| 1 | //===-- AVRRegisterInfo.cpp - AVR Register Information --------------------===// |
| 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 contains the AVR implementation of the TargetRegisterInfo class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "AVRRegisterInfo.h" |
| 14 | |
| 15 | #include "llvm/ADT/BitVector.h" |
| 16 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 17 | #include "llvm/CodeGen/MachineFunction.h" |
| 18 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 19 | #include "llvm/CodeGen/TargetFrameLowering.h" |
| 20 | |
| 21 | #include "AVR.h" |
| 22 | #include "AVRInstrInfo.h" |
| 23 | #include "AVRMachineFunctionInfo.h" |
| 24 | #include "AVRTargetMachine.h" |
| 25 | #include "MCTargetDesc/AVRMCTargetDesc.h" |
| 26 | |
| 27 | #define GET_REGINFO_TARGET_DESC |
| 28 | #include "AVRGenRegisterInfo.inc" |
| 29 | |
| 30 | namespace llvm { |
| 31 | |
| 32 | AVRRegisterInfo::AVRRegisterInfo() : AVRGenRegisterInfo(0) {} |
| 33 | |
| 34 | const uint16_t * |
| 35 | AVRRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { |
| 36 | const AVRMachineFunctionInfo *AFI = MF->getInfo<AVRMachineFunctionInfo>(); |
| 37 | const AVRSubtarget &STI = MF->getSubtarget<AVRSubtarget>(); |
| 38 | if (STI.hasTinyEncoding()) |
| 39 | return AFI->isInterruptOrSignalHandler() ? CSR_InterruptsTiny_SaveList |
| 40 | : CSR_NormalTiny_SaveList; |
| 41 | else |
| 42 | return AFI->isInterruptOrSignalHandler() ? CSR_Interrupts_SaveList |
| 43 | : CSR_Normal_SaveList; |
| 44 | } |
| 45 | |
| 46 | const uint32_t * |
| 47 | AVRRegisterInfo::getCallPreservedMask(const MachineFunction &MF, |
| 48 | CallingConv::ID CC) const { |
| 49 | const AVRSubtarget &STI = MF.getSubtarget<AVRSubtarget>(); |
| 50 | return STI.hasTinyEncoding() ? CSR_NormalTiny_RegMask : CSR_Normal_RegMask; |
| 51 | } |
| 52 | |
| 53 | BitVector AVRRegisterInfo::getReservedRegs(const MachineFunction &MF) const { |
| 54 | BitVector Reserved(getNumRegs()); |
| 55 | |
| 56 | // Reserve the intermediate result registers r1 and r2 |
| 57 | // The result of instructions like 'mul' is always stored here. |
| 58 | // R0/R1/R1R0 are always reserved on both avr and avrtiny. |
| 59 | Reserved.set(AVR::R0); |
| 60 | Reserved.set(AVR::R1); |
| 61 | Reserved.set(AVR::R1R0); |
| 62 | |
| 63 | // Reserve the stack pointer. |
| 64 | Reserved.set(AVR::SPL); |
| 65 | Reserved.set(AVR::SPH); |
| 66 | Reserved.set(AVR::SP); |
| 67 | |
| 68 | // Reserve R2~R17 only on avrtiny. |
| 69 | if (MF.getSubtarget<AVRSubtarget>().hasTinyEncoding()) { |
| 70 | // Reserve 8-bit registers R2~R15, Rtmp(R16) and Zero(R17). |
| 71 | for (unsigned Reg = AVR::R2; Reg <= AVR::R17; Reg++) |
| 72 | Reserved.set(Reg); |
| 73 | // Reserve 16-bit registers R3R2~R18R17. |
| 74 | for (unsigned Reg = AVR::R3R2; Reg <= AVR::R18R17; Reg++) |
| 75 | Reserved.set(Reg); |
| 76 | } |
| 77 | |
| 78 | // We tenatively reserve the frame pointer register r29:r28 because the |
| 79 | // function may require one, but we cannot tell until register allocation |
| 80 | // is complete, which can be too late. |
| 81 | // |
| 82 | // Instead we just unconditionally reserve the Y register. |
| 83 | // |
| 84 | // TODO: Write a pass to enumerate functions which reserved the Y register |
| 85 | // but didn't end up needing a frame pointer. In these, we can |
| 86 | // convert one or two of the spills inside to use the Y register. |
| 87 | Reserved.set(AVR::R28); |
| 88 | Reserved.set(AVR::R29); |
| 89 | Reserved.set(AVR::R29R28); |
| 90 | |
| 91 | return Reserved; |
| 92 | } |
| 93 | |
| 94 | const TargetRegisterClass * |
| 95 | AVRRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC, |
| 96 | const MachineFunction &MF) const { |
| 97 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
| 98 | if (TRI->isTypeLegalForClass(RC: *RC, T: MVT::i16)) { |
| 99 | return &AVR::DREGSRegClass; |
| 100 | } |
| 101 | |
| 102 | if (TRI->isTypeLegalForClass(RC: *RC, T: MVT::i8)) { |
| 103 | return &AVR::GPR8RegClass; |
| 104 | } |
| 105 | |
| 106 | llvm_unreachable("Invalid register size" ); |
| 107 | } |
| 108 | |
| 109 | /// Fold a frame offset shared between two add instructions into a single one. |
| 110 | static void foldFrameOffset(MachineBasicBlock::iterator &II, int &Offset, |
| 111 | Register DstReg) { |
| 112 | MachineInstr &MI = *II; |
| 113 | int Opcode = MI.getOpcode(); |
| 114 | |
| 115 | // Don't bother trying if the next instruction is not an add or a sub. |
| 116 | if ((Opcode != AVR::SUBIWRdK) && (Opcode != AVR::ADIWRdK)) { |
| 117 | return; |
| 118 | } |
| 119 | |
| 120 | // Check that DstReg matches with next instruction, otherwise the instruction |
| 121 | // is not related to stack address manipulation. |
| 122 | if (DstReg != MI.getOperand(i: 0).getReg()) { |
| 123 | return; |
| 124 | } |
| 125 | |
| 126 | // Add the offset in the next instruction to our offset. |
| 127 | switch (Opcode) { |
| 128 | case AVR::SUBIWRdK: |
| 129 | Offset += -MI.getOperand(i: 2).getImm(); |
| 130 | break; |
| 131 | case AVR::ADIWRdK: |
| 132 | Offset += MI.getOperand(i: 2).getImm(); |
| 133 | break; |
| 134 | } |
| 135 | |
| 136 | // Finally remove the instruction. |
| 137 | II++; |
| 138 | MI.eraseFromParent(); |
| 139 | } |
| 140 | |
| 141 | bool AVRRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, |
| 142 | int SPAdj, unsigned FIOperandNum, |
| 143 | RegScavenger *RS) const { |
| 144 | assert(SPAdj == 0 && "Unexpected SPAdj value" ); |
| 145 | |
| 146 | MachineInstr &MI = *II; |
| 147 | DebugLoc dl = MI.getDebugLoc(); |
| 148 | MachineBasicBlock &MBB = *MI.getParent(); |
| 149 | const MachineFunction &MF = *MBB.getParent(); |
| 150 | const AVRTargetMachine &TM = (const AVRTargetMachine &)MF.getTarget(); |
| 151 | const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); |
| 152 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 153 | const TargetFrameLowering *TFI = TM.getSubtargetImpl()->getFrameLowering(); |
| 154 | const AVRSubtarget &STI = MF.getSubtarget<AVRSubtarget>(); |
| 155 | int FrameIndex = MI.getOperand(i: FIOperandNum).getIndex(); |
| 156 | int Offset = MFI.getObjectOffset(ObjectIdx: FrameIndex); |
| 157 | |
| 158 | // Add one to the offset because SP points to an empty slot. |
| 159 | Offset += MFI.getStackSize() - TFI->getOffsetOfLocalArea() + 1; |
| 160 | // Fold incoming offset. |
| 161 | Offset += MI.getOperand(i: FIOperandNum + 1).getImm(); |
| 162 | |
| 163 | // This is actually "load effective address" of the stack slot |
| 164 | // instruction. We have only two-address instructions, thus we need to |
| 165 | // expand it into move + add. |
| 166 | if (MI.getOpcode() == AVR::FRMIDX) { |
| 167 | Register DstReg = MI.getOperand(i: 0).getReg(); |
| 168 | assert(DstReg != AVR::R29R28 && "Dest reg cannot be the frame pointer" ); |
| 169 | |
| 170 | // Copy the frame pointer. |
| 171 | if (STI.hasMOVW()) { |
| 172 | BuildMI(BB&: MBB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::MOVWRdRr), DestReg: DstReg).addReg(RegNo: AVR::R29R28); |
| 173 | } else { |
| 174 | Register DstLoReg, DstHiReg; |
| 175 | splitReg(Reg: DstReg, LoReg&: DstLoReg, HiReg&: DstHiReg); |
| 176 | BuildMI(BB&: MBB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::MOVRdRr), DestReg: DstLoReg).addReg(RegNo: AVR::R28); |
| 177 | BuildMI(BB&: MBB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::MOVRdRr), DestReg: DstHiReg).addReg(RegNo: AVR::R29); |
| 178 | } |
| 179 | |
| 180 | assert(Offset > 0 && "Invalid offset" ); |
| 181 | |
| 182 | // We need to materialize the offset via an add instruction. |
| 183 | unsigned Opcode; |
| 184 | |
| 185 | II++; // Skip over the FRMIDX instruction. |
| 186 | |
| 187 | // Generally, to load a frame address two add instructions are emitted that |
| 188 | // could get folded into a single one: |
| 189 | // movw r31:r30, r29:r28 |
| 190 | // adiw r31:r30, 29 |
| 191 | // adiw r31:r30, 16 |
| 192 | // to: |
| 193 | // movw r31:r30, r29:r28 |
| 194 | // adiw r31:r30, 45 |
| 195 | if (II != MBB.end()) |
| 196 | foldFrameOffset(II, Offset, DstReg); |
| 197 | |
| 198 | // Select the best opcode based on DstReg and the offset size. |
| 199 | switch (DstReg) { |
| 200 | case AVR::R25R24: |
| 201 | case AVR::R27R26: |
| 202 | case AVR::R31R30: { |
| 203 | if (isUInt<6>(x: Offset) && STI.hasADDSUBIW()) { |
| 204 | Opcode = AVR::ADIWRdK; |
| 205 | break; |
| 206 | } |
| 207 | [[fallthrough]]; |
| 208 | } |
| 209 | default: { |
| 210 | // This opcode will get expanded into a pair of subi/sbci. |
| 211 | Opcode = AVR::SUBIWRdK; |
| 212 | Offset = -Offset; |
| 213 | break; |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | MachineInstr *New = BuildMI(BB&: MBB, I: II, MIMD: dl, MCID: TII.get(Opcode), DestReg: DstReg) |
| 218 | .addReg(RegNo: DstReg, Flags: RegState::Kill) |
| 219 | .addImm(Val: Offset); |
| 220 | New->getOperand(i: 3).setIsDead(); |
| 221 | |
| 222 | MI.eraseFromParent(); // remove FRMIDX |
| 223 | |
| 224 | return false; |
| 225 | } |
| 226 | |
| 227 | // On most AVRs, we can use an offset up to 62 for load/store with |
| 228 | // displacement (63 for byte values, 62 for word values). However, the |
| 229 | // "reduced tiny" cores don't support load/store with displacement. So for |
| 230 | // them, we force an offset of 0 meaning that any positive offset will require |
| 231 | // adjusting the frame pointer. |
| 232 | int MaxOffset = STI.hasTinyEncoding() ? 0 : 62; |
| 233 | |
| 234 | // If the offset is too big we have to adjust and restore the frame pointer |
| 235 | // to materialize a valid load/store with displacement. |
| 236 | //: TODO: consider using only one adiw/sbiw chain for more than one frame |
| 237 | //: index |
| 238 | if (Offset > MaxOffset) { |
| 239 | unsigned AddOpc = AVR::ADIWRdK, SubOpc = AVR::SBIWRdK; |
| 240 | int AddOffset = Offset - MaxOffset; |
| 241 | |
| 242 | // For huge offsets where adiw/sbiw cannot be used use a pair of subi/sbci. |
| 243 | if ((Offset - MaxOffset) > 63 || !STI.hasADDSUBIW()) { |
| 244 | AddOpc = AVR::SUBIWRdK; |
| 245 | SubOpc = AVR::SUBIWRdK; |
| 246 | AddOffset = -AddOffset; |
| 247 | } |
| 248 | |
| 249 | // It is possible that the spiller places this frame instruction in between |
| 250 | // a compare and branch, invalidating the contents of SREG set by the |
| 251 | // compare instruction because of the add/sub pairs. Conservatively save and |
| 252 | // restore SREG before and after each add/sub pair. |
| 253 | BuildMI(BB&: MBB, I: II, MIMD: dl, MCID: TII.get(Opcode: AVR::INRdA), DestReg: STI.getTmpRegister()) |
| 254 | .addImm(Val: STI.getIORegSREG()); |
| 255 | |
| 256 | MachineInstr *New = BuildMI(BB&: MBB, I: II, MIMD: dl, MCID: TII.get(Opcode: AddOpc), DestReg: AVR::R29R28) |
| 257 | .addReg(RegNo: AVR::R29R28, Flags: RegState::Kill) |
| 258 | .addImm(Val: AddOffset); |
| 259 | New->getOperand(i: 3).setIsDead(); |
| 260 | |
| 261 | // Restore SREG. |
| 262 | BuildMI(BB&: MBB, I: std::next(x: II), MIMD: dl, MCID: TII.get(Opcode: AVR::OUTARr)) |
| 263 | .addImm(Val: STI.getIORegSREG()) |
| 264 | .addReg(RegNo: STI.getTmpRegister(), Flags: RegState::Kill); |
| 265 | |
| 266 | // No need to set SREG as dead here otherwise if the next instruction is a |
| 267 | // cond branch it will be using a dead register. |
| 268 | BuildMI(BB&: MBB, I: std::next(x: II), MIMD: dl, MCID: TII.get(Opcode: SubOpc), DestReg: AVR::R29R28) |
| 269 | .addReg(RegNo: AVR::R29R28, Flags: RegState::Kill) |
| 270 | .addImm(Val: Offset - MaxOffset); |
| 271 | |
| 272 | Offset = MaxOffset; |
| 273 | } |
| 274 | |
| 275 | MI.getOperand(i: FIOperandNum).ChangeToRegister(Reg: AVR::R29R28, isDef: false); |
| 276 | assert(isUInt<6>(Offset) && "Offset is out of range" ); |
| 277 | MI.getOperand(i: FIOperandNum + 1).ChangeToImmediate(ImmVal: Offset); |
| 278 | return false; |
| 279 | } |
| 280 | |
| 281 | Register AVRRegisterInfo::getFrameRegister(const MachineFunction &MF) const { |
| 282 | const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering(); |
| 283 | if (TFI->hasFP(MF)) { |
| 284 | // The Y pointer register |
| 285 | return AVR::R28; |
| 286 | } |
| 287 | |
| 288 | return AVR::SP; |
| 289 | } |
| 290 | |
| 291 | const TargetRegisterClass * |
| 292 | AVRRegisterInfo::getPointerRegClass(unsigned Kind) const { |
| 293 | // FIXME: Currently we're using avr-gcc as reference, so we restrict |
| 294 | // ptrs to Y and Z regs. Though avr-gcc has buggy implementation |
| 295 | // of memory constraint, so we can fix it and bit avr-gcc here ;-) |
| 296 | return &AVR::PTRDISPREGSRegClass; |
| 297 | } |
| 298 | |
| 299 | void AVRRegisterInfo::splitReg(Register Reg, Register &LoReg, |
| 300 | Register &HiReg) const { |
| 301 | assert(AVR::DREGSRegClass.contains(Reg) && "can only split 16-bit registers" ); |
| 302 | |
| 303 | LoReg = getSubReg(Reg, Idx: AVR::sub_lo); |
| 304 | HiReg = getSubReg(Reg, Idx: AVR::sub_hi); |
| 305 | } |
| 306 | |
| 307 | bool AVRRegisterInfo::shouldCoalesce( |
| 308 | MachineInstr *MI, const TargetRegisterClass *SrcRC, unsigned SubReg, |
| 309 | const TargetRegisterClass *DstRC, unsigned DstSubReg, |
| 310 | const TargetRegisterClass *NewRC, LiveIntervals &LIS) const { |
| 311 | if (this->getRegClass(i: AVR::PTRDISPREGSRegClassID)->hasSubClassEq(RC: NewRC)) { |
| 312 | return false; |
| 313 | } |
| 314 | |
| 315 | return TargetRegisterInfo::shouldCoalesce(MI, SrcRC, SubReg, DstRC, DstSubReg, |
| 316 | NewRC, LIS); |
| 317 | } |
| 318 | |
| 319 | } // end of namespace llvm |
| 320 | |