| 1 | //===----------------------- SIFrameLowering.cpp --------------------------===// |
| 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 | #include "SIFrameLowering.h" |
| 10 | #include "AMDGPU.h" |
| 11 | #include "AMDGPULaneMaskUtils.h" |
| 12 | #include "GCNSubtarget.h" |
| 13 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
| 14 | #include "SIMachineFunctionInfo.h" |
| 15 | #include "llvm/CodeGen/LiveRegUnits.h" |
| 16 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 17 | #include "llvm/CodeGen/RegisterScavenging.h" |
| 18 | #include "llvm/Target/TargetMachine.h" |
| 19 | |
| 20 | using namespace llvm; |
| 21 | |
| 22 | #define DEBUG_TYPE "frame-info" |
| 23 | |
| 24 | static cl::opt<bool> EnableSpillVGPRToAGPR( |
| 25 | "amdgpu-spill-vgpr-to-agpr" , |
| 26 | cl::desc("Enable spilling VGPRs to AGPRs" ), |
| 27 | cl::ReallyHidden, |
| 28 | cl::init(Val: true)); |
| 29 | |
| 30 | // Find a register matching \p RC from \p LiveUnits which is unused and |
| 31 | // available throughout the function. On failure, returns AMDGPU::NoRegister. |
| 32 | // TODO: Rewrite the loop here to iterate over MCRegUnits instead of |
| 33 | // MCRegisters. This should reduce the number of iterations and avoid redundant |
| 34 | // checking. |
| 35 | static MCRegister findUnusedRegister(MachineRegisterInfo &MRI, |
| 36 | const LiveRegUnits &LiveUnits, |
| 37 | const TargetRegisterClass &RC) { |
| 38 | for (MCRegister Reg : RC) { |
| 39 | if (!MRI.isPhysRegUsed(PhysReg: Reg) && LiveUnits.available(Reg) && |
| 40 | !MRI.isReserved(PhysReg: Reg)) |
| 41 | return Reg; |
| 42 | } |
| 43 | return MCRegister(); |
| 44 | } |
| 45 | |
| 46 | // Find a scratch register that we can use in the prologue. We avoid using |
| 47 | // callee-save registers since they may appear to be free when this is called |
| 48 | // from canUseAsPrologue (during shrink wrapping), but then no longer be free |
| 49 | // when this is called from emitPrologue. |
| 50 | static MCRegister findScratchNonCalleeSaveRegister( |
| 51 | MachineRegisterInfo &MRI, LiveRegUnits &LiveUnits, |
| 52 | const TargetRegisterClass &RC, bool Unused = false) { |
| 53 | // Mark callee saved registers as used so we will not choose them. |
| 54 | const MCPhysReg *CSRegs = MRI.getCalleeSavedRegs(); |
| 55 | for (unsigned i = 0; CSRegs[i]; ++i) |
| 56 | LiveUnits.addReg(Reg: CSRegs[i]); |
| 57 | |
| 58 | // We are looking for a register that can be used throughout the entire |
| 59 | // function, so any use is unacceptable. |
| 60 | if (Unused) |
| 61 | return findUnusedRegister(MRI, LiveUnits, RC); |
| 62 | |
| 63 | for (MCRegister Reg : RC) { |
| 64 | if (LiveUnits.available(Reg) && !MRI.isReserved(PhysReg: Reg)) |
| 65 | return Reg; |
| 66 | } |
| 67 | |
| 68 | return MCRegister(); |
| 69 | } |
| 70 | |
| 71 | /// Query target location for spilling SGPRs |
| 72 | /// \p IncludeScratchCopy : Also look for free scratch SGPRs |
| 73 | static void getVGPRSpillLaneOrTempRegister( |
| 74 | MachineFunction &MF, LiveRegUnits &LiveUnits, Register SGPR, |
| 75 | const TargetRegisterClass &RC = AMDGPU::SReg_32_XM0_XEXECRegClass, |
| 76 | bool IncludeScratchCopy = true) { |
| 77 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 78 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 79 | |
| 80 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 81 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 82 | unsigned Size = TRI->getSpillSize(RC); |
| 83 | Align Alignment = TRI->getSpillAlign(RC); |
| 84 | |
| 85 | // We need to save and restore the given SGPR. |
| 86 | |
| 87 | Register ScratchSGPR; |
| 88 | // 1: Try to save the given register into an unused scratch SGPR. The |
| 89 | // LiveUnits should have all the callee saved registers marked as used. For |
| 90 | // certain cases we skip copy to scratch SGPR. |
| 91 | if (IncludeScratchCopy) |
| 92 | ScratchSGPR = findUnusedRegister(MRI&: MF.getRegInfo(), LiveUnits, RC); |
| 93 | |
| 94 | if (!ScratchSGPR) { |
| 95 | int FI = FrameInfo.CreateStackObject(Size, Alignment, isSpillSlot: true, Alloca: nullptr, |
| 96 | ID: TargetStackID::SGPRSpill); |
| 97 | |
| 98 | if (TRI->spillSGPRToVGPR() && |
| 99 | MFI->allocateSGPRSpillToVGPRLane(MF, FI, /*SpillToPhysVGPRLane=*/true, |
| 100 | /*IsPrologEpilog=*/true)) { |
| 101 | // 2: There's no free lane to spill, and no free register to save the |
| 102 | // SGPR, so we're forced to take another VGPR to use for the spill. |
| 103 | MFI->addToPrologEpilogSGPRSpills( |
| 104 | Reg: SGPR, SI: PrologEpilogSGPRSaveRestoreInfo( |
| 105 | SGPRSaveKind::SPILL_TO_VGPR_LANE, FI)); |
| 106 | |
| 107 | LLVM_DEBUG(auto Spill = MFI->getSGPRSpillToPhysicalVGPRLanes(FI).front(); |
| 108 | dbgs() << printReg(SGPR, TRI) << " requires fallback spill to " |
| 109 | << printReg(Spill.VGPR, TRI) << ':' << Spill.Lane |
| 110 | << '\n';); |
| 111 | } else { |
| 112 | // Remove dead <FI> index |
| 113 | MF.getFrameInfo().RemoveStackObject(ObjectIdx: FI); |
| 114 | // 3: If all else fails, spill the register to memory. |
| 115 | FI = FrameInfo.CreateSpillStackObject(Size, Alignment); |
| 116 | MFI->addToPrologEpilogSGPRSpills( |
| 117 | Reg: SGPR, |
| 118 | SI: PrologEpilogSGPRSaveRestoreInfo(SGPRSaveKind::SPILL_TO_MEM, FI)); |
| 119 | LLVM_DEBUG(dbgs() << "Reserved FI " << FI << " for spilling " |
| 120 | << printReg(SGPR, TRI) << '\n'); |
| 121 | } |
| 122 | } else { |
| 123 | MFI->addToPrologEpilogSGPRSpills( |
| 124 | Reg: SGPR, SI: PrologEpilogSGPRSaveRestoreInfo( |
| 125 | SGPRSaveKind::COPY_TO_SCRATCH_SGPR, ScratchSGPR)); |
| 126 | LiveUnits.addReg(Reg: ScratchSGPR); |
| 127 | LLVM_DEBUG(dbgs() << "Saving " << printReg(SGPR, TRI) << " with copy to " |
| 128 | << printReg(ScratchSGPR, TRI) << '\n'); |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | // We need to specially emit stack operations here because a different frame |
| 133 | // register is used than in the rest of the function, as getFrameRegister would |
| 134 | // use. |
| 135 | static void buildPrologSpill(const GCNSubtarget &ST, const SIRegisterInfo &TRI, |
| 136 | const SIMachineFunctionInfo &FuncInfo, |
| 137 | LiveRegUnits &LiveUnits, MachineFunction &MF, |
| 138 | MachineBasicBlock &MBB, |
| 139 | MachineBasicBlock::iterator I, const DebugLoc &DL, |
| 140 | Register SpillReg, int FI, Register FrameReg, |
| 141 | int64_t DwordOff = 0) { |
| 142 | unsigned Opc = ST.hasFlatScratchEnabled() ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
| 143 | : AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
| 144 | |
| 145 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 146 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF, FI); |
| 147 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 148 | PtrInfo, F: MachineMemOperand::MOStore, Size: FrameInfo.getObjectSize(ObjectIdx: FI), |
| 149 | BaseAlignment: FrameInfo.getObjectAlign(ObjectIdx: FI)); |
| 150 | LiveUnits.addReg(Reg: SpillReg); |
| 151 | bool IsKill = !MBB.isLiveIn(Reg: SpillReg); |
| 152 | TRI.buildSpillLoadStore(MBB, MI: I, DL, LoadStoreOp: Opc, Index: FI, ValueReg: SpillReg, ValueIsKill: IsKill, ScratchOffsetReg: FrameReg, |
| 153 | InstrOffset: DwordOff, MMO, RS: nullptr, LiveUnits: &LiveUnits); |
| 154 | if (IsKill) |
| 155 | LiveUnits.removeReg(Reg: SpillReg); |
| 156 | } |
| 157 | |
| 158 | static void buildEpilogRestore(const GCNSubtarget &ST, |
| 159 | const SIRegisterInfo &TRI, |
| 160 | const SIMachineFunctionInfo &FuncInfo, |
| 161 | LiveRegUnits &LiveUnits, MachineFunction &MF, |
| 162 | MachineBasicBlock &MBB, |
| 163 | MachineBasicBlock::iterator I, |
| 164 | const DebugLoc &DL, Register SpillReg, int FI, |
| 165 | Register FrameReg, int64_t DwordOff = 0) { |
| 166 | unsigned Opc = ST.hasFlatScratchEnabled() ? AMDGPU::SCRATCH_LOAD_DWORD_SADDR |
| 167 | : AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
| 168 | |
| 169 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 170 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF, FI); |
| 171 | MachineMemOperand *MMO = MF.getMachineMemOperand( |
| 172 | PtrInfo, F: MachineMemOperand::MOLoad, Size: FrameInfo.getObjectSize(ObjectIdx: FI), |
| 173 | BaseAlignment: FrameInfo.getObjectAlign(ObjectIdx: FI)); |
| 174 | TRI.buildSpillLoadStore(MBB, MI: I, DL, LoadStoreOp: Opc, Index: FI, ValueReg: SpillReg, ValueIsKill: false, ScratchOffsetReg: FrameReg, |
| 175 | InstrOffset: DwordOff, MMO, RS: nullptr, LiveUnits: &LiveUnits); |
| 176 | } |
| 177 | |
| 178 | static void buildGitPtr(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, |
| 179 | const DebugLoc &DL, const SIInstrInfo *TII, |
| 180 | Register TargetReg) { |
| 181 | MachineFunction *MF = MBB.getParent(); |
| 182 | const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
| 183 | const SIRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 184 | const MCInstrDesc &SMovB32 = TII->get(Opcode: AMDGPU::S_MOV_B32); |
| 185 | Register TargetLo = TRI->getSubReg(Reg: TargetReg, Idx: AMDGPU::sub0); |
| 186 | Register TargetHi = TRI->getSubReg(Reg: TargetReg, Idx: AMDGPU::sub1); |
| 187 | |
| 188 | if (MFI->getGITPtrHigh() != 0xffffffff) { |
| 189 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: TargetHi) |
| 190 | .addImm(Val: MFI->getGITPtrHigh()) |
| 191 | .addReg(RegNo: TargetReg, Flags: RegState::ImplicitDefine); |
| 192 | } else { |
| 193 | const MCInstrDesc &GetPC64 = TII->get(Opcode: AMDGPU::S_GETPC_B64_pseudo); |
| 194 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: GetPC64, DestReg: TargetReg); |
| 195 | } |
| 196 | Register GitPtrLo = MFI->getGITPtrLoReg(MF: *MF); |
| 197 | MF->getRegInfo().addLiveIn(Reg: GitPtrLo); |
| 198 | MBB.addLiveIn(PhysReg: GitPtrLo); |
| 199 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: TargetLo) |
| 200 | .addReg(RegNo: GitPtrLo); |
| 201 | } |
| 202 | |
| 203 | static void initLiveUnits(LiveRegUnits &LiveUnits, const SIRegisterInfo &TRI, |
| 204 | const SIMachineFunctionInfo *FuncInfo, |
| 205 | MachineFunction &MF, MachineBasicBlock &MBB, |
| 206 | MachineBasicBlock::iterator MBBI, bool IsProlog) { |
| 207 | if (LiveUnits.empty()) { |
| 208 | LiveUnits.init(TRI); |
| 209 | if (IsProlog) { |
| 210 | LiveUnits.addLiveIns(MBB); |
| 211 | } else { |
| 212 | // In epilog. |
| 213 | LiveUnits.addLiveOuts(MBB); |
| 214 | LiveUnits.stepBackward(MI: *MBBI); |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | namespace llvm { |
| 220 | |
| 221 | // SpillBuilder to save/restore special SGPR spills like the one needed for FP, |
| 222 | // BP, etc. These spills are delayed until the current function's frame is |
| 223 | // finalized. For a given register, the builder uses the |
| 224 | // PrologEpilogSGPRSaveRestoreInfo to decide the spill method. |
| 225 | class PrologEpilogSGPRSpillBuilder { |
| 226 | MachineBasicBlock::iterator MI; |
| 227 | MachineBasicBlock &MBB; |
| 228 | MachineFunction &MF; |
| 229 | const GCNSubtarget &ST; |
| 230 | MachineFrameInfo &MFI; |
| 231 | SIMachineFunctionInfo *FuncInfo; |
| 232 | const SIInstrInfo *TII; |
| 233 | const SIRegisterInfo &TRI; |
| 234 | Register SuperReg; |
| 235 | const PrologEpilogSGPRSaveRestoreInfo SI; |
| 236 | LiveRegUnits &LiveUnits; |
| 237 | const DebugLoc &DL; |
| 238 | Register FrameReg; |
| 239 | ArrayRef<int16_t> SplitParts; |
| 240 | unsigned NumSubRegs; |
| 241 | unsigned EltSize = 4; |
| 242 | |
| 243 | void saveToMemory(const int FI) const { |
| 244 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 245 | assert(!MFI.isDeadObjectIndex(FI)); |
| 246 | |
| 247 | initLiveUnits(LiveUnits, TRI, FuncInfo, MF, MBB, MBBI: MI, /*IsProlog*/ true); |
| 248 | |
| 249 | MCPhysReg TmpVGPR = findScratchNonCalleeSaveRegister( |
| 250 | MRI, LiveUnits, RC: AMDGPU::VGPR_32RegClass); |
| 251 | if (!TmpVGPR) |
| 252 | report_fatal_error(reason: "failed to find free scratch register" ); |
| 253 | |
| 254 | for (unsigned I = 0, DwordOff = 0; I < NumSubRegs; ++I) { |
| 255 | Register SubReg = NumSubRegs == 1 |
| 256 | ? SuperReg |
| 257 | : Register(TRI.getSubReg(Reg: SuperReg, Idx: SplitParts[I])); |
| 258 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32), DestReg: TmpVGPR) |
| 259 | .addReg(RegNo: SubReg); |
| 260 | |
| 261 | buildPrologSpill(ST, TRI, FuncInfo: *FuncInfo, LiveUnits, MF, MBB, I: MI, DL, SpillReg: TmpVGPR, |
| 262 | FI, FrameReg, DwordOff); |
| 263 | DwordOff += 4; |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | void saveToVGPRLane(const int FI) const { |
| 268 | assert(!MFI.isDeadObjectIndex(FI)); |
| 269 | |
| 270 | assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill); |
| 271 | ArrayRef<SIRegisterInfo::SpilledReg> Spill = |
| 272 | FuncInfo->getSGPRSpillToPhysicalVGPRLanes(FrameIndex: FI); |
| 273 | assert(Spill.size() == NumSubRegs); |
| 274 | |
| 275 | for (unsigned I = 0; I < NumSubRegs; ++I) { |
| 276 | Register SubReg = NumSubRegs == 1 |
| 277 | ? SuperReg |
| 278 | : Register(TRI.getSubReg(Reg: SuperReg, Idx: SplitParts[I])); |
| 279 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::SI_SPILL_S32_TO_VGPR), |
| 280 | DestReg: Spill[I].VGPR) |
| 281 | .addReg(RegNo: SubReg) |
| 282 | .addImm(Val: Spill[I].Lane) |
| 283 | .addReg(RegNo: Spill[I].VGPR, Flags: RegState::Undef); |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | void copyToScratchSGPR(Register DstReg) const { |
| 288 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: DstReg) |
| 289 | .addReg(RegNo: SuperReg) |
| 290 | .setMIFlag(MachineInstr::FrameSetup); |
| 291 | } |
| 292 | |
| 293 | void restoreFromMemory(const int FI) { |
| 294 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 295 | |
| 296 | initLiveUnits(LiveUnits, TRI, FuncInfo, MF, MBB, MBBI: MI, /*IsProlog*/ false); |
| 297 | MCPhysReg TmpVGPR = findScratchNonCalleeSaveRegister( |
| 298 | MRI, LiveUnits, RC: AMDGPU::VGPR_32RegClass); |
| 299 | if (!TmpVGPR) |
| 300 | report_fatal_error(reason: "failed to find free scratch register" ); |
| 301 | |
| 302 | for (unsigned I = 0, DwordOff = 0; I < NumSubRegs; ++I) { |
| 303 | Register SubReg = NumSubRegs == 1 |
| 304 | ? SuperReg |
| 305 | : Register(TRI.getSubReg(Reg: SuperReg, Idx: SplitParts[I])); |
| 306 | |
| 307 | buildEpilogRestore(ST, TRI, FuncInfo: *FuncInfo, LiveUnits, MF, MBB, I: MI, DL, |
| 308 | SpillReg: TmpVGPR, FI, FrameReg, DwordOff); |
| 309 | assert(SubReg.isPhysical()); |
| 310 | |
| 311 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_READFIRSTLANE_B32), DestReg: SubReg) |
| 312 | .addReg(RegNo: TmpVGPR, Flags: RegState::Kill); |
| 313 | DwordOff += 4; |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | void restoreFromVGPRLane(const int FI) { |
| 318 | assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill); |
| 319 | ArrayRef<SIRegisterInfo::SpilledReg> Spill = |
| 320 | FuncInfo->getSGPRSpillToPhysicalVGPRLanes(FrameIndex: FI); |
| 321 | assert(Spill.size() == NumSubRegs); |
| 322 | |
| 323 | for (unsigned I = 0; I < NumSubRegs; ++I) { |
| 324 | Register SubReg = NumSubRegs == 1 |
| 325 | ? SuperReg |
| 326 | : Register(TRI.getSubReg(Reg: SuperReg, Idx: SplitParts[I])); |
| 327 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::SI_RESTORE_S32_FROM_VGPR), DestReg: SubReg) |
| 328 | .addReg(RegNo: Spill[I].VGPR) |
| 329 | .addImm(Val: Spill[I].Lane); |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | void copyFromScratchSGPR(Register SrcReg) const { |
| 334 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: SuperReg) |
| 335 | .addReg(RegNo: SrcReg) |
| 336 | .setMIFlag(MachineInstr::FrameDestroy); |
| 337 | } |
| 338 | |
| 339 | public: |
| 340 | PrologEpilogSGPRSpillBuilder(Register Reg, |
| 341 | const PrologEpilogSGPRSaveRestoreInfo SI, |
| 342 | MachineBasicBlock &MBB, |
| 343 | MachineBasicBlock::iterator MI, |
| 344 | const DebugLoc &DL, const SIInstrInfo *TII, |
| 345 | const SIRegisterInfo &TRI, |
| 346 | LiveRegUnits &LiveUnits, Register FrameReg) |
| 347 | : MI(MI), MBB(MBB), MF(*MBB.getParent()), |
| 348 | ST(MF.getSubtarget<GCNSubtarget>()), MFI(MF.getFrameInfo()), |
| 349 | FuncInfo(MF.getInfo<SIMachineFunctionInfo>()), TII(TII), TRI(TRI), |
| 350 | SuperReg(Reg), SI(SI), LiveUnits(LiveUnits), DL(DL), |
| 351 | FrameReg(FrameReg) { |
| 352 | const TargetRegisterClass *RC = TRI.getPhysRegBaseClass(Reg: SuperReg); |
| 353 | SplitParts = TRI.getRegSplitParts(RC, EltSize); |
| 354 | NumSubRegs = SplitParts.empty() ? 1 : SplitParts.size(); |
| 355 | |
| 356 | assert(SuperReg != AMDGPU::M0 && "m0 should never spill" ); |
| 357 | } |
| 358 | |
| 359 | void save() { |
| 360 | switch (SI.getKind()) { |
| 361 | case SGPRSaveKind::SPILL_TO_MEM: |
| 362 | return saveToMemory(FI: SI.getIndex()); |
| 363 | case SGPRSaveKind::SPILL_TO_VGPR_LANE: |
| 364 | return saveToVGPRLane(FI: SI.getIndex()); |
| 365 | case SGPRSaveKind::COPY_TO_SCRATCH_SGPR: |
| 366 | return copyToScratchSGPR(DstReg: SI.getReg()); |
| 367 | } |
| 368 | } |
| 369 | |
| 370 | void restore() { |
| 371 | switch (SI.getKind()) { |
| 372 | case SGPRSaveKind::SPILL_TO_MEM: |
| 373 | return restoreFromMemory(FI: SI.getIndex()); |
| 374 | case SGPRSaveKind::SPILL_TO_VGPR_LANE: |
| 375 | return restoreFromVGPRLane(FI: SI.getIndex()); |
| 376 | case SGPRSaveKind::COPY_TO_SCRATCH_SGPR: |
| 377 | return copyFromScratchSGPR(SrcReg: SI.getReg()); |
| 378 | } |
| 379 | } |
| 380 | }; |
| 381 | |
| 382 | } // namespace llvm |
| 383 | |
| 384 | // Emit flat scratch setup code, assuming `MFI->hasFlatScratchInit()` |
| 385 | void SIFrameLowering::emitEntryFunctionFlatScratchInit( |
| 386 | MachineFunction &MF, MachineBasicBlock &MBB, MachineBasicBlock::iterator I, |
| 387 | const DebugLoc &DL, Register ScratchWaveOffsetReg) const { |
| 388 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 389 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 390 | const SIRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 391 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 392 | |
| 393 | // We don't need this if we only have spills since there is no user facing |
| 394 | // scratch. |
| 395 | |
| 396 | // TODO: If we know we don't have flat instructions earlier, we can omit |
| 397 | // this from the input registers. |
| 398 | // |
| 399 | // TODO: We only need to know if we access scratch space through a flat |
| 400 | // pointer. Because we only detect if flat instructions are used at all, |
| 401 | // this will be used more often than necessary on VI. |
| 402 | |
| 403 | Register FlatScrInitLo; |
| 404 | Register FlatScrInitHi; |
| 405 | |
| 406 | if (ST.isAmdPalOS()) { |
| 407 | // Extract the scratch offset from the descriptor in the GIT |
| 408 | LiveRegUnits LiveUnits; |
| 409 | LiveUnits.init(TRI: *TRI); |
| 410 | LiveUnits.addLiveIns(MBB); |
| 411 | |
| 412 | // Find unused reg to load flat scratch init into |
| 413 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 414 | Register FlatScrInit = AMDGPU::NoRegister; |
| 415 | ArrayRef<MCPhysReg> AllSGPR64s = TRI->getAllSGPR64(MF); |
| 416 | unsigned NumPreloaded = (MFI->getNumPreloadedSGPRs() + 1) / 2; |
| 417 | AllSGPR64s = AllSGPR64s.slice( |
| 418 | N: std::min(a: static_cast<unsigned>(AllSGPR64s.size()), b: NumPreloaded)); |
| 419 | Register GITPtrLoReg = MFI->getGITPtrLoReg(MF); |
| 420 | for (MCPhysReg Reg : AllSGPR64s) { |
| 421 | if (LiveUnits.available(Reg) && !MRI.isReserved(PhysReg: Reg) && |
| 422 | MRI.isAllocatable(PhysReg: Reg) && !TRI->isSubRegisterEq(RegA: Reg, RegB: GITPtrLoReg)) { |
| 423 | FlatScrInit = Reg; |
| 424 | break; |
| 425 | } |
| 426 | } |
| 427 | assert(FlatScrInit && "Failed to find free register for scratch init" ); |
| 428 | |
| 429 | FlatScrInitLo = TRI->getSubReg(Reg: FlatScrInit, Idx: AMDGPU::sub0); |
| 430 | FlatScrInitHi = TRI->getSubReg(Reg: FlatScrInit, Idx: AMDGPU::sub1); |
| 431 | |
| 432 | buildGitPtr(MBB, I, DL, TII, TargetReg: FlatScrInit); |
| 433 | |
| 434 | // We now have the GIT ptr - now get the scratch descriptor from the entry |
| 435 | // at offset 0 (or offset 16 for a compute shader). |
| 436 | MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS); |
| 437 | const MCInstrDesc &LoadDwordX2 = TII->get(Opcode: AMDGPU::S_LOAD_DWORDX2_IMM); |
| 438 | auto *MMO = MF.getMachineMemOperand( |
| 439 | PtrInfo, |
| 440 | F: MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant | |
| 441 | MachineMemOperand::MODereferenceable, |
| 442 | Size: 8, BaseAlignment: Align(4)); |
| 443 | unsigned Offset = |
| 444 | MF.getFunction().getCallingConv() == CallingConv::AMDGPU_CS ? 16 : 0; |
| 445 | const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>(); |
| 446 | unsigned EncodedOffset = AMDGPU::convertSMRDOffsetUnits(ST: Subtarget, ByteOffset: Offset); |
| 447 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: LoadDwordX2, DestReg: FlatScrInit) |
| 448 | .addReg(RegNo: FlatScrInit) |
| 449 | .addImm(Val: EncodedOffset) // offset |
| 450 | .addImm(Val: 0) // cpol |
| 451 | .addMemOperand(MMO); |
| 452 | |
| 453 | // Mask the offset in [47:0] of the descriptor |
| 454 | const MCInstrDesc &SAndB32 = TII->get(Opcode: AMDGPU::S_AND_B32); |
| 455 | auto And = BuildMI(BB&: MBB, I, MIMD: DL, MCID: SAndB32, DestReg: FlatScrInitHi) |
| 456 | .addReg(RegNo: FlatScrInitHi) |
| 457 | .addImm(Val: 0xffff); |
| 458 | And->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 459 | } else { |
| 460 | Register FlatScratchInitReg = |
| 461 | MFI->getPreloadedReg(Value: AMDGPUFunctionArgInfo::FLAT_SCRATCH_INIT); |
| 462 | assert(FlatScratchInitReg); |
| 463 | |
| 464 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 465 | MRI.addLiveIn(Reg: FlatScratchInitReg); |
| 466 | MBB.addLiveIn(PhysReg: FlatScratchInitReg); |
| 467 | |
| 468 | FlatScrInitLo = TRI->getSubReg(Reg: FlatScratchInitReg, Idx: AMDGPU::sub0); |
| 469 | FlatScrInitHi = TRI->getSubReg(Reg: FlatScratchInitReg, Idx: AMDGPU::sub1); |
| 470 | } |
| 471 | |
| 472 | // Do a 64-bit pointer add. |
| 473 | if (ST.flatScratchIsPointer()) { |
| 474 | if (ST.getGeneration() >= AMDGPUSubtarget::GFX10) { |
| 475 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_U32), DestReg: FlatScrInitLo) |
| 476 | .addReg(RegNo: FlatScrInitLo) |
| 477 | .addReg(RegNo: ScratchWaveOffsetReg); |
| 478 | auto Addc = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADDC_U32), |
| 479 | DestReg: FlatScrInitHi) |
| 480 | .addReg(RegNo: FlatScrInitHi) |
| 481 | .addImm(Val: 0); |
| 482 | Addc->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 483 | |
| 484 | using namespace AMDGPU::Hwreg; |
| 485 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_SETREG_B32)) |
| 486 | .addReg(RegNo: FlatScrInitLo) |
| 487 | .addImm(Val: int16_t(HwregEncoding::encode(Values: ID_FLAT_SCR_LO, Values: 0, Values: 32))); |
| 488 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_SETREG_B32)) |
| 489 | .addReg(RegNo: FlatScrInitHi) |
| 490 | .addImm(Val: int16_t(HwregEncoding::encode(Values: ID_FLAT_SCR_HI, Values: 0, Values: 32))); |
| 491 | return; |
| 492 | } |
| 493 | |
| 494 | // For GFX9. |
| 495 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_U32), DestReg: AMDGPU::FLAT_SCR_LO) |
| 496 | .addReg(RegNo: FlatScrInitLo) |
| 497 | .addReg(RegNo: ScratchWaveOffsetReg); |
| 498 | auto Addc = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADDC_U32), |
| 499 | DestReg: AMDGPU::FLAT_SCR_HI) |
| 500 | .addReg(RegNo: FlatScrInitHi) |
| 501 | .addImm(Val: 0); |
| 502 | Addc->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 503 | |
| 504 | return; |
| 505 | } |
| 506 | |
| 507 | assert(ST.getGeneration() < AMDGPUSubtarget::GFX9); |
| 508 | |
| 509 | // Copy the size in bytes. |
| 510 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: AMDGPU::FLAT_SCR_LO) |
| 511 | .addReg(RegNo: FlatScrInitHi, Flags: RegState::Kill); |
| 512 | |
| 513 | // Add wave offset in bytes to private base offset. |
| 514 | // See comment in AMDKernelCodeT.h for enable_sgpr_flat_scratch_init. |
| 515 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: FlatScrInitLo) |
| 516 | .addReg(RegNo: FlatScrInitLo) |
| 517 | .addReg(RegNo: ScratchWaveOffsetReg); |
| 518 | |
| 519 | // Convert offset to 256-byte units. |
| 520 | auto LShr = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_LSHR_B32), |
| 521 | DestReg: AMDGPU::FLAT_SCR_HI) |
| 522 | .addReg(RegNo: FlatScrInitLo, Flags: RegState::Kill) |
| 523 | .addImm(Val: 8); |
| 524 | LShr->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 525 | } |
| 526 | |
| 527 | // Note SGPRSpill stack IDs should only be used for SGPR spilling to VGPRs, not |
| 528 | // memory. They should have been removed by now. |
| 529 | static bool allStackObjectsAreDead(const MachineFrameInfo &MFI) { |
| 530 | for (int I = MFI.getObjectIndexBegin(), E = MFI.getObjectIndexEnd(); |
| 531 | I != E; ++I) { |
| 532 | if (!MFI.isDeadObjectIndex(ObjectIdx: I)) |
| 533 | return false; |
| 534 | } |
| 535 | |
| 536 | return true; |
| 537 | } |
| 538 | |
| 539 | // Shift down registers reserved for the scratch RSRC. |
| 540 | Register SIFrameLowering::getEntryFunctionReservedScratchRsrcReg( |
| 541 | MachineFunction &MF) const { |
| 542 | |
| 543 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 544 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 545 | const SIRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 546 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 547 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 548 | |
| 549 | assert(MFI->isEntryFunction()); |
| 550 | |
| 551 | Register ScratchRsrcReg = MFI->getScratchRSrcReg(); |
| 552 | |
| 553 | if (!ScratchRsrcReg || (!MRI.isPhysRegUsed(PhysReg: ScratchRsrcReg) && |
| 554 | allStackObjectsAreDead(MFI: MF.getFrameInfo()))) |
| 555 | return Register(); |
| 556 | |
| 557 | if (ST.hasSGPRInitBug() || |
| 558 | ScratchRsrcReg != TRI->reservedPrivateSegmentBufferReg(MF)) |
| 559 | return ScratchRsrcReg; |
| 560 | |
| 561 | // We reserved the last registers for this. Shift it down to the end of those |
| 562 | // which were actually used. |
| 563 | // |
| 564 | // FIXME: It might be safer to use a pseudoregister before replacement. |
| 565 | |
| 566 | // FIXME: We should be able to eliminate unused input registers. We only |
| 567 | // cannot do this for the resources required for scratch access. For now we |
| 568 | // skip over user SGPRs and may leave unused holes. |
| 569 | |
| 570 | unsigned NumPreloaded = (MFI->getNumPreloadedSGPRs() + 3) / 4; |
| 571 | ArrayRef<MCPhysReg> AllSGPR128s = TRI->getAllSGPR128(MF); |
| 572 | AllSGPR128s = AllSGPR128s.slice(N: std::min(a: static_cast<unsigned>(AllSGPR128s.size()), b: NumPreloaded)); |
| 573 | |
| 574 | // Skip the last N reserved elements because they should have already been |
| 575 | // reserved for VCC etc. |
| 576 | Register GITPtrLoReg = MFI->getGITPtrLoReg(MF); |
| 577 | for (MCPhysReg Reg : AllSGPR128s) { |
| 578 | // Pick the first unallocated one. Make sure we don't clobber the other |
| 579 | // reserved input we needed. Also for PAL, make sure we don't clobber |
| 580 | // the GIT pointer passed in SGPR0 or SGPR8. |
| 581 | if (!MRI.isPhysRegUsed(PhysReg: Reg) && MRI.isAllocatable(PhysReg: Reg) && |
| 582 | (!GITPtrLoReg || !TRI->isSubRegisterEq(RegA: Reg, RegB: GITPtrLoReg))) { |
| 583 | MRI.replaceRegWith(FromReg: ScratchRsrcReg, ToReg: Reg); |
| 584 | MFI->setScratchRSrcReg(Reg); |
| 585 | MRI.reserveReg(PhysReg: Reg, TRI); |
| 586 | return Reg; |
| 587 | } |
| 588 | } |
| 589 | |
| 590 | return ScratchRsrcReg; |
| 591 | } |
| 592 | |
| 593 | static unsigned getScratchScaleFactor(const GCNSubtarget &ST) { |
| 594 | return ST.hasFlatScratchEnabled() ? 1 : ST.getWavefrontSize(); |
| 595 | } |
| 596 | |
| 597 | void SIFrameLowering::emitEntryFunctionPrologue(MachineFunction &MF, |
| 598 | MachineBasicBlock &MBB) const { |
| 599 | assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported" ); |
| 600 | |
| 601 | // FIXME: If we only have SGPR spills, we won't actually be using scratch |
| 602 | // memory since these spill to VGPRs. We should be cleaning up these unused |
| 603 | // SGPR spill frame indices somewhere. |
| 604 | |
| 605 | // FIXME: We still have implicit uses on SGPR spill instructions in case they |
| 606 | // need to spill to vector memory. It's likely that will not happen, but at |
| 607 | // this point it appears we need the setup. This part of the prolog should be |
| 608 | // emitted after frame indices are eliminated. |
| 609 | |
| 610 | // FIXME: Remove all of the isPhysRegUsed checks |
| 611 | |
| 612 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 613 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 614 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 615 | const SIRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 616 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 617 | const Function &F = MF.getFunction(); |
| 618 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 619 | |
| 620 | assert(MFI->isEntryFunction()); |
| 621 | |
| 622 | Register PreloadedScratchWaveOffsetReg = MFI->getPreloadedReg( |
| 623 | Value: AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET); |
| 624 | |
| 625 | // We need to do the replacement of the private segment buffer register even |
| 626 | // if there are no stack objects. There could be stores to undef or a |
| 627 | // constant without an associated object. |
| 628 | // |
| 629 | // This will return `Register()` in cases where there are no actual |
| 630 | // uses of the SRSRC. |
| 631 | Register ScratchRsrcReg; |
| 632 | if (!ST.hasFlatScratchEnabled()) |
| 633 | ScratchRsrcReg = getEntryFunctionReservedScratchRsrcReg(MF); |
| 634 | |
| 635 | // Make the selected register live throughout the function. |
| 636 | if (ScratchRsrcReg) { |
| 637 | for (MachineBasicBlock &OtherBB : MF) { |
| 638 | if (&OtherBB != &MBB) { |
| 639 | OtherBB.addLiveIn(PhysReg: ScratchRsrcReg); |
| 640 | } |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | // Now that we have fixed the reserved SRSRC we need to locate the |
| 645 | // (potentially) preloaded SRSRC. |
| 646 | Register PreloadedScratchRsrcReg; |
| 647 | if (ST.isAmdHsaOrMesa(F)) { |
| 648 | PreloadedScratchRsrcReg = |
| 649 | MFI->getPreloadedReg(Value: AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_BUFFER); |
| 650 | if (ScratchRsrcReg && PreloadedScratchRsrcReg) { |
| 651 | // We added live-ins during argument lowering, but since they were not |
| 652 | // used they were deleted. We're adding the uses now, so add them back. |
| 653 | MRI.addLiveIn(Reg: PreloadedScratchRsrcReg); |
| 654 | MBB.addLiveIn(PhysReg: PreloadedScratchRsrcReg); |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | // Debug location must be unknown since the first debug location is used to |
| 659 | // determine the end of the prologue. |
| 660 | DebugLoc DL; |
| 661 | MachineBasicBlock::iterator I = MBB.begin(); |
| 662 | |
| 663 | // We found the SRSRC first because it needs four registers and has an |
| 664 | // alignment requirement. If the SRSRC that we found is clobbering with |
| 665 | // the scratch wave offset, which may be in a fixed SGPR or a free SGPR |
| 666 | // chosen by SITargetLowering::allocateSystemSGPRs, COPY the scratch |
| 667 | // wave offset to a free SGPR. |
| 668 | Register ScratchWaveOffsetReg; |
| 669 | if (PreloadedScratchWaveOffsetReg && |
| 670 | TRI->isSubRegisterEq(RegA: ScratchRsrcReg, RegB: PreloadedScratchWaveOffsetReg)) { |
| 671 | ArrayRef<MCPhysReg> AllSGPRs = TRI->getAllSGPR32(MF); |
| 672 | unsigned NumPreloaded = MFI->getNumPreloadedSGPRs(); |
| 673 | AllSGPRs = AllSGPRs.slice( |
| 674 | N: std::min(a: static_cast<unsigned>(AllSGPRs.size()), b: NumPreloaded)); |
| 675 | Register GITPtrLoReg = MFI->getGITPtrLoReg(MF); |
| 676 | for (MCPhysReg Reg : AllSGPRs) { |
| 677 | if (!MRI.isPhysRegUsed(PhysReg: Reg) && MRI.isAllocatable(PhysReg: Reg) && |
| 678 | !TRI->isSubRegisterEq(RegA: ScratchRsrcReg, RegB: Reg) && GITPtrLoReg != Reg) { |
| 679 | ScratchWaveOffsetReg = Reg; |
| 680 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: ScratchWaveOffsetReg) |
| 681 | .addReg(RegNo: PreloadedScratchWaveOffsetReg, Flags: RegState::Kill); |
| 682 | break; |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | // FIXME: We can spill incoming arguments and restore at the end of the |
| 687 | // prolog. |
| 688 | if (!ScratchWaveOffsetReg) |
| 689 | report_fatal_error( |
| 690 | reason: "could not find temporary scratch offset register in prolog" ); |
| 691 | } else { |
| 692 | ScratchWaveOffsetReg = PreloadedScratchWaveOffsetReg; |
| 693 | } |
| 694 | assert(ScratchWaveOffsetReg || !PreloadedScratchWaveOffsetReg); |
| 695 | |
| 696 | unsigned Offset = FrameInfo.getStackSize() * getScratchScaleFactor(ST); |
| 697 | if (!mayReserveScratchForCWSR(MF)) { |
| 698 | if (hasFP(MF)) { |
| 699 | Register FPReg = MFI->getFrameOffsetReg(); |
| 700 | assert(FPReg != AMDGPU::FP_REG); |
| 701 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: FPReg).addImm(Val: 0); |
| 702 | } |
| 703 | |
| 704 | if (requiresStackPointerReference(MF)) { |
| 705 | Register SPReg = MFI->getStackPtrOffsetReg(); |
| 706 | assert(SPReg != AMDGPU::SP_REG); |
| 707 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: SPReg).addImm(Val: Offset); |
| 708 | } |
| 709 | } else { |
| 710 | // We need to check if we're on a compute queue - if we are, then the CWSR |
| 711 | // trap handler may need to store some VGPRs on the stack. The first VGPR |
| 712 | // block is saved separately, so we only need to allocate space for any |
| 713 | // additional VGPR blocks used. For now, we will make sure there's enough |
| 714 | // room for the theoretical maximum number of VGPRs that can be allocated. |
| 715 | // FIXME: Figure out if the shader uses fewer VGPRs in practice. |
| 716 | assert(hasFP(MF)); |
| 717 | Register FPReg = MFI->getFrameOffsetReg(); |
| 718 | assert(FPReg != AMDGPU::FP_REG); |
| 719 | unsigned VGPRSize = llvm::alignTo( |
| 720 | Size: (ST.getAddressableNumVGPRs(DynamicVGPRBlockSize: MFI->getDynamicVGPRBlockSize()) - |
| 721 | AMDGPU::IsaInfo::getVGPRAllocGranule(STI: &ST, |
| 722 | DynamicVGPRBlockSize: MFI->getDynamicVGPRBlockSize())) * |
| 723 | 4, |
| 724 | A: FrameInfo.getMaxAlign()); |
| 725 | MFI->setScratchReservedForDynamicVGPRs(VGPRSize); |
| 726 | |
| 727 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::GET_STACK_BASE), DestReg: FPReg); |
| 728 | if (requiresStackPointerReference(MF)) { |
| 729 | Register SPReg = MFI->getStackPtrOffsetReg(); |
| 730 | assert(SPReg != AMDGPU::SP_REG); |
| 731 | |
| 732 | // If at least one of the constants can be inlined, then we can use |
| 733 | // s_cselect. Otherwise, use a mov and cmovk. |
| 734 | if (AMDGPU::isInlinableLiteral32(Literal: Offset, HasInv2Pi: ST.hasInv2PiInlineImm()) || |
| 735 | AMDGPU::isInlinableLiteral32(Literal: Offset + VGPRSize, |
| 736 | HasInv2Pi: ST.hasInv2PiInlineImm())) { |
| 737 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CSELECT_B32), DestReg: SPReg) |
| 738 | .addImm(Val: Offset + VGPRSize) |
| 739 | .addImm(Val: Offset); |
| 740 | } else { |
| 741 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: SPReg).addImm(Val: Offset); |
| 742 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CMOVK_I32), DestReg: SPReg) |
| 743 | .addImm(Val: Offset + VGPRSize); |
| 744 | } |
| 745 | } |
| 746 | } |
| 747 | |
| 748 | bool NeedsFlatScratchInit = |
| 749 | MFI->getUserSGPRInfo().hasFlatScratchInit() && |
| 750 | (MRI.isPhysRegUsed(PhysReg: AMDGPU::FLAT_SCR) || FrameInfo.hasCalls() || |
| 751 | (!allStackObjectsAreDead(MFI: FrameInfo) && ST.hasFlatScratchEnabled())); |
| 752 | |
| 753 | if ((NeedsFlatScratchInit || ScratchRsrcReg) && |
| 754 | PreloadedScratchWaveOffsetReg && !ST.hasArchitectedFlatScratch()) { |
| 755 | MRI.addLiveIn(Reg: PreloadedScratchWaveOffsetReg); |
| 756 | MBB.addLiveIn(PhysReg: PreloadedScratchWaveOffsetReg); |
| 757 | } |
| 758 | |
| 759 | if (NeedsFlatScratchInit) { |
| 760 | emitEntryFunctionFlatScratchInit(MF, MBB, I, DL, ScratchWaveOffsetReg); |
| 761 | } |
| 762 | |
| 763 | if (ScratchRsrcReg) { |
| 764 | emitEntryFunctionScratchRsrcRegSetup(MF, MBB, I, DL, |
| 765 | PreloadedPrivateBufferReg: PreloadedScratchRsrcReg, |
| 766 | ScratchRsrcReg, ScratchWaveOffsetReg); |
| 767 | } |
| 768 | |
| 769 | if (ST.hasWaitXcnt()) { |
| 770 | // Set REPLAY_MODE (bit 25) in MODE register to enable multi-group XNACK |
| 771 | // replay. This aligns hardware behavior with the compiler's s_wait_xcnt |
| 772 | // insertion logic, which assumes multi-group mode by default. |
| 773 | unsigned RegEncoding = |
| 774 | AMDGPU::Hwreg::HwregEncoding::encode(Values: AMDGPU::Hwreg::ID_MODE, Values: 25, Values: 1); |
| 775 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_SETREG_IMM32_B32)) |
| 776 | .addImm(Val: 1) |
| 777 | .addImm(Val: RegEncoding); |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | // Emit scratch RSRC setup code, assuming `ScratchRsrcReg != AMDGPU::NoReg` |
| 782 | void SIFrameLowering::emitEntryFunctionScratchRsrcRegSetup( |
| 783 | MachineFunction &MF, MachineBasicBlock &MBB, MachineBasicBlock::iterator I, |
| 784 | const DebugLoc &DL, Register PreloadedScratchRsrcReg, |
| 785 | Register ScratchRsrcReg, Register ScratchWaveOffsetReg) const { |
| 786 | |
| 787 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 788 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 789 | const SIRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 790 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 791 | const Function &Fn = MF.getFunction(); |
| 792 | |
| 793 | if (ST.isAmdPalOS()) { |
| 794 | // The pointer to the GIT is formed from the offset passed in and either |
| 795 | // the amdgpu-git-ptr-high function attribute or the top part of the PC |
| 796 | Register Rsrc01 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub0_sub1); |
| 797 | Register Rsrc03 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub3); |
| 798 | |
| 799 | buildGitPtr(MBB, I, DL, TII, TargetReg: Rsrc01); |
| 800 | |
| 801 | // We now have the GIT ptr - now get the scratch descriptor from the entry |
| 802 | // at offset 0 (or offset 16 for a compute shader). |
| 803 | MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS); |
| 804 | const MCInstrDesc &LoadDwordX4 = TII->get(Opcode: AMDGPU::S_LOAD_DWORDX4_IMM); |
| 805 | auto *MMO = MF.getMachineMemOperand( |
| 806 | PtrInfo, |
| 807 | F: MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant | |
| 808 | MachineMemOperand::MODereferenceable, |
| 809 | Size: 16, BaseAlignment: Align(4)); |
| 810 | unsigned Offset = Fn.getCallingConv() == CallingConv::AMDGPU_CS ? 16 : 0; |
| 811 | const GCNSubtarget &Subtarget = MF.getSubtarget<GCNSubtarget>(); |
| 812 | unsigned EncodedOffset = AMDGPU::convertSMRDOffsetUnits(ST: Subtarget, ByteOffset: Offset); |
| 813 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: LoadDwordX4, DestReg: ScratchRsrcReg) |
| 814 | .addReg(RegNo: Rsrc01) |
| 815 | .addImm(Val: EncodedOffset) // offset |
| 816 | .addImm(Val: 0) // cpol |
| 817 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine) |
| 818 | .addMemOperand(MMO); |
| 819 | |
| 820 | // The driver will always set the SRD for wave 64 (bits 118:117 of |
| 821 | // descriptor / bits 22:21 of third sub-reg will be 0b11) |
| 822 | // If the shader is actually wave32 we have to modify the const_index_stride |
| 823 | // field of the descriptor 3rd sub-reg (bits 22:21) to 0b10 (stride=32). The |
| 824 | // reason the driver does this is that there can be cases where it presents |
| 825 | // 2 shaders with different wave size (e.g. VsFs). |
| 826 | // TODO: convert to using SCRATCH instructions or multiple SRD buffers |
| 827 | if (ST.isWave32()) { |
| 828 | const MCInstrDesc &SBitsetB32 = TII->get(Opcode: AMDGPU::S_BITSET0_B32); |
| 829 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SBitsetB32, DestReg: Rsrc03) |
| 830 | .addImm(Val: 21) |
| 831 | .addReg(RegNo: Rsrc03); |
| 832 | } |
| 833 | } else if (ST.isMesaGfxShader(F: Fn) || !PreloadedScratchRsrcReg) { |
| 834 | assert(!ST.isAmdHsaOrMesa(Fn)); |
| 835 | const MCInstrDesc &SMovB32 = TII->get(Opcode: AMDGPU::S_MOV_B32); |
| 836 | |
| 837 | Register Rsrc2 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub2); |
| 838 | Register Rsrc3 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub3); |
| 839 | |
| 840 | // Use relocations to get the pointer, and setup the other bits manually. |
| 841 | uint64_t Rsrc23 = TII->getScratchRsrcWords23(); |
| 842 | |
| 843 | if (MFI->getUserSGPRInfo().hasImplicitBufferPtr()) { |
| 844 | Register Rsrc01 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub0_sub1); |
| 845 | |
| 846 | if (AMDGPU::isCompute(CC: MF.getFunction().getCallingConv())) { |
| 847 | const MCInstrDesc &Mov64 = TII->get(Opcode: AMDGPU::S_MOV_B64); |
| 848 | |
| 849 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: Mov64, DestReg: Rsrc01) |
| 850 | .addReg(RegNo: MFI->getImplicitBufferPtrUserSGPR()) |
| 851 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 852 | } else { |
| 853 | const MCInstrDesc &LoadDwordX2 = TII->get(Opcode: AMDGPU::S_LOAD_DWORDX2_IMM); |
| 854 | |
| 855 | MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS); |
| 856 | auto *MMO = MF.getMachineMemOperand( |
| 857 | PtrInfo, |
| 858 | F: MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant | |
| 859 | MachineMemOperand::MODereferenceable, |
| 860 | Size: 8, BaseAlignment: Align(4)); |
| 861 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: LoadDwordX2, DestReg: Rsrc01) |
| 862 | .addReg(RegNo: MFI->getImplicitBufferPtrUserSGPR()) |
| 863 | .addImm(Val: 0) // offset |
| 864 | .addImm(Val: 0) // cpol |
| 865 | .addMemOperand(MMO) |
| 866 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 867 | |
| 868 | MF.getRegInfo().addLiveIn(Reg: MFI->getImplicitBufferPtrUserSGPR()); |
| 869 | MBB.addLiveIn(PhysReg: MFI->getImplicitBufferPtrUserSGPR()); |
| 870 | } |
| 871 | } else { |
| 872 | Register Rsrc0 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub0); |
| 873 | Register Rsrc1 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub1); |
| 874 | |
| 875 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: Rsrc0) |
| 876 | .addExternalSymbol(FnName: "SCRATCH_RSRC_DWORD0" ) |
| 877 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 878 | |
| 879 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: Rsrc1) |
| 880 | .addExternalSymbol(FnName: "SCRATCH_RSRC_DWORD1" ) |
| 881 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 882 | } |
| 883 | |
| 884 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: Rsrc2) |
| 885 | .addImm(Val: Lo_32(Value: Rsrc23)) |
| 886 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 887 | |
| 888 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: SMovB32, DestReg: Rsrc3) |
| 889 | .addImm(Val: Hi_32(Value: Rsrc23)) |
| 890 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 891 | } else if (ST.isAmdHsaOrMesa(F: Fn)) { |
| 892 | assert(PreloadedScratchRsrcReg); |
| 893 | |
| 894 | if (ScratchRsrcReg != PreloadedScratchRsrcReg) { |
| 895 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: ScratchRsrcReg) |
| 896 | .addReg(RegNo: PreloadedScratchRsrcReg, Flags: RegState::Kill); |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | // Add the scratch wave offset into the scratch RSRC. |
| 901 | // |
| 902 | // We only want to update the first 48 bits, which is the base address |
| 903 | // pointer, without touching the adjacent 16 bits of flags. We know this add |
| 904 | // cannot carry-out from bit 47, otherwise the scratch allocation would be |
| 905 | // impossible to fit in the 48-bit global address space. |
| 906 | // |
| 907 | // TODO: Evaluate if it is better to just construct an SRD using the flat |
| 908 | // scratch init and some constants rather than update the one we are passed. |
| 909 | Register ScratchRsrcSub0 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub0); |
| 910 | Register ScratchRsrcSub1 = TRI->getSubReg(Reg: ScratchRsrcReg, Idx: AMDGPU::sub1); |
| 911 | |
| 912 | // We cannot Kill ScratchWaveOffsetReg here because we allow it to be used in |
| 913 | // the kernel body via inreg arguments. |
| 914 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_U32), DestReg: ScratchRsrcSub0) |
| 915 | .addReg(RegNo: ScratchRsrcSub0) |
| 916 | .addReg(RegNo: ScratchWaveOffsetReg) |
| 917 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 918 | auto Addc = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADDC_U32), DestReg: ScratchRsrcSub1) |
| 919 | .addReg(RegNo: ScratchRsrcSub1) |
| 920 | .addImm(Val: 0) |
| 921 | .addReg(RegNo: ScratchRsrcReg, Flags: RegState::ImplicitDefine); |
| 922 | Addc->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 923 | } |
| 924 | |
| 925 | bool SIFrameLowering::isSupportedStackID(TargetStackID::Value ID) const { |
| 926 | switch (ID) { |
| 927 | case TargetStackID::Default: |
| 928 | case TargetStackID::NoAlloc: |
| 929 | case TargetStackID::SGPRSpill: |
| 930 | return true; |
| 931 | case TargetStackID::ScalableVector: |
| 932 | case TargetStackID::ScalablePredicateVector: |
| 933 | case TargetStackID::WasmLocal: |
| 934 | return false; |
| 935 | } |
| 936 | llvm_unreachable("Invalid TargetStackID::Value" ); |
| 937 | } |
| 938 | |
| 939 | // Activate only the inactive lanes when \p EnableInactiveLanes is true. |
| 940 | // Otherwise, activate all lanes. It returns the saved exec. |
| 941 | static Register buildScratchExecCopy(LiveRegUnits &LiveUnits, |
| 942 | MachineFunction &MF, |
| 943 | MachineBasicBlock &MBB, |
| 944 | MachineBasicBlock::iterator MBBI, |
| 945 | const DebugLoc &DL, bool IsProlog, |
| 946 | bool EnableInactiveLanes) { |
| 947 | Register ScratchExecCopy; |
| 948 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 949 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 950 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 951 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 952 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 953 | |
| 954 | initLiveUnits(LiveUnits, TRI, FuncInfo, MF, MBB, MBBI, IsProlog); |
| 955 | |
| 956 | if (FuncInfo->isWholeWaveFunction()) { |
| 957 | // Whole wave functions already have a copy of the original EXEC mask that |
| 958 | // we can use. |
| 959 | assert(IsProlog && "Epilog should look at return, not setup" ); |
| 960 | ScratchExecCopy = |
| 961 | TII->getWholeWaveFunctionSetup(MF)->getOperand(i: 0).getReg(); |
| 962 | assert(ScratchExecCopy && "Couldn't find copy of EXEC" ); |
| 963 | } else { |
| 964 | ScratchExecCopy = findScratchNonCalleeSaveRegister( |
| 965 | MRI, LiveUnits, RC: *TRI.getWaveMaskRegClass()); |
| 966 | } |
| 967 | |
| 968 | if (!ScratchExecCopy) |
| 969 | report_fatal_error(reason: "failed to find free scratch register" ); |
| 970 | |
| 971 | LiveUnits.addReg(Reg: ScratchExecCopy); |
| 972 | |
| 973 | const unsigned SaveExecOpc = |
| 974 | ST.isWave32() ? (EnableInactiveLanes ? AMDGPU::S_XOR_SAVEEXEC_B32 |
| 975 | : AMDGPU::S_OR_SAVEEXEC_B32) |
| 976 | : (EnableInactiveLanes ? AMDGPU::S_XOR_SAVEEXEC_B64 |
| 977 | : AMDGPU::S_OR_SAVEEXEC_B64); |
| 978 | auto SaveExec = |
| 979 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: SaveExecOpc), DestReg: ScratchExecCopy).addImm(Val: -1); |
| 980 | SaveExec->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 981 | |
| 982 | return ScratchExecCopy; |
| 983 | } |
| 984 | |
| 985 | void SIFrameLowering::emitCSRSpillStores( |
| 986 | MachineFunction &MF, MachineBasicBlock &MBB, |
| 987 | MachineBasicBlock::iterator MBBI, DebugLoc &DL, LiveRegUnits &LiveUnits, |
| 988 | Register FrameReg, Register FramePtrRegScratchCopy) const { |
| 989 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 990 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 991 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 992 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 993 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 994 | const AMDGPU::LaneMaskConstants &LMC = AMDGPU::LaneMaskConstants::get(ST); |
| 995 | |
| 996 | // Spill Whole-Wave Mode VGPRs. Save only the inactive lanes of the scratch |
| 997 | // registers. However, save all lanes of callee-saved VGPRs. Due to this, we |
| 998 | // might end up flipping the EXEC bits twice. |
| 999 | Register ScratchExecCopy; |
| 1000 | SmallVector<std::pair<Register, int>, 2> WWMCalleeSavedRegs, WWMScratchRegs; |
| 1001 | FuncInfo->splitWWMSpillRegisters(MF, CalleeSavedRegs&: WWMCalleeSavedRegs, ScratchRegs&: WWMScratchRegs); |
| 1002 | if (!WWMScratchRegs.empty()) |
| 1003 | ScratchExecCopy = |
| 1004 | buildScratchExecCopy(LiveUnits, MF, MBB, MBBI, DL, |
| 1005 | /*IsProlog*/ true, /*EnableInactiveLanes*/ true); |
| 1006 | |
| 1007 | auto StoreWWMRegisters = |
| 1008 | [&](SmallVectorImpl<std::pair<Register, int>> &WWMRegs) { |
| 1009 | for (const auto &Reg : WWMRegs) { |
| 1010 | Register VGPR = Reg.first; |
| 1011 | int FI = Reg.second; |
| 1012 | buildPrologSpill(ST, TRI, FuncInfo: *FuncInfo, LiveUnits, MF, MBB, I: MBBI, DL, |
| 1013 | SpillReg: VGPR, FI, FrameReg); |
| 1014 | } |
| 1015 | }; |
| 1016 | |
| 1017 | for (const Register Reg : make_first_range(c&: WWMScratchRegs)) { |
| 1018 | if (!MRI.isReserved(PhysReg: Reg)) { |
| 1019 | MRI.addLiveIn(Reg); |
| 1020 | MBB.addLiveIn(PhysReg: Reg); |
| 1021 | } |
| 1022 | } |
| 1023 | StoreWWMRegisters(WWMScratchRegs); |
| 1024 | |
| 1025 | auto EnableAllLanes = [&]() { |
| 1026 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.MovOpc), DestReg: LMC.ExecReg).addImm(Val: -1); |
| 1027 | }; |
| 1028 | |
| 1029 | if (!WWMCalleeSavedRegs.empty()) { |
| 1030 | if (ScratchExecCopy) { |
| 1031 | EnableAllLanes(); |
| 1032 | } else { |
| 1033 | ScratchExecCopy = buildScratchExecCopy(LiveUnits, MF, MBB, MBBI, DL, |
| 1034 | /*IsProlog*/ true, |
| 1035 | /*EnableInactiveLanes*/ false); |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | StoreWWMRegisters(WWMCalleeSavedRegs); |
| 1040 | if (FuncInfo->isWholeWaveFunction()) { |
| 1041 | // If we have already saved some WWM CSR registers, then the EXEC is already |
| 1042 | // -1 and we don't need to do anything else. Otherwise, set EXEC to -1 here. |
| 1043 | if (!ScratchExecCopy) |
| 1044 | buildScratchExecCopy(LiveUnits, MF, MBB, MBBI, DL, /*IsProlog*/ true, |
| 1045 | /*EnableInactiveLanes*/ true); |
| 1046 | else if (WWMCalleeSavedRegs.empty()) |
| 1047 | EnableAllLanes(); |
| 1048 | } else if (ScratchExecCopy) { |
| 1049 | // FIXME: Split block and make terminator. |
| 1050 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.MovOpc), DestReg: LMC.ExecReg) |
| 1051 | .addReg(RegNo: ScratchExecCopy, Flags: RegState::Kill); |
| 1052 | LiveUnits.addReg(Reg: ScratchExecCopy); |
| 1053 | } |
| 1054 | |
| 1055 | Register FramePtrReg = FuncInfo->getFrameOffsetReg(); |
| 1056 | |
| 1057 | for (const auto &Spill : FuncInfo->getPrologEpilogSGPRSpills()) { |
| 1058 | // Special handle FP spill: |
| 1059 | // Skip if FP is saved to a scratch SGPR, the save has already been emitted. |
| 1060 | // Otherwise, FP has been moved to a temporary register and spill it |
| 1061 | // instead. |
| 1062 | Register Reg = |
| 1063 | Spill.first == FramePtrReg ? FramePtrRegScratchCopy : Spill.first; |
| 1064 | if (!Reg) |
| 1065 | continue; |
| 1066 | |
| 1067 | PrologEpilogSGPRSpillBuilder SB(Reg, Spill.second, MBB, MBBI, DL, TII, TRI, |
| 1068 | LiveUnits, FrameReg); |
| 1069 | SB.save(); |
| 1070 | } |
| 1071 | |
| 1072 | // If a copy to scratch SGPR has been chosen for any of the SGPR spills, make |
| 1073 | // such scratch registers live throughout the function. |
| 1074 | SmallVector<Register, 1> ScratchSGPRs; |
| 1075 | FuncInfo->getAllScratchSGPRCopyDstRegs(Regs&: ScratchSGPRs); |
| 1076 | if (!ScratchSGPRs.empty()) { |
| 1077 | for (MachineBasicBlock &MBB : MF) { |
| 1078 | for (MCPhysReg Reg : ScratchSGPRs) |
| 1079 | MBB.addLiveIn(PhysReg: Reg); |
| 1080 | |
| 1081 | MBB.sortUniqueLiveIns(); |
| 1082 | } |
| 1083 | if (!LiveUnits.empty()) { |
| 1084 | for (MCPhysReg Reg : ScratchSGPRs) |
| 1085 | LiveUnits.addReg(Reg); |
| 1086 | } |
| 1087 | } |
| 1088 | } |
| 1089 | |
| 1090 | void SIFrameLowering::emitCSRSpillRestores( |
| 1091 | MachineFunction &MF, MachineBasicBlock &MBB, |
| 1092 | MachineBasicBlock::iterator MBBI, DebugLoc &DL, LiveRegUnits &LiveUnits, |
| 1093 | Register FrameReg, Register FramePtrRegScratchCopy) const { |
| 1094 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1095 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1096 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1097 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 1098 | const AMDGPU::LaneMaskConstants &LMC = AMDGPU::LaneMaskConstants::get(ST); |
| 1099 | Register FramePtrReg = FuncInfo->getFrameOffsetReg(); |
| 1100 | |
| 1101 | for (const auto &Spill : FuncInfo->getPrologEpilogSGPRSpills()) { |
| 1102 | // Special handle FP restore: |
| 1103 | // Skip if FP needs to be restored from the scratch SGPR. Otherwise, restore |
| 1104 | // the FP value to a temporary register. The frame pointer should be |
| 1105 | // overwritten only at the end when all other spills are restored from |
| 1106 | // current frame. |
| 1107 | Register Reg = |
| 1108 | Spill.first == FramePtrReg ? FramePtrRegScratchCopy : Spill.first; |
| 1109 | if (!Reg) |
| 1110 | continue; |
| 1111 | |
| 1112 | PrologEpilogSGPRSpillBuilder SB(Reg, Spill.second, MBB, MBBI, DL, TII, TRI, |
| 1113 | LiveUnits, FrameReg); |
| 1114 | SB.restore(); |
| 1115 | } |
| 1116 | |
| 1117 | // Restore Whole-Wave Mode VGPRs. Restore only the inactive lanes of the |
| 1118 | // scratch registers. However, restore all lanes of callee-saved VGPRs. Due to |
| 1119 | // this, we might end up flipping the EXEC bits twice. |
| 1120 | Register ScratchExecCopy; |
| 1121 | SmallVector<std::pair<Register, int>, 2> WWMCalleeSavedRegs, WWMScratchRegs; |
| 1122 | FuncInfo->splitWWMSpillRegisters(MF, CalleeSavedRegs&: WWMCalleeSavedRegs, ScratchRegs&: WWMScratchRegs); |
| 1123 | auto RestoreWWMRegisters = |
| 1124 | [&](SmallVectorImpl<std::pair<Register, int>> &WWMRegs) { |
| 1125 | for (const auto &Reg : WWMRegs) { |
| 1126 | Register VGPR = Reg.first; |
| 1127 | int FI = Reg.second; |
| 1128 | buildEpilogRestore(ST, TRI, FuncInfo: *FuncInfo, LiveUnits, MF, MBB, I: MBBI, DL, |
| 1129 | SpillReg: VGPR, FI, FrameReg); |
| 1130 | } |
| 1131 | }; |
| 1132 | |
| 1133 | if (FuncInfo->isWholeWaveFunction()) { |
| 1134 | // For whole wave functions, the EXEC is already -1 at this point. |
| 1135 | // Therefore, we can restore the CSR WWM registers right away. |
| 1136 | RestoreWWMRegisters(WWMCalleeSavedRegs); |
| 1137 | |
| 1138 | // The original EXEC is the first operand of the return instruction. |
| 1139 | MachineInstr &Return = MBB.instr_back(); |
| 1140 | unsigned Opcode = Return.getOpcode(); |
| 1141 | switch (Opcode) { |
| 1142 | case AMDGPU::SI_WHOLE_WAVE_FUNC_RETURN: |
| 1143 | Opcode = AMDGPU::SI_RETURN; |
| 1144 | break; |
| 1145 | case AMDGPU::SI_TCRETURN_GFX_WholeWave: |
| 1146 | Opcode = AMDGPU::SI_TCRETURN_GFX; |
| 1147 | break; |
| 1148 | default: |
| 1149 | llvm_unreachable("Unexpected return inst" ); |
| 1150 | } |
| 1151 | Register OrigExec = Return.getOperand(i: 0).getReg(); |
| 1152 | |
| 1153 | if (!WWMScratchRegs.empty()) { |
| 1154 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.XorOpc), DestReg: LMC.ExecReg) |
| 1155 | .addReg(RegNo: OrigExec) |
| 1156 | .addImm(Val: -1); |
| 1157 | RestoreWWMRegisters(WWMScratchRegs); |
| 1158 | } |
| 1159 | |
| 1160 | // Restore original EXEC. |
| 1161 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.MovOpc), DestReg: LMC.ExecReg).addReg(RegNo: OrigExec); |
| 1162 | |
| 1163 | // Drop the first operand and update the opcode. |
| 1164 | Return.removeOperand(OpNo: 0); |
| 1165 | Return.setDesc(TII->get(Opcode)); |
| 1166 | |
| 1167 | return; |
| 1168 | } |
| 1169 | |
| 1170 | if (!WWMScratchRegs.empty()) { |
| 1171 | ScratchExecCopy = |
| 1172 | buildScratchExecCopy(LiveUnits, MF, MBB, MBBI, DL, |
| 1173 | /*IsProlog=*/false, /*EnableInactiveLanes=*/true); |
| 1174 | } |
| 1175 | RestoreWWMRegisters(WWMScratchRegs); |
| 1176 | if (!WWMCalleeSavedRegs.empty()) { |
| 1177 | if (ScratchExecCopy) { |
| 1178 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.MovOpc), DestReg: LMC.ExecReg).addImm(Val: -1); |
| 1179 | } else { |
| 1180 | ScratchExecCopy = buildScratchExecCopy(LiveUnits, MF, MBB, MBBI, DL, |
| 1181 | /*IsProlog*/ false, |
| 1182 | /*EnableInactiveLanes*/ false); |
| 1183 | } |
| 1184 | } |
| 1185 | |
| 1186 | RestoreWWMRegisters(WWMCalleeSavedRegs); |
| 1187 | if (ScratchExecCopy) { |
| 1188 | // FIXME: Split block and make terminator. |
| 1189 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: LMC.MovOpc), DestReg: LMC.ExecReg) |
| 1190 | .addReg(RegNo: ScratchExecCopy, Flags: RegState::Kill); |
| 1191 | } |
| 1192 | } |
| 1193 | |
| 1194 | void SIFrameLowering::emitPrologue(MachineFunction &MF, |
| 1195 | MachineBasicBlock &MBB) const { |
| 1196 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1197 | if (FuncInfo->isEntryFunction()) { |
| 1198 | emitEntryFunctionPrologue(MF, MBB); |
| 1199 | return; |
| 1200 | } |
| 1201 | |
| 1202 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1203 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1204 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1205 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 1206 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1207 | |
| 1208 | Register StackPtrReg = FuncInfo->getStackPtrOffsetReg(); |
| 1209 | Register FramePtrReg = FuncInfo->getFrameOffsetReg(); |
| 1210 | Register BasePtrReg = |
| 1211 | TRI.hasBasePointer(MF) ? TRI.getBaseRegister() : Register(); |
| 1212 | LiveRegUnits LiveUnits; |
| 1213 | |
| 1214 | MachineBasicBlock::iterator MBBI = MBB.begin(); |
| 1215 | // DebugLoc must be unknown since the first instruction with DebugLoc is used |
| 1216 | // to determine the end of the prologue. |
| 1217 | DebugLoc DL; |
| 1218 | |
| 1219 | if (FuncInfo->isChainFunction()) { |
| 1220 | // Functions with the amdgpu_cs_chain[_preserve] CC don't receive a SP, but |
| 1221 | // are free to set one up if they need it. |
| 1222 | bool UseSP = requiresStackPointerReference(MF); |
| 1223 | if (UseSP) { |
| 1224 | assert(StackPtrReg != AMDGPU::SP_REG); |
| 1225 | |
| 1226 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: StackPtrReg) |
| 1227 | .addImm(Val: MFI.getStackSize() * getScratchScaleFactor(ST)); |
| 1228 | } |
| 1229 | } |
| 1230 | |
| 1231 | bool HasFP = false; |
| 1232 | bool HasBP = false; |
| 1233 | uint32_t NumBytes = MFI.getStackSize(); |
| 1234 | uint32_t RoundedSize = NumBytes; |
| 1235 | |
| 1236 | if (TRI.hasStackRealignment(MF)) |
| 1237 | HasFP = true; |
| 1238 | |
| 1239 | Register FramePtrRegScratchCopy; |
| 1240 | if (!HasFP && !hasFP(MF)) { |
| 1241 | // Emit the CSR spill stores with SP base register. |
| 1242 | emitCSRSpillStores(MF, MBB, MBBI, DL, LiveUnits, |
| 1243 | FrameReg: FuncInfo->isChainFunction() ? Register() : StackPtrReg, |
| 1244 | FramePtrRegScratchCopy); |
| 1245 | } else { |
| 1246 | // CSR spill stores will use FP as base register. |
| 1247 | Register SGPRForFPSaveRestoreCopy = |
| 1248 | FuncInfo->getScratchSGPRCopyDstReg(Reg: FramePtrReg); |
| 1249 | |
| 1250 | initLiveUnits(LiveUnits, TRI, FuncInfo, MF, MBB, MBBI, /*IsProlog*/ true); |
| 1251 | if (SGPRForFPSaveRestoreCopy) { |
| 1252 | // Copy FP to the scratch register now and emit the CFI entry. It avoids |
| 1253 | // the extra FP copy needed in the other two cases when FP is spilled to |
| 1254 | // memory or to a VGPR lane. |
| 1255 | PrologEpilogSGPRSpillBuilder SB( |
| 1256 | FramePtrReg, |
| 1257 | FuncInfo->getPrologEpilogSGPRSaveRestoreInfo(Reg: FramePtrReg), MBB, MBBI, |
| 1258 | DL, TII, TRI, LiveUnits, FramePtrReg); |
| 1259 | SB.save(); |
| 1260 | LiveUnits.addReg(Reg: SGPRForFPSaveRestoreCopy); |
| 1261 | } else { |
| 1262 | // Copy FP into a new scratch register so that its previous value can be |
| 1263 | // spilled after setting up the new frame. |
| 1264 | FramePtrRegScratchCopy = findScratchNonCalleeSaveRegister( |
| 1265 | MRI, LiveUnits, RC: AMDGPU::SReg_32_XM0_XEXECRegClass); |
| 1266 | if (!FramePtrRegScratchCopy) |
| 1267 | report_fatal_error(reason: "failed to find free scratch register" ); |
| 1268 | |
| 1269 | LiveUnits.addReg(Reg: FramePtrRegScratchCopy); |
| 1270 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: FramePtrRegScratchCopy) |
| 1271 | .addReg(RegNo: FramePtrReg); |
| 1272 | } |
| 1273 | } |
| 1274 | |
| 1275 | if (HasFP) { |
| 1276 | const unsigned Alignment = MFI.getMaxAlign().value(); |
| 1277 | |
| 1278 | RoundedSize += Alignment; |
| 1279 | if (LiveUnits.empty()) { |
| 1280 | LiveUnits.init(TRI); |
| 1281 | LiveUnits.addLiveIns(MBB); |
| 1282 | } |
| 1283 | |
| 1284 | // s_add_i32 s33, s32, NumBytes |
| 1285 | // s_and_b32 s33, s33, 0b111...0000 |
| 1286 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: FramePtrReg) |
| 1287 | .addReg(RegNo: StackPtrReg) |
| 1288 | .addImm(Val: (Alignment - 1) * getScratchScaleFactor(ST)) |
| 1289 | .setMIFlag(MachineInstr::FrameSetup); |
| 1290 | auto And = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_AND_B32), DestReg: FramePtrReg) |
| 1291 | .addReg(RegNo: FramePtrReg, Flags: RegState::Kill) |
| 1292 | .addImm(Val: -Alignment * getScratchScaleFactor(ST)) |
| 1293 | .setMIFlag(MachineInstr::FrameSetup); |
| 1294 | And->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 1295 | FuncInfo->setIsStackRealigned(true); |
| 1296 | } else if ((HasFP = hasFP(MF))) { |
| 1297 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: FramePtrReg) |
| 1298 | .addReg(RegNo: StackPtrReg) |
| 1299 | .setMIFlag(MachineInstr::FrameSetup); |
| 1300 | } |
| 1301 | |
| 1302 | // If FP is used, emit the CSR spills with FP base register. |
| 1303 | if (HasFP) { |
| 1304 | emitCSRSpillStores(MF, MBB, MBBI, DL, LiveUnits, FrameReg: FramePtrReg, |
| 1305 | FramePtrRegScratchCopy); |
| 1306 | if (FramePtrRegScratchCopy) |
| 1307 | LiveUnits.removeReg(Reg: FramePtrRegScratchCopy); |
| 1308 | } |
| 1309 | |
| 1310 | // If we need a base pointer, set it up here. It's whatever the value of |
| 1311 | // the stack pointer is at this point. Any variable size objects will be |
| 1312 | // allocated after this, so we can still use the base pointer to reference |
| 1313 | // the incoming arguments. |
| 1314 | if ((HasBP = TRI.hasBasePointer(MF))) { |
| 1315 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: BasePtrReg) |
| 1316 | .addReg(RegNo: StackPtrReg) |
| 1317 | .setMIFlag(MachineInstr::FrameSetup); |
| 1318 | } |
| 1319 | |
| 1320 | if (HasFP && RoundedSize != 0) { |
| 1321 | auto Add = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: StackPtrReg) |
| 1322 | .addReg(RegNo: StackPtrReg) |
| 1323 | .addImm(Val: RoundedSize * getScratchScaleFactor(ST)) |
| 1324 | .setMIFlag(MachineInstr::FrameSetup); |
| 1325 | Add->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 1326 | } |
| 1327 | |
| 1328 | bool FPSaved = FuncInfo->hasPrologEpilogSGPRSpillEntry(Reg: FramePtrReg); |
| 1329 | (void)FPSaved; |
| 1330 | assert((!HasFP || FPSaved) && |
| 1331 | "Needed to save FP but didn't save it anywhere" ); |
| 1332 | |
| 1333 | // If we allow spilling to AGPRs we may have saved FP but then spill |
| 1334 | // everything into AGPRs instead of the stack. |
| 1335 | assert((HasFP || !FPSaved || EnableSpillVGPRToAGPR) && |
| 1336 | "Saved FP but didn't need it" ); |
| 1337 | |
| 1338 | bool BPSaved = FuncInfo->hasPrologEpilogSGPRSpillEntry(Reg: BasePtrReg); |
| 1339 | (void)BPSaved; |
| 1340 | assert((!HasBP || BPSaved) && |
| 1341 | "Needed to save BP but didn't save it anywhere" ); |
| 1342 | |
| 1343 | assert((HasBP || !BPSaved) && "Saved BP but didn't need it" ); |
| 1344 | |
| 1345 | if (FuncInfo->isWholeWaveFunction()) { |
| 1346 | // SI_WHOLE_WAVE_FUNC_SETUP has outlived its purpose. |
| 1347 | TII->getWholeWaveFunctionSetup(MF)->eraseFromParent(); |
| 1348 | } |
| 1349 | } |
| 1350 | |
| 1351 | void SIFrameLowering::emitEpilogue(MachineFunction &MF, |
| 1352 | MachineBasicBlock &MBB) const { |
| 1353 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1354 | if (FuncInfo->isEntryFunction()) |
| 1355 | return; |
| 1356 | |
| 1357 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1358 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1359 | const SIRegisterInfo &TRI = TII->getRegisterInfo(); |
| 1360 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1361 | LiveRegUnits LiveUnits; |
| 1362 | // Get the insert location for the epilogue. If there were no terminators in |
| 1363 | // the block, get the last instruction. |
| 1364 | MachineBasicBlock::iterator MBBI = MBB.end(); |
| 1365 | DebugLoc DL; |
| 1366 | if (!MBB.empty()) { |
| 1367 | MBBI = MBB.getLastNonDebugInstr(); |
| 1368 | if (MBBI != MBB.end()) |
| 1369 | DL = MBBI->getDebugLoc(); |
| 1370 | |
| 1371 | MBBI = MBB.getFirstTerminator(); |
| 1372 | } |
| 1373 | |
| 1374 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1375 | uint32_t NumBytes = MFI.getStackSize(); |
| 1376 | uint32_t RoundedSize = FuncInfo->isStackRealigned() |
| 1377 | ? NumBytes + MFI.getMaxAlign().value() |
| 1378 | : NumBytes; |
| 1379 | const Register StackPtrReg = FuncInfo->getStackPtrOffsetReg(); |
| 1380 | Register FramePtrReg = FuncInfo->getFrameOffsetReg(); |
| 1381 | bool FPSaved = FuncInfo->hasPrologEpilogSGPRSpillEntry(Reg: FramePtrReg); |
| 1382 | |
| 1383 | if (RoundedSize != 0) { |
| 1384 | if (TRI.hasBasePointer(MF)) { |
| 1385 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: StackPtrReg) |
| 1386 | .addReg(RegNo: TRI.getBaseRegister()) |
| 1387 | .setMIFlag(MachineInstr::FrameDestroy); |
| 1388 | } else if (hasFP(MF)) { |
| 1389 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: StackPtrReg) |
| 1390 | .addReg(RegNo: FramePtrReg) |
| 1391 | .setMIFlag(MachineInstr::FrameDestroy); |
| 1392 | } |
| 1393 | } |
| 1394 | |
| 1395 | Register FramePtrRegScratchCopy; |
| 1396 | Register SGPRForFPSaveRestoreCopy = |
| 1397 | FuncInfo->getScratchSGPRCopyDstReg(Reg: FramePtrReg); |
| 1398 | if (FPSaved) { |
| 1399 | // CSR spill restores should use FP as base register. If |
| 1400 | // SGPRForFPSaveRestoreCopy is not true, restore the previous value of FP |
| 1401 | // into a new scratch register and copy to FP later when other registers are |
| 1402 | // restored from the current stack frame. |
| 1403 | initLiveUnits(LiveUnits, TRI, FuncInfo, MF, MBB, MBBI, /*IsProlog*/ false); |
| 1404 | if (SGPRForFPSaveRestoreCopy) { |
| 1405 | LiveUnits.addReg(Reg: SGPRForFPSaveRestoreCopy); |
| 1406 | } else { |
| 1407 | FramePtrRegScratchCopy = findScratchNonCalleeSaveRegister( |
| 1408 | MRI, LiveUnits, RC: AMDGPU::SReg_32_XM0_XEXECRegClass); |
| 1409 | if (!FramePtrRegScratchCopy) |
| 1410 | report_fatal_error(reason: "failed to find free scratch register" ); |
| 1411 | |
| 1412 | LiveUnits.addReg(Reg: FramePtrRegScratchCopy); |
| 1413 | } |
| 1414 | |
| 1415 | emitCSRSpillRestores(MF, MBB, MBBI, DL, LiveUnits, FrameReg: FramePtrReg, |
| 1416 | FramePtrRegScratchCopy); |
| 1417 | } |
| 1418 | |
| 1419 | if (FPSaved) { |
| 1420 | // Insert the copy to restore FP. |
| 1421 | Register SrcReg = SGPRForFPSaveRestoreCopy ? SGPRForFPSaveRestoreCopy |
| 1422 | : FramePtrRegScratchCopy; |
| 1423 | MachineInstrBuilder MIB = |
| 1424 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: FramePtrReg) |
| 1425 | .addReg(RegNo: SrcReg); |
| 1426 | if (SGPRForFPSaveRestoreCopy) |
| 1427 | MIB.setMIFlag(MachineInstr::FrameDestroy); |
| 1428 | } else { |
| 1429 | // Insert the CSR spill restores with SP as the base register. |
| 1430 | emitCSRSpillRestores(MF, MBB, MBBI, DL, LiveUnits, |
| 1431 | FrameReg: FuncInfo->isChainFunction() ? Register() : StackPtrReg, |
| 1432 | FramePtrRegScratchCopy); |
| 1433 | } |
| 1434 | } |
| 1435 | |
| 1436 | #ifndef NDEBUG |
| 1437 | static bool allSGPRSpillsAreDead(const MachineFunction &MF) { |
| 1438 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1439 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1440 | for (int I = MFI.getObjectIndexBegin(), E = MFI.getObjectIndexEnd(); |
| 1441 | I != E; ++I) { |
| 1442 | if (!MFI.isDeadObjectIndex(I) && |
| 1443 | MFI.getStackID(I) == TargetStackID::SGPRSpill && |
| 1444 | !FuncInfo->checkIndexInPrologEpilogSGPRSpills(I)) { |
| 1445 | return false; |
| 1446 | } |
| 1447 | } |
| 1448 | |
| 1449 | return true; |
| 1450 | } |
| 1451 | #endif |
| 1452 | |
| 1453 | StackOffset SIFrameLowering::getFrameIndexReference(const MachineFunction &MF, |
| 1454 | int FI, |
| 1455 | Register &FrameReg) const { |
| 1456 | const SIRegisterInfo *RI = MF.getSubtarget<GCNSubtarget>().getRegisterInfo(); |
| 1457 | |
| 1458 | FrameReg = RI->getFrameRegister(MF); |
| 1459 | return StackOffset::getFixed(Fixed: MF.getFrameInfo().getObjectOffset(ObjectIdx: FI)); |
| 1460 | } |
| 1461 | |
| 1462 | void SIFrameLowering::processFunctionBeforeFrameFinalized( |
| 1463 | MachineFunction &MF, |
| 1464 | RegScavenger *RS) const { |
| 1465 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1466 | |
| 1467 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1468 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1469 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1470 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1471 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1472 | |
| 1473 | const bool SpillVGPRToAGPR = ST.hasMAIInsts() && FuncInfo->hasSpilledVGPRs() |
| 1474 | && EnableSpillVGPRToAGPR; |
| 1475 | |
| 1476 | if (SpillVGPRToAGPR) { |
| 1477 | // To track the spill frame indices handled in this pass. |
| 1478 | BitVector SpillFIs(MFI.getObjectIndexEnd(), false); |
| 1479 | BitVector NonVGPRSpillFIs(MFI.getObjectIndexEnd(), false); |
| 1480 | |
| 1481 | bool SeenDbgInstr = false; |
| 1482 | |
| 1483 | for (MachineBasicBlock &MBB : MF) { |
| 1484 | for (MachineInstr &MI : llvm::make_early_inc_range(Range&: MBB)) { |
| 1485 | int FrameIndex; |
| 1486 | if (MI.isDebugInstr()) |
| 1487 | SeenDbgInstr = true; |
| 1488 | |
| 1489 | if (TII->isVGPRSpill(MI)) { |
| 1490 | // Try to eliminate stack used by VGPR spills before frame |
| 1491 | // finalization. |
| 1492 | unsigned FIOp = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), |
| 1493 | Name: AMDGPU::OpName::vaddr); |
| 1494 | int FI = MI.getOperand(i: FIOp).getIndex(); |
| 1495 | Register VReg = |
| 1496 | TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdata)->getReg(); |
| 1497 | if (FuncInfo->allocateVGPRSpillToAGPR(MF, FI, |
| 1498 | isAGPRtoVGPR: TRI->isAGPR(MRI, Reg: VReg))) { |
| 1499 | assert(RS != nullptr); |
| 1500 | RS->enterBasicBlockEnd(MBB); |
| 1501 | RS->backward(I: std::next(x: MI.getIterator())); |
| 1502 | TRI->eliminateFrameIndex(MI, SPAdj: 0, FIOperandNum: FIOp, RS); |
| 1503 | SpillFIs.set(FI); |
| 1504 | continue; |
| 1505 | } |
| 1506 | } else if (TII->isStoreToStackSlot(MI, FrameIndex) || |
| 1507 | TII->isLoadFromStackSlot(MI, FrameIndex)) |
| 1508 | if (!MFI.isFixedObjectIndex(ObjectIdx: FrameIndex)) |
| 1509 | NonVGPRSpillFIs.set(FrameIndex); |
| 1510 | } |
| 1511 | } |
| 1512 | |
| 1513 | // Stack slot coloring may assign different objects to the same stack slot. |
| 1514 | // If not, then the VGPR to AGPR spill slot is dead. |
| 1515 | for (unsigned FI : SpillFIs.set_bits()) |
| 1516 | if (!NonVGPRSpillFIs.test(Idx: FI)) |
| 1517 | FuncInfo->setVGPRToAGPRSpillDead(FI); |
| 1518 | |
| 1519 | for (MachineBasicBlock &MBB : MF) { |
| 1520 | for (MCPhysReg Reg : FuncInfo->getVGPRSpillAGPRs()) |
| 1521 | MBB.addLiveIn(PhysReg: Reg); |
| 1522 | |
| 1523 | for (MCPhysReg Reg : FuncInfo->getAGPRSpillVGPRs()) |
| 1524 | MBB.addLiveIn(PhysReg: Reg); |
| 1525 | |
| 1526 | MBB.sortUniqueLiveIns(); |
| 1527 | |
| 1528 | if (!SpillFIs.empty() && SeenDbgInstr) { |
| 1529 | // FIXME: The dead frame indices are replaced with a null register from |
| 1530 | // the debug value instructions. We should instead, update it with the |
| 1531 | // correct register value. But not sure the register value alone is |
| 1532 | for (MachineInstr &MI : MBB) { |
| 1533 | if (MI.isDebugValue()) { |
| 1534 | uint32_t StackOperandIdx = MI.isDebugValueList() ? 2 : 0; |
| 1535 | if (MI.getOperand(i: StackOperandIdx).isFI() && |
| 1536 | !MFI.isFixedObjectIndex( |
| 1537 | ObjectIdx: MI.getOperand(i: StackOperandIdx).getIndex()) && |
| 1538 | SpillFIs[MI.getOperand(i: StackOperandIdx).getIndex()]) { |
| 1539 | MI.getOperand(i: StackOperandIdx) |
| 1540 | .ChangeToRegister(Reg: Register(), isDef: false /*isDef*/); |
| 1541 | } |
| 1542 | } |
| 1543 | } |
| 1544 | } |
| 1545 | } |
| 1546 | } |
| 1547 | |
| 1548 | // At this point we've already allocated all spilled SGPRs to VGPRs if we |
| 1549 | // can. Any remaining SGPR spills will go to memory, so move them back to the |
| 1550 | // default stack. |
| 1551 | bool HaveSGPRToVMemSpill = |
| 1552 | FuncInfo->removeDeadFrameIndices(MFI, /*ResetSGPRSpillStackIDs*/ true); |
| 1553 | assert(allSGPRSpillsAreDead(MF) && |
| 1554 | "SGPR spill should have been removed in SILowerSGPRSpills" ); |
| 1555 | |
| 1556 | // FIXME: The other checks should be redundant with allStackObjectsAreDead, |
| 1557 | // but currently hasNonSpillStackObjects is set only from source |
| 1558 | // allocas. Stack temps produced from legalization are not counted currently. |
| 1559 | if (!allStackObjectsAreDead(MFI)) { |
| 1560 | assert(RS && "RegScavenger required if spilling" ); |
| 1561 | |
| 1562 | // Add an emergency spill slot |
| 1563 | RS->addScavengingFrameIndex(FI: FuncInfo->getScavengeFI(MFI, TRI: *TRI)); |
| 1564 | |
| 1565 | // If we are spilling SGPRs to memory with a large frame, we may need a |
| 1566 | // second VGPR emergency frame index. |
| 1567 | if (HaveSGPRToVMemSpill && |
| 1568 | allocateScavengingFrameIndexesNearIncomingSP(MF)) { |
| 1569 | RS->addScavengingFrameIndex(FI: MFI.CreateSpillStackObject(Size: 4, Alignment: Align(4))); |
| 1570 | } |
| 1571 | } |
| 1572 | } |
| 1573 | |
| 1574 | void SIFrameLowering::processFunctionBeforeFrameIndicesReplaced( |
| 1575 | MachineFunction &MF, RegScavenger *RS) const { |
| 1576 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1577 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1578 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1579 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1580 | |
| 1581 | if (ST.hasMAIInsts() && !ST.hasGFX90AInsts()) { |
| 1582 | // On gfx908, we had initially reserved highest available VGPR for AGPR |
| 1583 | // copy. Now since we are done with RA, check if there exist an unused VGPR |
| 1584 | // which is lower than the eariler reserved VGPR before RA. If one exist, |
| 1585 | // use it for AGPR copy instead of one reserved before RA. |
| 1586 | Register VGPRForAGPRCopy = FuncInfo->getVGPRForAGPRCopy(); |
| 1587 | Register UnusedLowVGPR = |
| 1588 | TRI->findUnusedRegister(MRI, RC: &AMDGPU::VGPR_32RegClass, MF); |
| 1589 | if (UnusedLowVGPR && (TRI->getHWRegIndex(Reg: UnusedLowVGPR) < |
| 1590 | TRI->getHWRegIndex(Reg: VGPRForAGPRCopy))) { |
| 1591 | // Reserve this newly identified VGPR (for AGPR copy) |
| 1592 | // reserved registers should already be frozen at this point |
| 1593 | // so we can avoid calling MRI.freezeReservedRegs and just use |
| 1594 | // MRI.reserveReg |
| 1595 | FuncInfo->setVGPRForAGPRCopy(UnusedLowVGPR); |
| 1596 | MRI.reserveReg(PhysReg: UnusedLowVGPR, TRI); |
| 1597 | } |
| 1598 | } |
| 1599 | // We initally reserved the highest available SGPR pair for long branches |
| 1600 | // now, after RA, we shift down to a lower unused one if one exists |
| 1601 | Register LongBranchReservedReg = FuncInfo->getLongBranchReservedReg(); |
| 1602 | Register UnusedLowSGPR = |
| 1603 | TRI->findUnusedRegister(MRI, RC: &AMDGPU::SGPR_64RegClass, MF); |
| 1604 | // If LongBranchReservedReg is null then we didn't find a long branch |
| 1605 | // and never reserved a register to begin with so there is nothing to |
| 1606 | // shift down. Then if UnusedLowSGPR is null, there isn't available lower |
| 1607 | // register to use so just keep the original one we set. |
| 1608 | if (LongBranchReservedReg && UnusedLowSGPR) { |
| 1609 | FuncInfo->setLongBranchReservedReg(UnusedLowSGPR); |
| 1610 | MRI.reserveReg(PhysReg: UnusedLowSGPR, TRI); |
| 1611 | } |
| 1612 | } |
| 1613 | |
| 1614 | // The special SGPR spills like the one needed for FP, BP or any reserved |
| 1615 | // registers delayed until frame lowering. |
| 1616 | void SIFrameLowering::determinePrologEpilogSGPRSaves( |
| 1617 | MachineFunction &MF, BitVector &SavedVGPRs, |
| 1618 | bool NeedExecCopyReservedReg) const { |
| 1619 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 1620 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1621 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 1622 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1623 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1624 | LiveRegUnits LiveUnits; |
| 1625 | LiveUnits.init(TRI: *TRI); |
| 1626 | // Initially mark callee saved registers as used so we will not choose them |
| 1627 | // while looking for scratch SGPRs. |
| 1628 | const MCPhysReg *CSRegs = MF.getRegInfo().getCalleeSavedRegs(); |
| 1629 | for (unsigned I = 0; CSRegs[I]; ++I) |
| 1630 | LiveUnits.addReg(Reg: CSRegs[I]); |
| 1631 | |
| 1632 | const TargetRegisterClass &RC = *TRI->getWaveMaskRegClass(); |
| 1633 | |
| 1634 | Register ReservedRegForExecCopy = MFI->getSGPRForEXECCopy(); |
| 1635 | if (NeedExecCopyReservedReg || |
| 1636 | (ReservedRegForExecCopy && |
| 1637 | MRI.isPhysRegUsed(PhysReg: ReservedRegForExecCopy, /*SkipRegMaskTest=*/true))) { |
| 1638 | MRI.reserveReg(PhysReg: ReservedRegForExecCopy, TRI); |
| 1639 | Register UnusedScratchReg = findUnusedRegister(MRI, LiveUnits, RC); |
| 1640 | if (UnusedScratchReg) { |
| 1641 | // If found any unused scratch SGPR, reserve the register itself for Exec |
| 1642 | // copy and there is no need for any spills in that case. |
| 1643 | MFI->setSGPRForEXECCopy(UnusedScratchReg); |
| 1644 | MRI.replaceRegWith(FromReg: ReservedRegForExecCopy, ToReg: UnusedScratchReg); |
| 1645 | LiveUnits.addReg(Reg: UnusedScratchReg); |
| 1646 | } else { |
| 1647 | // Needs spill. |
| 1648 | assert(!MFI->hasPrologEpilogSGPRSpillEntry(ReservedRegForExecCopy) && |
| 1649 | "Re-reserving spill slot for EXEC copy register" ); |
| 1650 | getVGPRSpillLaneOrTempRegister(MF, LiveUnits, SGPR: ReservedRegForExecCopy, RC, |
| 1651 | /*IncludeScratchCopy=*/false); |
| 1652 | } |
| 1653 | } else if (ReservedRegForExecCopy) { |
| 1654 | // Reset it at this point. There are no whole-wave copies and spills |
| 1655 | // encountered. |
| 1656 | MFI->setSGPRForEXECCopy(AMDGPU::NoRegister); |
| 1657 | } |
| 1658 | |
| 1659 | // hasFP only knows about stack objects that already exist. We're now |
| 1660 | // determining the stack slots that will be created, so we have to predict |
| 1661 | // them. Stack objects force FP usage with calls. |
| 1662 | // |
| 1663 | // Note a new VGPR CSR may be introduced if one is used for the spill, but we |
| 1664 | // don't want to report it here. |
| 1665 | // |
| 1666 | // FIXME: Is this really hasReservedCallFrame? |
| 1667 | const bool WillHaveFP = |
| 1668 | FrameInfo.hasCalls() && |
| 1669 | (SavedVGPRs.any() || !allStackObjectsAreDead(MFI: FrameInfo)); |
| 1670 | |
| 1671 | if (WillHaveFP || hasFP(MF)) { |
| 1672 | Register FramePtrReg = MFI->getFrameOffsetReg(); |
| 1673 | assert(!MFI->hasPrologEpilogSGPRSpillEntry(FramePtrReg) && |
| 1674 | "Re-reserving spill slot for FP" ); |
| 1675 | getVGPRSpillLaneOrTempRegister(MF, LiveUnits, SGPR: FramePtrReg); |
| 1676 | } |
| 1677 | |
| 1678 | if (TRI->hasBasePointer(MF)) { |
| 1679 | Register BasePtrReg = TRI->getBaseRegister(); |
| 1680 | assert(!MFI->hasPrologEpilogSGPRSpillEntry(BasePtrReg) && |
| 1681 | "Re-reserving spill slot for BP" ); |
| 1682 | getVGPRSpillLaneOrTempRegister(MF, LiveUnits, SGPR: BasePtrReg); |
| 1683 | } |
| 1684 | } |
| 1685 | |
| 1686 | // Only report VGPRs to generic code. |
| 1687 | void SIFrameLowering::determineCalleeSaves(MachineFunction &MF, |
| 1688 | BitVector &SavedVGPRs, |
| 1689 | RegScavenger *RS) const { |
| 1690 | SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 1691 | |
| 1692 | // If this is a function with the amdgpu_cs_chain[_preserve] calling |
| 1693 | // convention and it doesn't contain any calls to llvm.amdgcn.cs.chain, then |
| 1694 | // we don't need to save and restore anything. |
| 1695 | if (MFI->isChainFunction() && !MF.getFrameInfo().hasTailCall()) |
| 1696 | return; |
| 1697 | |
| 1698 | TargetFrameLowering::determineCalleeSaves(MF, SavedRegs&: SavedVGPRs, RS); |
| 1699 | |
| 1700 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1701 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1702 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1703 | bool NeedExecCopyReservedReg = false; |
| 1704 | |
| 1705 | MachineInstr *ReturnMI = nullptr; |
| 1706 | for (MachineBasicBlock &MBB : MF) { |
| 1707 | for (MachineInstr &MI : MBB) { |
| 1708 | // TODO: Walking through all MBBs here would be a bad heuristic. Better |
| 1709 | // handle them elsewhere. |
| 1710 | if (TII->isWWMRegSpillOpcode(Opcode: MI.getOpcode())) |
| 1711 | NeedExecCopyReservedReg = true; |
| 1712 | else if (MI.getOpcode() == AMDGPU::SI_RETURN || |
| 1713 | MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG || |
| 1714 | MI.getOpcode() == AMDGPU::SI_WHOLE_WAVE_FUNC_RETURN || |
| 1715 | (MFI->isChainFunction() && |
| 1716 | TII->isChainCallOpcode(Opcode: MI.getOpcode()))) { |
| 1717 | // We expect all return to be the same size. |
| 1718 | assert(!ReturnMI || |
| 1719 | (count_if(MI.operands(), [](auto Op) { return Op.isReg(); }) == |
| 1720 | count_if(ReturnMI->operands(), [](auto Op) { return Op.isReg(); }))); |
| 1721 | ReturnMI = &MI; |
| 1722 | } |
| 1723 | } |
| 1724 | } |
| 1725 | |
| 1726 | SmallVector<Register> SortedWWMVGPRs; |
| 1727 | for (Register Reg : MFI->getWWMReservedRegs()) { |
| 1728 | // The shift-back is needed only for the VGPRs used for SGPR spills and they |
| 1729 | // are of 32-bit size. SIPreAllocateWWMRegs pass can add tuples into WWM |
| 1730 | // reserved registers. |
| 1731 | const TargetRegisterClass *RC = TRI->getPhysRegBaseClass(Reg); |
| 1732 | if (TRI->getRegSizeInBits(RC: *RC) != 32) |
| 1733 | continue; |
| 1734 | SortedWWMVGPRs.push_back(Elt: Reg); |
| 1735 | } |
| 1736 | |
| 1737 | sort(C&: SortedWWMVGPRs, Comp: std::greater<Register>()); |
| 1738 | MFI->shiftWwmVGPRsToLowestRange(MF, WWMVGPRs&: SortedWWMVGPRs, SavedVGPRs); |
| 1739 | |
| 1740 | if (MFI->isEntryFunction()) |
| 1741 | return; |
| 1742 | |
| 1743 | if (MFI->isWholeWaveFunction()) { |
| 1744 | // In practice, all the VGPRs are WWM registers, and we will need to save at |
| 1745 | // least their inactive lanes. Add them to WWMReservedRegs. |
| 1746 | assert(!NeedExecCopyReservedReg && |
| 1747 | "Whole wave functions can use the reg mapped for their i1 argument" ); |
| 1748 | |
| 1749 | // FIXME: Be more efficient! |
| 1750 | unsigned NumArchVGPRs = ST.has1024AddressableVGPRs() ? 1024 : 256; |
| 1751 | for (MCRegister Reg : |
| 1752 | AMDGPU::VGPR_32RegClass.getRegisters().take_front(N: NumArchVGPRs)) |
| 1753 | if (MF.getRegInfo().isPhysRegModified(PhysReg: Reg)) { |
| 1754 | MFI->reserveWWMRegister(Reg); |
| 1755 | MF.begin()->addLiveIn(PhysReg: Reg); |
| 1756 | } |
| 1757 | MF.begin()->sortUniqueLiveIns(); |
| 1758 | } |
| 1759 | |
| 1760 | // Remove any VGPRs used in the return value because these do not need to be saved. |
| 1761 | // This prevents CSR restore from clobbering return VGPRs. |
| 1762 | if (ReturnMI) { |
| 1763 | for (auto &Op : ReturnMI->operands()) { |
| 1764 | if (Op.isReg()) |
| 1765 | SavedVGPRs.reset(Idx: Op.getReg()); |
| 1766 | } |
| 1767 | } |
| 1768 | |
| 1769 | // Create the stack objects for WWM registers now. |
| 1770 | for (Register Reg : MFI->getWWMReservedRegs()) { |
| 1771 | const TargetRegisterClass *RC = TRI->getPhysRegBaseClass(Reg); |
| 1772 | MFI->allocateWWMSpill(MF, VGPR: Reg, Size: TRI->getSpillSize(RC: *RC), |
| 1773 | Alignment: TRI->getSpillAlign(RC: *RC)); |
| 1774 | } |
| 1775 | |
| 1776 | // Ignore the SGPRs the default implementation found. |
| 1777 | SavedVGPRs.clearBitsNotInMask(Mask: TRI->getAllVectorRegMask()); |
| 1778 | |
| 1779 | // Do not save AGPRs prior to GFX90A because there was no easy way to do so. |
| 1780 | // In gfx908 there was do AGPR loads and stores and thus spilling also |
| 1781 | // require a temporary VGPR. |
| 1782 | if (!ST.hasGFX90AInsts()) |
| 1783 | SavedVGPRs.clearBitsInMask(Mask: TRI->getAllAGPRRegMask()); |
| 1784 | |
| 1785 | determinePrologEpilogSGPRSaves(MF, SavedVGPRs, NeedExecCopyReservedReg); |
| 1786 | |
| 1787 | // The Whole-Wave VGPRs need to be specially inserted in the prolog, so don't |
| 1788 | // allow the default insertion to handle them. |
| 1789 | for (auto &Reg : MFI->getWWMSpills()) |
| 1790 | SavedVGPRs.reset(Idx: Reg.first); |
| 1791 | } |
| 1792 | |
| 1793 | void SIFrameLowering::determineCalleeSavesSGPR(MachineFunction &MF, |
| 1794 | BitVector &SavedRegs, |
| 1795 | RegScavenger *RS) const { |
| 1796 | TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); |
| 1797 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 1798 | if (MFI->isEntryFunction()) |
| 1799 | return; |
| 1800 | |
| 1801 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1802 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1803 | |
| 1804 | // The SP is specifically managed and we don't want extra spills of it. |
| 1805 | SavedRegs.reset(Idx: MFI->getStackPtrOffsetReg()); |
| 1806 | |
| 1807 | const BitVector AllSavedRegs = SavedRegs; |
| 1808 | SavedRegs.clearBitsInMask(Mask: TRI->getAllVectorRegMask()); |
| 1809 | |
| 1810 | // We have to anticipate introducing CSR VGPR spills or spill of caller |
| 1811 | // save VGPR reserved for SGPR spills as we now always create stack entry |
| 1812 | // for it, if we don't have any stack objects already, since we require a FP |
| 1813 | // if there is a call and stack. We will allocate a VGPR for SGPR spills if |
| 1814 | // there are any SGPR spills. Whether they are CSR spills or otherwise. |
| 1815 | MachineFrameInfo &FrameInfo = MF.getFrameInfo(); |
| 1816 | const bool WillHaveFP = |
| 1817 | FrameInfo.hasCalls() && (AllSavedRegs.any() || MFI->hasSpilledSGPRs()); |
| 1818 | |
| 1819 | // FP will be specially managed like SP. |
| 1820 | if (WillHaveFP || hasFP(MF)) |
| 1821 | SavedRegs.reset(Idx: MFI->getFrameOffsetReg()); |
| 1822 | |
| 1823 | // Return address use with return instruction is hidden through the SI_RETURN |
| 1824 | // pseudo. Given that and since the IPRA computes actual register usage and |
| 1825 | // does not use CSR list, the clobbering of return address by function calls |
| 1826 | // (D117243) or otherwise (D120922) is ignored/not seen by the IPRA's register |
| 1827 | // usage collection. This will ensure save/restore of return address happens |
| 1828 | // in those scenarios. |
| 1829 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1830 | Register RetAddrReg = TRI->getReturnAddressReg(MF); |
| 1831 | if (!MFI->isEntryFunction() && |
| 1832 | (FrameInfo.hasCalls() || MRI.isPhysRegModified(PhysReg: RetAddrReg))) { |
| 1833 | SavedRegs.set(TRI->getSubReg(Reg: RetAddrReg, Idx: AMDGPU::sub0)); |
| 1834 | SavedRegs.set(TRI->getSubReg(Reg: RetAddrReg, Idx: AMDGPU::sub1)); |
| 1835 | } |
| 1836 | } |
| 1837 | |
| 1838 | static void assignSlotsUsingVGPRBlocks(MachineFunction &MF, |
| 1839 | const GCNSubtarget &ST, |
| 1840 | std::vector<CalleeSavedInfo> &CSI) { |
| 1841 | SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1842 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1843 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
| 1844 | |
| 1845 | assert( |
| 1846 | llvm::is_sorted(CSI, |
| 1847 | [](const CalleeSavedInfo &A, const CalleeSavedInfo &B) { |
| 1848 | return A.getReg() < B.getReg(); |
| 1849 | }) && |
| 1850 | "Callee saved registers not sorted" ); |
| 1851 | |
| 1852 | auto CanUseBlockOps = [&](const CalleeSavedInfo &CSI) { |
| 1853 | return !CSI.isSpilledToReg() && |
| 1854 | TRI->getPhysRegBaseClass(Reg: CSI.getReg()) == &AMDGPU::VGPR_32RegClass && |
| 1855 | !FuncInfo->isWWMReservedRegister(Reg: CSI.getReg()); |
| 1856 | }; |
| 1857 | |
| 1858 | auto CSEnd = CSI.end(); |
| 1859 | for (auto CSIt = CSI.begin(); CSIt != CSEnd; ++CSIt) { |
| 1860 | Register Reg = CSIt->getReg(); |
| 1861 | if (!CanUseBlockOps(*CSIt)) |
| 1862 | continue; |
| 1863 | |
| 1864 | // Find all the regs that will fit in a 32-bit mask starting at the current |
| 1865 | // reg and build said mask. It should have 1 for every register that's |
| 1866 | // included, with the current register as the least significant bit. |
| 1867 | uint32_t Mask = 1; |
| 1868 | CSEnd = std::remove_if( |
| 1869 | first: CSIt + 1, last: CSEnd, pred: [&](const CalleeSavedInfo &CSI) -> bool { |
| 1870 | if (CanUseBlockOps(CSI) && CSI.getReg() < Reg + 32) { |
| 1871 | Mask |= 1 << (CSI.getReg() - Reg); |
| 1872 | return true; |
| 1873 | } else { |
| 1874 | return false; |
| 1875 | } |
| 1876 | }); |
| 1877 | |
| 1878 | const TargetRegisterClass *BlockRegClass = TRI->getRegClassForBlockOp(MF); |
| 1879 | Register RegBlock = |
| 1880 | TRI->getMatchingSuperReg(Reg, SubIdx: AMDGPU::sub0, RC: BlockRegClass); |
| 1881 | if (!RegBlock) { |
| 1882 | // We couldn't find a super register for the block. This can happen if |
| 1883 | // the register we started with is too high (e.g. v232 if the maximum is |
| 1884 | // v255). We therefore try to get the last register block and figure out |
| 1885 | // the mask from there. |
| 1886 | Register LastBlockStart = |
| 1887 | AMDGPU::VGPR0 + alignDown(Value: Reg - AMDGPU::VGPR0, Align: 32); |
| 1888 | RegBlock = |
| 1889 | TRI->getMatchingSuperReg(Reg: LastBlockStart, SubIdx: AMDGPU::sub0, RC: BlockRegClass); |
| 1890 | assert(RegBlock && TRI->isSubRegister(RegBlock, Reg) && |
| 1891 | "Couldn't find super register" ); |
| 1892 | int RegDelta = Reg - LastBlockStart; |
| 1893 | assert(RegDelta > 0 && llvm::countl_zero(Mask) >= RegDelta && |
| 1894 | "Bad shift amount" ); |
| 1895 | Mask <<= RegDelta; |
| 1896 | } |
| 1897 | |
| 1898 | FuncInfo->setMaskForVGPRBlockOps(RegisterBlock: RegBlock, Mask); |
| 1899 | |
| 1900 | // The stack objects can be a bit smaller than the register block if we know |
| 1901 | // some of the high bits of Mask are 0. This may happen often with calling |
| 1902 | // conventions where the caller and callee-saved VGPRs are interleaved at |
| 1903 | // a small boundary (e.g. 8 or 16). |
| 1904 | int UnusedBits = llvm::countl_zero(Val: Mask); |
| 1905 | unsigned BlockSize = TRI->getSpillSize(RC: *BlockRegClass) - UnusedBits * 4; |
| 1906 | int FrameIdx = |
| 1907 | MFI.CreateStackObject(Size: BlockSize, Alignment: TRI->getSpillAlign(RC: *BlockRegClass), |
| 1908 | /*isSpillSlot=*/true); |
| 1909 | MFI.setIsCalleeSavedObjectIndex(ObjectIdx: FrameIdx, IsCalleeSaved: true); |
| 1910 | |
| 1911 | CSIt->setFrameIdx(FrameIdx); |
| 1912 | CSIt->setReg(RegBlock); |
| 1913 | } |
| 1914 | CSI.erase(first: CSEnd, last: CSI.end()); |
| 1915 | } |
| 1916 | |
| 1917 | bool SIFrameLowering::assignCalleeSavedSpillSlots( |
| 1918 | MachineFunction &MF, const TargetRegisterInfo *TRI, |
| 1919 | std::vector<CalleeSavedInfo> &CSI) const { |
| 1920 | if (CSI.empty()) |
| 1921 | return true; // Early exit if no callee saved registers are modified! |
| 1922 | |
| 1923 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1924 | bool UseVGPRBlocks = ST.useVGPRBlockOpsForCSR(); |
| 1925 | |
| 1926 | if (UseVGPRBlocks) |
| 1927 | assignSlotsUsingVGPRBlocks(MF, ST, CSI); |
| 1928 | |
| 1929 | return assignCalleeSavedSpillSlotsImpl(MF, TRI, CSI) || UseVGPRBlocks; |
| 1930 | } |
| 1931 | |
| 1932 | bool SIFrameLowering::assignCalleeSavedSpillSlotsImpl( |
| 1933 | MachineFunction &MF, const TargetRegisterInfo *TRI, |
| 1934 | std::vector<CalleeSavedInfo> &CSI) const { |
| 1935 | if (CSI.empty()) |
| 1936 | return true; // Early exit if no callee saved registers are modified! |
| 1937 | |
| 1938 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
| 1939 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1940 | const SIRegisterInfo *RI = ST.getRegisterInfo(); |
| 1941 | Register FramePtrReg = FuncInfo->getFrameOffsetReg(); |
| 1942 | Register BasePtrReg = RI->getBaseRegister(); |
| 1943 | Register SGPRForFPSaveRestoreCopy = |
| 1944 | FuncInfo->getScratchSGPRCopyDstReg(Reg: FramePtrReg); |
| 1945 | Register SGPRForBPSaveRestoreCopy = |
| 1946 | FuncInfo->getScratchSGPRCopyDstReg(Reg: BasePtrReg); |
| 1947 | if (!SGPRForFPSaveRestoreCopy && !SGPRForBPSaveRestoreCopy) |
| 1948 | return false; |
| 1949 | |
| 1950 | unsigned NumModifiedRegs = 0; |
| 1951 | |
| 1952 | if (SGPRForFPSaveRestoreCopy) |
| 1953 | NumModifiedRegs++; |
| 1954 | if (SGPRForBPSaveRestoreCopy) |
| 1955 | NumModifiedRegs++; |
| 1956 | |
| 1957 | for (auto &CS : CSI) { |
| 1958 | if (CS.getReg() == FramePtrReg.asMCReg() && SGPRForFPSaveRestoreCopy) { |
| 1959 | CS.setDstReg(SGPRForFPSaveRestoreCopy); |
| 1960 | if (--NumModifiedRegs) |
| 1961 | break; |
| 1962 | } else if (CS.getReg() == BasePtrReg.asMCReg() && |
| 1963 | SGPRForBPSaveRestoreCopy) { |
| 1964 | CS.setDstReg(SGPRForBPSaveRestoreCopy); |
| 1965 | if (--NumModifiedRegs) |
| 1966 | break; |
| 1967 | } |
| 1968 | } |
| 1969 | |
| 1970 | return false; |
| 1971 | } |
| 1972 | |
| 1973 | bool SIFrameLowering::allocateScavengingFrameIndexesNearIncomingSP( |
| 1974 | const MachineFunction &MF) const { |
| 1975 | |
| 1976 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 1977 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 1978 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 1979 | uint64_t EstStackSize = MFI.estimateStackSize(MF); |
| 1980 | uint64_t MaxOffset = EstStackSize - 1; |
| 1981 | |
| 1982 | // We need the emergency stack slots to be allocated in range of the |
| 1983 | // MUBUF/flat scratch immediate offset from the base register, so assign these |
| 1984 | // first at the incoming SP position. |
| 1985 | // |
| 1986 | // TODO: We could try sorting the objects to find a hole in the first bytes |
| 1987 | // rather than allocating as close to possible. This could save a lot of space |
| 1988 | // on frames with alignment requirements. |
| 1989 | if (ST.hasFlatScratchEnabled()) { |
| 1990 | if (TII->isLegalFLATOffset(Offset: MaxOffset, AddrSpace: AMDGPUAS::PRIVATE_ADDRESS, |
| 1991 | FlatVariant: SIInstrFlags::FlatScratch)) |
| 1992 | return false; |
| 1993 | } else { |
| 1994 | if (TII->isLegalMUBUFImmOffset(Imm: MaxOffset)) |
| 1995 | return false; |
| 1996 | } |
| 1997 | |
| 1998 | return true; |
| 1999 | } |
| 2000 | |
| 2001 | bool SIFrameLowering::spillCalleeSavedRegisters( |
| 2002 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
| 2003 | ArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
| 2004 | MachineFunction *MF = MBB.getParent(); |
| 2005 | const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>(); |
| 2006 | if (!ST.useVGPRBlockOpsForCSR()) |
| 2007 | return false; |
| 2008 | |
| 2009 | MachineFrameInfo &FrameInfo = MF->getFrameInfo(); |
| 2010 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
| 2011 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 2012 | SIMachineFunctionInfo *FuncInfo = MF->getInfo<SIMachineFunctionInfo>(); |
| 2013 | |
| 2014 | const TargetRegisterClass *BlockRegClass = |
| 2015 | static_cast<const SIRegisterInfo *>(TRI)->getRegClassForBlockOp(MF: *MF); |
| 2016 | for (const CalleeSavedInfo &CS : CSI) { |
| 2017 | Register Reg = CS.getReg(); |
| 2018 | if (!BlockRegClass->contains(Reg) || |
| 2019 | !FuncInfo->hasMaskForVGPRBlockOps(RegisterBlock: Reg)) { |
| 2020 | spillCalleeSavedRegister(SaveBlock&: MBB, MI, CS, TII, TRI); |
| 2021 | continue; |
| 2022 | } |
| 2023 | |
| 2024 | // Build a scratch block store. |
| 2025 | uint32_t Mask = FuncInfo->getMaskForVGPRBlockOps(RegisterBlock: Reg); |
| 2026 | int FrameIndex = CS.getFrameIdx(); |
| 2027 | MachinePointerInfo PtrInfo = |
| 2028 | MachinePointerInfo::getFixedStack(MF&: *MF, FI: FrameIndex); |
| 2029 | MachineMemOperand *MMO = |
| 2030 | MF->getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOStore, |
| 2031 | Size: FrameInfo.getObjectSize(ObjectIdx: FrameIndex), |
| 2032 | BaseAlignment: FrameInfo.getObjectAlign(ObjectIdx: FrameIndex)); |
| 2033 | |
| 2034 | BuildMI(BB&: MBB, I: MI, MIMD: MI->getDebugLoc(), |
| 2035 | MCID: TII->get(Opcode: AMDGPU::SI_BLOCK_SPILL_V1024_SAVE)) |
| 2036 | .addReg(RegNo: Reg, Flags: getKillRegState(B: false)) |
| 2037 | .addFrameIndex(Idx: FrameIndex) |
| 2038 | .addReg(RegNo: MFI->getStackPtrOffsetReg()) |
| 2039 | .addImm(Val: 0) |
| 2040 | .addImm(Val: Mask) |
| 2041 | .addMemOperand(MMO); |
| 2042 | |
| 2043 | FuncInfo->setHasSpilledVGPRs(); |
| 2044 | |
| 2045 | // Add the register to the liveins. This is necessary because if any of the |
| 2046 | // VGPRs in the register block is reserved (e.g. if it's a WWM register), |
| 2047 | // then the whole block will be marked as reserved and `updateLiveness` will |
| 2048 | // skip it. |
| 2049 | MBB.addLiveIn(PhysReg: Reg); |
| 2050 | } |
| 2051 | MBB.sortUniqueLiveIns(); |
| 2052 | |
| 2053 | return true; |
| 2054 | } |
| 2055 | |
| 2056 | bool SIFrameLowering::restoreCalleeSavedRegisters( |
| 2057 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
| 2058 | MutableArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const { |
| 2059 | MachineFunction *MF = MBB.getParent(); |
| 2060 | const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>(); |
| 2061 | if (!ST.useVGPRBlockOpsForCSR()) |
| 2062 | return false; |
| 2063 | |
| 2064 | SIMachineFunctionInfo *FuncInfo = MF->getInfo<SIMachineFunctionInfo>(); |
| 2065 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
| 2066 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 2067 | const SIRegisterInfo *SITRI = static_cast<const SIRegisterInfo *>(TRI); |
| 2068 | const TargetRegisterClass *BlockRegClass = SITRI->getRegClassForBlockOp(MF: *MF); |
| 2069 | for (const CalleeSavedInfo &CS : reverse(C&: CSI)) { |
| 2070 | Register Reg = CS.getReg(); |
| 2071 | if (!BlockRegClass->contains(Reg) || |
| 2072 | !FuncInfo->hasMaskForVGPRBlockOps(RegisterBlock: Reg)) { |
| 2073 | restoreCalleeSavedRegister(MBB, MI, CS, TII, TRI); |
| 2074 | continue; |
| 2075 | } |
| 2076 | |
| 2077 | // Build a scratch block load. |
| 2078 | uint32_t Mask = FuncInfo->getMaskForVGPRBlockOps(RegisterBlock: Reg); |
| 2079 | int FrameIndex = CS.getFrameIdx(); |
| 2080 | MachinePointerInfo PtrInfo = |
| 2081 | MachinePointerInfo::getFixedStack(MF&: *MF, FI: FrameIndex); |
| 2082 | MachineMemOperand *MMO = MF->getMachineMemOperand( |
| 2083 | PtrInfo, F: MachineMemOperand::MOLoad, Size: MFI.getObjectSize(ObjectIdx: FrameIndex), |
| 2084 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: FrameIndex)); |
| 2085 | |
| 2086 | auto MIB = BuildMI(BB&: MBB, I: MI, MIMD: MI->getDebugLoc(), |
| 2087 | MCID: TII->get(Opcode: AMDGPU::SI_BLOCK_SPILL_V1024_RESTORE), DestReg: Reg) |
| 2088 | .addFrameIndex(Idx: FrameIndex) |
| 2089 | .addReg(RegNo: FuncInfo->getStackPtrOffsetReg()) |
| 2090 | .addImm(Val: 0) |
| 2091 | .addImm(Val: Mask) |
| 2092 | .addMemOperand(MMO); |
| 2093 | SITRI->addImplicitUsesForBlockCSRLoad(MIB, BlockReg: Reg); |
| 2094 | |
| 2095 | // Add the register to the liveins. This is necessary because if any of the |
| 2096 | // VGPRs in the register block is reserved (e.g. if it's a WWM register), |
| 2097 | // then the whole block will be marked as reserved and `updateLiveness` will |
| 2098 | // skip it. |
| 2099 | MBB.addLiveIn(PhysReg: Reg); |
| 2100 | } |
| 2101 | |
| 2102 | MBB.sortUniqueLiveIns(); |
| 2103 | return true; |
| 2104 | } |
| 2105 | |
| 2106 | MachineBasicBlock::iterator SIFrameLowering::eliminateCallFramePseudoInstr( |
| 2107 | MachineFunction &MF, |
| 2108 | MachineBasicBlock &MBB, |
| 2109 | MachineBasicBlock::iterator I) const { |
| 2110 | int64_t Amount = I->getOperand(i: 0).getImm(); |
| 2111 | if (Amount == 0) |
| 2112 | return MBB.erase(I); |
| 2113 | |
| 2114 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
| 2115 | const SIInstrInfo *TII = ST.getInstrInfo(); |
| 2116 | const DebugLoc &DL = I->getDebugLoc(); |
| 2117 | unsigned Opc = I->getOpcode(); |
| 2118 | bool IsDestroy = Opc == TII->getCallFrameDestroyOpcode(); |
| 2119 | uint64_t CalleePopAmount = IsDestroy ? I->getOperand(i: 1).getImm() : 0; |
| 2120 | |
| 2121 | if (!hasReservedCallFrame(MF)) { |
| 2122 | Amount = alignTo(Size: Amount, A: getStackAlign()); |
| 2123 | assert(isUInt<32>(Amount) && "exceeded stack address space size" ); |
| 2124 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
| 2125 | Register SPReg = MFI->getStackPtrOffsetReg(); |
| 2126 | |
| 2127 | Amount *= getScratchScaleFactor(ST); |
| 2128 | if (IsDestroy) |
| 2129 | Amount = -Amount; |
| 2130 | auto Add = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: SPReg) |
| 2131 | .addReg(RegNo: SPReg) |
| 2132 | .addImm(Val: Amount); |
| 2133 | Add->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
| 2134 | } else if (CalleePopAmount != 0) { |
| 2135 | llvm_unreachable("is this used?" ); |
| 2136 | } |
| 2137 | |
| 2138 | return MBB.erase(I); |
| 2139 | } |
| 2140 | |
| 2141 | /// Returns true if the frame will require a reference to the stack pointer. |
| 2142 | /// |
| 2143 | /// This is the set of conditions common to setting up the stack pointer in a |
| 2144 | /// kernel, and for using a frame pointer in a callable function. |
| 2145 | /// |
| 2146 | /// FIXME: Should also check hasOpaqueSPAdjustment and if any inline asm |
| 2147 | /// references SP. |
| 2148 | static bool frameTriviallyRequiresSP(const MachineFrameInfo &MFI) { |
| 2149 | return MFI.hasVarSizedObjects() || MFI.hasStackMap() || MFI.hasPatchPoint(); |
| 2150 | } |
| 2151 | |
| 2152 | // The FP for kernels is always known 0, so we never really need to setup an |
| 2153 | // explicit register for it. However, DisableFramePointerElim will force us to |
| 2154 | // use a register for it. |
| 2155 | bool SIFrameLowering::hasFPImpl(const MachineFunction &MF) const { |
| 2156 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2157 | |
| 2158 | // For entry & chain functions we can use an immediate offset in most cases, |
| 2159 | // so the presence of calls doesn't imply we need a distinct frame pointer. |
| 2160 | if (MFI.hasCalls() && |
| 2161 | !MF.getInfo<SIMachineFunctionInfo>()->isEntryFunction() && |
| 2162 | !MF.getInfo<SIMachineFunctionInfo>()->isChainFunction()) { |
| 2163 | // All offsets are unsigned, so need to be addressed in the same direction |
| 2164 | // as stack growth. |
| 2165 | |
| 2166 | // FIXME: This function is pretty broken, since it can be called before the |
| 2167 | // frame layout is determined or CSR spills are inserted. |
| 2168 | return MFI.getStackSize() != 0; |
| 2169 | } |
| 2170 | |
| 2171 | return (frameTriviallyRequiresSP(MFI) && |
| 2172 | !MF.getInfo<SIMachineFunctionInfo>()->isChainFunction()) || |
| 2173 | MFI.isFrameAddressTaken() || |
| 2174 | MF.getSubtarget<GCNSubtarget>().getRegisterInfo()->hasStackRealignment( |
| 2175 | MF) || |
| 2176 | mayReserveScratchForCWSR(MF) || |
| 2177 | MF.getTarget().Options.DisableFramePointerElim(MF); |
| 2178 | } |
| 2179 | |
| 2180 | bool SIFrameLowering::mayReserveScratchForCWSR( |
| 2181 | const MachineFunction &MF) const { |
| 2182 | return MF.getInfo<SIMachineFunctionInfo>()->isDynamicVGPREnabled() && |
| 2183 | AMDGPU::isEntryFunctionCC(CC: MF.getFunction().getCallingConv()) && |
| 2184 | AMDGPU::isCompute(CC: MF.getFunction().getCallingConv()); |
| 2185 | } |
| 2186 | |
| 2187 | // This is essentially a reduced version of hasFP for entry functions. Since the |
| 2188 | // stack pointer is known 0 on entry to kernels, we never really need an FP |
| 2189 | // register. We may need to initialize the stack pointer depending on the frame |
| 2190 | // properties, which logically overlaps many of the cases where an ordinary |
| 2191 | // function would require an FP. |
| 2192 | // Also used for chain functions. While not technically entry functions, chain |
| 2193 | // functions may need to set up a stack pointer in some situations. |
| 2194 | bool SIFrameLowering::requiresStackPointerReference( |
| 2195 | const MachineFunction &MF) const { |
| 2196 | // Callable functions always require a stack pointer reference. |
| 2197 | assert((MF.getInfo<SIMachineFunctionInfo>()->isEntryFunction() || |
| 2198 | MF.getInfo<SIMachineFunctionInfo>()->isChainFunction()) && |
| 2199 | "only expected to call this for entry points and chain functions" ); |
| 2200 | |
| 2201 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 2202 | |
| 2203 | // Entry points ordinarily don't need to initialize SP. We have to set it up |
| 2204 | // for callees if there are any. Also note tail calls are impossible/don't |
| 2205 | // make any sense for kernels. |
| 2206 | if (MFI.hasCalls()) |
| 2207 | return true; |
| 2208 | |
| 2209 | // We still need to initialize the SP if we're doing anything weird that |
| 2210 | // references the SP, like variable sized stack objects. |
| 2211 | return frameTriviallyRequiresSP(MFI); |
| 2212 | } |
| 2213 | |