0) { |
| 977 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
| 978 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
| 979 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 980 | MemOpChains.push_back( |
| 981 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 982 | continue; |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | if (VA.getLocVT() == MVT::f64) { |
| 987 | // Move from the float value from float registers into the |
| 988 | // integer registers. |
| 989 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val&: Arg)) |
| 990 | Arg = bitcastConstantFPToInt(C, DL: dl, DAG); |
| 991 | else |
| 992 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::v2i32, Operand: Arg); |
| 993 | } |
| 994 | |
| 995 | SDValue Part0 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: MVT::i32, |
| 996 | N1: Arg, |
| 997 | N2: DAG.getConstant(Val: 0, DL: dl, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 998 | SDValue Part1 = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: MVT::i32, |
| 999 | N1: Arg, |
| 1000 | N2: DAG.getConstant(Val: 1, DL: dl, VT: getVectorIdxTy(DL: DAG.getDataLayout()))); |
| 1001 | |
| 1002 | if (VA.isRegLoc()) { |
| 1003 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Part0)); |
| 1004 | assert(i+1 != e); |
| 1005 | CCValAssign &NextVA = ArgLocs[++i]; |
| 1006 | if (NextVA.isRegLoc()) { |
| 1007 | RegsToPass.push_back(Elt: std::make_pair(x: NextVA.getLocReg(), y&: Part1)); |
| 1008 | } else { |
| 1009 | // Store the second part in stack. |
| 1010 | unsigned Offset = NextVA.getLocMemOffset() + StackOffset; |
| 1011 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
| 1012 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
| 1013 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 1014 | MemOpChains.push_back( |
| 1015 | Elt: DAG.getStore(Chain, dl, Val: Part1, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 1016 | } |
| 1017 | } else { |
| 1018 | unsigned Offset = VA.getLocMemOffset() + StackOffset; |
| 1019 | // Store the first part. |
| 1020 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
| 1021 | SDValue PtrOff = DAG.getIntPtrConstant(Val: Offset, DL: dl); |
| 1022 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 1023 | MemOpChains.push_back( |
| 1024 | Elt: DAG.getStore(Chain, dl, Val: Part0, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 1025 | // Store the second part. |
| 1026 | PtrOff = DAG.getIntPtrConstant(Val: Offset + 4, DL: dl); |
| 1027 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 1028 | MemOpChains.push_back( |
| 1029 | Elt: DAG.getStore(Chain, dl, Val: Part1, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 1030 | } |
| 1031 | continue; |
| 1032 | } |
| 1033 | |
| 1034 | if (VA.getLocInfo() == CCValAssign::Indirect) { |
| 1035 | // Store the argument in a stack slot and pass its address. |
| 1036 | unsigned ArgIndex = Outs[realArgIdx].OrigArgIndex; |
| 1037 | assert(Outs[realArgIdx].PartOffset == 0); |
| 1038 | |
| 1039 | EVT SlotVT; |
| 1040 | if (i + 1 != e && Outs[realArgIdx + 1].OrigArgIndex == ArgIndex) { |
| 1041 | Type *OrigArgType = CLI.Args[ArgIndex].Ty; |
| 1042 | EVT OrigArgVT = getValueType(DL: MF.getDataLayout(), Ty: OrigArgType); |
| 1043 | MVT PartVT = |
| 1044 | getRegisterTypeForCallingConv(Context&: Ctx, CC: CLI.CallConv, VT: OrigArgVT); |
| 1045 | unsigned N = |
| 1046 | getNumRegistersForCallingConv(Context&: Ctx, CC: CLI.CallConv, VT: OrigArgVT); |
| 1047 | SlotVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: PartVT.getSizeInBits() * N); |
| 1048 | } else { |
| 1049 | SlotVT = Outs[realArgIdx].VT; |
| 1050 | } |
| 1051 | |
| 1052 | SDValue SpillSlot = DAG.CreateStackTemporary(VT: SlotVT); |
| 1053 | int FI = cast<FrameIndexSDNode>(Val&: SpillSlot)->getIndex(); |
| 1054 | MemOpChains.push_back( |
| 1055 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: SpillSlot, |
| 1056 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
| 1057 | // If the original argument was split (e.g. f128), we need |
| 1058 | // to store all parts of it here (and pass just one address). |
| 1059 | while (i + 1 != e && Outs[realArgIdx + 1].OrigArgIndex == ArgIndex) { |
| 1060 | SDValue PartValue = OutVals[realArgIdx + 1]; |
| 1061 | unsigned PartOffset = Outs[realArgIdx + 1].PartOffset; |
| 1062 | SDValue Address = DAG.getMemBasePlusOffset( |
| 1063 | Base: DAG.getFrameIndex(FI, VT: PtrVT), Offset: TypeSize::getFixed(ExactSize: PartOffset), DL: dl); |
| 1064 | MemOpChains.push_back( |
| 1065 | Elt: DAG.getStore(Chain, dl, Val: PartValue, Ptr: Address, |
| 1066 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
| 1067 | assert((PartOffset + PartValue.getValueType().getStoreSize() <= |
| 1068 | SlotVT.getStoreSize()) && |
| 1069 | "Not enough space for argument part!" ); |
| 1070 | ++i; |
| 1071 | ++realArgIdx; |
| 1072 | } |
| 1073 | |
| 1074 | Arg = SpillSlot; |
| 1075 | } |
| 1076 | |
| 1077 | // Arguments that can be passed on register must be kept at |
| 1078 | // RegsToPass vector |
| 1079 | if (VA.isRegLoc()) { |
| 1080 | if (VA.getLocVT() != MVT::f32) { |
| 1081 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
| 1082 | continue; |
| 1083 | } |
| 1084 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i32, Operand: Arg); |
| 1085 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
| 1086 | continue; |
| 1087 | } |
| 1088 | |
| 1089 | assert(VA.isMemLoc()); |
| 1090 | |
| 1091 | // Create a store off the stack pointer for this argument. |
| 1092 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: MVT::i32); |
| 1093 | SDValue PtrOff = DAG.getIntPtrConstant(Val: VA.getLocMemOffset() + StackOffset, |
| 1094 | DL: dl); |
| 1095 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: StackPtr, N2: PtrOff); |
| 1096 | MemOpChains.push_back( |
| 1097 | Elt: DAG.getStore(Chain, dl, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 1098 | } |
| 1099 | |
| 1100 | |
| 1101 | // Emit all stores, make sure the occur before any copies into physregs. |
| 1102 | if (!MemOpChains.empty()) |
| 1103 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOpChains); |
| 1104 | |
| 1105 | // Build a sequence of copy-to-reg nodes chained together with token |
| 1106 | // chain and flag operands which copy the outgoing args into registers. |
| 1107 | // The InGlue in necessary since all emitted instructions must be |
| 1108 | // stuck together. |
| 1109 | SDValue InGlue; |
| 1110 | for (const auto &[OrigReg, N] : RegsToPass) { |
| 1111 | Register Reg = isTailCall ? OrigReg : toCallerWindow(Reg: OrigReg); |
| 1112 | Chain = DAG.getCopyToReg(Chain, dl, Reg, N, Glue: InGlue); |
| 1113 | InGlue = Chain.getValue(R: 1); |
| 1114 | } |
| 1115 | |
| 1116 | bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, Call: CLI.CB); |
| 1117 | |
| 1118 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
| 1119 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
| 1120 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
| 1121 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) |
| 1122 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL: dl, VT: MVT::i32, offset: 0); |
| 1123 | else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) |
| 1124 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: MVT::i32); |
| 1125 | |
| 1126 | // Returns a chain & a flag for retval copy to use |
| 1127 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 1128 | SmallVector<SDValue, 8> Ops; |
| 1129 | Ops.push_back(Elt: Chain); |
| 1130 | Ops.push_back(Elt: Callee); |
| 1131 | if (hasStructRetAttr) |
| 1132 | Ops.push_back(Elt: DAG.getTargetConstant(Val: SRetArgSize, DL: dl, VT: MVT::i32)); |
| 1133 | for (const auto &[OrigReg, N] : RegsToPass) { |
| 1134 | Register Reg = isTailCall ? OrigReg : toCallerWindow(Reg: OrigReg); |
| 1135 | Ops.push_back(Elt: DAG.getRegister(Reg, VT: N.getValueType())); |
| 1136 | } |
| 1137 | |
| 1138 | // Add a register mask operand representing the call-preserved registers. |
| 1139 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
| 1140 | const uint32_t *Mask = |
| 1141 | ((hasReturnsTwice) |
| 1142 | ? TRI->getRTCallPreservedMask(CC: CallConv) |
| 1143 | : TRI->getCallPreservedMask(MF: DAG.getMachineFunction(), CC: CallConv)); |
| 1144 | |
| 1145 | if (isAnyArgRegReserved(TRI, MF)) |
| 1146 | emitReservedArgRegCallError(MF); |
| 1147 | |
| 1148 | assert(Mask && "Missing call preserved mask for calling convention" ); |
| 1149 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask: Mask)); |
| 1150 | |
| 1151 | if (InGlue.getNode()) |
| 1152 | Ops.push_back(Elt: InGlue); |
| 1153 | |
| 1154 | if (isTailCall) { |
| 1155 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
| 1156 | return DAG.getNode(Opcode: SPISD::TAIL_CALL, DL: dl, VT: MVT::Other, Ops); |
| 1157 | } |
| 1158 | |
| 1159 | Chain = DAG.getNode(Opcode: SPISD::CALL, DL: dl, VTList: NodeTys, Ops); |
| 1160 | InGlue = Chain.getValue(R: 1); |
| 1161 | |
| 1162 | Chain = DAG.getCALLSEQ_END(Chain, Size1: ArgsSize, Size2: 0, Glue: InGlue, DL: dl); |
| 1163 | InGlue = Chain.getValue(R: 1); |
| 1164 | |
| 1165 | // Assign locations to each value returned by this call. |
| 1166 | SmallVector<CCValAssign, 16> RVLocs; |
| 1167 | CCState RVInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
| 1168 | *DAG.getContext()); |
| 1169 | |
| 1170 | RVInfo.AnalyzeCallResult(Ins, Fn: RetCC_Sparc32); |
| 1171 | |
| 1172 | // Copy all of the result registers out of their specified physreg. |
| 1173 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
| 1174 | assert(RVLocs[i].isRegLoc() && "Can only return in registers!" ); |
| 1175 | if (RVLocs[i].getLocVT() == MVT::v2i32) { |
| 1176 | SDValue Vec = DAG.getNode(Opcode: ISD::UNDEF, DL: dl, VT: MVT::v2i32); |
| 1177 | SDValue Lo = DAG.getCopyFromReg( |
| 1178 | Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i++].getLocReg()), VT: MVT::i32, Glue: InGlue); |
| 1179 | Chain = Lo.getValue(R: 1); |
| 1180 | InGlue = Lo.getValue(R: 2); |
| 1181 | Vec = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: dl, VT: MVT::v2i32, N1: Vec, N2: Lo, |
| 1182 | N3: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
| 1183 | SDValue Hi = DAG.getCopyFromReg( |
| 1184 | Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i].getLocReg()), VT: MVT::i32, Glue: InGlue); |
| 1185 | Chain = Hi.getValue(R: 1); |
| 1186 | InGlue = Hi.getValue(R: 2); |
| 1187 | Vec = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: dl, VT: MVT::v2i32, N1: Vec, N2: Hi, |
| 1188 | N3: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
| 1189 | InVals.push_back(Elt: Vec); |
| 1190 | } else { |
| 1191 | Chain = |
| 1192 | DAG.getCopyFromReg(Chain, dl, Reg: toCallerWindow(Reg: RVLocs[i].getLocReg()), |
| 1193 | VT: RVLocs[i].getValVT(), Glue: InGlue) |
| 1194 | .getValue(R: 1); |
| 1195 | InGlue = Chain.getValue(R: 2); |
| 1196 | InVals.push_back(Elt: Chain.getValue(R: 0)); |
| 1197 | } |
| 1198 | } |
| 1199 | |
| 1200 | return Chain; |
| 1201 | } |
| 1202 | |
| 1203 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
| 1204 | // this table could be generated automatically from RegInfo. |
| 1205 | Register SparcTargetLowering::getRegisterByName(const char* RegName, LLT VT, |
| 1206 | const MachineFunction &MF) const { |
| 1207 | Register Reg = StringSwitch<Register>(RegName) |
| 1208 | .Case(S: "i0" , Value: SP::I0).Case(S: "i1" , Value: SP::I1).Case(S: "i2" , Value: SP::I2).Case(S: "i3" , Value: SP::I3) |
| 1209 | .Case(S: "i4" , Value: SP::I4).Case(S: "i5" , Value: SP::I5).Case(S: "i6" , Value: SP::I6).Case(S: "i7" , Value: SP::I7) |
| 1210 | .Case(S: "o0" , Value: SP::O0).Case(S: "o1" , Value: SP::O1).Case(S: "o2" , Value: SP::O2).Case(S: "o3" , Value: SP::O3) |
| 1211 | .Case(S: "o4" , Value: SP::O4).Case(S: "o5" , Value: SP::O5).Case(S: "o6" , Value: SP::O6).Case(S: "o7" , Value: SP::O7) |
| 1212 | .Case(S: "l0" , Value: SP::L0).Case(S: "l1" , Value: SP::L1).Case(S: "l2" , Value: SP::L2).Case(S: "l3" , Value: SP::L3) |
| 1213 | .Case(S: "l4" , Value: SP::L4).Case(S: "l5" , Value: SP::L5).Case(S: "l6" , Value: SP::L6).Case(S: "l7" , Value: SP::L7) |
| 1214 | .Case(S: "g0" , Value: SP::G0).Case(S: "g1" , Value: SP::G1).Case(S: "g2" , Value: SP::G2).Case(S: "g3" , Value: SP::G3) |
| 1215 | .Case(S: "g4" , Value: SP::G4).Case(S: "g5" , Value: SP::G5).Case(S: "g6" , Value: SP::G6).Case(S: "g7" , Value: SP::G7) |
| 1216 | .Default(Value: 0); |
| 1217 | |
| 1218 | // If we're directly referencing register names |
| 1219 | // (e.g in GCC C extension `register int r asm("g1");`), |
| 1220 | // make sure that said register is in the reserve list. |
| 1221 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
| 1222 | if (!TRI->isReservedReg(MF, Reg)) |
| 1223 | Reg = Register(); |
| 1224 | |
| 1225 | return Reg; |
| 1226 | } |
| 1227 | |
| 1228 | // Fixup floating point arguments in the ... part of a varargs call. |
| 1229 | // |
| 1230 | // The SPARC v9 ABI requires that floating point arguments are treated the same |
| 1231 | // as integers when calling a varargs function. This does not apply to the |
| 1232 | // fixed arguments that are part of the function's prototype. |
| 1233 | // |
| 1234 | // This function post-processes a CCValAssign array created by |
| 1235 | // AnalyzeCallOperands(). |
| 1236 | static void fixupVariableFloatArgs(SmallVectorImpl<CCValAssign> &ArgLocs, |
| 1237 | ArrayRef<ISD::OutputArg> Outs) { |
| 1238 | for (CCValAssign &VA : ArgLocs) { |
| 1239 | MVT ValTy = VA.getLocVT(); |
| 1240 | // FIXME: What about f32 arguments? C promotes them to f64 when calling |
| 1241 | // varargs functions. |
| 1242 | if (!VA.isRegLoc() || (ValTy != MVT::f64 && ValTy != MVT::f128)) |
| 1243 | continue; |
| 1244 | // The fixed arguments to a varargs function still go in FP registers. |
| 1245 | if (!Outs[VA.getValNo()].Flags.isVarArg()) |
| 1246 | continue; |
| 1247 | |
| 1248 | // This floating point argument should be reassigned. |
| 1249 | // Determine the offset into the argument array. |
| 1250 | Register firstReg = (ValTy == MVT::f64) ? SP::D0 : SP::Q0; |
| 1251 | unsigned argSize = (ValTy == MVT::f64) ? 8 : 16; |
| 1252 | unsigned Offset = argSize * (VA.getLocReg() - firstReg); |
| 1253 | assert(Offset < 16*8 && "Offset out of range, bad register enum?" ); |
| 1254 | |
| 1255 | if (Offset < 6*8) { |
| 1256 | // This argument should go in %i0-%i5. |
| 1257 | unsigned IReg = SP::I0 + Offset/8; |
| 1258 | if (ValTy == MVT::f64) |
| 1259 | // Full register, just bitconvert into i64. |
| 1260 | VA = CCValAssign::getReg(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Reg: IReg, LocVT: MVT::i64, |
| 1261 | HTP: CCValAssign::BCvt); |
| 1262 | else { |
| 1263 | assert(ValTy == MVT::f128 && "Unexpected type!" ); |
| 1264 | // Full register, just bitconvert into i128 -- We will lower this into |
| 1265 | // two i64s in LowerCall_64. |
| 1266 | VA = CCValAssign::getCustomReg(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Reg: IReg, |
| 1267 | LocVT: MVT::i128, HTP: CCValAssign::BCvt); |
| 1268 | } |
| 1269 | } else { |
| 1270 | // This needs to go to memory, we're out of integer registers. |
| 1271 | VA = CCValAssign::getMem(ValNo: VA.getValNo(), ValVT: VA.getValVT(), Offset, |
| 1272 | LocVT: VA.getLocVT(), HTP: VA.getLocInfo()); |
| 1273 | } |
| 1274 | } |
| 1275 | } |
| 1276 | |
| 1277 | // Lower a call for the 64-bit ABI. |
| 1278 | SDValue |
| 1279 | SparcTargetLowering::LowerCall_64(TargetLowering::CallLoweringInfo &CLI, |
| 1280 | SmallVectorImpl<SDValue> &InVals) const { |
| 1281 | SelectionDAG &DAG = CLI.DAG; |
| 1282 | SDLoc DL = CLI.DL; |
| 1283 | SDValue Chain = CLI.Chain; |
| 1284 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 1285 | MachineFunction &MF = DAG.getMachineFunction(); |
| 1286 | |
| 1287 | // Analyze operands of the call, assigning locations to each operand. |
| 1288 | SmallVector<CCValAssign, 16> ArgLocs; |
| 1289 | CCState CCInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), ArgLocs, |
| 1290 | *DAG.getContext()); |
| 1291 | CCInfo.AnalyzeCallOperands(Outs: CLI.Outs, Fn: CC_Sparc64); |
| 1292 | |
| 1293 | CLI.IsTailCall = CLI.IsTailCall && IsEligibleForTailCallOptimization( |
| 1294 | CCInfo, CLI, MF&: DAG.getMachineFunction()); |
| 1295 | |
| 1296 | // Get the size of the outgoing arguments stack space requirement. |
| 1297 | // The stack offset computed by CC_Sparc64 includes all arguments. |
| 1298 | // Called functions expect 6 argument words to exist in the stack frame, used |
| 1299 | // or not. |
| 1300 | unsigned StackReserved = 6 * 8u; |
| 1301 | unsigned ArgsSize = std::max<unsigned>(a: StackReserved, b: CCInfo.getStackSize()); |
| 1302 | |
| 1303 | // Keep stack frames 16-byte aligned. |
| 1304 | ArgsSize = alignTo(Value: ArgsSize, Align: 16); |
| 1305 | |
| 1306 | // Varargs calls require special treatment. |
| 1307 | if (CLI.IsVarArg) |
| 1308 | fixupVariableFloatArgs(ArgLocs, Outs: CLI.Outs); |
| 1309 | |
| 1310 | assert(!CLI.IsTailCall || ArgsSize == StackReserved); |
| 1311 | |
| 1312 | // Adjust the stack pointer to make room for the arguments. |
| 1313 | // FIXME: Use hasReservedCallFrame to avoid %sp adjustments around all calls |
| 1314 | // with more than 6 arguments. |
| 1315 | if (!CLI.IsTailCall) |
| 1316 | Chain = DAG.getCALLSEQ_START(Chain, InSize: ArgsSize, OutSize: 0, DL); |
| 1317 | |
| 1318 | // Collect the set of registers to pass to the function and their values. |
| 1319 | // This will be emitted as a sequence of CopyToReg nodes glued to the call |
| 1320 | // instruction. |
| 1321 | SmallVector<std::pair<Register, SDValue>, 8> RegsToPass; |
| 1322 | |
| 1323 | // Collect chains from all the memory opeations that copy arguments to the |
| 1324 | // stack. They must follow the stack pointer adjustment above and precede the |
| 1325 | // call instruction itself. |
| 1326 | SmallVector<SDValue, 8> MemOpChains; |
| 1327 | |
| 1328 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
| 1329 | const CCValAssign &VA = ArgLocs[i]; |
| 1330 | SDValue Arg = CLI.OutVals[i]; |
| 1331 | |
| 1332 | // Promote the value if needed. |
| 1333 | switch (VA.getLocInfo()) { |
| 1334 | default: |
| 1335 | llvm_unreachable("Unknown location info!" ); |
| 1336 | case CCValAssign::Full: |
| 1337 | break; |
| 1338 | case CCValAssign::SExt: |
| 1339 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
| 1340 | break; |
| 1341 | case CCValAssign::ZExt: |
| 1342 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
| 1343 | break; |
| 1344 | case CCValAssign::AExt: |
| 1345 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: Arg); |
| 1346 | break; |
| 1347 | case CCValAssign::BCvt: |
| 1348 | // fixupVariableFloatArgs() may create bitcasts from f128 to i128. But |
| 1349 | // SPARC does not support i128 natively. Lower it into two i64, see below. |
| 1350 | if (!VA.needsCustom() || VA.getValVT() != MVT::f128 |
| 1351 | || VA.getLocVT() != MVT::i128) |
| 1352 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getLocVT(), Operand: Arg); |
| 1353 | break; |
| 1354 | } |
| 1355 | |
| 1356 | if (VA.isRegLoc()) { |
| 1357 | if (VA.needsCustom() && VA.getValVT() == MVT::f128 |
| 1358 | && VA.getLocVT() == MVT::i128) { |
| 1359 | // Store and reload into the integer register reg and reg+1. |
| 1360 | unsigned Offset = 8 * (VA.getLocReg() - SP::I0); |
| 1361 | unsigned StackOffset = Offset + Subtarget->getStackPointerBias() + 128; |
| 1362 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: PtrVT); |
| 1363 | SDValue HiPtrOff = DAG.getIntPtrConstant(Val: StackOffset, DL); |
| 1364 | HiPtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: HiPtrOff); |
| 1365 | SDValue LoPtrOff = DAG.getIntPtrConstant(Val: StackOffset + 8, DL); |
| 1366 | LoPtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: LoPtrOff); |
| 1367 | |
| 1368 | // Store to %sp+BIAS+128+Offset |
| 1369 | SDValue Store = |
| 1370 | DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: HiPtrOff, PtrInfo: MachinePointerInfo()); |
| 1371 | // Load into Reg and Reg+1 |
| 1372 | SDValue Hi64 = |
| 1373 | DAG.getLoad(VT: MVT::i64, dl: DL, Chain: Store, Ptr: HiPtrOff, PtrInfo: MachinePointerInfo()); |
| 1374 | SDValue Lo64 = |
| 1375 | DAG.getLoad(VT: MVT::i64, dl: DL, Chain: Store, Ptr: LoPtrOff, PtrInfo: MachinePointerInfo()); |
| 1376 | |
| 1377 | Register HiReg = VA.getLocReg(); |
| 1378 | Register LoReg = VA.getLocReg() + 1; |
| 1379 | if (!CLI.IsTailCall) { |
| 1380 | HiReg = toCallerWindow(Reg: HiReg); |
| 1381 | LoReg = toCallerWindow(Reg: LoReg); |
| 1382 | } |
| 1383 | |
| 1384 | RegsToPass.push_back(Elt: std::make_pair(x&: HiReg, y&: Hi64)); |
| 1385 | RegsToPass.push_back(Elt: std::make_pair(x&: LoReg, y&: Lo64)); |
| 1386 | continue; |
| 1387 | } |
| 1388 | |
| 1389 | // The custom bit on an i32 return value indicates that it should be |
| 1390 | // passed in the high bits of the register. |
| 1391 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) { |
| 1392 | Arg = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: Arg, |
| 1393 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
| 1394 | |
| 1395 | // The next value may go in the low bits of the same register. |
| 1396 | // Handle both at once. |
| 1397 | if (i+1 < ArgLocs.size() && ArgLocs[i+1].isRegLoc() && |
| 1398 | ArgLocs[i+1].getLocReg() == VA.getLocReg()) { |
| 1399 | SDValue NV = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: MVT::i64, |
| 1400 | Operand: CLI.OutVals[i+1]); |
| 1401 | Arg = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i64, N1: Arg, N2: NV); |
| 1402 | // Skip the next value, it's already done. |
| 1403 | ++i; |
| 1404 | } |
| 1405 | } |
| 1406 | |
| 1407 | Register Reg = VA.getLocReg(); |
| 1408 | if (!CLI.IsTailCall) |
| 1409 | Reg = toCallerWindow(Reg); |
| 1410 | RegsToPass.push_back(Elt: std::make_pair(x&: Reg, y&: Arg)); |
| 1411 | continue; |
| 1412 | } |
| 1413 | |
| 1414 | assert(VA.isMemLoc()); |
| 1415 | |
| 1416 | // Create a store off the stack pointer for this argument. |
| 1417 | SDValue StackPtr = DAG.getRegister(Reg: SP::O6, VT: PtrVT); |
| 1418 | // The argument area starts at %fp+BIAS+128 in the callee frame, |
| 1419 | // %sp+BIAS+128 in ours. |
| 1420 | SDValue PtrOff = DAG.getIntPtrConstant(Val: VA.getLocMemOffset() + |
| 1421 | Subtarget->getStackPointerBias() + |
| 1422 | 128, DL); |
| 1423 | PtrOff = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: StackPtr, N2: PtrOff); |
| 1424 | MemOpChains.push_back( |
| 1425 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo())); |
| 1426 | } |
| 1427 | |
| 1428 | // Emit all stores, make sure they occur before the call. |
| 1429 | if (!MemOpChains.empty()) |
| 1430 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: MemOpChains); |
| 1431 | |
| 1432 | // Build a sequence of CopyToReg nodes glued together with token chain and |
| 1433 | // glue operands which copy the outgoing args into registers. The InGlue is |
| 1434 | // necessary since all emitted instructions must be stuck together in order |
| 1435 | // to pass the live physical registers. |
| 1436 | SDValue InGlue; |
| 1437 | for (const auto &[Reg, N] : RegsToPass) { |
| 1438 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg, N, Glue: InGlue); |
| 1439 | InGlue = Chain.getValue(R: 1); |
| 1440 | } |
| 1441 | |
| 1442 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
| 1443 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
| 1444 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
| 1445 | SDValue Callee = CLI.Callee; |
| 1446 | bool hasReturnsTwice = hasReturnsTwiceAttr(DAG, Callee, Call: CLI.CB); |
| 1447 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) |
| 1448 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL, VT: PtrVT, offset: 0); |
| 1449 | else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) |
| 1450 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: PtrVT); |
| 1451 | |
| 1452 | // Build the operands for the call instruction itself. |
| 1453 | SmallVector<SDValue, 8> Ops; |
| 1454 | Ops.push_back(Elt: Chain); |
| 1455 | Ops.push_back(Elt: Callee); |
| 1456 | for (const auto &[Reg, N] : RegsToPass) |
| 1457 | Ops.push_back(Elt: DAG.getRegister(Reg, VT: N.getValueType())); |
| 1458 | |
| 1459 | // Add a register mask operand representing the call-preserved registers. |
| 1460 | const SparcRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
| 1461 | const uint32_t *Mask = |
| 1462 | ((hasReturnsTwice) ? TRI->getRTCallPreservedMask(CC: CLI.CallConv) |
| 1463 | : TRI->getCallPreservedMask(MF: DAG.getMachineFunction(), |
| 1464 | CC: CLI.CallConv)); |
| 1465 | |
| 1466 | if (isAnyArgRegReserved(TRI, MF)) |
| 1467 | emitReservedArgRegCallError(MF); |
| 1468 | |
| 1469 | assert(Mask && "Missing call preserved mask for calling convention" ); |
| 1470 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask: Mask)); |
| 1471 | |
| 1472 | // Make sure the CopyToReg nodes are glued to the call instruction which |
| 1473 | // consumes the registers. |
| 1474 | if (InGlue.getNode()) |
| 1475 | Ops.push_back(Elt: InGlue); |
| 1476 | |
| 1477 | // Now the call itself. |
| 1478 | if (CLI.IsTailCall) { |
| 1479 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
| 1480 | return DAG.getNode(Opcode: SPISD::TAIL_CALL, DL, VT: MVT::Other, Ops); |
| 1481 | } |
| 1482 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 1483 | Chain = DAG.getNode(Opcode: SPISD::CALL, DL, VTList: NodeTys, Ops); |
| 1484 | InGlue = Chain.getValue(R: 1); |
| 1485 | |
| 1486 | // Revert the stack pointer immediately after the call. |
| 1487 | Chain = DAG.getCALLSEQ_END(Chain, Size1: ArgsSize, Size2: 0, Glue: InGlue, DL); |
| 1488 | InGlue = Chain.getValue(R: 1); |
| 1489 | |
| 1490 | // Now extract the return values. This is more or less the same as |
| 1491 | // LowerFormalArguments_64. |
| 1492 | |
| 1493 | // Assign locations to each value returned by this call. |
| 1494 | SmallVector<CCValAssign, 16> RVLocs; |
| 1495 | CCState RVInfo(CLI.CallConv, CLI.IsVarArg, DAG.getMachineFunction(), RVLocs, |
| 1496 | *DAG.getContext()); |
| 1497 | |
| 1498 | // Set inreg flag manually for codegen generated library calls that |
| 1499 | // return float. |
| 1500 | if (CLI.Ins.size() == 1 && CLI.Ins[0].VT == MVT::f32 && !CLI.CB) |
| 1501 | CLI.Ins[0].Flags.setInReg(); |
| 1502 | |
| 1503 | RVInfo.AnalyzeCallResult(Ins: CLI.Ins, Fn: RetCC_Sparc64); |
| 1504 | |
| 1505 | // Copy all of the result registers out of their specified physreg. |
| 1506 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
| 1507 | CCValAssign &VA = RVLocs[i]; |
| 1508 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
| 1509 | unsigned Reg = toCallerWindow(Reg: VA.getLocReg()); |
| 1510 | |
| 1511 | // When returning 'inreg {i32, i32 }', two consecutive i32 arguments can |
| 1512 | // reside in the same register in the high and low bits. Reuse the |
| 1513 | // CopyFromReg previous node to avoid duplicate copies. |
| 1514 | SDValue RV; |
| 1515 | if (RegisterSDNode *SrcReg = dyn_cast<RegisterSDNode>(Val: Chain.getOperand(i: 1))) |
| 1516 | if (SrcReg->getReg() == Reg && Chain->getOpcode() == ISD::CopyFromReg) |
| 1517 | RV = Chain.getValue(R: 0); |
| 1518 | |
| 1519 | // But usually we'll create a new CopyFromReg for a different register. |
| 1520 | if (!RV.getNode()) { |
| 1521 | RV = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RVLocs[i].getLocVT(), Glue: InGlue); |
| 1522 | Chain = RV.getValue(R: 1); |
| 1523 | InGlue = Chain.getValue(R: 2); |
| 1524 | } |
| 1525 | |
| 1526 | // Get the high bits for i32 struct elements. |
| 1527 | if (VA.getValVT() == MVT::i32 && VA.needsCustom()) |
| 1528 | RV = DAG.getNode(Opcode: ISD::SRL, DL, VT: VA.getLocVT(), N1: RV, |
| 1529 | N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
| 1530 | |
| 1531 | // The callee promoted the return value, so insert an Assert?ext SDNode so |
| 1532 | // we won't promote the value again in this function. |
| 1533 | switch (VA.getLocInfo()) { |
| 1534 | case CCValAssign::SExt: |
| 1535 | RV = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: RV, |
| 1536 | N2: DAG.getValueType(VA.getValVT())); |
| 1537 | break; |
| 1538 | case CCValAssign::ZExt: |
| 1539 | RV = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: RV, |
| 1540 | N2: DAG.getValueType(VA.getValVT())); |
| 1541 | break; |
| 1542 | default: |
| 1543 | break; |
| 1544 | } |
| 1545 | |
| 1546 | // Truncate the register down to the return value type. |
| 1547 | if (VA.isExtInLoc()) |
| 1548 | RV = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: RV); |
| 1549 | |
| 1550 | InVals.push_back(Elt: RV); |
| 1551 | } |
| 1552 | |
| 1553 | return Chain; |
| 1554 | } |
| 1555 | |
| 1556 | //===----------------------------------------------------------------------===// |
| 1557 | // TargetLowering Implementation |
| 1558 | //===----------------------------------------------------------------------===// |
| 1559 | |
| 1560 | TargetLowering::AtomicExpansionKind |
| 1561 | SparcTargetLowering::shouldExpandAtomicRMWInIR(const AtomicRMWInst *AI) const { |
| 1562 | if (AI->getOperation() == AtomicRMWInst::Xchg && |
| 1563 | AI->getType()->getPrimitiveSizeInBits() == 32) |
| 1564 | return AtomicExpansionKind::None; // Uses xchg instruction |
| 1565 | |
| 1566 | return AtomicExpansionKind::CmpXChg; |
| 1567 | } |
| 1568 | |
| 1569 | /// intCondCCodeToRcond - Convert a DAG integer condition code to a SPARC |
| 1570 | /// rcond condition. |
| 1571 | static SPCC::CondCodes intCondCCodeToRcond(ISD::CondCode CC) { |
| 1572 | switch (CC) { |
| 1573 | default: |
| 1574 | llvm_unreachable("Unknown/unsigned integer condition code!" ); |
| 1575 | case ISD::SETEQ: |
| 1576 | return SPCC::REG_Z; |
| 1577 | case ISD::SETNE: |
| 1578 | return SPCC::REG_NZ; |
| 1579 | case ISD::SETLT: |
| 1580 | return SPCC::REG_LZ; |
| 1581 | case ISD::SETGT: |
| 1582 | return SPCC::REG_GZ; |
| 1583 | case ISD::SETLE: |
| 1584 | return SPCC::REG_LEZ; |
| 1585 | case ISD::SETGE: |
| 1586 | return SPCC::REG_GEZ; |
| 1587 | } |
| 1588 | } |
| 1589 | |
| 1590 | /// IntCondCCodeToICC - Convert a DAG integer condition code to a SPARC ICC |
| 1591 | /// condition. |
| 1592 | static SPCC::CondCodes IntCondCCodeToICC(ISD::CondCode CC) { |
| 1593 | switch (CC) { |
| 1594 | default: llvm_unreachable("Unknown integer condition code!" ); |
| 1595 | case ISD::SETEQ: return SPCC::ICC_E; |
| 1596 | case ISD::SETNE: return SPCC::ICC_NE; |
| 1597 | case ISD::SETLT: return SPCC::ICC_L; |
| 1598 | case ISD::SETGT: return SPCC::ICC_G; |
| 1599 | case ISD::SETLE: return SPCC::ICC_LE; |
| 1600 | case ISD::SETGE: return SPCC::ICC_GE; |
| 1601 | case ISD::SETULT: return SPCC::ICC_CS; |
| 1602 | case ISD::SETULE: return SPCC::ICC_LEU; |
| 1603 | case ISD::SETUGT: return SPCC::ICC_GU; |
| 1604 | case ISD::SETUGE: return SPCC::ICC_CC; |
| 1605 | } |
| 1606 | } |
| 1607 | |
| 1608 | /// FPCondCCodeToFCC - Convert a DAG floatingp oint condition code to a SPARC |
| 1609 | /// FCC condition. |
| 1610 | static SPCC::CondCodes FPCondCCodeToFCC(ISD::CondCode CC) { |
| 1611 | switch (CC) { |
| 1612 | default: llvm_unreachable("Unknown fp condition code!" ); |
| 1613 | case ISD::SETEQ: |
| 1614 | case ISD::SETOEQ: return SPCC::FCC_E; |
| 1615 | case ISD::SETNE: |
| 1616 | case ISD::SETUNE: return SPCC::FCC_NE; |
| 1617 | case ISD::SETLT: |
| 1618 | case ISD::SETOLT: return SPCC::FCC_L; |
| 1619 | case ISD::SETGT: |
| 1620 | case ISD::SETOGT: return SPCC::FCC_G; |
| 1621 | case ISD::SETLE: |
| 1622 | case ISD::SETOLE: return SPCC::FCC_LE; |
| 1623 | case ISD::SETGE: |
| 1624 | case ISD::SETOGE: return SPCC::FCC_GE; |
| 1625 | case ISD::SETULT: return SPCC::FCC_UL; |
| 1626 | case ISD::SETULE: return SPCC::FCC_ULE; |
| 1627 | case ISD::SETUGT: return SPCC::FCC_UG; |
| 1628 | case ISD::SETUGE: return SPCC::FCC_UGE; |
| 1629 | case ISD::SETUO: return SPCC::FCC_U; |
| 1630 | case ISD::SETO: return SPCC::FCC_O; |
| 1631 | case ISD::SETONE: return SPCC::FCC_LG; |
| 1632 | case ISD::SETUEQ: return SPCC::FCC_UE; |
| 1633 | } |
| 1634 | } |
| 1635 | |
| 1636 | SparcTargetLowering::SparcTargetLowering(const TargetMachine &TM, |
| 1637 | const SparcSubtarget &STI) |
| 1638 | : TargetLowering(TM, STI), Subtarget(&STI) { |
| 1639 | MVT PtrVT = MVT::getIntegerVT(BitWidth: TM.getPointerSizeInBits(AS: 0)); |
| 1640 | |
| 1641 | // Instructions which use registers as conditionals examine all the |
| 1642 | // bits (as does the pseudo SELECT_CC expansion). I don't think it |
| 1643 | // matters much whether it's ZeroOrOneBooleanContent, or |
| 1644 | // ZeroOrNegativeOneBooleanContent, so, arbitrarily choose the |
| 1645 | // former. |
| 1646 | setBooleanContents(ZeroOrOneBooleanContent); |
| 1647 | setBooleanVectorContents(ZeroOrOneBooleanContent); |
| 1648 | |
| 1649 | // Set up the register classes. |
| 1650 | addRegisterClass(VT: MVT::i32, RC: &SP::IntRegsRegClass); |
| 1651 | if (!Subtarget->useSoftFloat()) { |
| 1652 | addRegisterClass(VT: MVT::f32, RC: &SP::FPRegsRegClass); |
| 1653 | addRegisterClass(VT: MVT::f64, RC: &SP::DFPRegsRegClass); |
| 1654 | addRegisterClass(VT: MVT::f128, RC: &SP::QFPRegsRegClass); |
| 1655 | } |
| 1656 | if (Subtarget->is64Bit()) { |
| 1657 | addRegisterClass(VT: MVT::i64, RC: &SP::I64RegsRegClass); |
| 1658 | } else { |
| 1659 | // On 32bit sparc, we define a double-register 32bit register |
| 1660 | // class, as well. This is modeled in LLVM as a 2-vector of i32. |
| 1661 | addRegisterClass(VT: MVT::v2i32, RC: &SP::IntPairRegClass); |
| 1662 | |
| 1663 | // ...but almost all operations must be expanded, so set that as |
| 1664 | // the default. |
| 1665 | for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) { |
| 1666 | setOperationAction(Op, VT: MVT::v2i32, Action: Expand); |
| 1667 | } |
| 1668 | // Truncating/extending stores/loads are also not supported. |
| 1669 | for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) { |
| 1670 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
| 1671 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
| 1672 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
| 1673 | |
| 1674 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
| 1675 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
| 1676 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
| 1677 | |
| 1678 | setTruncStoreAction(ValVT: VT, MemVT: MVT::v2i32, Action: Expand); |
| 1679 | setTruncStoreAction(ValVT: MVT::v2i32, MemVT: VT, Action: Expand); |
| 1680 | } |
| 1681 | // However, load and store *are* legal. |
| 1682 | setOperationAction(Op: ISD::LOAD, VT: MVT::v2i32, Action: Legal); |
| 1683 | setOperationAction(Op: ISD::STORE, VT: MVT::v2i32, Action: Legal); |
| 1684 | setOperationAction(Op: ISD::EXTRACT_VECTOR_ELT, VT: MVT::v2i32, Action: Legal); |
| 1685 | setOperationAction(Op: ISD::BUILD_VECTOR, VT: MVT::v2i32, Action: Legal); |
| 1686 | |
| 1687 | // And we need to promote i64 loads/stores into vector load/store |
| 1688 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Custom); |
| 1689 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Custom); |
| 1690 | |
| 1691 | // Sadly, this doesn't work: |
| 1692 | // AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32); |
| 1693 | // AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32); |
| 1694 | } |
| 1695 | |
| 1696 | // Turn FP extload into load/fpextend |
| 1697 | for (MVT VT : MVT::fp_valuetypes()) { |
| 1698 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f16, Action: Expand); |
| 1699 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f32, Action: Expand); |
| 1700 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f64, Action: Expand); |
| 1701 | } |
| 1702 | |
| 1703 | // Sparc doesn't have i1 sign extending load |
| 1704 | for (MVT VT : MVT::integer_valuetypes()) |
| 1705 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
| 1706 | |
| 1707 | // Turn FP truncstore into trunc + store. |
| 1708 | setTruncStoreAction(ValVT: MVT::f32, MemVT: MVT::f16, Action: Expand); |
| 1709 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f16, Action: Expand); |
| 1710 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
| 1711 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f16, Action: Expand); |
| 1712 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f32, Action: Expand); |
| 1713 | setTruncStoreAction(ValVT: MVT::f128, MemVT: MVT::f64, Action: Expand); |
| 1714 | |
| 1715 | // Custom legalize GlobalAddress nodes into LO/HI parts. |
| 1716 | setOperationAction(Op: ISD::GlobalAddress, VT: PtrVT, Action: Custom); |
| 1717 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: PtrVT, Action: Custom); |
| 1718 | setOperationAction(Op: ISD::ConstantPool, VT: PtrVT, Action: Custom); |
| 1719 | setOperationAction(Op: ISD::BlockAddress, VT: PtrVT, Action: Custom); |
| 1720 | |
| 1721 | // Sparc doesn't have sext_inreg, replace them with shl/sra |
| 1722 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i16, Action: Expand); |
| 1723 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i8 , Action: Expand); |
| 1724 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1 , Action: Expand); |
| 1725 | |
| 1726 | // Sparc has no REM or DIVREM operations. |
| 1727 | setOperationAction(Op: ISD::UREM, VT: MVT::i32, Action: Expand); |
| 1728 | setOperationAction(Op: ISD::SREM, VT: MVT::i32, Action: Expand); |
| 1729 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i32, Action: Expand); |
| 1730 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i32, Action: Expand); |
| 1731 | |
| 1732 | // ... nor does SparcV9. |
| 1733 | if (Subtarget->is64Bit()) { |
| 1734 | setOperationAction(Op: ISD::UREM, VT: MVT::i64, Action: Expand); |
| 1735 | setOperationAction(Op: ISD::SREM, VT: MVT::i64, Action: Expand); |
| 1736 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i64, Action: Expand); |
| 1737 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i64, Action: Expand); |
| 1738 | } |
| 1739 | |
| 1740 | // Custom expand fp<->sint |
| 1741 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i32, Action: Custom); |
| 1742 | setOperationAction(Op: ISD::SINT_TO_FP, VT: MVT::i32, Action: Custom); |
| 1743 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i64, Action: Custom); |
| 1744 | setOperationAction(Op: ISD::SINT_TO_FP, VT: MVT::i64, Action: Custom); |
| 1745 | |
| 1746 | // Custom Expand fp<->uint |
| 1747 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i32, Action: Custom); |
| 1748 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i32, Action: Custom); |
| 1749 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i64, Action: Custom); |
| 1750 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i64, Action: Custom); |
| 1751 | |
| 1752 | // Lower f16 conversion operations into library calls |
| 1753 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f32, Action: Expand); |
| 1754 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f32, Action: Expand); |
| 1755 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f64, Action: Expand); |
| 1756 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f64, Action: Expand); |
| 1757 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f128, Action: Expand); |
| 1758 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f128, Action: Expand); |
| 1759 | |
| 1760 | setOperationAction(Op: ISD::BITCAST, VT: MVT::f32, |
| 1761 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1762 | setOperationAction(Op: ISD::BITCAST, VT: MVT::i32, |
| 1763 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1764 | |
| 1765 | // Sparc has no select or setcc: expand to SELECT_CC. |
| 1766 | setOperationAction(Op: ISD::SELECT, VT: MVT::i32, Action: Expand); |
| 1767 | setOperationAction(Op: ISD::SELECT, VT: MVT::f32, Action: Expand); |
| 1768 | setOperationAction(Op: ISD::SELECT, VT: MVT::f64, Action: Expand); |
| 1769 | setOperationAction(Op: ISD::SELECT, VT: MVT::f128, Action: Expand); |
| 1770 | |
| 1771 | setOperationAction(Op: ISD::SETCC, VT: MVT::i32, Action: Expand); |
| 1772 | setOperationAction(Op: ISD::SETCC, VT: MVT::f32, Action: Expand); |
| 1773 | setOperationAction(Op: ISD::SETCC, VT: MVT::f64, Action: Expand); |
| 1774 | setOperationAction(Op: ISD::SETCC, VT: MVT::f128, Action: Expand); |
| 1775 | |
| 1776 | // Sparc doesn't have BRCOND either, it has BR_CC. |
| 1777 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Expand); |
| 1778 | setOperationAction(Op: ISD::BRIND, VT: MVT::Other, Action: Expand); |
| 1779 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
| 1780 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Custom); |
| 1781 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f32, Action: Custom); |
| 1782 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f64, Action: Custom); |
| 1783 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f128, Action: Custom); |
| 1784 | |
| 1785 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Custom); |
| 1786 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f32, Action: Custom); |
| 1787 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f64, Action: Custom); |
| 1788 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f128, Action: Custom); |
| 1789 | |
| 1790 | setOperationAction(Op: ISD::ADDC, VT: MVT::i32, Action: Legal); |
| 1791 | setOperationAction(Op: ISD::ADDE, VT: MVT::i32, Action: Legal); |
| 1792 | setOperationAction(Op: ISD::SUBC, VT: MVT::i32, Action: Legal); |
| 1793 | setOperationAction(Op: ISD::SUBE, VT: MVT::i32, Action: Legal); |
| 1794 | |
| 1795 | if (Subtarget->isVIS3()) { |
| 1796 | setOperationAction(Op: ISD::ADDC, VT: MVT::i64, Action: Legal); |
| 1797 | setOperationAction(Op: ISD::ADDE, VT: MVT::i64, Action: Legal); |
| 1798 | } |
| 1799 | |
| 1800 | if (Subtarget->is64Bit()) { |
| 1801 | setOperationAction(Op: ISD::BITCAST, VT: MVT::f64, |
| 1802 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1803 | setOperationAction(Op: ISD::BITCAST, VT: MVT::i64, |
| 1804 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1805 | setOperationAction(Op: ISD::SELECT, VT: MVT::i64, Action: Expand); |
| 1806 | setOperationAction(Op: ISD::SETCC, VT: MVT::i64, Action: Expand); |
| 1807 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Custom); |
| 1808 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i64, Action: Custom); |
| 1809 | |
| 1810 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, |
| 1811 | Action: Subtarget->usePopc() ? Legal : Expand); |
| 1812 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i64, Action: Custom); |
| 1813 | setOperationAction(Op: ISD::ROTL , VT: MVT::i64, Action: Expand); |
| 1814 | setOperationAction(Op: ISD::ROTR , VT: MVT::i64, Action: Expand); |
| 1815 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i64, Action: Custom); |
| 1816 | } |
| 1817 | |
| 1818 | // ATOMICs. |
| 1819 | // Atomics are supported on SparcV9. 32-bit atomics are also |
| 1820 | // supported by some Leon SparcV8 variants. Otherwise, atomics |
| 1821 | // are unsupported. |
| 1822 | if (Subtarget->isV9()) { |
| 1823 | // TODO: we _ought_ to be able to support 64-bit atomics on 32-bit sparcv9, |
| 1824 | // but it hasn't been implemented in the backend yet. |
| 1825 | if (Subtarget->is64Bit()) |
| 1826 | setMaxAtomicSizeInBitsSupported(64); |
| 1827 | else |
| 1828 | setMaxAtomicSizeInBitsSupported(32); |
| 1829 | } else if (Subtarget->hasLeonCasa()) |
| 1830 | setMaxAtomicSizeInBitsSupported(32); |
| 1831 | else |
| 1832 | setMaxAtomicSizeInBitsSupported(0); |
| 1833 | |
| 1834 | setMinCmpXchgSizeInBits(32); |
| 1835 | |
| 1836 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT: MVT::i32, Action: Legal); |
| 1837 | |
| 1838 | setOperationAction(Op: ISD::ATOMIC_FENCE, VT: MVT::Other, Action: Legal); |
| 1839 | |
| 1840 | // Custom Lower Atomic LOAD/STORE |
| 1841 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i32, Action: Custom); |
| 1842 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i32, Action: Custom); |
| 1843 | |
| 1844 | if (Subtarget->is64Bit()) { |
| 1845 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP, VT: MVT::i64, Action: Legal); |
| 1846 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT: MVT::i64, Action: Legal); |
| 1847 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i64, Action: Custom); |
| 1848 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i64, Action: Custom); |
| 1849 | } |
| 1850 | |
| 1851 | if (!Subtarget->isV9()) { |
| 1852 | // SparcV8 does not have FNEGD and FABSD. |
| 1853 | setOperationAction(Op: ISD::FNEG, VT: MVT::f64, Action: Custom); |
| 1854 | setOperationAction(Op: ISD::FABS, VT: MVT::f64, Action: Custom); |
| 1855 | } |
| 1856 | |
| 1857 | setOperationAction(Op: ISD::FSIN , VT: MVT::f128, Action: Expand); |
| 1858 | setOperationAction(Op: ISD::FCOS , VT: MVT::f128, Action: Expand); |
| 1859 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f128, Action: Expand); |
| 1860 | setOperationAction(Op: ISD::FREM, VT: MVT::f128, Action: LibCall); |
| 1861 | setOperationAction(Op: ISD::FMA , VT: MVT::f128, Action: Expand); |
| 1862 | setOperationAction(Op: ISD::FSIN , VT: MVT::f64, Action: Expand); |
| 1863 | setOperationAction(Op: ISD::FCOS , VT: MVT::f64, Action: Expand); |
| 1864 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f64, Action: Expand); |
| 1865 | setOperationAction(Op: ISD::FREM, VT: MVT::f64, Action: LibCall); |
| 1866 | setOperationAction(Op: ISD::FMA, VT: MVT::f64, |
| 1867 | Action: Subtarget->isUA2007() ? Legal : Expand); |
| 1868 | setOperationAction(Op: ISD::FSIN , VT: MVT::f32, Action: Expand); |
| 1869 | setOperationAction(Op: ISD::FCOS , VT: MVT::f32, Action: Expand); |
| 1870 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f32, Action: Expand); |
| 1871 | setOperationAction(Op: ISD::FREM, VT: MVT::f32, Action: LibCall); |
| 1872 | setOperationAction(Op: ISD::FMA, VT: MVT::f32, |
| 1873 | Action: Subtarget->isUA2007() ? Legal : Expand); |
| 1874 | setOperationAction(Op: ISD::ROTL , VT: MVT::i32, Action: Expand); |
| 1875 | setOperationAction(Op: ISD::ROTR , VT: MVT::i32, Action: Expand); |
| 1876 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Subtarget->isV9() ? Custom : Expand); |
| 1877 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f128, Action: Expand); |
| 1878 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f64, Action: Expand); |
| 1879 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f32, Action: Expand); |
| 1880 | setOperationAction(Op: ISD::FPOW , VT: MVT::f128, Action: Expand); |
| 1881 | setOperationAction(Op: ISD::FPOW , VT: MVT::f64, Action: Expand); |
| 1882 | setOperationAction(Op: ISD::FPOW , VT: MVT::f32, Action: Expand); |
| 1883 | |
| 1884 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i32, Action: Expand); |
| 1885 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i32, Action: Expand); |
| 1886 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i32, Action: Expand); |
| 1887 | |
| 1888 | // Expands to [SU]MUL_LOHI. |
| 1889 | setOperationAction(Op: ISD::MULHU, VT: MVT::i32, Action: Expand); |
| 1890 | setOperationAction(Op: ISD::MULHS, VT: MVT::i32, Action: Expand); |
| 1891 | setOperationAction(Op: ISD::MUL, VT: MVT::i32, Action: Expand); |
| 1892 | |
| 1893 | if (Subtarget->useSoftMulDiv()) { |
| 1894 | // .umul works for both signed and unsigned |
| 1895 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i32, Action: Expand); |
| 1896 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i32, Action: Expand); |
| 1897 | setOperationAction(Op: ISD::SDIV, VT: MVT::i32, Action: Expand); |
| 1898 | setOperationAction(Op: ISD::UDIV, VT: MVT::i32, Action: Expand); |
| 1899 | } |
| 1900 | |
| 1901 | if (Subtarget->is64Bit()) { |
| 1902 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i64, Action: Expand); |
| 1903 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i64, Action: Expand); |
| 1904 | setOperationAction(Op: ISD::MULHU, VT: MVT::i64, |
| 1905 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1906 | setOperationAction(Op: ISD::MULHS, VT: MVT::i64, |
| 1907 | Action: Subtarget->isVIS3() ? Legal : Expand); |
| 1908 | |
| 1909 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i64, Action: Expand); |
| 1910 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i64, Action: Expand); |
| 1911 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i64, Action: Expand); |
| 1912 | } |
| 1913 | |
| 1914 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex. |
| 1915 | setOperationAction(Op: ISD::VASTART , VT: MVT::Other, Action: Custom); |
| 1916 | // VAARG needs to be lowered to not do unaligned accesses for doubles. |
| 1917 | setOperationAction(Op: ISD::VAARG , VT: MVT::Other, Action: Custom); |
| 1918 | |
| 1919 | setOperationAction(Op: ISD::TRAP , VT: MVT::Other, Action: Legal); |
| 1920 | setOperationAction(Op: ISD::DEBUGTRAP , VT: MVT::Other, Action: Legal); |
| 1921 | |
| 1922 | // Use the default implementation. |
| 1923 | setOperationAction(Op: ISD::VACOPY , VT: MVT::Other, Action: Expand); |
| 1924 | setOperationAction(Op: ISD::VAEND , VT: MVT::Other, Action: Expand); |
| 1925 | setOperationAction(Op: ISD::STACKSAVE , VT: MVT::Other, Action: Expand); |
| 1926 | setOperationAction(Op: ISD::STACKRESTORE , VT: MVT::Other, Action: Expand); |
| 1927 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i32 , Action: Custom); |
| 1928 | setOperationAction(Op: ISD::STACKADDRESS, VT: MVT::Other, Action: Custom); |
| 1929 | |
| 1930 | setStackPointerRegisterToSaveRestore(SP::O6); |
| 1931 | |
| 1932 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, |
| 1933 | Action: Subtarget->usePopc() ? Legal : Expand); |
| 1934 | |
| 1935 | if (Subtarget->isV9() && Subtarget->hasHardQuad()) { |
| 1936 | setOperationAction(Op: ISD::LOAD, VT: MVT::f128, Action: Legal); |
| 1937 | setOperationAction(Op: ISD::STORE, VT: MVT::f128, Action: Legal); |
| 1938 | } else { |
| 1939 | setOperationAction(Op: ISD::LOAD, VT: MVT::f128, Action: Custom); |
| 1940 | setOperationAction(Op: ISD::STORE, VT: MVT::f128, Action: Custom); |
| 1941 | } |
| 1942 | |
| 1943 | if (Subtarget->hasHardQuad()) { |
| 1944 | setOperationAction(Op: ISD::FADD, VT: MVT::f128, Action: Legal); |
| 1945 | setOperationAction(Op: ISD::FSUB, VT: MVT::f128, Action: Legal); |
| 1946 | setOperationAction(Op: ISD::FMUL, VT: MVT::f128, Action: Legal); |
| 1947 | setOperationAction(Op: ISD::FDIV, VT: MVT::f128, Action: Legal); |
| 1948 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f128, Action: Legal); |
| 1949 | setOperationAction(Op: ISD::FP_EXTEND, VT: MVT::f128, Action: Legal); |
| 1950 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f64, Action: Legal); |
| 1951 | if (Subtarget->isV9()) { |
| 1952 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Legal); |
| 1953 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Legal); |
| 1954 | } else { |
| 1955 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Custom); |
| 1956 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Custom); |
| 1957 | } |
| 1958 | } else { |
| 1959 | // Custom legalize f128 operations. |
| 1960 | |
| 1961 | setOperationAction(Op: ISD::FADD, VT: MVT::f128, Action: Custom); |
| 1962 | setOperationAction(Op: ISD::FSUB, VT: MVT::f128, Action: Custom); |
| 1963 | setOperationAction(Op: ISD::FMUL, VT: MVT::f128, Action: Custom); |
| 1964 | setOperationAction(Op: ISD::FDIV, VT: MVT::f128, Action: Custom); |
| 1965 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f128, Action: Custom); |
| 1966 | setOperationAction(Op: ISD::FNEG, VT: MVT::f128, Action: Custom); |
| 1967 | setOperationAction(Op: ISD::FABS, VT: MVT::f128, Action: Custom); |
| 1968 | |
| 1969 | setOperationAction(Op: ISD::FP_EXTEND, VT: MVT::f128, Action: Custom); |
| 1970 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f64, Action: Custom); |
| 1971 | setOperationAction(Op: ISD::FP_ROUND, VT: MVT::f32, Action: Custom); |
| 1972 | } |
| 1973 | |
| 1974 | if (Subtarget->fixAllFDIVSQRT()) { |
| 1975 | // Promote FDIVS and FSQRTS to FDIVD and FSQRTD instructions instead as |
| 1976 | // the former instructions generate errata on LEON processors. |
| 1977 | setOperationAction(Op: ISD::FDIV, VT: MVT::f32, Action: Promote); |
| 1978 | setOperationAction(Op: ISD::FSQRT, VT: MVT::f32, Action: Promote); |
| 1979 | } |
| 1980 | |
| 1981 | if (Subtarget->hasNoFMULS()) { |
| 1982 | setOperationAction(Op: ISD::FMUL, VT: MVT::f32, Action: Promote); |
| 1983 | } |
| 1984 | |
| 1985 | // Custom combine bitcast between f64 and v2i32 |
| 1986 | if (!Subtarget->is64Bit()) |
| 1987 | setTargetDAGCombine(ISD::BITCAST); |
| 1988 | |
| 1989 | if (Subtarget->isV9()) |
| 1990 | setTargetDAGCombine({ISD::BSWAP, ISD::STORE}); |
| 1991 | |
| 1992 | if (Subtarget->hasLeonCycleCounter()) |
| 1993 | setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Custom); |
| 1994 | |
| 1995 | if (Subtarget->isVIS3()) { |
| 1996 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Legal); |
| 1997 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Legal); |
| 1998 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i32, Action: Legal); |
| 1999 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i64, Action: Legal); |
| 2000 | |
| 2001 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, |
| 2002 | Action: Subtarget->is64Bit() ? Promote : Expand); |
| 2003 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
| 2004 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i32, |
| 2005 | Action: Subtarget->is64Bit() ? Promote : Expand); |
| 2006 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i64, Action: Expand); |
| 2007 | } else if (Subtarget->usePopc()) { |
| 2008 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
| 2009 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
| 2010 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i32, Action: Expand); |
| 2011 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i64, Action: Expand); |
| 2012 | |
| 2013 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
| 2014 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
| 2015 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i32, Action: Expand); |
| 2016 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i64, Action: Expand); |
| 2017 | } else { |
| 2018 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
| 2019 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
| 2020 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i32, |
| 2021 | Action: Subtarget->is64Bit() ? Promote : LibCall); |
| 2022 | setOperationAction(Op: ISD::CTLZ_ZERO_POISON, VT: MVT::i64, Action: LibCall); |
| 2023 | |
| 2024 | // FIXME here we don't have any ISA extensions that could help us, so to |
| 2025 | // prevent large expansions those should be made into LibCalls. |
| 2026 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
| 2027 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
| 2028 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i32, Action: Expand); |
| 2029 | setOperationAction(Op: ISD::CTTZ_ZERO_POISON, VT: MVT::i64, Action: Expand); |
| 2030 | } |
| 2031 | |
| 2032 | setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom); |
| 2033 | |
| 2034 | // Some processors have no branch predictor and have pipelines longer than |
| 2035 | // what can be covered by the delay slot. This results in a stall, so mark |
| 2036 | // branches to be expensive on those processors. |
| 2037 | setJumpIsExpensive(Subtarget->hasNoPredictor()); |
| 2038 | // The high cost of branching means that using conditional moves will |
| 2039 | // still be profitable even if the condition is predictable. |
| 2040 | PredictableSelectIsExpensive = !isJumpExpensive(); |
| 2041 | |
| 2042 | setMinFunctionAlignment(Align(4)); |
| 2043 | |
| 2044 | computeRegisterProperties(TRI: Subtarget->getRegisterInfo()); |
| 2045 | } |
| 2046 | |
| 2047 | bool SparcTargetLowering::useSoftFloat() const { |
| 2048 | return Subtarget->useSoftFloat(); |
| 2049 | } |
| 2050 | |
| 2051 | EVT SparcTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, |
| 2052 | EVT VT) const { |
| 2053 | if (!VT.isVector()) |
| 2054 | return MVT::i32; |
| 2055 | return VT.changeVectorElementTypeToInteger(); |
| 2056 | } |
| 2057 | |
| 2058 | /// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to |
| 2059 | /// be zero. Op is expected to be a target specific node. Used by DAG |
| 2060 | /// combiner. |
| 2061 | void SparcTargetLowering::computeKnownBitsForTargetNode |
| 2062 | (const SDValue Op, |
| 2063 | KnownBits &Known, |
| 2064 | const APInt &DemandedElts, |
| 2065 | const SelectionDAG &DAG, |
| 2066 | unsigned Depth) const { |
| 2067 | KnownBits Known2; |
| 2068 | Known.resetAll(); |
| 2069 | |
| 2070 | switch (Op.getOpcode()) { |
| 2071 | default: break; |
| 2072 | case SPISD::SELECT_ICC: |
| 2073 | case SPISD::SELECT_XCC: |
| 2074 | case SPISD::SELECT_FCC: |
| 2075 | Known = DAG.computeKnownBits(Op: Op.getOperand(i: 1), Depth: Depth + 1); |
| 2076 | Known2 = DAG.computeKnownBits(Op: Op.getOperand(i: 0), Depth: Depth + 1); |
| 2077 | |
| 2078 | // Only known if known in both the LHS and RHS. |
| 2079 | Known = Known.intersectWith(RHS: Known2); |
| 2080 | break; |
| 2081 | } |
| 2082 | } |
| 2083 | |
| 2084 | // Look at LHS/RHS/CC and see if they are a lowered setcc instruction. If so |
| 2085 | // set LHS/RHS and SPCC to the LHS/RHS of the setcc and SPCC to the condition. |
| 2086 | static void LookThroughSetCC(SDValue &LHS, SDValue &RHS, |
| 2087 | ISD::CondCode CC, unsigned &SPCC) { |
| 2088 | if (isNullConstant(V: RHS) && CC == ISD::SETNE && |
| 2089 | (((LHS.getOpcode() == SPISD::SELECT_ICC || |
| 2090 | LHS.getOpcode() == SPISD::SELECT_XCC) && |
| 2091 | LHS.getOperand(i: 3).getOpcode() == SPISD::CMPICC) || |
| 2092 | (LHS.getOpcode() == SPISD::SELECT_FCC && |
| 2093 | (LHS.getOperand(i: 3).getOpcode() == SPISD::CMPFCC || |
| 2094 | LHS.getOperand(i: 3).getOpcode() == SPISD::CMPFCC_V9))) && |
| 2095 | isOneConstant(V: LHS.getOperand(i: 0)) && isNullConstant(V: LHS.getOperand(i: 1))) { |
| 2096 | SDValue CMPCC = LHS.getOperand(i: 3); |
| 2097 | SPCC = LHS.getConstantOperandVal(i: 2); |
| 2098 | LHS = CMPCC.getOperand(i: 0); |
| 2099 | RHS = CMPCC.getOperand(i: 1); |
| 2100 | } |
| 2101 | } |
| 2102 | |
| 2103 | // Convert to a target node and set target flags. |
| 2104 | SDValue SparcTargetLowering::withTargetFlags(SDValue Op, unsigned TF, |
| 2105 | SelectionDAG &DAG) const { |
| 2106 | if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) |
| 2107 | return DAG.getTargetGlobalAddress(GV: GA->getGlobal(), |
| 2108 | DL: SDLoc(GA), |
| 2109 | VT: GA->getValueType(ResNo: 0), |
| 2110 | offset: GA->getOffset(), TargetFlags: TF); |
| 2111 | |
| 2112 | if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Val&: Op)) |
| 2113 | return DAG.getTargetConstantPool(C: CP->getConstVal(), VT: CP->getValueType(ResNo: 0), |
| 2114 | Align: CP->getAlign(), Offset: CP->getOffset(), TargetFlags: TF); |
| 2115 | |
| 2116 | if (const BlockAddressSDNode *BA = dyn_cast<BlockAddressSDNode>(Val&: Op)) |
| 2117 | return DAG.getTargetBlockAddress(BA: BA->getBlockAddress(), |
| 2118 | VT: Op.getValueType(), |
| 2119 | Offset: 0, |
| 2120 | TargetFlags: TF); |
| 2121 | |
| 2122 | if (const ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Val&: Op)) |
| 2123 | return DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), |
| 2124 | VT: ES->getValueType(ResNo: 0), TargetFlags: TF); |
| 2125 | |
| 2126 | llvm_unreachable("Unhandled address SDNode" ); |
| 2127 | } |
| 2128 | |
| 2129 | // Split Op into high and low parts according to HiTF and LoTF. |
| 2130 | // Return an ADD node combining the parts. |
| 2131 | SDValue SparcTargetLowering::makeHiLoPair(SDValue Op, |
| 2132 | unsigned HiTF, unsigned LoTF, |
| 2133 | SelectionDAG &DAG) const { |
| 2134 | SDLoc DL(Op); |
| 2135 | EVT VT = Op.getValueType(); |
| 2136 | SDValue Hi = DAG.getNode(Opcode: SPISD::Hi, DL, VT, Operand: withTargetFlags(Op, TF: HiTF, DAG)); |
| 2137 | SDValue Lo = DAG.getNode(Opcode: SPISD::Lo, DL, VT, Operand: withTargetFlags(Op, TF: LoTF, DAG)); |
| 2138 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: Hi, N2: Lo); |
| 2139 | } |
| 2140 | |
| 2141 | // Build SDNodes for producing an address from a GlobalAddress, ConstantPool, |
| 2142 | // or ExternalSymbol SDNode. |
| 2143 | SDValue SparcTargetLowering::makeAddress(SDValue Op, SelectionDAG &DAG) const { |
| 2144 | SDLoc DL(Op); |
| 2145 | EVT VT = getPointerTy(DL: DAG.getDataLayout()); |
| 2146 | |
| 2147 | // Handle PIC mode first. SPARC needs a got load for every variable! |
| 2148 | if (isPositionIndependent()) { |
| 2149 | const Module *M = DAG.getMachineFunction().getFunction().getParent(); |
| 2150 | PICLevel::Level picLevel = M->getPICLevel(); |
| 2151 | SDValue Idx; |
| 2152 | |
| 2153 | if (picLevel == PICLevel::SmallPIC) { |
| 2154 | // This is the pic13 code model, the GOT is known to be smaller than 8KiB. |
| 2155 | Idx = DAG.getNode(Opcode: SPISD::Lo, DL, VT: Op.getValueType(), |
| 2156 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_GOT13, DAG)); |
| 2157 | } else { |
| 2158 | // This is the pic32 code model, the GOT is known to be smaller than 4GB. |
| 2159 | Idx = makeHiLoPair(Op, HiTF: ELF::R_SPARC_GOT22, LoTF: ELF::R_SPARC_GOT10, DAG); |
| 2160 | } |
| 2161 | |
| 2162 | SDValue GlobalBase = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT); |
| 2163 | SDValue AbsAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: GlobalBase, N2: Idx); |
| 2164 | // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this |
| 2165 | // function has calls. |
| 2166 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2167 | MFI.setHasCalls(true); |
| 2168 | return DAG.getLoad(VT, dl: DL, Chain: DAG.getEntryNode(), Ptr: AbsAddr, |
| 2169 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
| 2170 | } |
| 2171 | |
| 2172 | // This is one of the absolute code models. |
| 2173 | switch(getTargetMachine().getCodeModel()) { |
| 2174 | default: |
| 2175 | llvm_unreachable("Unsupported absolute code model" ); |
| 2176 | case CodeModel::Small: |
| 2177 | // abs32. |
| 2178 | return makeHiLoPair(Op, HiTF: ELF::R_SPARC_HI22, LoTF: ELF::R_SPARC_LO10, DAG); |
| 2179 | case CodeModel::Medium: { |
| 2180 | // abs44. |
| 2181 | SDValue H44 = makeHiLoPair(Op, HiTF: ELF::R_SPARC_H44, LoTF: ELF::R_SPARC_M44, DAG); |
| 2182 | H44 = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: H44, N2: DAG.getConstant(Val: 12, DL, VT: MVT::i32)); |
| 2183 | SDValue L44 = withTargetFlags(Op, TF: ELF::R_SPARC_L44, DAG); |
| 2184 | L44 = DAG.getNode(Opcode: SPISD::Lo, DL, VT, Operand: L44); |
| 2185 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: H44, N2: L44); |
| 2186 | } |
| 2187 | case CodeModel::Large: { |
| 2188 | // abs64. |
| 2189 | SDValue Hi = makeHiLoPair(Op, HiTF: ELF::R_SPARC_HH22, LoTF: ELF::R_SPARC_HM10, DAG); |
| 2190 | Hi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, N2: DAG.getConstant(Val: 32, DL, VT: MVT::i32)); |
| 2191 | SDValue Lo = makeHiLoPair(Op, HiTF: ELF::R_SPARC_HI22, LoTF: ELF::R_SPARC_LO10, DAG); |
| 2192 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: Hi, N2: Lo); |
| 2193 | } |
| 2194 | } |
| 2195 | } |
| 2196 | |
| 2197 | SDValue SparcTargetLowering::LowerGlobalAddress(SDValue Op, |
| 2198 | SelectionDAG &DAG) const { |
| 2199 | return makeAddress(Op, DAG); |
| 2200 | } |
| 2201 | |
| 2202 | SDValue SparcTargetLowering::LowerConstantPool(SDValue Op, |
| 2203 | SelectionDAG &DAG) const { |
| 2204 | return makeAddress(Op, DAG); |
| 2205 | } |
| 2206 | |
| 2207 | SDValue SparcTargetLowering::LowerBlockAddress(SDValue Op, |
| 2208 | SelectionDAG &DAG) const { |
| 2209 | return makeAddress(Op, DAG); |
| 2210 | } |
| 2211 | |
| 2212 | SDValue SparcTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
| 2213 | SelectionDAG &DAG) const { |
| 2214 | |
| 2215 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Op); |
| 2216 | if (DAG.getTarget().useEmulatedTLS()) |
| 2217 | return LowerToTLSEmulatedModel(GA, DAG); |
| 2218 | |
| 2219 | SDLoc DL(GA); |
| 2220 | const GlobalValue *GV = GA->getGlobal(); |
| 2221 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2222 | |
| 2223 | TLSModel::Model model = getTargetMachine().getTLSModel(GV); |
| 2224 | |
| 2225 | if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { |
| 2226 | unsigned HiTF = |
| 2227 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_HI22 |
| 2228 | : ELF::R_SPARC_TLS_LDM_HI22); |
| 2229 | unsigned LoTF = |
| 2230 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_LO10 |
| 2231 | : ELF::R_SPARC_TLS_LDM_LO10); |
| 2232 | unsigned addTF = |
| 2233 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_ADD |
| 2234 | : ELF::R_SPARC_TLS_LDM_ADD); |
| 2235 | unsigned callTF = |
| 2236 | ((model == TLSModel::GeneralDynamic) ? ELF::R_SPARC_TLS_GD_CALL |
| 2237 | : ELF::R_SPARC_TLS_LDM_CALL); |
| 2238 | |
| 2239 | SDValue HiLo = makeHiLoPair(Op, HiTF, LoTF, DAG); |
| 2240 | SDValue Base = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT: PtrVT); |
| 2241 | SDValue Argument = DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, N1: Base, N2: HiLo, |
| 2242 | N3: withTargetFlags(Op, TF: addTF, DAG)); |
| 2243 | |
| 2244 | SDValue Chain = DAG.getEntryNode(); |
| 2245 | SDValue InGlue; |
| 2246 | |
| 2247 | Chain = DAG.getCALLSEQ_START(Chain, InSize: 0, OutSize: 0, DL); |
| 2248 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: SP::O0, N: Argument, Glue: InGlue); |
| 2249 | InGlue = Chain.getValue(R: 1); |
| 2250 | SDValue Callee = DAG.getTargetExternalSymbol(Sym: "__tls_get_addr" , VT: PtrVT); |
| 2251 | SDValue Symbol = withTargetFlags(Op, TF: callTF, DAG); |
| 2252 | |
| 2253 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
| 2254 | const uint32_t *Mask = Subtarget->getRegisterInfo()->getCallPreservedMask( |
| 2255 | MF: DAG.getMachineFunction(), CC: CallingConv::C); |
| 2256 | assert(Mask && "Missing call preserved mask for calling convention" ); |
| 2257 | SDValue Ops[] = {Chain, |
| 2258 | Callee, |
| 2259 | Symbol, |
| 2260 | DAG.getRegister(Reg: SP::O0, VT: PtrVT), |
| 2261 | DAG.getRegisterMask(RegMask: Mask), |
| 2262 | InGlue}; |
| 2263 | Chain = DAG.getNode(Opcode: SPISD::TLS_CALL, DL, VTList: NodeTys, Ops); |
| 2264 | InGlue = Chain.getValue(R: 1); |
| 2265 | Chain = DAG.getCALLSEQ_END(Chain, Size1: 0, Size2: 0, Glue: InGlue, DL); |
| 2266 | InGlue = Chain.getValue(R: 1); |
| 2267 | SDValue Ret = DAG.getCopyFromReg(Chain, dl: DL, Reg: SP::O0, VT: PtrVT, Glue: InGlue); |
| 2268 | |
| 2269 | if (model != TLSModel::LocalDynamic) |
| 2270 | return Ret; |
| 2271 | |
| 2272 | SDValue Hi = |
| 2273 | DAG.getNode(Opcode: SPISD::Hi, DL, VT: PtrVT, |
| 2274 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_HIX22, DAG)); |
| 2275 | SDValue Lo = |
| 2276 | DAG.getNode(Opcode: SPISD::Lo, DL, VT: PtrVT, |
| 2277 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_LOX10, DAG)); |
| 2278 | HiLo = DAG.getNode(Opcode: ISD::XOR, DL, VT: PtrVT, N1: Hi, N2: Lo); |
| 2279 | return DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, N1: Ret, N2: HiLo, |
| 2280 | N3: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LDO_ADD, DAG)); |
| 2281 | } |
| 2282 | |
| 2283 | if (model == TLSModel::InitialExec) { |
| 2284 | unsigned ldTF = ((PtrVT == MVT::i64) ? ELF::R_SPARC_TLS_IE_LDX |
| 2285 | : ELF::R_SPARC_TLS_IE_LD); |
| 2286 | |
| 2287 | SDValue Base = DAG.getNode(Opcode: SPISD::GLOBAL_BASE_REG, DL, VT: PtrVT); |
| 2288 | |
| 2289 | // GLOBAL_BASE_REG codegen'ed with call. Inform MFI that this |
| 2290 | // function has calls. |
| 2291 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2292 | MFI.setHasCalls(true); |
| 2293 | |
| 2294 | SDValue TGA = makeHiLoPair(Op, HiTF: ELF::R_SPARC_TLS_IE_HI22, |
| 2295 | LoTF: ELF::R_SPARC_TLS_IE_LO10, DAG); |
| 2296 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Base, N2: TGA); |
| 2297 | SDValue Offset = DAG.getNode(Opcode: SPISD::TLS_LD, |
| 2298 | DL, VT: PtrVT, N1: Ptr, |
| 2299 | N2: withTargetFlags(Op, TF: ldTF, DAG)); |
| 2300 | return DAG.getNode(Opcode: SPISD::TLS_ADD, DL, VT: PtrVT, |
| 2301 | N1: DAG.getRegister(Reg: SP::G7, VT: PtrVT), N2: Offset, |
| 2302 | N3: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_IE_ADD, DAG)); |
| 2303 | } |
| 2304 | |
| 2305 | assert(model == TLSModel::LocalExec); |
| 2306 | SDValue Hi = DAG.getNode(Opcode: SPISD::Hi, DL, VT: PtrVT, |
| 2307 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LE_HIX22, DAG)); |
| 2308 | SDValue Lo = DAG.getNode(Opcode: SPISD::Lo, DL, VT: PtrVT, |
| 2309 | Operand: withTargetFlags(Op, TF: ELF::R_SPARC_TLS_LE_LOX10, DAG)); |
| 2310 | SDValue Offset = DAG.getNode(Opcode: ISD::XOR, DL, VT: PtrVT, N1: Hi, N2: Lo); |
| 2311 | |
| 2312 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, |
| 2313 | N1: DAG.getRegister(Reg: SP::G7, VT: PtrVT), N2: Offset); |
| 2314 | } |
| 2315 | |
| 2316 | SDValue SparcTargetLowering::LowerF128_LibCallArg(SDValue Chain, |
| 2317 | ArgListTy &Args, SDValue Arg, |
| 2318 | const SDLoc &DL, |
| 2319 | SelectionDAG &DAG) const { |
| 2320 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2321 | EVT ArgVT = Arg.getValueType(); |
| 2322 | Type *ArgTy = ArgVT.getTypeForEVT(Context&: *DAG.getContext()); |
| 2323 | |
| 2324 | if (ArgTy->isFP128Ty()) { |
| 2325 | // Create a stack object and pass the pointer to the library function. |
| 2326 | int FI = MFI.CreateStackObject(Size: 16, Alignment: Align(8), isSpillSlot: false); |
| 2327 | SDValue FIPtr = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
| 2328 | Chain = DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: FIPtr, PtrInfo: MachinePointerInfo(), Alignment: Align(8)); |
| 2329 | Args.emplace_back(args&: FIPtr, args: PointerType::getUnqual(C&: ArgTy->getContext())); |
| 2330 | } else { |
| 2331 | Args.emplace_back(args&: Arg, args&: ArgTy); |
| 2332 | } |
| 2333 | return Chain; |
| 2334 | } |
| 2335 | |
| 2336 | SDValue SparcTargetLowering::LowerF128Op(SDValue Op, SelectionDAG &DAG, |
| 2337 | RTLIB::Libcall LibFunc, |
| 2338 | unsigned numArgs) const { |
| 2339 | RTLIB::LibcallImpl LibFuncImpl = DAG.getLibcalls().getLibcallImpl(Call: LibFunc); |
| 2340 | if (LibFuncImpl == RTLIB::Unsupported) |
| 2341 | return SDValue(); |
| 2342 | |
| 2343 | ArgListTy Args; |
| 2344 | |
| 2345 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2346 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2347 | |
| 2348 | SDValue Callee = DAG.getExternalSymbol(LCImpl: LibFuncImpl, VT: PtrVT); |
| 2349 | Type *RetTy = Op.getValueType().getTypeForEVT(Context&: *DAG.getContext()); |
| 2350 | Type *RetTyABI = RetTy; |
| 2351 | SDValue Chain = DAG.getEntryNode(); |
| 2352 | SDValue RetPtr; |
| 2353 | |
| 2354 | if (RetTy->isFP128Ty()) { |
| 2355 | // Create a Stack Object to receive the return value of type f128. |
| 2356 | int RetFI = MFI.CreateStackObject(Size: 16, Alignment: Align(8), isSpillSlot: false); |
| 2357 | RetPtr = DAG.getFrameIndex(FI: RetFI, VT: PtrVT); |
| 2358 | ArgListEntry Entry(RetPtr, PointerType::getUnqual(C&: RetTy->getContext())); |
| 2359 | if (!Subtarget->is64Bit()) { |
| 2360 | Entry.IsSRet = true; |
| 2361 | Entry.IndirectType = RetTy; |
| 2362 | } |
| 2363 | Entry.IsReturned = false; |
| 2364 | Args.push_back(x: Entry); |
| 2365 | RetTyABI = Type::getVoidTy(C&: *DAG.getContext()); |
| 2366 | } |
| 2367 | |
| 2368 | assert(Op->getNumOperands() >= numArgs && "Not enough operands!" ); |
| 2369 | for (unsigned i = 0, e = numArgs; i != e; ++i) { |
| 2370 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: Op.getOperand(i), DL: SDLoc(Op), DAG); |
| 2371 | } |
| 2372 | |
| 2373 | CallingConv::ID CC = DAG.getLibcalls().getLibcallImplCallingConv(Call: LibFuncImpl); |
| 2374 | TargetLowering::CallLoweringInfo CLI(DAG); |
| 2375 | CLI.setDebugLoc(SDLoc(Op)).setChain(Chain).setCallee(CC, ResultType: RetTyABI, Target: Callee, |
| 2376 | ArgsList: std::move(Args)); |
| 2377 | |
| 2378 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
| 2379 | |
| 2380 | // chain is in second result. |
| 2381 | if (RetTyABI == RetTy) |
| 2382 | return CallInfo.first; |
| 2383 | |
| 2384 | assert (RetTy->isFP128Ty() && "Unexpected return type!" ); |
| 2385 | |
| 2386 | Chain = CallInfo.second; |
| 2387 | |
| 2388 | // Load RetPtr to get the return value. |
| 2389 | return DAG.getLoad(VT: Op.getValueType(), dl: SDLoc(Op), Chain, Ptr: RetPtr, |
| 2390 | PtrInfo: MachinePointerInfo(), Alignment: Align(8)); |
| 2391 | } |
| 2392 | |
| 2393 | SDValue SparcTargetLowering::LowerF128Compare(SDValue LHS, SDValue RHS, |
| 2394 | unsigned &SPCC, const SDLoc &DL, |
| 2395 | SelectionDAG &DAG) const { |
| 2396 | |
| 2397 | const char *LibCall = nullptr; |
| 2398 | bool is64Bit = Subtarget->is64Bit(); |
| 2399 | switch(SPCC) { |
| 2400 | default: llvm_unreachable("Unhandled conditional code!" ); |
| 2401 | case SPCC::FCC_E : LibCall = is64Bit? "_Qp_feq" : "_Q_feq" ; break; |
| 2402 | case SPCC::FCC_NE : LibCall = is64Bit? "_Qp_fne" : "_Q_fne" ; break; |
| 2403 | case SPCC::FCC_L : LibCall = is64Bit? "_Qp_flt" : "_Q_flt" ; break; |
| 2404 | case SPCC::FCC_G : LibCall = is64Bit? "_Qp_fgt" : "_Q_fgt" ; break; |
| 2405 | case SPCC::FCC_LE : LibCall = is64Bit? "_Qp_fle" : "_Q_fle" ; break; |
| 2406 | case SPCC::FCC_GE : LibCall = is64Bit? "_Qp_fge" : "_Q_fge" ; break; |
| 2407 | case SPCC::FCC_UL : |
| 2408 | case SPCC::FCC_ULE: |
| 2409 | case SPCC::FCC_UG : |
| 2410 | case SPCC::FCC_UGE: |
| 2411 | case SPCC::FCC_U : |
| 2412 | case SPCC::FCC_O : |
| 2413 | case SPCC::FCC_LG : |
| 2414 | case SPCC::FCC_UE : LibCall = is64Bit? "_Qp_cmp" : "_Q_cmp" ; break; |
| 2415 | } |
| 2416 | |
| 2417 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 2418 | SDValue Callee = DAG.getExternalSymbol(Sym: LibCall, VT: PtrVT); |
| 2419 | Type *RetTy = Type::getInt32Ty(C&: *DAG.getContext()); |
| 2420 | ArgListTy Args; |
| 2421 | SDValue Chain = DAG.getEntryNode(); |
| 2422 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: LHS, DL, DAG); |
| 2423 | Chain = LowerF128_LibCallArg(Chain, Args, Arg: RHS, DL, DAG); |
| 2424 | |
| 2425 | TargetLowering::CallLoweringInfo CLI(DAG); |
| 2426 | CLI.setDebugLoc(DL).setChain(Chain) |
| 2427 | .setCallee(CC: CallingConv::C, ResultType: RetTy, Target: Callee, ArgsList: std::move(Args)); |
| 2428 | |
| 2429 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
| 2430 | |
| 2431 | // result is in first, and chain is in second result. |
| 2432 | SDValue Result = CallInfo.first; |
| 2433 | |
| 2434 | switch(SPCC) { |
| 2435 | default: { |
| 2436 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
| 2437 | SPCC = SPCC::ICC_NE; |
| 2438 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2439 | } |
| 2440 | case SPCC::FCC_UL : { |
| 2441 | SDValue Mask = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
| 2442 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
| 2443 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
| 2444 | SPCC = SPCC::ICC_NE; |
| 2445 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2446 | } |
| 2447 | case SPCC::FCC_ULE: { |
| 2448 | SDValue RHS = DAG.getConstant(Val: 2, DL, VT: Result.getValueType()); |
| 2449 | SPCC = SPCC::ICC_NE; |
| 2450 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2451 | } |
| 2452 | case SPCC::FCC_UG : { |
| 2453 | SDValue RHS = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
| 2454 | SPCC = SPCC::ICC_G; |
| 2455 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2456 | } |
| 2457 | case SPCC::FCC_UGE: { |
| 2458 | SDValue RHS = DAG.getConstant(Val: 1, DL, VT: Result.getValueType()); |
| 2459 | SPCC = SPCC::ICC_NE; |
| 2460 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2461 | } |
| 2462 | |
| 2463 | case SPCC::FCC_U : { |
| 2464 | SDValue RHS = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
| 2465 | SPCC = SPCC::ICC_E; |
| 2466 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2467 | } |
| 2468 | case SPCC::FCC_O : { |
| 2469 | SDValue RHS = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
| 2470 | SPCC = SPCC::ICC_NE; |
| 2471 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2472 | } |
| 2473 | case SPCC::FCC_LG : { |
| 2474 | SDValue Mask = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
| 2475 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
| 2476 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
| 2477 | SPCC = SPCC::ICC_NE; |
| 2478 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2479 | } |
| 2480 | case SPCC::FCC_UE : { |
| 2481 | SDValue Mask = DAG.getConstant(Val: 3, DL, VT: Result.getValueType()); |
| 2482 | Result = DAG.getNode(Opcode: ISD::AND, DL, VT: Result.getValueType(), N1: Result, N2: Mask); |
| 2483 | SDValue RHS = DAG.getConstant(Val: 0, DL, VT: Result.getValueType()); |
| 2484 | SPCC = SPCC::ICC_E; |
| 2485 | return DAG.getNode(Opcode: SPISD::CMPICC, DL, VT: MVT::Glue, N1: Result, N2: RHS); |
| 2486 | } |
| 2487 | } |
| 2488 | } |
| 2489 | |
| 2490 | static SDValue |
| 2491 | LowerF128_FPEXTEND(SDValue Op, SelectionDAG &DAG, |
| 2492 | const SparcTargetLowering &TLI) { |
| 2493 | |
| 2494 | if (Op.getOperand(i: 0).getValueType() == MVT::f64) |
| 2495 | return TLI.LowerF128Op(Op, DAG, LibFunc: RTLIB::FPEXT_F64_F128, numArgs: 1); |
| 2496 | |
| 2497 | if (Op.getOperand(i: 0).getValueType() == MVT::f32) |
| 2498 | return TLI.LowerF128Op(Op, DAG, LibFunc: RTLIB::FPEXT_F32_F128, numArgs: 1); |
| 2499 | |
| 2500 | llvm_unreachable("fpextend with non-float operand!" ); |
| 2501 | return SDValue(); |
| 2502 | } |
| 2503 | |
| 2504 | static SDValue |
| 2505 | LowerF128_FPROUND(SDValue Op, SelectionDAG &DAG, |
| 2506 | const SparcTargetLowering &TLI) { |
| 2507 | // FP_ROUND on f64 and f32 are legal. |
| 2508 | if (Op.getOperand(i: 0).getValueType() != MVT::f128) |
| 2509 | return Op; |
| 2510 | |
| 2511 | if (Op.getValueType() == MVT::f64) |
| 2512 | return TLI.LowerF128Op(Op, DAG, LibFunc: RTLIB::FPROUND_F128_F64, numArgs: 1); |
| 2513 | if (Op.getValueType() == MVT::f32) |
| 2514 | return TLI.LowerF128Op(Op, DAG, LibFunc: RTLIB::FPROUND_F128_F32, numArgs: 1); |
| 2515 | |
| 2516 | llvm_unreachable("fpround to non-float!" ); |
| 2517 | return SDValue(); |
| 2518 | } |
| 2519 | |
| 2520 | static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG, |
| 2521 | const SparcTargetLowering &TLI, |
| 2522 | bool hasHardQuad) { |
| 2523 | SDLoc dl(Op); |
| 2524 | EVT VT = Op.getValueType(); |
| 2525 | assert(VT == MVT::i32 || VT == MVT::i64); |
| 2526 | |
| 2527 | // Expand f128 operations to fp128 abi calls. |
| 2528 | if (Op.getOperand(i: 0).getValueType() == MVT::f128 |
| 2529 | && (!hasHardQuad || !TLI.isTypeLegal(VT))) { |
| 2530 | RTLIB::Libcall LibFunc = |
| 2531 | VT == MVT::i32 ? RTLIB::FPTOSINT_F128_I32 : RTLIB::FPTOSINT_F128_I64; |
| 2532 | return TLI.LowerF128Op(Op, DAG, LibFunc, numArgs: 1); |
| 2533 | } |
| 2534 | |
| 2535 | // Expand if the resulting type is illegal. |
| 2536 | if (!TLI.isTypeLegal(VT)) |
| 2537 | return SDValue(); |
| 2538 | |
| 2539 | // Otherwise, Convert the fp value to integer in an FP register. |
| 2540 | if (VT == MVT::i32) |
| 2541 | Op = DAG.getNode(Opcode: SPISD::FTOI, DL: dl, VT: MVT::f32, Operand: Op.getOperand(i: 0)); |
| 2542 | else |
| 2543 | Op = DAG.getNode(Opcode: SPISD::FTOX, DL: dl, VT: MVT::f64, Operand: Op.getOperand(i: 0)); |
| 2544 | |
| 2545 | return DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: Op); |
| 2546 | } |
| 2547 | |
| 2548 | static SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG, |
| 2549 | const SparcTargetLowering &TLI, |
| 2550 | bool hasHardQuad) { |
| 2551 | SDLoc dl(Op); |
| 2552 | EVT OpVT = Op.getOperand(i: 0).getValueType(); |
| 2553 | assert(OpVT == MVT::i32 || (OpVT == MVT::i64)); |
| 2554 | |
| 2555 | EVT floatVT = (OpVT == MVT::i32) ? MVT::f32 : MVT::f64; |
| 2556 | |
| 2557 | // Expand f128 operations to fp128 ABI calls. |
| 2558 | if (Op.getValueType() == MVT::f128 |
| 2559 | && (!hasHardQuad || !TLI.isTypeLegal(VT: OpVT))) { |
| 2560 | RTLIB::Libcall LibFunc = |
| 2561 | OpVT == MVT::i32 ? RTLIB::SINTTOFP_I32_F128 : RTLIB::SINTTOFP_I64_F128; |
| 2562 | return TLI.LowerF128Op(Op, DAG, LibFunc, numArgs: 1); |
| 2563 | } |
| 2564 | |
| 2565 | // Expand if the operand type is illegal. |
| 2566 | if (!TLI.isTypeLegal(VT: OpVT)) |
| 2567 | return SDValue(); |
| 2568 | |
| 2569 | // Otherwise, Convert the int value to FP in an FP register. |
| 2570 | SDValue Tmp = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: floatVT, Operand: Op.getOperand(i: 0)); |
| 2571 | unsigned opcode = (OpVT == MVT::i32)? SPISD::ITOF : SPISD::XTOF; |
| 2572 | return DAG.getNode(Opcode: opcode, DL: dl, VT: Op.getValueType(), Operand: Tmp); |
| 2573 | } |
| 2574 | |
| 2575 | static SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG, |
| 2576 | const SparcTargetLowering &TLI, |
| 2577 | bool hasHardQuad) { |
| 2578 | EVT VT = Op.getValueType(); |
| 2579 | |
| 2580 | // Expand if it does not involve f128 or the target has support for |
| 2581 | // quad floating point instructions and the resulting type is legal. |
| 2582 | if (Op.getOperand(i: 0).getValueType() != MVT::f128 || |
| 2583 | (hasHardQuad && TLI.isTypeLegal(VT))) |
| 2584 | return SDValue(); |
| 2585 | |
| 2586 | assert(VT == MVT::i32 || VT == MVT::i64); |
| 2587 | |
| 2588 | return TLI.LowerF128Op( |
| 2589 | Op, DAG, |
| 2590 | LibFunc: VT == MVT::i32 ? RTLIB::FPTOUINT_F128_I32 : RTLIB::FPTOUINT_F128_I64, numArgs: 1); |
| 2591 | } |
| 2592 | |
| 2593 | static SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG, |
| 2594 | const SparcTargetLowering &TLI, |
| 2595 | bool hasHardQuad) { |
| 2596 | EVT OpVT = Op.getOperand(i: 0).getValueType(); |
| 2597 | assert(OpVT == MVT::i32 || OpVT == MVT::i64); |
| 2598 | |
| 2599 | // Expand if it does not involve f128 or the target has support for |
| 2600 | // quad floating point instructions and the operand type is legal. |
| 2601 | if (Op.getValueType() != MVT::f128 || (hasHardQuad && TLI.isTypeLegal(VT: OpVT))) |
| 2602 | return SDValue(); |
| 2603 | |
| 2604 | return TLI.LowerF128Op(Op, DAG, |
| 2605 | LibFunc: OpVT == MVT::i32 ? RTLIB::UINTTOFP_I32_F128 |
| 2606 | : RTLIB::UINTTOFP_I64_F128, |
| 2607 | numArgs: 1); |
| 2608 | } |
| 2609 | |
| 2610 | static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG, |
| 2611 | const SparcTargetLowering &TLI, bool hasHardQuad, |
| 2612 | bool isV9, bool is64Bit) { |
| 2613 | SDValue Chain = Op.getOperand(i: 0); |
| 2614 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 1))->get(); |
| 2615 | SDValue LHS = Op.getOperand(i: 2); |
| 2616 | SDValue RHS = Op.getOperand(i: 3); |
| 2617 | SDValue Dest = Op.getOperand(i: 4); |
| 2618 | SDLoc dl(Op); |
| 2619 | unsigned Opc, SPCC = ~0U; |
| 2620 | |
| 2621 | // If this is a br_cc of a "setcc", and if the setcc got lowered into |
| 2622 | // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. |
| 2623 | LookThroughSetCC(LHS, RHS, CC, SPCC); |
| 2624 | assert(LHS.getValueType() == RHS.getValueType()); |
| 2625 | |
| 2626 | // Get the condition flag. |
| 2627 | SDValue CompareFlag; |
| 2628 | if (LHS.getValueType().isInteger()) { |
| 2629 | // On V9 processors running in 64-bit mode, if CC compares two `i64`s |
| 2630 | // and the RHS is zero we might be able to use a specialized branch. |
| 2631 | if (is64Bit && isV9 && LHS.getValueType() == MVT::i64 && |
| 2632 | isNullConstant(V: RHS) && !ISD::isUnsignedIntSetCC(Code: CC)) |
| 2633 | return DAG.getNode(Opcode: SPISD::BR_REG, DL: dl, VT: MVT::Other, N1: Chain, N2: Dest, |
| 2634 | N3: DAG.getConstant(Val: intCondCCodeToRcond(CC), DL: dl, VT: MVT::i32), |
| 2635 | N4: LHS); |
| 2636 | |
| 2637 | CompareFlag = DAG.getNode(Opcode: SPISD::CMPICC, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
| 2638 | if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); |
| 2639 | if (isV9) |
| 2640 | // 32-bit compares use the icc flags, 64-bit uses the xcc flags. |
| 2641 | Opc = LHS.getValueType() == MVT::i32 ? SPISD::BPICC : SPISD::BPXCC; |
| 2642 | else |
| 2643 | // Non-v9 targets don't have xcc. |
| 2644 | Opc = SPISD::BRICC; |
| 2645 | } else { |
| 2646 | if (!hasHardQuad && LHS.getValueType() == MVT::f128) { |
| 2647 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
| 2648 | CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, DL: dl, DAG); |
| 2649 | Opc = isV9 ? SPISD::BPICC : SPISD::BRICC; |
| 2650 | } else { |
| 2651 | unsigned CmpOpc = isV9 ? SPISD::CMPFCC_V9 : SPISD::CMPFCC; |
| 2652 | CompareFlag = DAG.getNode(Opcode: CmpOpc, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
| 2653 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
| 2654 | Opc = isV9 ? SPISD::BRFCC_V9 : SPISD::BRFCC; |
| 2655 | } |
| 2656 | } |
| 2657 | return DAG.getNode(Opcode: Opc, DL: dl, VT: MVT::Other, N1: Chain, N2: Dest, |
| 2658 | N3: DAG.getConstant(Val: SPCC, DL: dl, VT: MVT::i32), N4: CompareFlag); |
| 2659 | } |
| 2660 | |
| 2661 | static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG, |
| 2662 | const SparcTargetLowering &TLI, bool hasHardQuad, |
| 2663 | bool isV9, bool is64Bit) { |
| 2664 | SDValue LHS = Op.getOperand(i: 0); |
| 2665 | SDValue RHS = Op.getOperand(i: 1); |
| 2666 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 4))->get(); |
| 2667 | SDValue TrueVal = Op.getOperand(i: 2); |
| 2668 | SDValue FalseVal = Op.getOperand(i: 3); |
| 2669 | SDLoc dl(Op); |
| 2670 | unsigned Opc, SPCC = ~0U; |
| 2671 | |
| 2672 | // If this is a select_cc of a "setcc", and if the setcc got lowered into |
| 2673 | // an CMP[IF]CC/SELECT_[IF]CC pair, find the original compared values. |
| 2674 | LookThroughSetCC(LHS, RHS, CC, SPCC); |
| 2675 | assert(LHS.getValueType() == RHS.getValueType()); |
| 2676 | |
| 2677 | SDValue CompareFlag; |
| 2678 | if (LHS.getValueType().isInteger()) { |
| 2679 | // On V9 processors running in 64-bit mode, if CC compares two `i64`s |
| 2680 | // and the RHS is zero we might be able to use a specialized select. |
| 2681 | // All SELECT_CC between any two scalar integer types are eligible for |
| 2682 | // lowering to specialized instructions. Additionally, f32 and f64 types |
| 2683 | // are also eligible, but for f128 we can only use the specialized |
| 2684 | // instruction when we have hardquad. |
| 2685 | EVT ValType = TrueVal.getValueType(); |
| 2686 | bool IsEligibleType = ValType.isScalarInteger() || ValType == MVT::f32 || |
| 2687 | ValType == MVT::f64 || |
| 2688 | (ValType == MVT::f128 && hasHardQuad); |
| 2689 | if (is64Bit && isV9 && LHS.getValueType() == MVT::i64 && |
| 2690 | isNullConstant(V: RHS) && !ISD::isUnsignedIntSetCC(Code: CC) && IsEligibleType) |
| 2691 | return DAG.getNode( |
| 2692 | Opcode: SPISD::SELECT_REG, DL: dl, VT: TrueVal.getValueType(), N1: TrueVal, N2: FalseVal, |
| 2693 | N3: DAG.getConstant(Val: intCondCCodeToRcond(CC), DL: dl, VT: MVT::i32), N4: LHS); |
| 2694 | |
| 2695 | CompareFlag = DAG.getNode(Opcode: SPISD::CMPICC, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
| 2696 | Opc = LHS.getValueType() == MVT::i32 ? |
| 2697 | SPISD::SELECT_ICC : SPISD::SELECT_XCC; |
| 2698 | if (SPCC == ~0U) SPCC = IntCondCCodeToICC(CC); |
| 2699 | } else { |
| 2700 | if (!hasHardQuad && LHS.getValueType() == MVT::f128) { |
| 2701 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
| 2702 | CompareFlag = TLI.LowerF128Compare(LHS, RHS, SPCC, DL: dl, DAG); |
| 2703 | Opc = SPISD::SELECT_ICC; |
| 2704 | } else { |
| 2705 | unsigned CmpOpc = isV9 ? SPISD::CMPFCC_V9 : SPISD::CMPFCC; |
| 2706 | CompareFlag = DAG.getNode(Opcode: CmpOpc, DL: dl, VT: MVT::Glue, N1: LHS, N2: RHS); |
| 2707 | Opc = SPISD::SELECT_FCC; |
| 2708 | if (SPCC == ~0U) SPCC = FPCondCCodeToFCC(CC); |
| 2709 | } |
| 2710 | } |
| 2711 | return DAG.getNode(Opcode: Opc, DL: dl, VT: TrueVal.getValueType(), N1: TrueVal, N2: FalseVal, |
| 2712 | N3: DAG.getConstant(Val: SPCC, DL: dl, VT: MVT::i32), N4: CompareFlag); |
| 2713 | } |
| 2714 | |
| 2715 | static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, |
| 2716 | const SparcTargetLowering &TLI) { |
| 2717 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2718 | SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>(); |
| 2719 | auto PtrVT = TLI.getPointerTy(DL: DAG.getDataLayout()); |
| 2720 | |
| 2721 | // Need frame address to find the address of VarArgsFrameIndex. |
| 2722 | MF.getFrameInfo().setFrameAddressIsTaken(true); |
| 2723 | |
| 2724 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
| 2725 | // memory location argument. |
| 2726 | SDLoc DL(Op); |
| 2727 | SDValue Offset = |
| 2728 | DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: DAG.getRegister(Reg: SP::I6, VT: PtrVT), |
| 2729 | N2: DAG.getIntPtrConstant(Val: FuncInfo->getVarArgsFrameOffset(), DL)); |
| 2730 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
| 2731 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: Offset, Ptr: Op.getOperand(i: 1), |
| 2732 | PtrInfo: MachinePointerInfo(SV)); |
| 2733 | } |
| 2734 | |
| 2735 | static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) { |
| 2736 | SDNode *Node = Op.getNode(); |
| 2737 | EVT VT = Node->getValueType(ResNo: 0); |
| 2738 | SDValue InChain = Node->getOperand(Num: 0); |
| 2739 | SDValue VAListPtr = Node->getOperand(Num: 1); |
| 2740 | EVT PtrVT = VAListPtr.getValueType(); |
| 2741 | const Value *SV = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue(); |
| 2742 | SDLoc DL(Node); |
| 2743 | SDValue VAList = |
| 2744 | DAG.getLoad(VT: PtrVT, dl: DL, Chain: InChain, Ptr: VAListPtr, PtrInfo: MachinePointerInfo(SV)); |
| 2745 | // Increment the pointer, VAList, to the next vaarg. |
| 2746 | SDValue NextPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: VAList, |
| 2747 | N2: DAG.getIntPtrConstant(Val: VT.getSizeInBits()/8, |
| 2748 | DL)); |
| 2749 | // Store the incremented VAList to the legalized pointer. |
| 2750 | InChain = DAG.getStore(Chain: VAList.getValue(R: 1), dl: DL, Val: NextPtr, Ptr: VAListPtr, |
| 2751 | PtrInfo: MachinePointerInfo(SV)); |
| 2752 | // Load the actual argument out of the pointer VAList. |
| 2753 | // We can't count on greater alignment than the word size. |
| 2754 | return DAG.getLoad( |
| 2755 | VT, dl: DL, Chain: InChain, Ptr: VAList, PtrInfo: MachinePointerInfo(), |
| 2756 | Alignment: Align(std::min(a: PtrVT.getFixedSizeInBits(), b: VT.getFixedSizeInBits()) / 8)); |
| 2757 | } |
| 2758 | |
| 2759 | static SDValue LowerSTACKADDRESS(SDValue Op, SelectionDAG &DAG, |
| 2760 | const SparcSubtarget &Subtarget) { |
| 2761 | SDValue Chain = Op.getOperand(i: 0); |
| 2762 | EVT VT = Op->getValueType(ResNo: 0); |
| 2763 | SDLoc DL(Op); |
| 2764 | |
| 2765 | MCRegister SPReg = SP::O6; |
| 2766 | SDValue SP = DAG.getCopyFromReg(Chain, dl: DL, Reg: SPReg, VT); |
| 2767 | |
| 2768 | // Unbias the stack pointer register. |
| 2769 | unsigned OffsetToStackStart = Subtarget.getStackPointerBias(); |
| 2770 | // Move past the register save area: 8 in registers + 8 local registers. |
| 2771 | OffsetToStackStart += 16 * (Subtarget.is64Bit() ? 8 : 4); |
| 2772 | // Move past the struct return address slot (4 bytes) on SPARC 32-bit. |
| 2773 | if (!Subtarget.is64Bit()) |
| 2774 | OffsetToStackStart += 4; |
| 2775 | |
| 2776 | SDValue StackAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: SP, |
| 2777 | N2: DAG.getConstant(Val: OffsetToStackStart, DL, VT)); |
| 2778 | return DAG.getMergeValues(Ops: {StackAddr, Chain}, dl: DL); |
| 2779 | } |
| 2780 | |
| 2781 | static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG, |
| 2782 | const SparcSubtarget *Subtarget) { |
| 2783 | SDValue Chain = Op.getOperand(i: 0); |
| 2784 | SDValue Size = Op.getOperand(i: 1); |
| 2785 | SDValue Alignment = Op.getOperand(i: 2); |
| 2786 | MaybeAlign MaybeAlignment = |
| 2787 | cast<ConstantSDNode>(Val&: Alignment)->getMaybeAlignValue(); |
| 2788 | EVT VT = Size->getValueType(ResNo: 0); |
| 2789 | SDLoc dl(Op); |
| 2790 | |
| 2791 | unsigned SPReg = SP::O6; |
| 2792 | SDValue SP = DAG.getCopyFromReg(Chain, dl, Reg: SPReg, VT); |
| 2793 | |
| 2794 | // The resultant pointer needs to be above the register spill area |
| 2795 | // at the bottom of the stack. |
| 2796 | unsigned regSpillArea; |
| 2797 | if (Subtarget->is64Bit()) { |
| 2798 | regSpillArea = 128; |
| 2799 | } else { |
| 2800 | // On Sparc32, the size of the spill area is 92. Unfortunately, |
| 2801 | // that's only 4-byte aligned, not 8-byte aligned (the stack |
| 2802 | // pointer is 8-byte aligned). So, if the user asked for an 8-byte |
| 2803 | // aligned dynamic allocation, we actually need to add 96 to the |
| 2804 | // bottom of the stack, instead of 92, to ensure 8-byte alignment. |
| 2805 | |
| 2806 | // That also means adding 4 to the size of the allocation -- |
| 2807 | // before applying the 8-byte rounding. Unfortunately, we the |
| 2808 | // value we get here has already had rounding applied. So, we need |
| 2809 | // to add 8, instead, wasting a bit more memory. |
| 2810 | |
| 2811 | // Further, this only actually needs to be done if the required |
| 2812 | // alignment is > 4, but, we've lost that info by this point, too, |
| 2813 | // so we always apply it. |
| 2814 | |
| 2815 | // (An alternative approach would be to always reserve 96 bytes |
| 2816 | // instead of the required 92, but then we'd waste 4 extra bytes |
| 2817 | // in every frame, not just those with dynamic stack allocations) |
| 2818 | |
| 2819 | // TODO: modify code in SelectionDAGBuilder to make this less sad. |
| 2820 | |
| 2821 | Size = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: Size, |
| 2822 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
| 2823 | regSpillArea = 96; |
| 2824 | } |
| 2825 | |
| 2826 | int64_t Bias = Subtarget->getStackPointerBias(); |
| 2827 | |
| 2828 | // Debias and increment SP past the reserved spill area. |
| 2829 | // We need the SP to point to the first usable region before calculating |
| 2830 | // anything to prevent any of the pointers from becoming out of alignment when |
| 2831 | // we rebias the SP later on. |
| 2832 | SDValue StartOfUsableStack = DAG.getNode( |
| 2833 | Opcode: ISD::ADD, DL: dl, VT, N1: SP, N2: DAG.getConstant(Val: regSpillArea + Bias, DL: dl, VT)); |
| 2834 | SDValue AllocatedPtr = |
| 2835 | DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: StartOfUsableStack, N2: Size); |
| 2836 | |
| 2837 | bool IsOveraligned = MaybeAlignment.has_value(); |
| 2838 | SDValue AlignedPtr = |
| 2839 | IsOveraligned |
| 2840 | ? DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: AllocatedPtr, |
| 2841 | N2: DAG.getSignedConstant(Val: -MaybeAlignment->value(), DL: dl, VT)) |
| 2842 | : AllocatedPtr; |
| 2843 | |
| 2844 | // Now that we are done, restore the bias and reserved spill area. |
| 2845 | SDValue NewSP = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: AlignedPtr, |
| 2846 | N2: DAG.getConstant(Val: regSpillArea + Bias, DL: dl, VT)); |
| 2847 | Chain = DAG.getCopyToReg(Chain: SP.getValue(R: 1), dl, Reg: SPReg, N: NewSP); |
| 2848 | SDValue Ops[2] = {AlignedPtr, Chain}; |
| 2849 | return DAG.getMergeValues(Ops, dl); |
| 2850 | } |
| 2851 | |
| 2852 | |
| 2853 | static SDValue getFLUSHW(SDValue Op, SelectionDAG &DAG) { |
| 2854 | SDLoc dl(Op); |
| 2855 | SDValue Chain = DAG.getNode(Opcode: SPISD::FLUSHW, |
| 2856 | DL: dl, VT: MVT::Other, Operand: DAG.getEntryNode()); |
| 2857 | return Chain; |
| 2858 | } |
| 2859 | |
| 2860 | static SDValue getFRAMEADDR(uint64_t depth, SDValue Op, SelectionDAG &DAG, |
| 2861 | const SparcSubtarget *Subtarget, |
| 2862 | bool AlwaysFlush = false) { |
| 2863 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
| 2864 | MFI.setFrameAddressIsTaken(true); |
| 2865 | |
| 2866 | EVT VT = Op.getValueType(); |
| 2867 | SDLoc dl(Op); |
| 2868 | unsigned FrameReg = SP::I6; |
| 2869 | unsigned stackBias = Subtarget->getStackPointerBias(); |
| 2870 | |
| 2871 | SDValue FrameAddr; |
| 2872 | SDValue Chain; |
| 2873 | |
| 2874 | // flush first to make sure the windowed registers' values are in stack |
| 2875 | Chain = (depth || AlwaysFlush) ? getFLUSHW(Op, DAG) : DAG.getEntryNode(); |
| 2876 | |
| 2877 | FrameAddr = DAG.getCopyFromReg(Chain, dl, Reg: FrameReg, VT); |
| 2878 | |
| 2879 | unsigned Offset = (Subtarget->is64Bit()) ? (stackBias + 112) : 56; |
| 2880 | |
| 2881 | while (depth--) { |
| 2882 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: FrameAddr, |
| 2883 | N2: DAG.getIntPtrConstant(Val: Offset, DL: dl)); |
| 2884 | FrameAddr = DAG.getLoad(VT, dl, Chain, Ptr, PtrInfo: MachinePointerInfo()); |
| 2885 | } |
| 2886 | if (Subtarget->is64Bit()) |
| 2887 | FrameAddr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: FrameAddr, |
| 2888 | N2: DAG.getIntPtrConstant(Val: stackBias, DL: dl)); |
| 2889 | return FrameAddr; |
| 2890 | } |
| 2891 | |
| 2892 | |
| 2893 | static SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG, |
| 2894 | const SparcSubtarget *Subtarget) { |
| 2895 | |
| 2896 | uint64_t depth = Op.getConstantOperandVal(i: 0); |
| 2897 | |
| 2898 | return getFRAMEADDR(depth, Op, DAG, Subtarget); |
| 2899 | |
| 2900 | } |
| 2901 | |
| 2902 | static SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG, |
| 2903 | const SparcTargetLowering &TLI, |
| 2904 | const SparcSubtarget *Subtarget) { |
| 2905 | MachineFunction &MF = DAG.getMachineFunction(); |
| 2906 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2907 | MFI.setReturnAddressIsTaken(true); |
| 2908 | |
| 2909 | EVT VT = Op.getValueType(); |
| 2910 | SDLoc dl(Op); |
| 2911 | uint64_t depth = Op.getConstantOperandVal(i: 0); |
| 2912 | |
| 2913 | SDValue RetAddr; |
| 2914 | if (depth == 0) { |
| 2915 | auto PtrVT = TLI.getPointerTy(DL: DAG.getDataLayout()); |
| 2916 | Register RetReg = MF.addLiveIn(PReg: SP::I7, RC: TLI.getRegClassFor(VT: PtrVT)); |
| 2917 | RetAddr = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl, Reg: RetReg, VT); |
| 2918 | return RetAddr; |
| 2919 | } |
| 2920 | |
| 2921 | // Need frame address to find return address of the caller. |
| 2922 | SDValue FrameAddr = getFRAMEADDR(depth: depth - 1, Op, DAG, Subtarget, AlwaysFlush: true); |
| 2923 | |
| 2924 | unsigned Offset = (Subtarget->is64Bit()) ? 120 : 60; |
| 2925 | SDValue Ptr = DAG.getNode(Opcode: ISD::ADD, |
| 2926 | DL: dl, VT, |
| 2927 | N1: FrameAddr, |
| 2928 | N2: DAG.getIntPtrConstant(Val: Offset, DL: dl)); |
| 2929 | RetAddr = DAG.getLoad(VT, dl, Chain: DAG.getEntryNode(), Ptr, PtrInfo: MachinePointerInfo()); |
| 2930 | |
| 2931 | return RetAddr; |
| 2932 | } |
| 2933 | |
| 2934 | static SDValue LowerF64Op(SDValue SrcReg64, const SDLoc &dl, SelectionDAG &DAG, |
| 2935 | unsigned opcode) { |
| 2936 | assert(SrcReg64.getValueType() == MVT::f64 && "LowerF64Op called on non-double!" ); |
| 2937 | assert(opcode == ISD::FNEG || opcode == ISD::FABS); |
| 2938 | |
| 2939 | // Lower fneg/fabs on f64 to fneg/fabs on f32. |
| 2940 | // fneg f64 => fneg f32:sub_even, fmov f32:sub_odd. |
| 2941 | // fabs f64 => fabs f32:sub_even, fmov f32:sub_odd. |
| 2942 | |
| 2943 | // Note: in little-endian, the floating-point value is stored in the |
| 2944 | // registers are in the opposite order, so the subreg with the sign |
| 2945 | // bit is the highest-numbered (odd), rather than the |
| 2946 | // lowest-numbered (even). |
| 2947 | |
| 2948 | SDValue Hi32 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_even, DL: dl, VT: MVT::f32, |
| 2949 | Operand: SrcReg64); |
| 2950 | SDValue Lo32 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_odd, DL: dl, VT: MVT::f32, |
| 2951 | Operand: SrcReg64); |
| 2952 | |
| 2953 | if (DAG.getDataLayout().isLittleEndian()) |
| 2954 | Lo32 = DAG.getNode(Opcode: opcode, DL: dl, VT: MVT::f32, Operand: Lo32); |
| 2955 | else |
| 2956 | Hi32 = DAG.getNode(Opcode: opcode, DL: dl, VT: MVT::f32, Operand: Hi32); |
| 2957 | |
| 2958 | SDValue DstReg64 = SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
| 2959 | dl, VT: MVT::f64), 0); |
| 2960 | DstReg64 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_even, DL: dl, VT: MVT::f64, |
| 2961 | Operand: DstReg64, Subreg: Hi32); |
| 2962 | DstReg64 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_odd, DL: dl, VT: MVT::f64, |
| 2963 | Operand: DstReg64, Subreg: Lo32); |
| 2964 | return DstReg64; |
| 2965 | } |
| 2966 | |
| 2967 | // Lower a f128 load into two f64 loads. |
| 2968 | static SDValue LowerF128Load(SDValue Op, SelectionDAG &DAG) |
| 2969 | { |
| 2970 | SDLoc dl(Op); |
| 2971 | LoadSDNode *LdNode = cast<LoadSDNode>(Val: Op.getNode()); |
| 2972 | assert(LdNode->getOffset().isUndef() && "Unexpected node type" ); |
| 2973 | |
| 2974 | Align Alignment = commonAlignment(A: LdNode->getBaseAlign(), Offset: 8); |
| 2975 | |
| 2976 | SDValue Hi64 = |
| 2977 | DAG.getLoad(VT: MVT::f64, dl, Chain: LdNode->getChain(), Ptr: LdNode->getBasePtr(), |
| 2978 | PtrInfo: LdNode->getPointerInfo(), Alignment); |
| 2979 | EVT addrVT = LdNode->getBasePtr().getValueType(); |
| 2980 | SDValue LoPtr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: addrVT, |
| 2981 | N1: LdNode->getBasePtr(), |
| 2982 | N2: DAG.getConstant(Val: 8, DL: dl, VT: addrVT)); |
| 2983 | SDValue Lo64 = DAG.getLoad(VT: MVT::f64, dl, Chain: LdNode->getChain(), Ptr: LoPtr, |
| 2984 | PtrInfo: LdNode->getPointerInfo().getWithOffset(O: 8), |
| 2985 | Alignment); |
| 2986 | |
| 2987 | SDValue SubRegEven = DAG.getTargetConstant(Val: SP::sub_even64, DL: dl, VT: MVT::i32); |
| 2988 | SDValue SubRegOdd = DAG.getTargetConstant(Val: SP::sub_odd64, DL: dl, VT: MVT::i32); |
| 2989 | |
| 2990 | SDNode *InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
| 2991 | dl, VT: MVT::f128); |
| 2992 | InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl, |
| 2993 | VT: MVT::f128, |
| 2994 | Op1: SDValue(InFP128, 0), |
| 2995 | Op2: Hi64, |
| 2996 | Op3: SubRegEven); |
| 2997 | InFP128 = DAG.getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl, |
| 2998 | VT: MVT::f128, |
| 2999 | Op1: SDValue(InFP128, 0), |
| 3000 | Op2: Lo64, |
| 3001 | Op3: SubRegOdd); |
| 3002 | SDValue OutChains[2] = { SDValue(Hi64.getNode(), 1), |
| 3003 | SDValue(Lo64.getNode(), 1) }; |
| 3004 | SDValue OutChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
| 3005 | SDValue Ops[2] = {SDValue(InFP128,0), OutChain}; |
| 3006 | return DAG.getMergeValues(Ops, dl); |
| 3007 | } |
| 3008 | |
| 3009 | SDValue SparcTargetLowering::LowerBSWAP(SDValue Op, SelectionDAG &DAG) const { |
| 3010 | // We don't have an in-register bswap, so expand bswap(x) into |
| 3011 | // load(store-swapped(x)). The reason the swap is done during the store is |
| 3012 | // that on some implementations (mainly older ones) ASI-tagged memory |
| 3013 | // operations are not pipelined, and generally stores finish faster than |
| 3014 | // loads. |
| 3015 | |
| 3016 | MachineFunction &MF = DAG.getMachineFunction(); |
| 3017 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 3018 | MVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 3019 | SDValue Chain = DAG.getEntryNode(); |
| 3020 | bool IsLittleEndian = DAG.getDataLayout().isLittleEndian(); |
| 3021 | SDLoc DL(Op); |
| 3022 | |
| 3023 | SDValue BSwapOp = Op.getOperand(i: 0); |
| 3024 | EVT VT = BSwapOp.getValueType(); |
| 3025 | Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext()); |
| 3026 | Align Al = DAG.getDataLayout().getPrefTypeAlign(Ty); |
| 3027 | |
| 3028 | // Create a stack object to serve as temporary storage. |
| 3029 | int TmpFI = MFI.CreateStackObject(Size: VT.getStoreSize(), Alignment: Al, isSpillSlot: false); |
| 3030 | SDValue TmpPtr = DAG.getFrameIndex(FI: TmpFI, VT: PtrVT); |
| 3031 | |
| 3032 | // Store-swap the value, then load it back. |
| 3033 | SDValue Ops[] = {Chain, BSwapOp, TmpPtr, DAG.getValueType(VT)}; |
| 3034 | SDValue ST = DAG.getMemIntrinsicNode( |
| 3035 | Opcode: IsLittleEndian ? SPISD::STORE_BIG : SPISD::STORE_LITTLE, dl: DL, |
| 3036 | VTList: DAG.getVTList(VT: MVT::Other), Ops, MemVT: VT, |
| 3037 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI: TmpFI), Alignment: std::nullopt, |
| 3038 | Flags: MachineMemOperand::MOStore); |
| 3039 | return DAG.getLoad(VT, dl: DL, Chain: ST, Ptr: TmpPtr, |
| 3040 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI: TmpFI)); |
| 3041 | } |
| 3042 | |
| 3043 | static SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) |
| 3044 | { |
| 3045 | LoadSDNode *LdNode = cast<LoadSDNode>(Val: Op.getNode()); |
| 3046 | |
| 3047 | EVT MemVT = LdNode->getMemoryVT(); |
| 3048 | if (MemVT == MVT::f128) |
| 3049 | return LowerF128Load(Op, DAG); |
| 3050 | |
| 3051 | return Op; |
| 3052 | } |
| 3053 | |
| 3054 | // Lower a f128 store into two f64 stores. |
| 3055 | static SDValue LowerF128Store(SDValue Op, SelectionDAG &DAG) { |
| 3056 | SDLoc dl(Op); |
| 3057 | StoreSDNode *StNode = cast<StoreSDNode>(Val: Op.getNode()); |
| 3058 | assert(StNode->getOffset().isUndef() && "Unexpected node type" ); |
| 3059 | |
| 3060 | SDValue SubRegEven = DAG.getTargetConstant(Val: SP::sub_even64, DL: dl, VT: MVT::i32); |
| 3061 | SDValue SubRegOdd = DAG.getTargetConstant(Val: SP::sub_odd64, DL: dl, VT: MVT::i32); |
| 3062 | |
| 3063 | SDNode *Hi64 = DAG.getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
| 3064 | dl, |
| 3065 | VT: MVT::f64, |
| 3066 | Op1: StNode->getValue(), |
| 3067 | Op2: SubRegEven); |
| 3068 | SDNode *Lo64 = DAG.getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
| 3069 | dl, |
| 3070 | VT: MVT::f64, |
| 3071 | Op1: StNode->getValue(), |
| 3072 | Op2: SubRegOdd); |
| 3073 | |
| 3074 | Align Alignment = commonAlignment(A: StNode->getBaseAlign(), Offset: 8); |
| 3075 | |
| 3076 | SDValue OutChains[2]; |
| 3077 | OutChains[0] = |
| 3078 | DAG.getStore(Chain: StNode->getChain(), dl, Val: SDValue(Hi64, 0), |
| 3079 | Ptr: StNode->getBasePtr(), PtrInfo: StNode->getPointerInfo(), |
| 3080 | Alignment); |
| 3081 | EVT addrVT = StNode->getBasePtr().getValueType(); |
| 3082 | SDValue LoPtr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: addrVT, |
| 3083 | N1: StNode->getBasePtr(), |
| 3084 | N2: DAG.getConstant(Val: 8, DL: dl, VT: addrVT)); |
| 3085 | OutChains[1] = DAG.getStore(Chain: StNode->getChain(), dl, Val: SDValue(Lo64, 0), Ptr: LoPtr, |
| 3086 | PtrInfo: StNode->getPointerInfo().getWithOffset(O: 8), |
| 3087 | Alignment); |
| 3088 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
| 3089 | } |
| 3090 | |
| 3091 | static SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) |
| 3092 | { |
| 3093 | SDLoc dl(Op); |
| 3094 | StoreSDNode *St = cast<StoreSDNode>(Val: Op.getNode()); |
| 3095 | |
| 3096 | EVT MemVT = St->getMemoryVT(); |
| 3097 | if (MemVT == MVT::f128) |
| 3098 | return LowerF128Store(Op, DAG); |
| 3099 | |
| 3100 | if (MemVT == MVT::i64) { |
| 3101 | // Custom handling for i64 stores: turn it into a bitcast and a |
| 3102 | // v2i32 store. |
| 3103 | SDValue Val = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::v2i32, Operand: St->getValue()); |
| 3104 | SDValue Chain = DAG.getStore( |
| 3105 | Chain: St->getChain(), dl, Val, Ptr: St->getBasePtr(), PtrInfo: St->getPointerInfo(), |
| 3106 | Alignment: St->getBaseAlign(), MMOFlags: St->getMemOperand()->getFlags(), AAInfo: St->getAAInfo()); |
| 3107 | return Chain; |
| 3108 | } |
| 3109 | |
| 3110 | return SDValue(); |
| 3111 | } |
| 3112 | |
| 3113 | static SDValue LowerFNEGorFABS(SDValue Op, SelectionDAG &DAG, bool isV9) { |
| 3114 | assert((Op.getOpcode() == ISD::FNEG || Op.getOpcode() == ISD::FABS) |
| 3115 | && "invalid opcode" ); |
| 3116 | |
| 3117 | SDLoc dl(Op); |
| 3118 | |
| 3119 | if (Op.getValueType() == MVT::f64) |
| 3120 | return LowerF64Op(SrcReg64: Op.getOperand(i: 0), dl, DAG, opcode: Op.getOpcode()); |
| 3121 | if (Op.getValueType() != MVT::f128) |
| 3122 | return Op; |
| 3123 | |
| 3124 | // Lower fabs/fneg on f128 to fabs/fneg on f64 |
| 3125 | // fabs/fneg f128 => fabs/fneg f64:sub_even64, fmov f64:sub_odd64 |
| 3126 | // (As with LowerF64Op, on little-endian, we need to negate the odd |
| 3127 | // subreg) |
| 3128 | |
| 3129 | SDValue SrcReg128 = Op.getOperand(i: 0); |
| 3130 | SDValue Hi64 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_even64, DL: dl, VT: MVT::f64, |
| 3131 | Operand: SrcReg128); |
| 3132 | SDValue Lo64 = DAG.getTargetExtractSubreg(SRIdx: SP::sub_odd64, DL: dl, VT: MVT::f64, |
| 3133 | Operand: SrcReg128); |
| 3134 | |
| 3135 | if (DAG.getDataLayout().isLittleEndian()) { |
| 3136 | if (isV9) |
| 3137 | Lo64 = DAG.getNode(Opcode: Op.getOpcode(), DL: dl, VT: MVT::f64, Operand: Lo64); |
| 3138 | else |
| 3139 | Lo64 = LowerF64Op(SrcReg64: Lo64, dl, DAG, opcode: Op.getOpcode()); |
| 3140 | } else { |
| 3141 | if (isV9) |
| 3142 | Hi64 = DAG.getNode(Opcode: Op.getOpcode(), DL: dl, VT: MVT::f64, Operand: Hi64); |
| 3143 | else |
| 3144 | Hi64 = LowerF64Op(SrcReg64: Hi64, dl, DAG, opcode: Op.getOpcode()); |
| 3145 | } |
| 3146 | |
| 3147 | SDValue DstReg128 = SDValue(DAG.getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, |
| 3148 | dl, VT: MVT::f128), 0); |
| 3149 | DstReg128 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_even64, DL: dl, VT: MVT::f128, |
| 3150 | Operand: DstReg128, Subreg: Hi64); |
| 3151 | DstReg128 = DAG.getTargetInsertSubreg(SRIdx: SP::sub_odd64, DL: dl, VT: MVT::f128, |
| 3152 | Operand: DstReg128, Subreg: Lo64); |
| 3153 | return DstReg128; |
| 3154 | } |
| 3155 | |
| 3156 | static SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) { |
| 3157 | if (isStrongerThanMonotonic(AO: cast<AtomicSDNode>(Val&: Op)->getSuccessOrdering())) { |
| 3158 | // Expand with a fence. |
| 3159 | return SDValue(); |
| 3160 | } |
| 3161 | |
| 3162 | // Monotonic load/stores are legal. |
| 3163 | return Op; |
| 3164 | } |
| 3165 | |
| 3166 | SDValue SparcTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, |
| 3167 | SelectionDAG &DAG) const { |
| 3168 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
| 3169 | switch (IntNo) { |
| 3170 | default: return SDValue(); // Don't custom lower most intrinsics. |
| 3171 | case Intrinsic::thread_pointer: { |
| 3172 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
| 3173 | return DAG.getRegister(Reg: SP::G7, VT: PtrVT); |
| 3174 | } |
| 3175 | } |
| 3176 | } |
| 3177 | |
| 3178 | SDValue SparcTargetLowering:: |
| 3179 | LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
| 3180 | |
| 3181 | bool hasHardQuad = Subtarget->hasHardQuad(); |
| 3182 | bool isV9 = Subtarget->isV9(); |
| 3183 | bool is64Bit = Subtarget->is64Bit(); |
| 3184 | |
| 3185 | switch (Op.getOpcode()) { |
| 3186 | default: llvm_unreachable("Should not custom lower this!" ); |
| 3187 | |
| 3188 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG, TLI: *this, |
| 3189 | Subtarget); |
| 3190 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG, |
| 3191 | Subtarget); |
| 3192 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
| 3193 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
| 3194 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
| 3195 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
| 3196 | case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG, TLI: *this, |
| 3197 | hasHardQuad); |
| 3198 | case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG, TLI: *this, |
| 3199 | hasHardQuad); |
| 3200 | case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG, TLI: *this, |
| 3201 | hasHardQuad); |
| 3202 | case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG, TLI: *this, |
| 3203 | hasHardQuad); |
| 3204 | case ISD::BR_CC: |
| 3205 | return LowerBR_CC(Op, DAG, TLI: *this, hasHardQuad, isV9, is64Bit); |
| 3206 | case ISD::SELECT_CC: |
| 3207 | return LowerSELECT_CC(Op, DAG, TLI: *this, hasHardQuad, isV9, is64Bit); |
| 3208 | case ISD::VASTART: return LowerVASTART(Op, DAG, TLI: *this); |
| 3209 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
| 3210 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG, |
| 3211 | Subtarget); |
| 3212 | case ISD::STACKADDRESS: |
| 3213 | return LowerSTACKADDRESS(Op, DAG, Subtarget: *Subtarget); |
| 3214 | |
| 3215 | case ISD::BSWAP: |
| 3216 | return LowerBSWAP(Op, DAG); |
| 3217 | |
| 3218 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
| 3219 | case ISD::STORE: return LowerSTORE(Op, DAG); |
| 3220 | case ISD::FADD: |
| 3221 | return LowerF128Op(Op, DAG, LibFunc: RTLIB::ADD_F128, numArgs: 2); |
| 3222 | case ISD::FSUB: |
| 3223 | return LowerF128Op(Op, DAG, LibFunc: RTLIB::SUB_F128, numArgs: 2); |
| 3224 | case ISD::FMUL: |
| 3225 | return LowerF128Op(Op, DAG, LibFunc: RTLIB::MUL_F128, numArgs: 2); |
| 3226 | case ISD::FDIV: |
| 3227 | return LowerF128Op(Op, DAG, LibFunc: RTLIB::DIV_F128, numArgs: 2); |
| 3228 | case ISD::FSQRT: |
| 3229 | return LowerF128Op(Op, DAG, LibFunc: RTLIB::SQRT_F128, numArgs: 1); |
| 3230 | case ISD::FABS: |
| 3231 | case ISD::FNEG: return LowerFNEGorFABS(Op, DAG, isV9); |
| 3232 | case ISD::FP_EXTEND: return LowerF128_FPEXTEND(Op, DAG, TLI: *this); |
| 3233 | case ISD::FP_ROUND: return LowerF128_FPROUND(Op, DAG, TLI: *this); |
| 3234 | case ISD::ATOMIC_LOAD: |
| 3235 | case ISD::ATOMIC_STORE: return LowerATOMIC_LOAD_STORE(Op, DAG); |
| 3236 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
| 3237 | } |
| 3238 | } |
| 3239 | |
| 3240 | SDValue SparcTargetLowering::bitcastConstantFPToInt(ConstantFPSDNode *C, |
| 3241 | const SDLoc &DL, |
| 3242 | SelectionDAG &DAG) const { |
| 3243 | APInt V = C->getValueAPF().bitcastToAPInt(); |
| 3244 | SDValue Lo = DAG.getConstant(Val: V.zextOrTrunc(width: 32), DL, VT: MVT::i32); |
| 3245 | SDValue Hi = DAG.getConstant(Val: V.lshr(shiftAmt: 32).zextOrTrunc(width: 32), DL, VT: MVT::i32); |
| 3246 | if (DAG.getDataLayout().isLittleEndian()) |
| 3247 | std::swap(a&: Lo, b&: Hi); |
| 3248 | return DAG.getBuildVector(VT: MVT::v2i32, DL, Ops: {Hi, Lo}); |
| 3249 | } |
| 3250 | |
| 3251 | SDValue SparcTargetLowering::PerformBITCASTCombine(SDNode *N, |
| 3252 | DAGCombinerInfo &DCI) const { |
| 3253 | SDLoc dl(N); |
| 3254 | SDValue Src = N->getOperand(Num: 0); |
| 3255 | |
| 3256 | if (isa<ConstantFPSDNode>(Val: Src) && N->getSimpleValueType(ResNo: 0) == MVT::v2i32 && |
| 3257 | Src.getSimpleValueType() == MVT::f64) |
| 3258 | return bitcastConstantFPToInt(C: cast<ConstantFPSDNode>(Val&: Src), DL: dl, DAG&: DCI.DAG); |
| 3259 | |
| 3260 | return SDValue(); |
| 3261 | } |
| 3262 | |
| 3263 | SDValue SparcTargetLowering::PerformBSWAPCombine(SDNode *N, |
| 3264 | DAGCombinerInfo &DCI) const { |
| 3265 | SDLoc DL(N); |
| 3266 | SelectionDAG &DAG = DCI.DAG; |
| 3267 | SDValue Op = N->getOperand(Num: 0); |
| 3268 | EVT VT = N->getValueType(ResNo: 0); |
| 3269 | auto *LN = dyn_cast<LoadSDNode>(Val: Op.getNode()); |
| 3270 | |
| 3271 | bool IsLittleEndian = DAG.getDataLayout().isLittleEndian(); |
| 3272 | bool IsAlignedLoad = LN && ISD::isNormalLoad(N: Op.getNode()) && |
| 3273 | LN->getAlign() >= VT.getScalarStoreSize(); |
| 3274 | |
| 3275 | // Turn BSWAP (aligned-LOAD) -> ld*a #ASI_P(_L) on V9. |
| 3276 | if (Subtarget->isV9() && IsAlignedLoad && Op.getNode()->hasOneUse() && |
| 3277 | (VT == MVT::i16 || VT == MVT::i32 || |
| 3278 | (Subtarget->is64Bit() && VT == MVT::i64))) { |
| 3279 | SDValue Load = Op; |
| 3280 | auto *LD = cast<LoadSDNode>(Val&: Load); |
| 3281 | |
| 3282 | // Create the byte-swapping load. |
| 3283 | SDValue Ops[] = {LD->getChain(), LD->getBasePtr(), DAG.getValueType(VT)}; |
| 3284 | |
| 3285 | SDValue BSLoad = DAG.getMemIntrinsicNode( |
| 3286 | Opcode: IsLittleEndian ? SPISD::LOAD_BIG : SPISD::LOAD_LITTLE, dl: DL, |
| 3287 | VTList: DAG.getVTList(VT1: VT == MVT::i64 ? MVT::i64 : MVT::i32, VT2: MVT::Other), Ops, |
| 3288 | MemVT: LD->getMemoryVT(), MMO: LD->getMemOperand()); |
| 3289 | |
| 3290 | // If this is an i16 load, insert the truncate. |
| 3291 | SDValue ResVal = BSLoad; |
| 3292 | if (VT == MVT::i16) |
| 3293 | ResVal = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i16, Operand: BSLoad); |
| 3294 | |
| 3295 | return DCI.CombineTo(N, Res: ResVal); |
| 3296 | } |
| 3297 | |
| 3298 | return SDValue(); |
| 3299 | } |
| 3300 | |
| 3301 | SDValue SparcTargetLowering::PerformSTORECombine(SDNode *N, |
| 3302 | DAGCombinerInfo &DCI) const { |
| 3303 | SDLoc DL(N); |
| 3304 | SelectionDAG &DAG = DCI.DAG; |
| 3305 | SDValue Op = N->getOperand(Num: 1); |
| 3306 | EVT VT = Op.getValueType(); |
| 3307 | EVT MemVT = cast<StoreSDNode>(Val: N)->getMemoryVT(); |
| 3308 | unsigned Opcode = Op.getOpcode(); |
| 3309 | auto *SN = dyn_cast<StoreSDNode>(Val: N); |
| 3310 | |
| 3311 | bool IsLittleEndian = DAG.getDataLayout().isLittleEndian(); |
| 3312 | bool IsAlignedStore = SN && SN->getAlign() >= MemVT.getScalarStoreSize(); |
| 3313 | |
| 3314 | // Turn aligned-STORE (BSWAP) -> st*a #ASI_P(_L) on V9. |
| 3315 | if (Subtarget->isV9() && Opcode == ISD::BSWAP && Op.getNode()->hasOneUse() && |
| 3316 | IsAlignedStore && |
| 3317 | (VT == MVT::i16 || VT == MVT::i32 || |
| 3318 | (Subtarget->is64Bit() && VT == MVT::i64))) { |
| 3319 | |
| 3320 | // st*a can only handle simple types and it makes no sense to store less |
| 3321 | // than two bytes in byte-reversed order. |
| 3322 | if (MemVT.getSizeInBits() < 16) |
| 3323 | return SDValue(); |
| 3324 | |
| 3325 | SDValue BSwapOp = Op.getOperand(i: 0); |
| 3326 | // Do an any-extend to 32-bits if this is a half-word input. |
| 3327 | if (BSwapOp.getValueType() == MVT::i16) |
| 3328 | BSwapOp = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: MVT::i32, Operand: BSwapOp); |
| 3329 | |
| 3330 | // If the type of BSWAP operand is wider than stored memory width |
| 3331 | // it needs to be shifted to the right side before st*a. |
| 3332 | if (VT.bitsGT(VT: MemVT)) { |
| 3333 | unsigned Shift = VT.getSizeInBits() - MemVT.getSizeInBits(); |
| 3334 | BSwapOp = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: BSwapOp, |
| 3335 | N2: DAG.getShiftAmountConstant(Val: Shift, VT, DL)); |
| 3336 | } |
| 3337 | |
| 3338 | SDValue Ops[] = {N->getOperand(Num: 0), BSwapOp, N->getOperand(Num: 2), |
| 3339 | DAG.getValueType(MemVT)}; |
| 3340 | return DAG.getMemIntrinsicNode( |
| 3341 | Opcode: IsLittleEndian ? SPISD::STORE_BIG : SPISD::STORE_LITTLE, dl: DL, |
| 3342 | VTList: DAG.getVTList(VT: MVT::Other), Ops, MemVT: cast<StoreSDNode>(Val: N)->getMemoryVT(), |
| 3343 | MMO: cast<StoreSDNode>(Val: N)->getMemOperand()); |
| 3344 | } |
| 3345 | |
| 3346 | return SDValue(); |
| 3347 | } |
| 3348 | |
| 3349 | SDValue SparcTargetLowering::PerformDAGCombine(SDNode *N, |
| 3350 | DAGCombinerInfo &DCI) const { |
| 3351 | switch (N->getOpcode()) { |
| 3352 | default: |
| 3353 | break; |
| 3354 | case ISD::BITCAST: |
| 3355 | return PerformBITCASTCombine(N, DCI); |
| 3356 | case ISD::BSWAP: |
| 3357 | return PerformBSWAPCombine(N, DCI); |
| 3358 | case ISD::STORE: |
| 3359 | return PerformSTORECombine(N, DCI); |
| 3360 | } |
| 3361 | return SDValue(); |
| 3362 | } |
| 3363 | |
| 3364 | MachineBasicBlock * |
| 3365 | SparcTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
| 3366 | MachineBasicBlock *BB) const { |
| 3367 | switch (MI.getOpcode()) { |
| 3368 | default: llvm_unreachable("Unknown SELECT_CC!" ); |
| 3369 | case SP::SELECT_CC_Int_ICC: |
| 3370 | case SP::SELECT_CC_FP_ICC: |
| 3371 | case SP::SELECT_CC_DFP_ICC: |
| 3372 | case SP::SELECT_CC_QFP_ICC: |
| 3373 | if (Subtarget->isV9()) |
| 3374 | return expandSelectCC(MI, BB, BROpcode: SP::BPICC); |
| 3375 | return expandSelectCC(MI, BB, BROpcode: SP::BCOND); |
| 3376 | case SP::SELECT_CC_Int_XCC: |
| 3377 | case SP::SELECT_CC_FP_XCC: |
| 3378 | case SP::SELECT_CC_DFP_XCC: |
| 3379 | case SP::SELECT_CC_QFP_XCC: |
| 3380 | return expandSelectCC(MI, BB, BROpcode: SP::BPXCC); |
| 3381 | case SP::SELECT_CC_Int_FCC: |
| 3382 | case SP::SELECT_CC_FP_FCC: |
| 3383 | case SP::SELECT_CC_DFP_FCC: |
| 3384 | case SP::SELECT_CC_QFP_FCC: |
| 3385 | if (Subtarget->isV9()) |
| 3386 | return expandSelectCC(MI, BB, BROpcode: SP::FBCOND_V9); |
| 3387 | return expandSelectCC(MI, BB, BROpcode: SP::FBCOND); |
| 3388 | } |
| 3389 | } |
| 3390 | |
| 3391 | MachineBasicBlock * |
| 3392 | SparcTargetLowering::expandSelectCC(MachineInstr &MI, MachineBasicBlock *BB, |
| 3393 | unsigned BROpcode) const { |
| 3394 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
| 3395 | DebugLoc dl = MI.getDebugLoc(); |
| 3396 | unsigned CC = (SPCC::CondCodes)MI.getOperand(i: 3).getImm(); |
| 3397 | |
| 3398 | // To "insert" a SELECT_CC instruction, we actually have to insert the |
| 3399 | // triangle control-flow pattern. The incoming instruction knows the |
| 3400 | // destination vreg to set, the condition code register to branch on, the |
| 3401 | // true/false values to select between, and the condition code for the branch. |
| 3402 | // |
| 3403 | // We produce the following control flow: |
| 3404 | // ThisMBB |
| 3405 | // | \ |
| 3406 | // | IfFalseMBB |
| 3407 | // | / |
| 3408 | // SinkMBB |
| 3409 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
| 3410 | MachineFunction::iterator It = ++BB->getIterator(); |
| 3411 | |
| 3412 | MachineBasicBlock *ThisMBB = BB; |
| 3413 | MachineFunction *F = BB->getParent(); |
| 3414 | MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 3415 | MachineBasicBlock *SinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
| 3416 | F->insert(MBBI: It, MBB: IfFalseMBB); |
| 3417 | F->insert(MBBI: It, MBB: SinkMBB); |
| 3418 | |
| 3419 | // Transfer the remainder of ThisMBB and its successor edges to SinkMBB. |
| 3420 | SinkMBB->splice(Where: SinkMBB->begin(), Other: ThisMBB, |
| 3421 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: ThisMBB->end()); |
| 3422 | SinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: ThisMBB); |
| 3423 | |
| 3424 | // Set the new successors for ThisMBB. |
| 3425 | ThisMBB->addSuccessor(Succ: IfFalseMBB); |
| 3426 | ThisMBB->addSuccessor(Succ: SinkMBB); |
| 3427 | |
| 3428 | BuildMI(BB: ThisMBB, MIMD: dl, MCID: TII.get(Opcode: BROpcode)) |
| 3429 | .addMBB(MBB: SinkMBB) |
| 3430 | .addImm(Val: CC); |
| 3431 | |
| 3432 | // IfFalseMBB just falls through to SinkMBB. |
| 3433 | IfFalseMBB->addSuccessor(Succ: SinkMBB); |
| 3434 | |
| 3435 | // %Result = phi [ %TrueValue, ThisMBB ], [ %FalseValue, IfFalseMBB ] |
| 3436 | BuildMI(BB&: *SinkMBB, I: SinkMBB->begin(), MIMD: dl, MCID: TII.get(Opcode: SP::PHI), |
| 3437 | DestReg: MI.getOperand(i: 0).getReg()) |
| 3438 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
| 3439 | .addMBB(MBB: ThisMBB) |
| 3440 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
| 3441 | .addMBB(MBB: IfFalseMBB); |
| 3442 | |
| 3443 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
| 3444 | return SinkMBB; |
| 3445 | } |
| 3446 | |
| 3447 | //===----------------------------------------------------------------------===// |
| 3448 | // Sparc Inline Assembly Support |
| 3449 | //===----------------------------------------------------------------------===// |
| 3450 | |
| 3451 | /// getConstraintType - Given a constraint letter, return the type of |
| 3452 | /// constraint it is for this target. |
| 3453 | SparcTargetLowering::ConstraintType |
| 3454 | SparcTargetLowering::getConstraintType(StringRef Constraint) const { |
| 3455 | if (Constraint.size() == 1) { |
| 3456 | switch (Constraint[0]) { |
| 3457 | default: break; |
| 3458 | case 'r': |
| 3459 | case 'f': |
| 3460 | case 'e': |
| 3461 | return C_RegisterClass; |
| 3462 | case 'I': // SIMM13 |
| 3463 | return C_Immediate; |
| 3464 | } |
| 3465 | } |
| 3466 | |
| 3467 | return TargetLowering::getConstraintType(Constraint); |
| 3468 | } |
| 3469 | |
| 3470 | TargetLowering::ConstraintWeight SparcTargetLowering:: |
| 3471 | getSingleConstraintMatchWeight(AsmOperandInfo &info, |
| 3472 | const char *constraint) const { |
| 3473 | ConstraintWeight weight = CW_Invalid; |
| 3474 | Value *CallOperandVal = info.CallOperandVal; |
| 3475 | // If we don't have a value, we can't do a match, |
| 3476 | // but allow it at the lowest weight. |
| 3477 | if (!CallOperandVal) |
| 3478 | return CW_Default; |
| 3479 | |
| 3480 | // Look at the constraint type. |
| 3481 | switch (*constraint) { |
| 3482 | default: |
| 3483 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
| 3484 | break; |
| 3485 | case 'I': // SIMM13 |
| 3486 | if (ConstantInt *C = dyn_cast<ConstantInt>(Val: info.CallOperandVal)) { |
| 3487 | if (isInt<13>(x: C->getSExtValue())) |
| 3488 | weight = CW_Constant; |
| 3489 | } |
| 3490 | break; |
| 3491 | } |
| 3492 | return weight; |
| 3493 | } |
| 3494 | |
| 3495 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
| 3496 | /// vector. If it is invalid, don't add anything to Ops. |
| 3497 | void SparcTargetLowering::LowerAsmOperandForConstraint( |
| 3498 | SDValue Op, StringRef Constraint, std::vector<SDValue> &Ops, |
| 3499 | SelectionDAG &DAG) const { |
| 3500 | SDValue Result; |
| 3501 | |
| 3502 | // Only support length 1 constraints for now. |
| 3503 | if (Constraint.size() > 1) |
| 3504 | return; |
| 3505 | |
| 3506 | char ConstraintLetter = Constraint[0]; |
| 3507 | switch (ConstraintLetter) { |
| 3508 | default: break; |
| 3509 | case 'I': |
| 3510 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
| 3511 | if (isInt<13>(x: C->getSExtValue())) { |
| 3512 | Result = DAG.getSignedTargetConstant(Val: C->getSExtValue(), DL: SDLoc(Op), |
| 3513 | VT: Op.getValueType()); |
| 3514 | break; |
| 3515 | } |
| 3516 | return; |
| 3517 | } |
| 3518 | } |
| 3519 | |
| 3520 | if (Result.getNode()) { |
| 3521 | Ops.push_back(x: Result); |
| 3522 | return; |
| 3523 | } |
| 3524 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
| 3525 | } |
| 3526 | |
| 3527 | std::pair<unsigned, const TargetRegisterClass *> |
| 3528 | SparcTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
| 3529 | StringRef Constraint, |
| 3530 | MVT VT) const { |
| 3531 | if (Constraint.empty()) |
| 3532 | return std::make_pair(x: 0U, y: nullptr); |
| 3533 | |
| 3534 | if (Constraint.size() == 1) { |
| 3535 | switch (Constraint[0]) { |
| 3536 | case 'r': |
| 3537 | if (VT == MVT::v2i32) |
| 3538 | return std::make_pair(x: 0U, y: &SP::IntPairRegClass); |
| 3539 | else if (Subtarget->is64Bit()) |
| 3540 | return std::make_pair(x: 0U, y: &SP::I64RegsRegClass); |
| 3541 | else |
| 3542 | return std::make_pair(x: 0U, y: &SP::IntRegsRegClass); |
| 3543 | case 'f': |
| 3544 | if (VT == MVT::f32 || VT == MVT::i32) |
| 3545 | return std::make_pair(x: 0U, y: &SP::FPRegsRegClass); |
| 3546 | else if (VT == MVT::f64 || VT == MVT::i64) |
| 3547 | return std::make_pair(x: 0U, y: &SP::LowDFPRegsRegClass); |
| 3548 | else if (VT == MVT::f128) |
| 3549 | return std::make_pair(x: 0U, y: &SP::LowQFPRegsRegClass); |
| 3550 | // This will generate an error message |
| 3551 | return std::make_pair(x: 0U, y: nullptr); |
| 3552 | case 'e': |
| 3553 | if (VT == MVT::f32 || VT == MVT::i32) |
| 3554 | return std::make_pair(x: 0U, y: &SP::FPRegsRegClass); |
| 3555 | else if (VT == MVT::f64 || VT == MVT::i64 ) |
| 3556 | return std::make_pair(x: 0U, y: &SP::DFPRegsRegClass); |
| 3557 | else if (VT == MVT::f128) |
| 3558 | return std::make_pair(x: 0U, y: &SP::QFPRegsRegClass); |
| 3559 | // This will generate an error message |
| 3560 | return std::make_pair(x: 0U, y: nullptr); |
| 3561 | } |
| 3562 | } |
| 3563 | |
| 3564 | if (Constraint.front() != '{') |
| 3565 | return std::make_pair(x: 0U, y: nullptr); |
| 3566 | |
| 3567 | assert(Constraint.back() == '}' && "Not a brace enclosed constraint?" ); |
| 3568 | StringRef RegName(Constraint.data() + 1, Constraint.size() - 2); |
| 3569 | if (RegName.empty()) |
| 3570 | return std::make_pair(x: 0U, y: nullptr); |
| 3571 | |
| 3572 | unsigned long long RegNo; |
| 3573 | // Handle numbered register aliases. |
| 3574 | if (RegName[0] == 'r' && |
| 3575 | getAsUnsignedInteger(Str: RegName.begin() + 1, Radix: 10, Result&: RegNo)) { |
| 3576 | // r0-r7 -> g0-g7 |
| 3577 | // r8-r15 -> o0-o7 |
| 3578 | // r16-r23 -> l0-l7 |
| 3579 | // r24-r31 -> i0-i7 |
| 3580 | if (RegNo > 31) |
| 3581 | return std::make_pair(x: 0U, y: nullptr); |
| 3582 | const char RegTypes[] = {'g', 'o', 'l', 'i'}; |
| 3583 | char RegType = RegTypes[RegNo / 8]; |
| 3584 | char RegIndex = '0' + (RegNo % 8); |
| 3585 | char Tmp[] = {'{', RegType, RegIndex, '}', 0}; |
| 3586 | return getRegForInlineAsmConstraint(TRI, Constraint: Tmp, VT); |
| 3587 | } |
| 3588 | |
| 3589 | // Rewrite the fN constraint according to the value type if needed. |
| 3590 | if (VT != MVT::f32 && VT != MVT::Other && RegName[0] == 'f' && |
| 3591 | getAsUnsignedInteger(Str: RegName.begin() + 1, Radix: 10, Result&: RegNo)) { |
| 3592 | if (VT == MVT::f64 && (RegNo % 2 == 0)) { |
| 3593 | return getRegForInlineAsmConstraint( |
| 3594 | TRI, Constraint: StringRef("{d" + utostr(X: RegNo / 2) + "}" ), VT); |
| 3595 | } else if (VT == MVT::f128 && (RegNo % 4 == 0)) { |
| 3596 | return getRegForInlineAsmConstraint( |
| 3597 | TRI, Constraint: StringRef("{q" + utostr(X: RegNo / 4) + "}" ), VT); |
| 3598 | } else { |
| 3599 | return std::make_pair(x: 0U, y: nullptr); |
| 3600 | } |
| 3601 | } |
| 3602 | |
| 3603 | auto ResultPair = |
| 3604 | TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
| 3605 | if (!ResultPair.second) |
| 3606 | return std::make_pair(x: 0U, y: nullptr); |
| 3607 | |
| 3608 | // Force the use of I64Regs over IntRegs for 64-bit values. |
| 3609 | if (Subtarget->is64Bit() && VT == MVT::i64) { |
| 3610 | assert(ResultPair.second == &SP::IntRegsRegClass && |
| 3611 | "Unexpected register class" ); |
| 3612 | return std::make_pair(x&: ResultPair.first, y: &SP::I64RegsRegClass); |
| 3613 | } |
| 3614 | |
| 3615 | return ResultPair; |
| 3616 | } |
| 3617 | |
| 3618 | bool |
| 3619 | SparcTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
| 3620 | // The Sparc target isn't yet aware of offsets. |
| 3621 | return false; |
| 3622 | } |
| 3623 | |
| 3624 | void SparcTargetLowering::ReplaceNodeResults(SDNode *N, |
| 3625 | SmallVectorImpl<SDValue>& Results, |
| 3626 | SelectionDAG &DAG) const { |
| 3627 | |
| 3628 | SDLoc dl(N); |
| 3629 | |
| 3630 | RTLIB::Libcall libCall = RTLIB::UNKNOWN_LIBCALL; |
| 3631 | |
| 3632 | switch (N->getOpcode()) { |
| 3633 | default: |
| 3634 | llvm_unreachable("Do not know how to custom type legalize this operation!" ); |
| 3635 | |
| 3636 | case ISD::FP_TO_SINT: |
| 3637 | case ISD::FP_TO_UINT: |
| 3638 | // Custom lower only if it involves f128 or i64. |
| 3639 | if (N->getOperand(Num: 0).getValueType() != MVT::f128 |
| 3640 | || N->getValueType(ResNo: 0) != MVT::i64) |
| 3641 | return; |
| 3642 | libCall = ((N->getOpcode() == ISD::FP_TO_SINT) |
| 3643 | ? RTLIB::FPTOSINT_F128_I64 |
| 3644 | : RTLIB::FPTOUINT_F128_I64); |
| 3645 | |
| 3646 | Results.push_back(Elt: LowerF128Op(Op: SDValue(N, 0), DAG, LibFunc: libCall, numArgs: 1)); |
| 3647 | return; |
| 3648 | case ISD::READCYCLECOUNTER: { |
| 3649 | assert(Subtarget->hasLeonCycleCounter()); |
| 3650 | SDValue Lo = DAG.getCopyFromReg(Chain: N->getOperand(Num: 0), dl, Reg: SP::ASR23, VT: MVT::i32); |
| 3651 | SDValue Hi = DAG.getCopyFromReg(Chain: Lo, dl, Reg: SP::G0, VT: MVT::i32); |
| 3652 | SDValue Ops[] = { Lo, Hi }; |
| 3653 | SDValue Pair = DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, Ops); |
| 3654 | Results.push_back(Elt: Pair); |
| 3655 | Results.push_back(Elt: N->getOperand(Num: 0)); |
| 3656 | return; |
| 3657 | } |
| 3658 | case ISD::SINT_TO_FP: |
| 3659 | case ISD::UINT_TO_FP: |
| 3660 | // Custom lower only if it involves f128 or i64. |
| 3661 | if (N->getValueType(ResNo: 0) != MVT::f128 |
| 3662 | || N->getOperand(Num: 0).getValueType() != MVT::i64) |
| 3663 | return; |
| 3664 | |
| 3665 | libCall = ((N->getOpcode() == ISD::SINT_TO_FP) |
| 3666 | ? RTLIB::SINTTOFP_I64_F128 |
| 3667 | : RTLIB::UINTTOFP_I64_F128); |
| 3668 | |
| 3669 | Results.push_back(Elt: LowerF128Op(Op: SDValue(N, 0), DAG, LibFunc: libCall, numArgs: 1)); |
| 3670 | return; |
| 3671 | case ISD::LOAD: { |
| 3672 | LoadSDNode *Ld = cast<LoadSDNode>(Val: N); |
| 3673 | // Custom handling only for i64: turn i64 load into a v2i32 load, |
| 3674 | // and a bitcast. |
| 3675 | if (Ld->getValueType(ResNo: 0) != MVT::i64 || Ld->getMemoryVT() != MVT::i64) |
| 3676 | return; |
| 3677 | |
| 3678 | SDLoc dl(N); |
| 3679 | SDValue LoadRes = DAG.getExtLoad( |
| 3680 | ExtType: Ld->getExtensionType(), dl, VT: MVT::v2i32, Chain: Ld->getChain(), |
| 3681 | Ptr: Ld->getBasePtr(), PtrInfo: Ld->getPointerInfo(), MemVT: MVT::v2i32, Alignment: Ld->getBaseAlign(), |
| 3682 | MMOFlags: Ld->getMemOperand()->getFlags(), AAInfo: Ld->getAAInfo()); |
| 3683 | |
| 3684 | SDValue Res = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i64, Operand: LoadRes); |
| 3685 | Results.push_back(Elt: Res); |
| 3686 | Results.push_back(Elt: LoadRes.getValue(R: 1)); |
| 3687 | return; |
| 3688 | } |
| 3689 | } |
| 3690 | } |
| 3691 | |
| 3692 | // Override to enable LOAD_STACK_GUARD lowering on Linux. |
| 3693 | bool SparcTargetLowering::useLoadStackGuardNode(const Module &M) const { |
| 3694 | if (!Subtarget->getTargetTriple().isOSLinux()) |
| 3695 | return TargetLowering::useLoadStackGuardNode(M); |
| 3696 | return true; |
| 3697 | } |
| 3698 | |
| 3699 | bool SparcTargetLowering::isFNegFree(EVT VT) const { |
| 3700 | if (Subtarget->isVIS3()) |
| 3701 | return VT == MVT::f32 || VT == MVT::f64; |
| 3702 | return false; |
| 3703 | } |
| 3704 | |
| 3705 | bool SparcTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
| 3706 | bool ForCodeSize) const { |
| 3707 | if (VT != MVT::f32 && VT != MVT::f64) |
| 3708 | return false; |
| 3709 | if (Subtarget->isVIS() && Imm.isZero()) |
| 3710 | return true; |
| 3711 | if (Subtarget->isVIS3()) |
| 3712 | return Imm.isExactlyValue(V: +0.5) || Imm.isExactlyValue(V: -0.5) || |
| 3713 | Imm.getExactLog2Abs() == -1; |
| 3714 | return false; |
| 3715 | } |
| 3716 | |
| 3717 | bool SparcTargetLowering::isCtlzFast() const { return Subtarget->isVIS3(); } |
| 3718 | |
| 3719 | bool SparcTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
| 3720 | // We lack native cttz, however, |
| 3721 | // On 64-bit targets it is cheap to implement it in terms of popc. |
| 3722 | if (Subtarget->is64Bit() && Subtarget->usePopc()) |
| 3723 | return true; |
| 3724 | // Otherwise, implementing cttz in terms of ctlz is still cheap. |
| 3725 | return isCheapToSpeculateCtlz(Ty); |
| 3726 | } |
| 3727 | |
| 3728 | bool SparcTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, |
| 3729 | EVT VT) const { |
| 3730 | return Subtarget->isUA2007() && !Subtarget->useSoftFloat(); |
| 3731 | } |
| 3732 | |
| 3733 | void SparcTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, |
| 3734 | SDNode *Node) const { |
| 3735 | assert(MI.getOpcode() == SP::SUBCCrr || MI.getOpcode() == SP::SUBCCri); |
| 3736 | // If the result is dead, replace it with %g0. |
| 3737 | if (!Node->hasAnyUseOfValue(Value: 0)) |
| 3738 | MI.getOperand(i: 0).setReg(SP::G0); |
| 3739 | } |
| 3740 | |
| 3741 | Instruction *SparcTargetLowering::emitLeadingFence(IRBuilderBase &Builder, |
| 3742 | Instruction *Inst, |
| 3743 | AtomicOrdering Ord) const { |
| 3744 | bool HasStoreSemantics = |
| 3745 | isa<AtomicCmpXchgInst, AtomicRMWInst, StoreInst>(Val: Inst); |
| 3746 | if (HasStoreSemantics && isReleaseOrStronger(AO: Ord)) |
| 3747 | return Builder.CreateFence(Ordering: AtomicOrdering::Release); |
| 3748 | return nullptr; |
| 3749 | } |
| 3750 | |
| 3751 | Instruction *SparcTargetLowering::emitTrailingFence(IRBuilderBase &Builder, |
| 3752 | Instruction *Inst, |
| 3753 | AtomicOrdering Ord) const { |
| 3754 | // V8 loads already come with implicit acquire barrier so there's no need to |
| 3755 | // emit it again. |
| 3756 | bool HasLoadSemantics = isa<AtomicCmpXchgInst, AtomicRMWInst, LoadInst>(Val: Inst); |
| 3757 | if (Subtarget->isV9() && HasLoadSemantics && isAcquireOrStronger(AO: Ord)) |
| 3758 | return Builder.CreateFence(Ordering: AtomicOrdering::Acquire); |
| 3759 | |
| 3760 | // SC plain stores would need a trailing full barrier. |
| 3761 | if (isa<StoreInst>(Val: Inst) && Ord == AtomicOrdering::SequentiallyConsistent) |
| 3762 | return Builder.CreateFence(Ordering: Ord); |
| 3763 | return nullptr; |
| 3764 | } |
| 3765 | |