1 | //===-- SIRegisterInfo.cpp - SI Register Information ---------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file |
10 | /// SI implementation of the TargetRegisterInfo class. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "AMDGPU.h" |
15 | #include "AMDGPURegisterBankInfo.h" |
16 | #include "GCNSubtarget.h" |
17 | #include "MCTargetDesc/AMDGPUInstPrinter.h" |
18 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
19 | #include "SIMachineFunctionInfo.h" |
20 | #include "SIRegisterInfo.h" |
21 | #include "llvm/CodeGen/LiveIntervals.h" |
22 | #include "llvm/CodeGen/LiveRegUnits.h" |
23 | #include "llvm/CodeGen/MachineDominators.h" |
24 | #include "llvm/CodeGen/MachineFrameInfo.h" |
25 | #include "llvm/CodeGen/RegisterScavenging.h" |
26 | |
27 | using namespace llvm; |
28 | |
29 | #define GET_REGINFO_TARGET_DESC |
30 | #include "AMDGPUGenRegisterInfo.inc" |
31 | |
32 | static cl::opt<bool> EnableSpillSGPRToVGPR( |
33 | "amdgpu-spill-sgpr-to-vgpr" , |
34 | cl::desc("Enable spilling SGPRs to VGPRs" ), |
35 | cl::ReallyHidden, |
36 | cl::init(Val: true)); |
37 | |
38 | std::array<std::vector<int16_t>, 16> SIRegisterInfo::RegSplitParts; |
39 | std::array<std::array<uint16_t, 32>, 9> SIRegisterInfo::SubRegFromChannelTable; |
40 | |
41 | // Map numbers of DWORDs to indexes in SubRegFromChannelTable. |
42 | // Valid indexes are shifted 1, such that a 0 mapping means unsupported. |
43 | // e.g. for 8 DWORDs (256-bit), SubRegFromChannelTableWidthMap[8] = 8, |
44 | // meaning index 7 in SubRegFromChannelTable. |
45 | static const std::array<unsigned, 17> SubRegFromChannelTableWidthMap = { |
46 | 0, 1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 9}; |
47 | |
48 | namespace llvm { |
49 | |
50 | // A temporary struct to spill SGPRs. |
51 | // This is mostly to spill SGPRs to memory. Spilling SGPRs into VGPR lanes emits |
52 | // just v_writelane and v_readlane. |
53 | // |
54 | // When spilling to memory, the SGPRs are written into VGPR lanes and the VGPR |
55 | // is saved to scratch (or the other way around for loads). |
56 | // For this, a VGPR is required where the needed lanes can be clobbered. The |
57 | // RegScavenger can provide a VGPR where currently active lanes can be |
58 | // clobbered, but we still need to save inactive lanes. |
59 | // The high-level steps are: |
60 | // - Try to scavenge SGPR(s) to save exec |
61 | // - Try to scavenge VGPR |
62 | // - Save needed, all or inactive lanes of a TmpVGPR |
63 | // - Spill/Restore SGPRs using TmpVGPR |
64 | // - Restore TmpVGPR |
65 | // |
66 | // To save all lanes of TmpVGPR, exec needs to be saved and modified. If we |
67 | // cannot scavenge temporary SGPRs to save exec, we use the following code: |
68 | // buffer_store_dword TmpVGPR ; only if active lanes need to be saved |
69 | // s_not exec, exec |
70 | // buffer_store_dword TmpVGPR ; save inactive lanes |
71 | // s_not exec, exec |
72 | struct SGPRSpillBuilder { |
73 | struct PerVGPRData { |
74 | unsigned PerVGPR; |
75 | unsigned NumVGPRs; |
76 | int64_t VGPRLanes; |
77 | }; |
78 | |
79 | // The SGPR to save |
80 | Register SuperReg; |
81 | MachineBasicBlock::iterator MI; |
82 | ArrayRef<int16_t> SplitParts; |
83 | unsigned NumSubRegs; |
84 | bool IsKill; |
85 | const DebugLoc &DL; |
86 | |
87 | /* When spilling to stack */ |
88 | // The SGPRs are written into this VGPR, which is then written to scratch |
89 | // (or vice versa for loads). |
90 | Register TmpVGPR = AMDGPU::NoRegister; |
91 | // Temporary spill slot to save TmpVGPR to. |
92 | int TmpVGPRIndex = 0; |
93 | // If TmpVGPR is live before the spill or if it is scavenged. |
94 | bool TmpVGPRLive = false; |
95 | // Scavenged SGPR to save EXEC. |
96 | Register SavedExecReg = AMDGPU::NoRegister; |
97 | // Stack index to write the SGPRs to. |
98 | int Index; |
99 | unsigned EltSize = 4; |
100 | |
101 | RegScavenger *RS; |
102 | MachineBasicBlock *MBB; |
103 | MachineFunction &MF; |
104 | SIMachineFunctionInfo &MFI; |
105 | const SIInstrInfo &TII; |
106 | const SIRegisterInfo &TRI; |
107 | bool IsWave32; |
108 | Register ExecReg; |
109 | unsigned MovOpc; |
110 | unsigned NotOpc; |
111 | |
112 | SGPRSpillBuilder(const SIRegisterInfo &TRI, const SIInstrInfo &TII, |
113 | bool IsWave32, MachineBasicBlock::iterator MI, int Index, |
114 | RegScavenger *RS) |
115 | : SGPRSpillBuilder(TRI, TII, IsWave32, MI, MI->getOperand(i: 0).getReg(), |
116 | MI->getOperand(i: 0).isKill(), Index, RS) {} |
117 | |
118 | SGPRSpillBuilder(const SIRegisterInfo &TRI, const SIInstrInfo &TII, |
119 | bool IsWave32, MachineBasicBlock::iterator MI, Register Reg, |
120 | bool IsKill, int Index, RegScavenger *RS) |
121 | : SuperReg(Reg), MI(MI), IsKill(IsKill), DL(MI->getDebugLoc()), |
122 | Index(Index), RS(RS), MBB(MI->getParent()), MF(*MBB->getParent()), |
123 | MFI(*MF.getInfo<SIMachineFunctionInfo>()), TII(TII), TRI(TRI), |
124 | IsWave32(IsWave32) { |
125 | const TargetRegisterClass *RC = TRI.getPhysRegBaseClass(Reg: SuperReg); |
126 | SplitParts = TRI.getRegSplitParts(RC, EltSize); |
127 | NumSubRegs = SplitParts.empty() ? 1 : SplitParts.size(); |
128 | |
129 | if (IsWave32) { |
130 | ExecReg = AMDGPU::EXEC_LO; |
131 | MovOpc = AMDGPU::S_MOV_B32; |
132 | NotOpc = AMDGPU::S_NOT_B32; |
133 | } else { |
134 | ExecReg = AMDGPU::EXEC; |
135 | MovOpc = AMDGPU::S_MOV_B64; |
136 | NotOpc = AMDGPU::S_NOT_B64; |
137 | } |
138 | |
139 | assert(SuperReg != AMDGPU::M0 && "m0 should never spill" ); |
140 | assert(SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI && |
141 | SuperReg != AMDGPU::EXEC && "exec should never spill" ); |
142 | } |
143 | |
144 | PerVGPRData getPerVGPRData() { |
145 | PerVGPRData Data; |
146 | Data.PerVGPR = IsWave32 ? 32 : 64; |
147 | Data.NumVGPRs = (NumSubRegs + (Data.PerVGPR - 1)) / Data.PerVGPR; |
148 | Data.VGPRLanes = (1LL << std::min(a: Data.PerVGPR, b: NumSubRegs)) - 1LL; |
149 | return Data; |
150 | } |
151 | |
152 | // Tries to scavenge SGPRs to save EXEC and a VGPR. Uses v0 if no VGPR is |
153 | // free. |
154 | // Writes these instructions if an SGPR can be scavenged: |
155 | // s_mov_b64 s[6:7], exec ; Save exec |
156 | // s_mov_b64 exec, 3 ; Wanted lanemask |
157 | // buffer_store_dword v1 ; Write scavenged VGPR to emergency slot |
158 | // |
159 | // Writes these instructions if no SGPR can be scavenged: |
160 | // buffer_store_dword v0 ; Only if no free VGPR was found |
161 | // s_not_b64 exec, exec |
162 | // buffer_store_dword v0 ; Save inactive lanes |
163 | // ; exec stays inverted, it is flipped back in |
164 | // ; restore. |
165 | void prepare() { |
166 | // Scavenged temporary VGPR to use. It must be scavenged once for any number |
167 | // of spilled subregs. |
168 | // FIXME: The liveness analysis is limited and does not tell if a register |
169 | // is in use in lanes that are currently inactive. We can never be sure if |
170 | // a register as actually in use in another lane, so we need to save all |
171 | // used lanes of the chosen VGPR. |
172 | assert(RS && "Cannot spill SGPR to memory without RegScavenger" ); |
173 | TmpVGPR = RS->scavengeRegisterBackwards(RC: AMDGPU::VGPR_32RegClass, To: MI, RestoreAfter: false, |
174 | SPAdj: 0, AllowSpill: false); |
175 | |
176 | // Reserve temporary stack slot |
177 | TmpVGPRIndex = MFI.getScavengeFI(MFI&: MF.getFrameInfo(), TRI); |
178 | if (TmpVGPR) { |
179 | // Found a register that is dead in the currently active lanes, we only |
180 | // need to spill inactive lanes. |
181 | TmpVGPRLive = false; |
182 | } else { |
183 | // Pick v0 because it doesn't make a difference. |
184 | TmpVGPR = AMDGPU::VGPR0; |
185 | TmpVGPRLive = true; |
186 | } |
187 | |
188 | if (TmpVGPRLive) { |
189 | // We need to inform the scavenger that this index is already in use until |
190 | // we're done with the custom emergency spill. |
191 | RS->assignRegToScavengingIndex(FI: TmpVGPRIndex, Reg: TmpVGPR); |
192 | } |
193 | |
194 | // We may end up recursively calling the scavenger, and don't want to re-use |
195 | // the same register. |
196 | RS->setRegUsed(Reg: TmpVGPR); |
197 | |
198 | // Try to scavenge SGPRs to save exec |
199 | assert(!SavedExecReg && "Exec is already saved, refuse to save again" ); |
200 | const TargetRegisterClass &RC = |
201 | IsWave32 ? AMDGPU::SGPR_32RegClass : AMDGPU::SGPR_64RegClass; |
202 | RS->setRegUsed(Reg: SuperReg); |
203 | SavedExecReg = RS->scavengeRegisterBackwards(RC, To: MI, RestoreAfter: false, SPAdj: 0, AllowSpill: false); |
204 | |
205 | int64_t VGPRLanes = getPerVGPRData().VGPRLanes; |
206 | |
207 | if (SavedExecReg) { |
208 | RS->setRegUsed(Reg: SavedExecReg); |
209 | // Set exec to needed lanes |
210 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: MovOpc), DestReg: SavedExecReg).addReg(RegNo: ExecReg); |
211 | auto I = |
212 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: MovOpc), DestReg: ExecReg).addImm(Val: VGPRLanes); |
213 | if (!TmpVGPRLive) |
214 | I.addReg(RegNo: TmpVGPR, flags: RegState::ImplicitDefine); |
215 | // Spill needed lanes |
216 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ false); |
217 | } else { |
218 | // The modify and restore of exec clobber SCC, which we would have to save |
219 | // and restore. FIXME: We probably would need to reserve a register for |
220 | // this. |
221 | if (RS->isRegUsed(Reg: AMDGPU::SCC)) |
222 | MI->emitError(Msg: "unhandled SGPR spill to memory" ); |
223 | |
224 | // Spill active lanes |
225 | if (TmpVGPRLive) |
226 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ false, |
227 | /*IsKill*/ false); |
228 | // Spill inactive lanes |
229 | auto I = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: NotOpc), DestReg: ExecReg).addReg(RegNo: ExecReg); |
230 | if (!TmpVGPRLive) |
231 | I.addReg(RegNo: TmpVGPR, flags: RegState::ImplicitDefine); |
232 | I->getOperand(i: 2).setIsDead(); // Mark SCC as dead. |
233 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ false); |
234 | } |
235 | } |
236 | |
237 | // Writes these instructions if an SGPR can be scavenged: |
238 | // buffer_load_dword v1 ; Write scavenged VGPR to emergency slot |
239 | // s_waitcnt vmcnt(0) ; If a free VGPR was found |
240 | // s_mov_b64 exec, s[6:7] ; Save exec |
241 | // |
242 | // Writes these instructions if no SGPR can be scavenged: |
243 | // buffer_load_dword v0 ; Restore inactive lanes |
244 | // s_waitcnt vmcnt(0) ; If a free VGPR was found |
245 | // s_not_b64 exec, exec |
246 | // buffer_load_dword v0 ; Only if no free VGPR was found |
247 | void restore() { |
248 | if (SavedExecReg) { |
249 | // Restore used lanes |
250 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ true, |
251 | /*IsKill*/ false); |
252 | // Restore exec |
253 | auto I = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: MovOpc), DestReg: ExecReg) |
254 | .addReg(RegNo: SavedExecReg, flags: RegState::Kill); |
255 | // Add an implicit use of the load so it is not dead. |
256 | // FIXME This inserts an unnecessary waitcnt |
257 | if (!TmpVGPRLive) { |
258 | I.addReg(RegNo: TmpVGPR, flags: RegState::ImplicitKill); |
259 | } |
260 | } else { |
261 | // Restore inactive lanes |
262 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ true, |
263 | /*IsKill*/ false); |
264 | auto I = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: NotOpc), DestReg: ExecReg).addReg(RegNo: ExecReg); |
265 | if (!TmpVGPRLive) |
266 | I.addReg(RegNo: TmpVGPR, flags: RegState::ImplicitKill); |
267 | I->getOperand(i: 2).setIsDead(); // Mark SCC as dead. |
268 | |
269 | // Restore active lanes |
270 | if (TmpVGPRLive) |
271 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index: TmpVGPRIndex, Offset: 0, /*IsLoad*/ true); |
272 | } |
273 | |
274 | // Inform the scavenger where we're releasing our custom scavenged register. |
275 | if (TmpVGPRLive) { |
276 | MachineBasicBlock::iterator RestorePt = std::prev(x: MI); |
277 | RS->assignRegToScavengingIndex(FI: TmpVGPRIndex, Reg: TmpVGPR, Restore: &*RestorePt); |
278 | } |
279 | } |
280 | |
281 | // Write TmpVGPR to memory or read TmpVGPR from memory. |
282 | // Either using a single buffer_load/store if exec is set to the needed mask |
283 | // or using |
284 | // buffer_load |
285 | // s_not exec, exec |
286 | // buffer_load |
287 | // s_not exec, exec |
288 | void readWriteTmpVGPR(unsigned Offset, bool IsLoad) { |
289 | if (SavedExecReg) { |
290 | // Spill needed lanes |
291 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index, Offset, IsLoad); |
292 | } else { |
293 | // The modify and restore of exec clobber SCC, which we would have to save |
294 | // and restore. FIXME: We probably would need to reserve a register for |
295 | // this. |
296 | if (RS->isRegUsed(Reg: AMDGPU::SCC)) |
297 | MI->emitError(Msg: "unhandled SGPR spill to memory" ); |
298 | |
299 | // Spill active lanes |
300 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index, Offset, IsLoad, |
301 | /*IsKill*/ false); |
302 | // Spill inactive lanes |
303 | auto Not0 = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: NotOpc), DestReg: ExecReg).addReg(RegNo: ExecReg); |
304 | Not0->getOperand(i: 2).setIsDead(); // Mark SCC as dead. |
305 | TRI.buildVGPRSpillLoadStore(SB&: *this, Index, Offset, IsLoad); |
306 | auto Not1 = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: NotOpc), DestReg: ExecReg).addReg(RegNo: ExecReg); |
307 | Not1->getOperand(i: 2).setIsDead(); // Mark SCC as dead. |
308 | } |
309 | } |
310 | |
311 | void setMI(MachineBasicBlock *NewMBB, MachineBasicBlock::iterator NewMI) { |
312 | assert(MBB->getParent() == &MF); |
313 | MI = NewMI; |
314 | MBB = NewMBB; |
315 | } |
316 | }; |
317 | |
318 | } // namespace llvm |
319 | |
320 | SIRegisterInfo::SIRegisterInfo(const GCNSubtarget &ST) |
321 | : AMDGPUGenRegisterInfo(AMDGPU::PC_REG, ST.getAMDGPUDwarfFlavour(), |
322 | ST.getAMDGPUDwarfFlavour()), |
323 | ST(ST), SpillSGPRToVGPR(EnableSpillSGPRToVGPR), isWave32(ST.isWave32()) { |
324 | |
325 | assert(getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && |
326 | getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && |
327 | (getSubRegIndexLaneMask(AMDGPU::lo16) | |
328 | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == |
329 | getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && |
330 | "getNumCoveredRegs() will not work with generated subreg masks!" ); |
331 | |
332 | RegPressureIgnoredUnits.resize(N: getNumRegUnits()); |
333 | RegPressureIgnoredUnits.set(*regunits(Reg: MCRegister::from(Val: AMDGPU::M0)).begin()); |
334 | for (auto Reg : AMDGPU::VGPR_16RegClass) { |
335 | if (AMDGPU::isHi(Reg, MRI: *this)) |
336 | RegPressureIgnoredUnits.set(*regunits(Reg).begin()); |
337 | } |
338 | |
339 | // HACK: Until this is fully tablegen'd. |
340 | static llvm::once_flag InitializeRegSplitPartsFlag; |
341 | |
342 | static auto InitializeRegSplitPartsOnce = [this]() { |
343 | for (unsigned Idx = 1, E = getNumSubRegIndices() - 1; Idx < E; ++Idx) { |
344 | unsigned Size = getSubRegIdxSize(Idx); |
345 | if (Size & 31) |
346 | continue; |
347 | std::vector<int16_t> &Vec = RegSplitParts[Size / 32 - 1]; |
348 | unsigned Pos = getSubRegIdxOffset(Idx); |
349 | if (Pos % Size) |
350 | continue; |
351 | Pos /= Size; |
352 | if (Vec.empty()) { |
353 | unsigned MaxNumParts = 1024 / Size; // Maximum register is 1024 bits. |
354 | Vec.resize(new_size: MaxNumParts); |
355 | } |
356 | Vec[Pos] = Idx; |
357 | } |
358 | }; |
359 | |
360 | static llvm::once_flag InitializeSubRegFromChannelTableFlag; |
361 | |
362 | static auto InitializeSubRegFromChannelTableOnce = [this]() { |
363 | for (auto &Row : SubRegFromChannelTable) |
364 | Row.fill(u: AMDGPU::NoSubRegister); |
365 | for (unsigned Idx = 1; Idx < getNumSubRegIndices(); ++Idx) { |
366 | unsigned Width = getSubRegIdxSize(Idx) / 32; |
367 | unsigned Offset = getSubRegIdxOffset(Idx) / 32; |
368 | assert(Width < SubRegFromChannelTableWidthMap.size()); |
369 | Width = SubRegFromChannelTableWidthMap[Width]; |
370 | if (Width == 0) |
371 | continue; |
372 | unsigned TableIdx = Width - 1; |
373 | assert(TableIdx < SubRegFromChannelTable.size()); |
374 | assert(Offset < SubRegFromChannelTable[TableIdx].size()); |
375 | SubRegFromChannelTable[TableIdx][Offset] = Idx; |
376 | } |
377 | }; |
378 | |
379 | llvm::call_once(flag&: InitializeRegSplitPartsFlag, F&: InitializeRegSplitPartsOnce); |
380 | llvm::call_once(flag&: InitializeSubRegFromChannelTableFlag, |
381 | F&: InitializeSubRegFromChannelTableOnce); |
382 | } |
383 | |
384 | void SIRegisterInfo::reserveRegisterTuples(BitVector &Reserved, |
385 | MCRegister Reg) const { |
386 | for (MCRegAliasIterator R(Reg, this, true); R.isValid(); ++R) |
387 | Reserved.set(*R); |
388 | } |
389 | |
390 | // Forced to be here by one .inc |
391 | const MCPhysReg *SIRegisterInfo::getCalleeSavedRegs( |
392 | const MachineFunction *MF) const { |
393 | CallingConv::ID CC = MF->getFunction().getCallingConv(); |
394 | switch (CC) { |
395 | case CallingConv::C: |
396 | case CallingConv::Fast: |
397 | case CallingConv::Cold: |
398 | return ST.hasGFX90AInsts() ? CSR_AMDGPU_GFX90AInsts_SaveList |
399 | : CSR_AMDGPU_SaveList; |
400 | case CallingConv::AMDGPU_Gfx: |
401 | return ST.hasGFX90AInsts() ? CSR_AMDGPU_SI_Gfx_GFX90AInsts_SaveList |
402 | : CSR_AMDGPU_SI_Gfx_SaveList; |
403 | case CallingConv::AMDGPU_CS_ChainPreserve: |
404 | return CSR_AMDGPU_CS_ChainPreserve_SaveList; |
405 | default: { |
406 | // Dummy to not crash RegisterClassInfo. |
407 | static const MCPhysReg NoCalleeSavedReg = AMDGPU::NoRegister; |
408 | return &NoCalleeSavedReg; |
409 | } |
410 | } |
411 | } |
412 | |
413 | const MCPhysReg * |
414 | SIRegisterInfo::getCalleeSavedRegsViaCopy(const MachineFunction *MF) const { |
415 | return nullptr; |
416 | } |
417 | |
418 | const uint32_t *SIRegisterInfo::getCallPreservedMask(const MachineFunction &MF, |
419 | CallingConv::ID CC) const { |
420 | switch (CC) { |
421 | case CallingConv::C: |
422 | case CallingConv::Fast: |
423 | case CallingConv::Cold: |
424 | return ST.hasGFX90AInsts() ? CSR_AMDGPU_GFX90AInsts_RegMask |
425 | : CSR_AMDGPU_RegMask; |
426 | case CallingConv::AMDGPU_Gfx: |
427 | return ST.hasGFX90AInsts() ? CSR_AMDGPU_SI_Gfx_GFX90AInsts_RegMask |
428 | : CSR_AMDGPU_SI_Gfx_RegMask; |
429 | case CallingConv::AMDGPU_CS_Chain: |
430 | case CallingConv::AMDGPU_CS_ChainPreserve: |
431 | // Calls to these functions never return, so we can pretend everything is |
432 | // preserved. |
433 | return AMDGPU_AllVGPRs_RegMask; |
434 | default: |
435 | return nullptr; |
436 | } |
437 | } |
438 | |
439 | const uint32_t *SIRegisterInfo::getNoPreservedMask() const { |
440 | return CSR_AMDGPU_NoRegs_RegMask; |
441 | } |
442 | |
443 | bool SIRegisterInfo::isChainScratchRegister(Register VGPR) { |
444 | return VGPR >= AMDGPU::VGPR0 && VGPR < AMDGPU::VGPR8; |
445 | } |
446 | |
447 | const TargetRegisterClass * |
448 | SIRegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC, |
449 | const MachineFunction &MF) const { |
450 | // FIXME: Should have a helper function like getEquivalentVGPRClass to get the |
451 | // equivalent AV class. If used one, the verifier will crash after |
452 | // RegBankSelect in the GISel flow. The aligned regclasses are not fully given |
453 | // until Instruction selection. |
454 | if (ST.hasMAIInsts() && (isVGPRClass(RC) || isAGPRClass(RC))) { |
455 | if (RC == &AMDGPU::VGPR_32RegClass || RC == &AMDGPU::AGPR_32RegClass) |
456 | return &AMDGPU::AV_32RegClass; |
457 | if (RC == &AMDGPU::VReg_64RegClass || RC == &AMDGPU::AReg_64RegClass) |
458 | return &AMDGPU::AV_64RegClass; |
459 | if (RC == &AMDGPU::VReg_64_Align2RegClass || |
460 | RC == &AMDGPU::AReg_64_Align2RegClass) |
461 | return &AMDGPU::AV_64_Align2RegClass; |
462 | if (RC == &AMDGPU::VReg_96RegClass || RC == &AMDGPU::AReg_96RegClass) |
463 | return &AMDGPU::AV_96RegClass; |
464 | if (RC == &AMDGPU::VReg_96_Align2RegClass || |
465 | RC == &AMDGPU::AReg_96_Align2RegClass) |
466 | return &AMDGPU::AV_96_Align2RegClass; |
467 | if (RC == &AMDGPU::VReg_128RegClass || RC == &AMDGPU::AReg_128RegClass) |
468 | return &AMDGPU::AV_128RegClass; |
469 | if (RC == &AMDGPU::VReg_128_Align2RegClass || |
470 | RC == &AMDGPU::AReg_128_Align2RegClass) |
471 | return &AMDGPU::AV_128_Align2RegClass; |
472 | if (RC == &AMDGPU::VReg_160RegClass || RC == &AMDGPU::AReg_160RegClass) |
473 | return &AMDGPU::AV_160RegClass; |
474 | if (RC == &AMDGPU::VReg_160_Align2RegClass || |
475 | RC == &AMDGPU::AReg_160_Align2RegClass) |
476 | return &AMDGPU::AV_160_Align2RegClass; |
477 | if (RC == &AMDGPU::VReg_192RegClass || RC == &AMDGPU::AReg_192RegClass) |
478 | return &AMDGPU::AV_192RegClass; |
479 | if (RC == &AMDGPU::VReg_192_Align2RegClass || |
480 | RC == &AMDGPU::AReg_192_Align2RegClass) |
481 | return &AMDGPU::AV_192_Align2RegClass; |
482 | if (RC == &AMDGPU::VReg_256RegClass || RC == &AMDGPU::AReg_256RegClass) |
483 | return &AMDGPU::AV_256RegClass; |
484 | if (RC == &AMDGPU::VReg_256_Align2RegClass || |
485 | RC == &AMDGPU::AReg_256_Align2RegClass) |
486 | return &AMDGPU::AV_256_Align2RegClass; |
487 | if (RC == &AMDGPU::VReg_512RegClass || RC == &AMDGPU::AReg_512RegClass) |
488 | return &AMDGPU::AV_512RegClass; |
489 | if (RC == &AMDGPU::VReg_512_Align2RegClass || |
490 | RC == &AMDGPU::AReg_512_Align2RegClass) |
491 | return &AMDGPU::AV_512_Align2RegClass; |
492 | if (RC == &AMDGPU::VReg_1024RegClass || RC == &AMDGPU::AReg_1024RegClass) |
493 | return &AMDGPU::AV_1024RegClass; |
494 | if (RC == &AMDGPU::VReg_1024_Align2RegClass || |
495 | RC == &AMDGPU::AReg_1024_Align2RegClass) |
496 | return &AMDGPU::AV_1024_Align2RegClass; |
497 | } |
498 | |
499 | return TargetRegisterInfo::getLargestLegalSuperClass(RC, MF); |
500 | } |
501 | |
502 | Register SIRegisterInfo::getFrameRegister(const MachineFunction &MF) const { |
503 | const SIFrameLowering *TFI = ST.getFrameLowering(); |
504 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
505 | // During ISel lowering we always reserve the stack pointer in entry and chain |
506 | // functions, but never actually want to reference it when accessing our own |
507 | // frame. If we need a frame pointer we use it, but otherwise we can just use |
508 | // an immediate "0" which we represent by returning NoRegister. |
509 | if (FuncInfo->isBottomOfStack()) { |
510 | return TFI->hasFP(MF) ? FuncInfo->getFrameOffsetReg() : Register(); |
511 | } |
512 | return TFI->hasFP(MF) ? FuncInfo->getFrameOffsetReg() |
513 | : FuncInfo->getStackPtrOffsetReg(); |
514 | } |
515 | |
516 | bool SIRegisterInfo::hasBasePointer(const MachineFunction &MF) const { |
517 | // When we need stack realignment, we can't reference off of the |
518 | // stack pointer, so we reserve a base pointer. |
519 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
520 | return MFI.getNumFixedObjects() && shouldRealignStack(MF); |
521 | } |
522 | |
523 | Register SIRegisterInfo::getBaseRegister() const { return AMDGPU::SGPR34; } |
524 | |
525 | const uint32_t *SIRegisterInfo::getAllVGPRRegMask() const { |
526 | return AMDGPU_AllVGPRs_RegMask; |
527 | } |
528 | |
529 | const uint32_t *SIRegisterInfo::getAllAGPRRegMask() const { |
530 | return AMDGPU_AllAGPRs_RegMask; |
531 | } |
532 | |
533 | const uint32_t *SIRegisterInfo::getAllVectorRegMask() const { |
534 | return AMDGPU_AllVectorRegs_RegMask; |
535 | } |
536 | |
537 | const uint32_t *SIRegisterInfo::getAllAllocatableSRegMask() const { |
538 | return AMDGPU_AllAllocatableSRegs_RegMask; |
539 | } |
540 | |
541 | unsigned SIRegisterInfo::getSubRegFromChannel(unsigned Channel, |
542 | unsigned NumRegs) { |
543 | assert(NumRegs < SubRegFromChannelTableWidthMap.size()); |
544 | unsigned NumRegIndex = SubRegFromChannelTableWidthMap[NumRegs]; |
545 | assert(NumRegIndex && "Not implemented" ); |
546 | assert(Channel < SubRegFromChannelTable[NumRegIndex - 1].size()); |
547 | return SubRegFromChannelTable[NumRegIndex - 1][Channel]; |
548 | } |
549 | |
550 | MCRegister |
551 | SIRegisterInfo::getAlignedHighSGPRForRC(const MachineFunction &MF, |
552 | const unsigned Align, |
553 | const TargetRegisterClass *RC) const { |
554 | unsigned BaseIdx = alignDown(Value: ST.getMaxNumSGPRs(MF), Align) - Align; |
555 | MCRegister BaseReg(AMDGPU::SGPR_32RegClass.getRegister(i: BaseIdx)); |
556 | return getMatchingSuperReg(Reg: BaseReg, SubIdx: AMDGPU::sub0, RC); |
557 | } |
558 | |
559 | MCRegister SIRegisterInfo::reservedPrivateSegmentBufferReg( |
560 | const MachineFunction &MF) const { |
561 | return getAlignedHighSGPRForRC(MF, /*Align=*/4, RC: &AMDGPU::SGPR_128RegClass); |
562 | } |
563 | |
564 | BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const { |
565 | BitVector Reserved(getNumRegs()); |
566 | Reserved.set(AMDGPU::MODE); |
567 | |
568 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
569 | |
570 | // Reserve special purpose registers. |
571 | // |
572 | // EXEC_LO and EXEC_HI could be allocated and used as regular register, but |
573 | // this seems likely to result in bugs, so I'm marking them as reserved. |
574 | reserveRegisterTuples(Reserved, Reg: AMDGPU::EXEC); |
575 | reserveRegisterTuples(Reserved, Reg: AMDGPU::FLAT_SCR); |
576 | |
577 | // M0 has to be reserved so that llvm accepts it as a live-in into a block. |
578 | reserveRegisterTuples(Reserved, Reg: AMDGPU::M0); |
579 | |
580 | // Reserve src_vccz, src_execz, src_scc. |
581 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_VCCZ); |
582 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_EXECZ); |
583 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_SCC); |
584 | |
585 | // Reserve the memory aperture registers |
586 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_SHARED_BASE); |
587 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_SHARED_LIMIT); |
588 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_PRIVATE_BASE); |
589 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_PRIVATE_LIMIT); |
590 | |
591 | // Reserve src_pops_exiting_wave_id - support is not implemented in Codegen. |
592 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SRC_POPS_EXITING_WAVE_ID); |
593 | |
594 | // Reserve xnack_mask registers - support is not implemented in Codegen. |
595 | reserveRegisterTuples(Reserved, Reg: AMDGPU::XNACK_MASK); |
596 | |
597 | // Reserve lds_direct register - support is not implemented in Codegen. |
598 | reserveRegisterTuples(Reserved, Reg: AMDGPU::LDS_DIRECT); |
599 | |
600 | // Reserve Trap Handler registers - support is not implemented in Codegen. |
601 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TBA); |
602 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TMA); |
603 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP0_TTMP1); |
604 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP2_TTMP3); |
605 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP4_TTMP5); |
606 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP6_TTMP7); |
607 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP8_TTMP9); |
608 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP10_TTMP11); |
609 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP12_TTMP13); |
610 | reserveRegisterTuples(Reserved, Reg: AMDGPU::TTMP14_TTMP15); |
611 | |
612 | // Reserve null register - it shall never be allocated |
613 | reserveRegisterTuples(Reserved, Reg: AMDGPU::SGPR_NULL64); |
614 | |
615 | // Reserve SGPRs. |
616 | // |
617 | unsigned MaxNumSGPRs = ST.getMaxNumSGPRs(MF); |
618 | unsigned TotalNumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs(); |
619 | for (const TargetRegisterClass *RC : regclasses()) { |
620 | if (RC->isBaseClass() && isSGPRClass(RC)) { |
621 | unsigned NumRegs = divideCeil(Numerator: getRegSizeInBits(RC: *RC), Denominator: 32); |
622 | for (MCPhysReg Reg : *RC) { |
623 | unsigned Index = getHWRegIndex(Reg); |
624 | if (Index + NumRegs > MaxNumSGPRs && Index < TotalNumSGPRs) |
625 | Reserved.set(Reg); |
626 | } |
627 | } |
628 | } |
629 | |
630 | Register ScratchRSrcReg = MFI->getScratchRSrcReg(); |
631 | if (ScratchRSrcReg != AMDGPU::NoRegister) { |
632 | // Reserve 4 SGPRs for the scratch buffer resource descriptor in case we |
633 | // need to spill. |
634 | // TODO: May need to reserve a VGPR if doing LDS spilling. |
635 | reserveRegisterTuples(Reserved, Reg: ScratchRSrcReg); |
636 | } |
637 | |
638 | Register LongBranchReservedReg = MFI->getLongBranchReservedReg(); |
639 | if (LongBranchReservedReg) |
640 | reserveRegisterTuples(Reserved, Reg: LongBranchReservedReg); |
641 | |
642 | // We have to assume the SP is needed in case there are calls in the function, |
643 | // which is detected after the function is lowered. If we aren't really going |
644 | // to need SP, don't bother reserving it. |
645 | MCRegister StackPtrReg = MFI->getStackPtrOffsetReg(); |
646 | if (StackPtrReg) { |
647 | reserveRegisterTuples(Reserved, Reg: StackPtrReg); |
648 | assert(!isSubRegister(ScratchRSrcReg, StackPtrReg)); |
649 | } |
650 | |
651 | MCRegister FrameReg = MFI->getFrameOffsetReg(); |
652 | if (FrameReg) { |
653 | reserveRegisterTuples(Reserved, Reg: FrameReg); |
654 | assert(!isSubRegister(ScratchRSrcReg, FrameReg)); |
655 | } |
656 | |
657 | if (hasBasePointer(MF)) { |
658 | MCRegister BasePtrReg = getBaseRegister(); |
659 | reserveRegisterTuples(Reserved, Reg: BasePtrReg); |
660 | assert(!isSubRegister(ScratchRSrcReg, BasePtrReg)); |
661 | } |
662 | |
663 | // FIXME: Use same reserved register introduced in D149775 |
664 | // SGPR used to preserve EXEC MASK around WWM spill/copy instructions. |
665 | Register ExecCopyReg = MFI->getSGPRForEXECCopy(); |
666 | if (ExecCopyReg) |
667 | reserveRegisterTuples(Reserved, Reg: ExecCopyReg); |
668 | |
669 | // Reserve VGPRs/AGPRs. |
670 | // |
671 | unsigned MaxNumVGPRs = ST.getMaxNumVGPRs(MF); |
672 | unsigned MaxNumAGPRs = MaxNumVGPRs; |
673 | unsigned TotalNumVGPRs = AMDGPU::VGPR_32RegClass.getNumRegs(); |
674 | |
675 | // On GFX90A, the number of VGPRs and AGPRs need not be equal. Theoretically, |
676 | // a wave may have up to 512 total vector registers combining together both |
677 | // VGPRs and AGPRs. Hence, in an entry function without calls and without |
678 | // AGPRs used within it, it is possible to use the whole vector register |
679 | // budget for VGPRs. |
680 | // |
681 | // TODO: it shall be possible to estimate maximum AGPR/VGPR pressure and split |
682 | // register file accordingly. |
683 | if (ST.hasGFX90AInsts()) { |
684 | if (MFI->usesAGPRs(MF)) { |
685 | MaxNumVGPRs /= 2; |
686 | MaxNumAGPRs = MaxNumVGPRs; |
687 | } else { |
688 | if (MaxNumVGPRs > TotalNumVGPRs) { |
689 | MaxNumAGPRs = MaxNumVGPRs - TotalNumVGPRs; |
690 | MaxNumVGPRs = TotalNumVGPRs; |
691 | } else |
692 | MaxNumAGPRs = 0; |
693 | } |
694 | } |
695 | |
696 | for (const TargetRegisterClass *RC : regclasses()) { |
697 | if (RC->isBaseClass() && isVGPRClass(RC)) { |
698 | unsigned NumRegs = divideCeil(Numerator: getRegSizeInBits(RC: *RC), Denominator: 32); |
699 | for (MCPhysReg Reg : *RC) { |
700 | unsigned Index = getHWRegIndex(Reg); |
701 | if (Index + NumRegs > MaxNumVGPRs) |
702 | Reserved.set(Reg); |
703 | } |
704 | } |
705 | } |
706 | |
707 | // Reserve all the AGPRs if there are no instructions to use it. |
708 | if (!ST.hasMAIInsts()) |
709 | MaxNumAGPRs = 0; |
710 | for (const TargetRegisterClass *RC : regclasses()) { |
711 | if (RC->isBaseClass() && isAGPRClass(RC)) { |
712 | unsigned NumRegs = divideCeil(Numerator: getRegSizeInBits(RC: *RC), Denominator: 32); |
713 | for (MCPhysReg Reg : *RC) { |
714 | unsigned Index = getHWRegIndex(Reg); |
715 | if (Index + NumRegs > MaxNumAGPRs) |
716 | Reserved.set(Reg); |
717 | } |
718 | } |
719 | } |
720 | |
721 | // On GFX908, in order to guarantee copying between AGPRs, we need a scratch |
722 | // VGPR available at all times. |
723 | if (ST.hasMAIInsts() && !ST.hasGFX90AInsts()) { |
724 | reserveRegisterTuples(Reserved, Reg: MFI->getVGPRForAGPRCopy()); |
725 | } |
726 | |
727 | for (Register Reg : MFI->getWWMReservedRegs()) |
728 | reserveRegisterTuples(Reserved, Reg); |
729 | |
730 | // FIXME: Stop using reserved registers for this. |
731 | for (MCPhysReg Reg : MFI->getAGPRSpillVGPRs()) |
732 | reserveRegisterTuples(Reserved, Reg); |
733 | |
734 | for (MCPhysReg Reg : MFI->getVGPRSpillAGPRs()) |
735 | reserveRegisterTuples(Reserved, Reg); |
736 | |
737 | return Reserved; |
738 | } |
739 | |
740 | bool SIRegisterInfo::isAsmClobberable(const MachineFunction &MF, |
741 | MCRegister PhysReg) const { |
742 | return !MF.getRegInfo().isReserved(PhysReg); |
743 | } |
744 | |
745 | bool SIRegisterInfo::shouldRealignStack(const MachineFunction &MF) const { |
746 | const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); |
747 | // On entry or in chain functions, the base address is 0, so it can't possibly |
748 | // need any more alignment. |
749 | |
750 | // FIXME: Should be able to specify the entry frame alignment per calling |
751 | // convention instead. |
752 | if (Info->isBottomOfStack()) |
753 | return false; |
754 | |
755 | return TargetRegisterInfo::shouldRealignStack(MF); |
756 | } |
757 | |
758 | bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const { |
759 | const SIMachineFunctionInfo *Info = Fn.getInfo<SIMachineFunctionInfo>(); |
760 | if (Info->isEntryFunction()) { |
761 | const MachineFrameInfo &MFI = Fn.getFrameInfo(); |
762 | return MFI.hasStackObjects() || MFI.hasCalls(); |
763 | } |
764 | |
765 | // May need scavenger for dealing with callee saved registers. |
766 | return true; |
767 | } |
768 | |
769 | bool SIRegisterInfo::requiresFrameIndexScavenging( |
770 | const MachineFunction &MF) const { |
771 | // Do not use frame virtual registers. They used to be used for SGPRs, but |
772 | // once we reach PrologEpilogInserter, we can no longer spill SGPRs. If the |
773 | // scavenger fails, we can increment/decrement the necessary SGPRs to avoid a |
774 | // spill. |
775 | return false; |
776 | } |
777 | |
778 | bool SIRegisterInfo::requiresFrameIndexReplacementScavenging( |
779 | const MachineFunction &MF) const { |
780 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
781 | return MFI.hasStackObjects(); |
782 | } |
783 | |
784 | bool SIRegisterInfo::requiresVirtualBaseRegisters( |
785 | const MachineFunction &) const { |
786 | // There are no special dedicated stack or frame pointers. |
787 | return true; |
788 | } |
789 | |
790 | int64_t SIRegisterInfo::getScratchInstrOffset(const MachineInstr *MI) const { |
791 | assert(SIInstrInfo::isMUBUF(*MI) || SIInstrInfo::isFLATScratch(*MI)); |
792 | |
793 | int OffIdx = AMDGPU::getNamedOperandIdx(Opcode: MI->getOpcode(), |
794 | NamedIdx: AMDGPU::OpName::offset); |
795 | return MI->getOperand(i: OffIdx).getImm(); |
796 | } |
797 | |
798 | int64_t SIRegisterInfo::getFrameIndexInstrOffset(const MachineInstr *MI, |
799 | int Idx) const { |
800 | if (!SIInstrInfo::isMUBUF(MI: *MI) && !SIInstrInfo::isFLATScratch(MI: *MI)) |
801 | return 0; |
802 | |
803 | assert((Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), |
804 | AMDGPU::OpName::vaddr) || |
805 | (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), |
806 | AMDGPU::OpName::saddr))) && |
807 | "Should never see frame index on non-address operand" ); |
808 | |
809 | return getScratchInstrOffset(MI); |
810 | } |
811 | |
812 | bool SIRegisterInfo::needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { |
813 | if (!SIInstrInfo::isMUBUF(MI: *MI) && !SIInstrInfo::isFLATScratch(MI: *MI)) |
814 | return false; |
815 | |
816 | int64_t FullOffset = Offset + getScratchInstrOffset(MI); |
817 | |
818 | const SIInstrInfo *TII = ST.getInstrInfo(); |
819 | if (SIInstrInfo::isMUBUF(MI: *MI)) |
820 | return !TII->isLegalMUBUFImmOffset(Imm: FullOffset); |
821 | |
822 | return !TII->isLegalFLATOffset(Offset: FullOffset, AddrSpace: AMDGPUAS::PRIVATE_ADDRESS, |
823 | FlatVariant: SIInstrFlags::FlatScratch); |
824 | } |
825 | |
826 | Register SIRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, |
827 | int FrameIdx, |
828 | int64_t Offset) const { |
829 | MachineBasicBlock::iterator Ins = MBB->begin(); |
830 | DebugLoc DL; // Defaults to "unknown" |
831 | |
832 | if (Ins != MBB->end()) |
833 | DL = Ins->getDebugLoc(); |
834 | |
835 | MachineFunction *MF = MBB->getParent(); |
836 | const SIInstrInfo *TII = ST.getInstrInfo(); |
837 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
838 | unsigned MovOpc = ST.enableFlatScratch() ? AMDGPU::S_MOV_B32 |
839 | : AMDGPU::V_MOV_B32_e32; |
840 | |
841 | Register BaseReg = MRI.createVirtualRegister( |
842 | RegClass: ST.enableFlatScratch() ? &AMDGPU::SReg_32_XEXEC_HIRegClass |
843 | : &AMDGPU::VGPR_32RegClass); |
844 | |
845 | if (Offset == 0) { |
846 | BuildMI(BB&: *MBB, I: Ins, MIMD: DL, MCID: TII->get(Opcode: MovOpc), DestReg: BaseReg) |
847 | .addFrameIndex(Idx: FrameIdx); |
848 | return BaseReg; |
849 | } |
850 | |
851 | Register OffsetReg = MRI.createVirtualRegister(RegClass: &AMDGPU::SReg_32_XM0RegClass); |
852 | |
853 | Register FIReg = MRI.createVirtualRegister( |
854 | RegClass: ST.enableFlatScratch() ? &AMDGPU::SReg_32_XM0RegClass |
855 | : &AMDGPU::VGPR_32RegClass); |
856 | |
857 | BuildMI(BB&: *MBB, I: Ins, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: OffsetReg) |
858 | .addImm(Val: Offset); |
859 | BuildMI(BB&: *MBB, I: Ins, MIMD: DL, MCID: TII->get(Opcode: MovOpc), DestReg: FIReg) |
860 | .addFrameIndex(Idx: FrameIdx); |
861 | |
862 | if (ST.enableFlatScratch() ) { |
863 | BuildMI(BB&: *MBB, I: Ins, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: BaseReg) |
864 | .addReg(RegNo: OffsetReg, flags: RegState::Kill) |
865 | .addReg(RegNo: FIReg); |
866 | return BaseReg; |
867 | } |
868 | |
869 | TII->getAddNoCarry(MBB&: *MBB, I: Ins, DL, DestReg: BaseReg) |
870 | .addReg(RegNo: OffsetReg, flags: RegState::Kill) |
871 | .addReg(RegNo: FIReg) |
872 | .addImm(Val: 0); // clamp bit |
873 | |
874 | return BaseReg; |
875 | } |
876 | |
877 | void SIRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, |
878 | int64_t Offset) const { |
879 | const SIInstrInfo *TII = ST.getInstrInfo(); |
880 | bool IsFlat = TII->isFLATScratch(MI); |
881 | |
882 | #ifndef NDEBUG |
883 | // FIXME: Is it possible to be storing a frame index to itself? |
884 | bool SeenFI = false; |
885 | for (const MachineOperand &MO: MI.operands()) { |
886 | if (MO.isFI()) { |
887 | if (SeenFI) |
888 | llvm_unreachable("should not see multiple frame indices" ); |
889 | |
890 | SeenFI = true; |
891 | } |
892 | } |
893 | #endif |
894 | |
895 | MachineOperand *FIOp = |
896 | TII->getNamedOperand(MI, OperandName: IsFlat ? AMDGPU::OpName::saddr |
897 | : AMDGPU::OpName::vaddr); |
898 | |
899 | MachineOperand *OffsetOp = TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::offset); |
900 | int64_t NewOffset = OffsetOp->getImm() + Offset; |
901 | |
902 | assert(FIOp && FIOp->isFI() && "frame index must be address operand" ); |
903 | assert(TII->isMUBUF(MI) || TII->isFLATScratch(MI)); |
904 | |
905 | if (IsFlat) { |
906 | assert(TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, |
907 | SIInstrFlags::FlatScratch) && |
908 | "offset should be legal" ); |
909 | FIOp->ChangeToRegister(Reg: BaseReg, isDef: false); |
910 | OffsetOp->setImm(NewOffset); |
911 | return; |
912 | } |
913 | |
914 | #ifndef NDEBUG |
915 | MachineOperand *SOffset = TII->getNamedOperand(MI, AMDGPU::OpName::soffset); |
916 | assert(SOffset->isImm() && SOffset->getImm() == 0); |
917 | #endif |
918 | |
919 | assert(TII->isLegalMUBUFImmOffset(NewOffset) && "offset should be legal" ); |
920 | |
921 | FIOp->ChangeToRegister(Reg: BaseReg, isDef: false); |
922 | OffsetOp->setImm(NewOffset); |
923 | } |
924 | |
925 | bool SIRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, |
926 | Register BaseReg, |
927 | int64_t Offset) const { |
928 | if (!SIInstrInfo::isMUBUF(MI: *MI) && !SIInstrInfo::isFLATScratch(MI: *MI)) |
929 | return false; |
930 | |
931 | int64_t NewOffset = Offset + getScratchInstrOffset(MI); |
932 | |
933 | const SIInstrInfo *TII = ST.getInstrInfo(); |
934 | if (SIInstrInfo::isMUBUF(MI: *MI)) |
935 | return TII->isLegalMUBUFImmOffset(Imm: NewOffset); |
936 | |
937 | return TII->isLegalFLATOffset(Offset: NewOffset, AddrSpace: AMDGPUAS::PRIVATE_ADDRESS, |
938 | FlatVariant: SIInstrFlags::FlatScratch); |
939 | } |
940 | |
941 | const TargetRegisterClass *SIRegisterInfo::getPointerRegClass( |
942 | const MachineFunction &MF, unsigned Kind) const { |
943 | // This is inaccurate. It depends on the instruction and address space. The |
944 | // only place where we should hit this is for dealing with frame indexes / |
945 | // private accesses, so this is correct in that case. |
946 | return &AMDGPU::VGPR_32RegClass; |
947 | } |
948 | |
949 | const TargetRegisterClass * |
950 | SIRegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { |
951 | if (isAGPRClass(RC) && !ST.hasGFX90AInsts()) |
952 | return getEquivalentVGPRClass(SRC: RC); |
953 | if (RC == &AMDGPU::SCC_CLASSRegClass) |
954 | return getWaveMaskRegClass(); |
955 | |
956 | return RC; |
957 | } |
958 | |
959 | static unsigned getNumSubRegsForSpillOp(unsigned Op) { |
960 | |
961 | switch (Op) { |
962 | case AMDGPU::SI_SPILL_S1024_SAVE: |
963 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
964 | case AMDGPU::SI_SPILL_V1024_SAVE: |
965 | case AMDGPU::SI_SPILL_V1024_RESTORE: |
966 | case AMDGPU::SI_SPILL_A1024_SAVE: |
967 | case AMDGPU::SI_SPILL_A1024_RESTORE: |
968 | case AMDGPU::SI_SPILL_AV1024_SAVE: |
969 | case AMDGPU::SI_SPILL_AV1024_RESTORE: |
970 | return 32; |
971 | case AMDGPU::SI_SPILL_S512_SAVE: |
972 | case AMDGPU::SI_SPILL_S512_RESTORE: |
973 | case AMDGPU::SI_SPILL_V512_SAVE: |
974 | case AMDGPU::SI_SPILL_V512_RESTORE: |
975 | case AMDGPU::SI_SPILL_A512_SAVE: |
976 | case AMDGPU::SI_SPILL_A512_RESTORE: |
977 | case AMDGPU::SI_SPILL_AV512_SAVE: |
978 | case AMDGPU::SI_SPILL_AV512_RESTORE: |
979 | return 16; |
980 | case AMDGPU::SI_SPILL_S384_SAVE: |
981 | case AMDGPU::SI_SPILL_S384_RESTORE: |
982 | case AMDGPU::SI_SPILL_V384_SAVE: |
983 | case AMDGPU::SI_SPILL_V384_RESTORE: |
984 | case AMDGPU::SI_SPILL_A384_SAVE: |
985 | case AMDGPU::SI_SPILL_A384_RESTORE: |
986 | case AMDGPU::SI_SPILL_AV384_SAVE: |
987 | case AMDGPU::SI_SPILL_AV384_RESTORE: |
988 | return 12; |
989 | case AMDGPU::SI_SPILL_S352_SAVE: |
990 | case AMDGPU::SI_SPILL_S352_RESTORE: |
991 | case AMDGPU::SI_SPILL_V352_SAVE: |
992 | case AMDGPU::SI_SPILL_V352_RESTORE: |
993 | case AMDGPU::SI_SPILL_A352_SAVE: |
994 | case AMDGPU::SI_SPILL_A352_RESTORE: |
995 | case AMDGPU::SI_SPILL_AV352_SAVE: |
996 | case AMDGPU::SI_SPILL_AV352_RESTORE: |
997 | return 11; |
998 | case AMDGPU::SI_SPILL_S320_SAVE: |
999 | case AMDGPU::SI_SPILL_S320_RESTORE: |
1000 | case AMDGPU::SI_SPILL_V320_SAVE: |
1001 | case AMDGPU::SI_SPILL_V320_RESTORE: |
1002 | case AMDGPU::SI_SPILL_A320_SAVE: |
1003 | case AMDGPU::SI_SPILL_A320_RESTORE: |
1004 | case AMDGPU::SI_SPILL_AV320_SAVE: |
1005 | case AMDGPU::SI_SPILL_AV320_RESTORE: |
1006 | return 10; |
1007 | case AMDGPU::SI_SPILL_S288_SAVE: |
1008 | case AMDGPU::SI_SPILL_S288_RESTORE: |
1009 | case AMDGPU::SI_SPILL_V288_SAVE: |
1010 | case AMDGPU::SI_SPILL_V288_RESTORE: |
1011 | case AMDGPU::SI_SPILL_A288_SAVE: |
1012 | case AMDGPU::SI_SPILL_A288_RESTORE: |
1013 | case AMDGPU::SI_SPILL_AV288_SAVE: |
1014 | case AMDGPU::SI_SPILL_AV288_RESTORE: |
1015 | return 9; |
1016 | case AMDGPU::SI_SPILL_S256_SAVE: |
1017 | case AMDGPU::SI_SPILL_S256_RESTORE: |
1018 | case AMDGPU::SI_SPILL_V256_SAVE: |
1019 | case AMDGPU::SI_SPILL_V256_RESTORE: |
1020 | case AMDGPU::SI_SPILL_A256_SAVE: |
1021 | case AMDGPU::SI_SPILL_A256_RESTORE: |
1022 | case AMDGPU::SI_SPILL_AV256_SAVE: |
1023 | case AMDGPU::SI_SPILL_AV256_RESTORE: |
1024 | return 8; |
1025 | case AMDGPU::SI_SPILL_S224_SAVE: |
1026 | case AMDGPU::SI_SPILL_S224_RESTORE: |
1027 | case AMDGPU::SI_SPILL_V224_SAVE: |
1028 | case AMDGPU::SI_SPILL_V224_RESTORE: |
1029 | case AMDGPU::SI_SPILL_A224_SAVE: |
1030 | case AMDGPU::SI_SPILL_A224_RESTORE: |
1031 | case AMDGPU::SI_SPILL_AV224_SAVE: |
1032 | case AMDGPU::SI_SPILL_AV224_RESTORE: |
1033 | return 7; |
1034 | case AMDGPU::SI_SPILL_S192_SAVE: |
1035 | case AMDGPU::SI_SPILL_S192_RESTORE: |
1036 | case AMDGPU::SI_SPILL_V192_SAVE: |
1037 | case AMDGPU::SI_SPILL_V192_RESTORE: |
1038 | case AMDGPU::SI_SPILL_A192_SAVE: |
1039 | case AMDGPU::SI_SPILL_A192_RESTORE: |
1040 | case AMDGPU::SI_SPILL_AV192_SAVE: |
1041 | case AMDGPU::SI_SPILL_AV192_RESTORE: |
1042 | return 6; |
1043 | case AMDGPU::SI_SPILL_S160_SAVE: |
1044 | case AMDGPU::SI_SPILL_S160_RESTORE: |
1045 | case AMDGPU::SI_SPILL_V160_SAVE: |
1046 | case AMDGPU::SI_SPILL_V160_RESTORE: |
1047 | case AMDGPU::SI_SPILL_A160_SAVE: |
1048 | case AMDGPU::SI_SPILL_A160_RESTORE: |
1049 | case AMDGPU::SI_SPILL_AV160_SAVE: |
1050 | case AMDGPU::SI_SPILL_AV160_RESTORE: |
1051 | return 5; |
1052 | case AMDGPU::SI_SPILL_S128_SAVE: |
1053 | case AMDGPU::SI_SPILL_S128_RESTORE: |
1054 | case AMDGPU::SI_SPILL_V128_SAVE: |
1055 | case AMDGPU::SI_SPILL_V128_RESTORE: |
1056 | case AMDGPU::SI_SPILL_A128_SAVE: |
1057 | case AMDGPU::SI_SPILL_A128_RESTORE: |
1058 | case AMDGPU::SI_SPILL_AV128_SAVE: |
1059 | case AMDGPU::SI_SPILL_AV128_RESTORE: |
1060 | return 4; |
1061 | case AMDGPU::SI_SPILL_S96_SAVE: |
1062 | case AMDGPU::SI_SPILL_S96_RESTORE: |
1063 | case AMDGPU::SI_SPILL_V96_SAVE: |
1064 | case AMDGPU::SI_SPILL_V96_RESTORE: |
1065 | case AMDGPU::SI_SPILL_A96_SAVE: |
1066 | case AMDGPU::SI_SPILL_A96_RESTORE: |
1067 | case AMDGPU::SI_SPILL_AV96_SAVE: |
1068 | case AMDGPU::SI_SPILL_AV96_RESTORE: |
1069 | return 3; |
1070 | case AMDGPU::SI_SPILL_S64_SAVE: |
1071 | case AMDGPU::SI_SPILL_S64_RESTORE: |
1072 | case AMDGPU::SI_SPILL_V64_SAVE: |
1073 | case AMDGPU::SI_SPILL_V64_RESTORE: |
1074 | case AMDGPU::SI_SPILL_A64_SAVE: |
1075 | case AMDGPU::SI_SPILL_A64_RESTORE: |
1076 | case AMDGPU::SI_SPILL_AV64_SAVE: |
1077 | case AMDGPU::SI_SPILL_AV64_RESTORE: |
1078 | return 2; |
1079 | case AMDGPU::SI_SPILL_S32_SAVE: |
1080 | case AMDGPU::SI_SPILL_S32_RESTORE: |
1081 | case AMDGPU::SI_SPILL_V32_SAVE: |
1082 | case AMDGPU::SI_SPILL_V32_RESTORE: |
1083 | case AMDGPU::SI_SPILL_A32_SAVE: |
1084 | case AMDGPU::SI_SPILL_A32_RESTORE: |
1085 | case AMDGPU::SI_SPILL_AV32_SAVE: |
1086 | case AMDGPU::SI_SPILL_AV32_RESTORE: |
1087 | case AMDGPU::SI_SPILL_WWM_V32_SAVE: |
1088 | case AMDGPU::SI_SPILL_WWM_V32_RESTORE: |
1089 | case AMDGPU::SI_SPILL_WWM_AV32_SAVE: |
1090 | case AMDGPU::SI_SPILL_WWM_AV32_RESTORE: |
1091 | return 1; |
1092 | default: llvm_unreachable("Invalid spill opcode" ); |
1093 | } |
1094 | } |
1095 | |
1096 | static int getOffsetMUBUFStore(unsigned Opc) { |
1097 | switch (Opc) { |
1098 | case AMDGPU::BUFFER_STORE_DWORD_OFFEN: |
1099 | return AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
1100 | case AMDGPU::BUFFER_STORE_BYTE_OFFEN: |
1101 | return AMDGPU::BUFFER_STORE_BYTE_OFFSET; |
1102 | case AMDGPU::BUFFER_STORE_SHORT_OFFEN: |
1103 | return AMDGPU::BUFFER_STORE_SHORT_OFFSET; |
1104 | case AMDGPU::BUFFER_STORE_DWORDX2_OFFEN: |
1105 | return AMDGPU::BUFFER_STORE_DWORDX2_OFFSET; |
1106 | case AMDGPU::BUFFER_STORE_DWORDX3_OFFEN: |
1107 | return AMDGPU::BUFFER_STORE_DWORDX3_OFFSET; |
1108 | case AMDGPU::BUFFER_STORE_DWORDX4_OFFEN: |
1109 | return AMDGPU::BUFFER_STORE_DWORDX4_OFFSET; |
1110 | case AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFEN: |
1111 | return AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFSET; |
1112 | case AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFEN: |
1113 | return AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFSET; |
1114 | default: |
1115 | return -1; |
1116 | } |
1117 | } |
1118 | |
1119 | static int getOffsetMUBUFLoad(unsigned Opc) { |
1120 | switch (Opc) { |
1121 | case AMDGPU::BUFFER_LOAD_DWORD_OFFEN: |
1122 | return AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
1123 | case AMDGPU::BUFFER_LOAD_UBYTE_OFFEN: |
1124 | return AMDGPU::BUFFER_LOAD_UBYTE_OFFSET; |
1125 | case AMDGPU::BUFFER_LOAD_SBYTE_OFFEN: |
1126 | return AMDGPU::BUFFER_LOAD_SBYTE_OFFSET; |
1127 | case AMDGPU::BUFFER_LOAD_USHORT_OFFEN: |
1128 | return AMDGPU::BUFFER_LOAD_USHORT_OFFSET; |
1129 | case AMDGPU::BUFFER_LOAD_SSHORT_OFFEN: |
1130 | return AMDGPU::BUFFER_LOAD_SSHORT_OFFSET; |
1131 | case AMDGPU::BUFFER_LOAD_DWORDX2_OFFEN: |
1132 | return AMDGPU::BUFFER_LOAD_DWORDX2_OFFSET; |
1133 | case AMDGPU::BUFFER_LOAD_DWORDX3_OFFEN: |
1134 | return AMDGPU::BUFFER_LOAD_DWORDX3_OFFSET; |
1135 | case AMDGPU::BUFFER_LOAD_DWORDX4_OFFEN: |
1136 | return AMDGPU::BUFFER_LOAD_DWORDX4_OFFSET; |
1137 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFEN: |
1138 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFSET; |
1139 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFEN: |
1140 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFSET; |
1141 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFEN: |
1142 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFSET; |
1143 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFEN: |
1144 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFSET; |
1145 | case AMDGPU::BUFFER_LOAD_SHORT_D16_OFFEN: |
1146 | return AMDGPU::BUFFER_LOAD_SHORT_D16_OFFSET; |
1147 | case AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFEN: |
1148 | return AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFSET; |
1149 | default: |
1150 | return -1; |
1151 | } |
1152 | } |
1153 | |
1154 | static int getOffenMUBUFStore(unsigned Opc) { |
1155 | switch (Opc) { |
1156 | case AMDGPU::BUFFER_STORE_DWORD_OFFSET: |
1157 | return AMDGPU::BUFFER_STORE_DWORD_OFFEN; |
1158 | case AMDGPU::BUFFER_STORE_BYTE_OFFSET: |
1159 | return AMDGPU::BUFFER_STORE_BYTE_OFFEN; |
1160 | case AMDGPU::BUFFER_STORE_SHORT_OFFSET: |
1161 | return AMDGPU::BUFFER_STORE_SHORT_OFFEN; |
1162 | case AMDGPU::BUFFER_STORE_DWORDX2_OFFSET: |
1163 | return AMDGPU::BUFFER_STORE_DWORDX2_OFFEN; |
1164 | case AMDGPU::BUFFER_STORE_DWORDX3_OFFSET: |
1165 | return AMDGPU::BUFFER_STORE_DWORDX3_OFFEN; |
1166 | case AMDGPU::BUFFER_STORE_DWORDX4_OFFSET: |
1167 | return AMDGPU::BUFFER_STORE_DWORDX4_OFFEN; |
1168 | case AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFSET: |
1169 | return AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFEN; |
1170 | case AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFSET: |
1171 | return AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFEN; |
1172 | default: |
1173 | return -1; |
1174 | } |
1175 | } |
1176 | |
1177 | static int getOffenMUBUFLoad(unsigned Opc) { |
1178 | switch (Opc) { |
1179 | case AMDGPU::BUFFER_LOAD_DWORD_OFFSET: |
1180 | return AMDGPU::BUFFER_LOAD_DWORD_OFFEN; |
1181 | case AMDGPU::BUFFER_LOAD_UBYTE_OFFSET: |
1182 | return AMDGPU::BUFFER_LOAD_UBYTE_OFFEN; |
1183 | case AMDGPU::BUFFER_LOAD_SBYTE_OFFSET: |
1184 | return AMDGPU::BUFFER_LOAD_SBYTE_OFFEN; |
1185 | case AMDGPU::BUFFER_LOAD_USHORT_OFFSET: |
1186 | return AMDGPU::BUFFER_LOAD_USHORT_OFFEN; |
1187 | case AMDGPU::BUFFER_LOAD_SSHORT_OFFSET: |
1188 | return AMDGPU::BUFFER_LOAD_SSHORT_OFFEN; |
1189 | case AMDGPU::BUFFER_LOAD_DWORDX2_OFFSET: |
1190 | return AMDGPU::BUFFER_LOAD_DWORDX2_OFFEN; |
1191 | case AMDGPU::BUFFER_LOAD_DWORDX3_OFFSET: |
1192 | return AMDGPU::BUFFER_LOAD_DWORDX3_OFFEN; |
1193 | case AMDGPU::BUFFER_LOAD_DWORDX4_OFFSET: |
1194 | return AMDGPU::BUFFER_LOAD_DWORDX4_OFFEN; |
1195 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFSET: |
1196 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFEN; |
1197 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFSET: |
1198 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFEN; |
1199 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFSET: |
1200 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFEN; |
1201 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFSET: |
1202 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFEN; |
1203 | case AMDGPU::BUFFER_LOAD_SHORT_D16_OFFSET: |
1204 | return AMDGPU::BUFFER_LOAD_SHORT_D16_OFFEN; |
1205 | case AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFSET: |
1206 | return AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFEN; |
1207 | default: |
1208 | return -1; |
1209 | } |
1210 | } |
1211 | |
1212 | static MachineInstrBuilder spillVGPRtoAGPR(const GCNSubtarget &ST, |
1213 | MachineBasicBlock &MBB, |
1214 | MachineBasicBlock::iterator MI, |
1215 | int Index, unsigned Lane, |
1216 | unsigned ValueReg, bool IsKill) { |
1217 | MachineFunction *MF = MBB.getParent(); |
1218 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
1219 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1220 | |
1221 | MCPhysReg Reg = MFI->getVGPRToAGPRSpill(FrameIndex: Index, Lane); |
1222 | |
1223 | if (Reg == AMDGPU::NoRegister) |
1224 | return MachineInstrBuilder(); |
1225 | |
1226 | bool IsStore = MI->mayStore(); |
1227 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
1228 | auto *TRI = static_cast<const SIRegisterInfo*>(MRI.getTargetRegisterInfo()); |
1229 | |
1230 | unsigned Dst = IsStore ? Reg : ValueReg; |
1231 | unsigned Src = IsStore ? ValueReg : Reg; |
1232 | bool IsVGPR = TRI->isVGPR(MRI, Reg); |
1233 | DebugLoc DL = MI->getDebugLoc(); |
1234 | if (IsVGPR == TRI->isVGPR(MRI, Reg: ValueReg)) { |
1235 | // Spiller during regalloc may restore a spilled register to its superclass. |
1236 | // It could result in AGPR spills restored to VGPRs or the other way around, |
1237 | // making the src and dst with identical regclasses at this point. It just |
1238 | // needs a copy in such cases. |
1239 | auto CopyMIB = BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: Dst) |
1240 | .addReg(RegNo: Src, flags: getKillRegState(B: IsKill)); |
1241 | CopyMIB->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1242 | return CopyMIB; |
1243 | } |
1244 | unsigned Opc = (IsStore ^ IsVGPR) ? AMDGPU::V_ACCVGPR_WRITE_B32_e64 |
1245 | : AMDGPU::V_ACCVGPR_READ_B32_e64; |
1246 | |
1247 | auto MIB = BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: Opc), DestReg: Dst) |
1248 | .addReg(RegNo: Src, flags: getKillRegState(B: IsKill)); |
1249 | MIB->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1250 | return MIB; |
1251 | } |
1252 | |
1253 | // This differs from buildSpillLoadStore by only scavenging a VGPR. It does not |
1254 | // need to handle the case where an SGPR may need to be spilled while spilling. |
1255 | static bool buildMUBUFOffsetLoadStore(const GCNSubtarget &ST, |
1256 | MachineFrameInfo &MFI, |
1257 | MachineBasicBlock::iterator MI, |
1258 | int Index, |
1259 | int64_t Offset) { |
1260 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1261 | MachineBasicBlock *MBB = MI->getParent(); |
1262 | const DebugLoc &DL = MI->getDebugLoc(); |
1263 | bool IsStore = MI->mayStore(); |
1264 | |
1265 | unsigned Opc = MI->getOpcode(); |
1266 | int LoadStoreOp = IsStore ? |
1267 | getOffsetMUBUFStore(Opc) : getOffsetMUBUFLoad(Opc); |
1268 | if (LoadStoreOp == -1) |
1269 | return false; |
1270 | |
1271 | const MachineOperand *Reg = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::vdata); |
1272 | if (spillVGPRtoAGPR(ST, MBB&: *MBB, MI, Index, Lane: 0, ValueReg: Reg->getReg(), IsKill: false).getInstr()) |
1273 | return true; |
1274 | |
1275 | MachineInstrBuilder NewMI = |
1276 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: LoadStoreOp)) |
1277 | .add(MO: *Reg) |
1278 | .add(MO: *TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::srsrc)) |
1279 | .add(MO: *TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::soffset)) |
1280 | .addImm(Val: Offset) |
1281 | .addImm(Val: 0) // cpol |
1282 | .addImm(Val: 0) // swz |
1283 | .cloneMemRefs(OtherMI: *MI); |
1284 | |
1285 | const MachineOperand *VDataIn = TII->getNamedOperand(MI&: *MI, |
1286 | OperandName: AMDGPU::OpName::vdata_in); |
1287 | if (VDataIn) |
1288 | NewMI.add(MO: *VDataIn); |
1289 | return true; |
1290 | } |
1291 | |
1292 | static unsigned getFlatScratchSpillOpcode(const SIInstrInfo *TII, |
1293 | unsigned LoadStoreOp, |
1294 | unsigned EltSize) { |
1295 | bool IsStore = TII->get(Opcode: LoadStoreOp).mayStore(); |
1296 | bool HasVAddr = AMDGPU::hasNamedOperand(Opcode: LoadStoreOp, NamedIdx: AMDGPU::OpName::vaddr); |
1297 | bool UseST = |
1298 | !HasVAddr && !AMDGPU::hasNamedOperand(Opcode: LoadStoreOp, NamedIdx: AMDGPU::OpName::saddr); |
1299 | |
1300 | switch (EltSize) { |
1301 | case 4: |
1302 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
1303 | : AMDGPU::SCRATCH_LOAD_DWORD_SADDR; |
1304 | break; |
1305 | case 8: |
1306 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX2_SADDR |
1307 | : AMDGPU::SCRATCH_LOAD_DWORDX2_SADDR; |
1308 | break; |
1309 | case 12: |
1310 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX3_SADDR |
1311 | : AMDGPU::SCRATCH_LOAD_DWORDX3_SADDR; |
1312 | break; |
1313 | case 16: |
1314 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX4_SADDR |
1315 | : AMDGPU::SCRATCH_LOAD_DWORDX4_SADDR; |
1316 | break; |
1317 | default: |
1318 | llvm_unreachable("Unexpected spill load/store size!" ); |
1319 | } |
1320 | |
1321 | if (HasVAddr) |
1322 | LoadStoreOp = AMDGPU::getFlatScratchInstSVfromSS(Opcode: LoadStoreOp); |
1323 | else if (UseST) |
1324 | LoadStoreOp = AMDGPU::getFlatScratchInstSTfromSS(Opcode: LoadStoreOp); |
1325 | |
1326 | return LoadStoreOp; |
1327 | } |
1328 | |
1329 | void SIRegisterInfo::buildSpillLoadStore( |
1330 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, const DebugLoc &DL, |
1331 | unsigned LoadStoreOp, int Index, Register ValueReg, bool IsKill, |
1332 | MCRegister ScratchOffsetReg, int64_t InstOffset, MachineMemOperand *MMO, |
1333 | RegScavenger *RS, LiveRegUnits *LiveUnits) const { |
1334 | assert((!RS || !LiveUnits) && "Only RS or LiveUnits can be set but not both" ); |
1335 | |
1336 | MachineFunction *MF = MBB.getParent(); |
1337 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1338 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
1339 | const SIMachineFunctionInfo *FuncInfo = MF->getInfo<SIMachineFunctionInfo>(); |
1340 | |
1341 | const MCInstrDesc *Desc = &TII->get(Opcode: LoadStoreOp); |
1342 | bool IsStore = Desc->mayStore(); |
1343 | bool IsFlat = TII->isFLATScratch(Opcode: LoadStoreOp); |
1344 | |
1345 | bool CanClobberSCC = false; |
1346 | bool Scavenged = false; |
1347 | MCRegister SOffset = ScratchOffsetReg; |
1348 | |
1349 | const TargetRegisterClass *RC = getRegClassForReg(MRI: MF->getRegInfo(), Reg: ValueReg); |
1350 | // On gfx90a+ AGPR is a regular VGPR acceptable for loads and stores. |
1351 | const bool IsAGPR = !ST.hasGFX90AInsts() && isAGPRClass(RC); |
1352 | const unsigned RegWidth = AMDGPU::getRegBitWidth(RC: *RC) / 8; |
1353 | |
1354 | // Always use 4 byte operations for AGPRs because we need to scavenge |
1355 | // a temporary VGPR. |
1356 | unsigned EltSize = (IsFlat && !IsAGPR) ? std::min(a: RegWidth, b: 16u) : 4u; |
1357 | unsigned NumSubRegs = RegWidth / EltSize; |
1358 | unsigned Size = NumSubRegs * EltSize; |
1359 | unsigned RemSize = RegWidth - Size; |
1360 | unsigned NumRemSubRegs = RemSize ? 1 : 0; |
1361 | int64_t Offset = InstOffset + MFI.getObjectOffset(ObjectIdx: Index); |
1362 | int64_t MaterializedOffset = Offset; |
1363 | |
1364 | int64_t MaxOffset = Offset + Size + RemSize - EltSize; |
1365 | int64_t ScratchOffsetRegDelta = 0; |
1366 | |
1367 | if (IsFlat && EltSize > 4) { |
1368 | LoadStoreOp = getFlatScratchSpillOpcode(TII, LoadStoreOp, EltSize); |
1369 | Desc = &TII->get(Opcode: LoadStoreOp); |
1370 | } |
1371 | |
1372 | Align Alignment = MFI.getObjectAlign(ObjectIdx: Index); |
1373 | const MachinePointerInfo &BasePtrInfo = MMO->getPointerInfo(); |
1374 | |
1375 | assert((IsFlat || ((Offset % EltSize) == 0)) && |
1376 | "unexpected VGPR spill offset" ); |
1377 | |
1378 | // Track a VGPR to use for a constant offset we need to materialize. |
1379 | Register TmpOffsetVGPR; |
1380 | |
1381 | // Track a VGPR to use as an intermediate value. |
1382 | Register TmpIntermediateVGPR; |
1383 | bool UseVGPROffset = false; |
1384 | |
1385 | // Materialize a VGPR offset required for the given SGPR/VGPR/Immediate |
1386 | // combination. |
1387 | auto MaterializeVOffset = [&](Register SGPRBase, Register TmpVGPR, |
1388 | int64_t VOffset) { |
1389 | // We are using a VGPR offset |
1390 | if (IsFlat && SGPRBase) { |
1391 | // We only have 1 VGPR offset, or 1 SGPR offset. We don't have a free |
1392 | // SGPR, so perform the add as vector. |
1393 | // We don't need a base SGPR in the kernel. |
1394 | |
1395 | if (ST.getConstantBusLimit(Opcode: AMDGPU::V_ADD_U32_e64) >= 2) { |
1396 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_ADD_U32_e64), DestReg: TmpVGPR) |
1397 | .addReg(RegNo: SGPRBase) |
1398 | .addImm(Val: VOffset) |
1399 | .addImm(Val: 0); // clamp |
1400 | } else { |
1401 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32), DestReg: TmpVGPR) |
1402 | .addReg(RegNo: SGPRBase); |
1403 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_ADD_U32_e32), DestReg: TmpVGPR) |
1404 | .addImm(Val: VOffset) |
1405 | .addReg(RegNo: TmpOffsetVGPR); |
1406 | } |
1407 | } else { |
1408 | assert(TmpOffsetVGPR); |
1409 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32), DestReg: TmpVGPR) |
1410 | .addImm(Val: VOffset); |
1411 | } |
1412 | }; |
1413 | |
1414 | bool IsOffsetLegal = |
1415 | IsFlat ? TII->isLegalFLATOffset(Offset: MaxOffset, AddrSpace: AMDGPUAS::PRIVATE_ADDRESS, |
1416 | FlatVariant: SIInstrFlags::FlatScratch) |
1417 | : TII->isLegalMUBUFImmOffset(Imm: MaxOffset); |
1418 | if (!IsOffsetLegal || (IsFlat && !SOffset && !ST.hasFlatScratchSTMode())) { |
1419 | SOffset = MCRegister(); |
1420 | |
1421 | // We don't have access to the register scavenger if this function is called |
1422 | // during PEI::scavengeFrameVirtualRegs() so use LiveUnits in this case. |
1423 | // TODO: Clobbering SCC is not necessary for scratch instructions in the |
1424 | // entry. |
1425 | if (RS) { |
1426 | SOffset = RS->scavengeRegisterBackwards(RC: AMDGPU::SGPR_32RegClass, To: MI, RestoreAfter: false, SPAdj: 0, AllowSpill: false); |
1427 | |
1428 | // Piggy back on the liveness scan we just did see if SCC is dead. |
1429 | CanClobberSCC = !RS->isRegUsed(Reg: AMDGPU::SCC); |
1430 | } else if (LiveUnits) { |
1431 | CanClobberSCC = LiveUnits->available(Reg: AMDGPU::SCC); |
1432 | for (MCRegister Reg : AMDGPU::SGPR_32RegClass) { |
1433 | if (LiveUnits->available(Reg) && !MF->getRegInfo().isReserved(PhysReg: Reg)) { |
1434 | SOffset = Reg; |
1435 | break; |
1436 | } |
1437 | } |
1438 | } |
1439 | |
1440 | if (ScratchOffsetReg != AMDGPU::NoRegister && !CanClobberSCC) |
1441 | SOffset = Register(); |
1442 | |
1443 | if (!SOffset) { |
1444 | UseVGPROffset = true; |
1445 | |
1446 | if (RS) { |
1447 | TmpOffsetVGPR = RS->scavengeRegisterBackwards(RC: AMDGPU::VGPR_32RegClass, To: MI, RestoreAfter: false, SPAdj: 0); |
1448 | } else { |
1449 | assert(LiveUnits); |
1450 | for (MCRegister Reg : AMDGPU::VGPR_32RegClass) { |
1451 | if (LiveUnits->available(Reg) && !MF->getRegInfo().isReserved(PhysReg: Reg)) { |
1452 | TmpOffsetVGPR = Reg; |
1453 | break; |
1454 | } |
1455 | } |
1456 | } |
1457 | |
1458 | assert(TmpOffsetVGPR); |
1459 | } else if (!SOffset && CanClobberSCC) { |
1460 | // There are no free SGPRs, and since we are in the process of spilling |
1461 | // VGPRs too. Since we need a VGPR in order to spill SGPRs (this is true |
1462 | // on SI/CI and on VI it is true until we implement spilling using scalar |
1463 | // stores), we have no way to free up an SGPR. Our solution here is to |
1464 | // add the offset directly to the ScratchOffset or StackPtrOffset |
1465 | // register, and then subtract the offset after the spill to return the |
1466 | // register to it's original value. |
1467 | |
1468 | // TODO: If we don't have to do an emergency stack slot spill, converting |
1469 | // to use the VGPR offset is fewer instructions. |
1470 | if (!ScratchOffsetReg) |
1471 | ScratchOffsetReg = FuncInfo->getStackPtrOffsetReg(); |
1472 | SOffset = ScratchOffsetReg; |
1473 | ScratchOffsetRegDelta = Offset; |
1474 | } else { |
1475 | Scavenged = true; |
1476 | } |
1477 | |
1478 | // We currently only support spilling VGPRs to EltSize boundaries, meaning |
1479 | // we can simplify the adjustment of Offset here to just scale with |
1480 | // WavefrontSize. |
1481 | if (!IsFlat && !UseVGPROffset) |
1482 | Offset *= ST.getWavefrontSize(); |
1483 | |
1484 | if (!UseVGPROffset && !SOffset) |
1485 | report_fatal_error(reason: "could not scavenge SGPR to spill in entry function" ); |
1486 | |
1487 | if (UseVGPROffset) { |
1488 | // We are using a VGPR offset |
1489 | MaterializeVOffset(ScratchOffsetReg, TmpOffsetVGPR, Offset); |
1490 | } else if (ScratchOffsetReg == AMDGPU::NoRegister) { |
1491 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: SOffset).addImm(Val: Offset); |
1492 | } else { |
1493 | assert(Offset != 0); |
1494 | auto Add = BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: SOffset) |
1495 | .addReg(RegNo: ScratchOffsetReg) |
1496 | .addImm(Val: Offset); |
1497 | Add->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
1498 | } |
1499 | |
1500 | Offset = 0; |
1501 | } |
1502 | |
1503 | if (IsFlat && SOffset == AMDGPU::NoRegister) { |
1504 | assert(AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::vaddr) < 0 |
1505 | && "Unexpected vaddr for flat scratch with a FI operand" ); |
1506 | |
1507 | if (UseVGPROffset) { |
1508 | LoadStoreOp = AMDGPU::getFlatScratchInstSVfromSS(Opcode: LoadStoreOp); |
1509 | } else { |
1510 | assert(ST.hasFlatScratchSTMode()); |
1511 | LoadStoreOp = AMDGPU::getFlatScratchInstSTfromSS(Opcode: LoadStoreOp); |
1512 | } |
1513 | |
1514 | Desc = &TII->get(Opcode: LoadStoreOp); |
1515 | } |
1516 | |
1517 | for (unsigned i = 0, e = NumSubRegs + NumRemSubRegs, RegOffset = 0; i != e; |
1518 | ++i, RegOffset += EltSize) { |
1519 | if (i == NumSubRegs) { |
1520 | EltSize = RemSize; |
1521 | LoadStoreOp = getFlatScratchSpillOpcode(TII, LoadStoreOp, EltSize); |
1522 | } |
1523 | Desc = &TII->get(Opcode: LoadStoreOp); |
1524 | |
1525 | if (!IsFlat && UseVGPROffset) { |
1526 | int NewLoadStoreOp = IsStore ? getOffenMUBUFStore(Opc: LoadStoreOp) |
1527 | : getOffenMUBUFLoad(Opc: LoadStoreOp); |
1528 | Desc = &TII->get(Opcode: NewLoadStoreOp); |
1529 | } |
1530 | |
1531 | if (UseVGPROffset && TmpOffsetVGPR == TmpIntermediateVGPR) { |
1532 | // If we are spilling an AGPR beyond the range of the memory instruction |
1533 | // offset and need to use a VGPR offset, we ideally have at least 2 |
1534 | // scratch VGPRs. If we don't have a second free VGPR without spilling, |
1535 | // recycle the VGPR used for the offset which requires resetting after |
1536 | // each subregister. |
1537 | |
1538 | MaterializeVOffset(ScratchOffsetReg, TmpOffsetVGPR, MaterializedOffset); |
1539 | } |
1540 | |
1541 | unsigned NumRegs = EltSize / 4; |
1542 | Register SubReg = e == 1 |
1543 | ? ValueReg |
1544 | : Register(getSubReg(Reg: ValueReg, |
1545 | Idx: getSubRegFromChannel(Channel: RegOffset / 4, NumRegs))); |
1546 | |
1547 | unsigned SOffsetRegState = 0; |
1548 | unsigned SrcDstRegState = getDefRegState(B: !IsStore); |
1549 | const bool IsLastSubReg = i + 1 == e; |
1550 | const bool IsFirstSubReg = i == 0; |
1551 | if (IsLastSubReg) { |
1552 | SOffsetRegState |= getKillRegState(B: Scavenged); |
1553 | // The last implicit use carries the "Kill" flag. |
1554 | SrcDstRegState |= getKillRegState(B: IsKill); |
1555 | } |
1556 | |
1557 | // Make sure the whole register is defined if there are undef components by |
1558 | // adding an implicit def of the super-reg on the first instruction. |
1559 | bool NeedSuperRegDef = e > 1 && IsStore && IsFirstSubReg; |
1560 | bool NeedSuperRegImpOperand = e > 1; |
1561 | |
1562 | // Remaining element size to spill into memory after some parts of it |
1563 | // spilled into either AGPRs or VGPRs. |
1564 | unsigned RemEltSize = EltSize; |
1565 | |
1566 | // AGPRs to spill VGPRs and vice versa are allocated in a reverse order, |
1567 | // starting from the last lane. In case if a register cannot be completely |
1568 | // spilled into another register that will ensure its alignment does not |
1569 | // change. For targets with VGPR alignment requirement this is important |
1570 | // in case of flat scratch usage as we might get a scratch_load or |
1571 | // scratch_store of an unaligned register otherwise. |
1572 | for (int LaneS = (RegOffset + EltSize) / 4 - 1, Lane = LaneS, |
1573 | LaneE = RegOffset / 4; |
1574 | Lane >= LaneE; --Lane) { |
1575 | bool IsSubReg = e > 1 || EltSize > 4; |
1576 | Register Sub = IsSubReg |
1577 | ? Register(getSubReg(Reg: ValueReg, Idx: getSubRegFromChannel(Channel: Lane))) |
1578 | : ValueReg; |
1579 | auto MIB = spillVGPRtoAGPR(ST, MBB, MI, Index, Lane, ValueReg: Sub, IsKill); |
1580 | if (!MIB.getInstr()) |
1581 | break; |
1582 | if (NeedSuperRegDef || (IsSubReg && IsStore && Lane == LaneS && IsFirstSubReg)) { |
1583 | MIB.addReg(RegNo: ValueReg, flags: RegState::ImplicitDefine); |
1584 | NeedSuperRegDef = false; |
1585 | } |
1586 | if ((IsSubReg || NeedSuperRegImpOperand) && (IsFirstSubReg || IsLastSubReg)) { |
1587 | NeedSuperRegImpOperand = true; |
1588 | unsigned State = SrcDstRegState; |
1589 | if (!IsLastSubReg || (Lane != LaneE)) |
1590 | State &= ~RegState::Kill; |
1591 | if (!IsFirstSubReg || (Lane != LaneS)) |
1592 | State &= ~RegState::Define; |
1593 | MIB.addReg(RegNo: ValueReg, flags: RegState::Implicit | State); |
1594 | } |
1595 | RemEltSize -= 4; |
1596 | } |
1597 | |
1598 | if (!RemEltSize) // Fully spilled into AGPRs. |
1599 | continue; |
1600 | |
1601 | if (RemEltSize != EltSize) { // Partially spilled to AGPRs |
1602 | assert(IsFlat && EltSize > 4); |
1603 | |
1604 | unsigned NumRegs = RemEltSize / 4; |
1605 | SubReg = Register(getSubReg(Reg: ValueReg, |
1606 | Idx: getSubRegFromChannel(Channel: RegOffset / 4, NumRegs))); |
1607 | unsigned Opc = getFlatScratchSpillOpcode(TII, LoadStoreOp, EltSize: RemEltSize); |
1608 | Desc = &TII->get(Opcode: Opc); |
1609 | } |
1610 | |
1611 | unsigned FinalReg = SubReg; |
1612 | |
1613 | if (IsAGPR) { |
1614 | assert(EltSize == 4); |
1615 | |
1616 | if (!TmpIntermediateVGPR) { |
1617 | TmpIntermediateVGPR = FuncInfo->getVGPRForAGPRCopy(); |
1618 | assert(MF->getRegInfo().isReserved(TmpIntermediateVGPR)); |
1619 | } |
1620 | if (IsStore) { |
1621 | auto AccRead = BuildMI(BB&: MBB, I: MI, MIMD: DL, |
1622 | MCID: TII->get(Opcode: AMDGPU::V_ACCVGPR_READ_B32_e64), |
1623 | DestReg: TmpIntermediateVGPR) |
1624 | .addReg(RegNo: SubReg, flags: getKillRegState(B: IsKill)); |
1625 | if (NeedSuperRegDef) |
1626 | AccRead.addReg(RegNo: ValueReg, flags: RegState::ImplicitDefine); |
1627 | AccRead->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1628 | } |
1629 | SubReg = TmpIntermediateVGPR; |
1630 | } else if (UseVGPROffset) { |
1631 | if (!TmpOffsetVGPR) { |
1632 | TmpOffsetVGPR = RS->scavengeRegisterBackwards(RC: AMDGPU::VGPR_32RegClass, |
1633 | To: MI, RestoreAfter: false, SPAdj: 0); |
1634 | RS->setRegUsed(Reg: TmpOffsetVGPR); |
1635 | } |
1636 | } |
1637 | |
1638 | MachinePointerInfo PInfo = BasePtrInfo.getWithOffset(O: RegOffset); |
1639 | MachineMemOperand *NewMMO = |
1640 | MF->getMachineMemOperand(PtrInfo: PInfo, F: MMO->getFlags(), Size: RemEltSize, |
1641 | BaseAlignment: commonAlignment(A: Alignment, Offset: RegOffset)); |
1642 | |
1643 | auto MIB = |
1644 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: *Desc) |
1645 | .addReg(RegNo: SubReg, flags: getDefRegState(B: !IsStore) | getKillRegState(B: IsKill)); |
1646 | |
1647 | if (UseVGPROffset) { |
1648 | // For an AGPR spill, we reuse the same temp VGPR for the offset and the |
1649 | // intermediate accvgpr_write. |
1650 | MIB.addReg(RegNo: TmpOffsetVGPR, flags: getKillRegState(B: IsLastSubReg && !IsAGPR)); |
1651 | } |
1652 | |
1653 | if (!IsFlat) |
1654 | MIB.addReg(RegNo: FuncInfo->getScratchRSrcReg()); |
1655 | |
1656 | if (SOffset == AMDGPU::NoRegister) { |
1657 | if (!IsFlat) { |
1658 | if (UseVGPROffset && ScratchOffsetReg) { |
1659 | MIB.addReg(RegNo: ScratchOffsetReg); |
1660 | } else { |
1661 | assert(FuncInfo->isBottomOfStack()); |
1662 | MIB.addImm(Val: 0); |
1663 | } |
1664 | } |
1665 | } else { |
1666 | MIB.addReg(RegNo: SOffset, flags: SOffsetRegState); |
1667 | } |
1668 | |
1669 | MIB.addImm(Val: Offset + RegOffset); |
1670 | |
1671 | bool LastUse = MMO->getFlags() & MOLastUse; |
1672 | MIB.addImm(Val: LastUse ? AMDGPU::CPol::TH_LU : 0); // cpol |
1673 | |
1674 | if (!IsFlat) |
1675 | MIB.addImm(Val: 0); // swz |
1676 | MIB.addMemOperand(MMO: NewMMO); |
1677 | |
1678 | if (!IsAGPR && NeedSuperRegDef) |
1679 | MIB.addReg(RegNo: ValueReg, flags: RegState::ImplicitDefine); |
1680 | |
1681 | if (!IsStore && IsAGPR && TmpIntermediateVGPR != AMDGPU::NoRegister) { |
1682 | MIB = BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_ACCVGPR_WRITE_B32_e64), |
1683 | DestReg: FinalReg) |
1684 | .addReg(RegNo: TmpIntermediateVGPR, flags: RegState::Kill); |
1685 | MIB->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1686 | } |
1687 | |
1688 | if (NeedSuperRegImpOperand && (IsFirstSubReg || IsLastSubReg)) |
1689 | MIB.addReg(RegNo: ValueReg, flags: RegState::Implicit | SrcDstRegState); |
1690 | |
1691 | // The epilog restore of a wwm-scratch register can cause undesired |
1692 | // optimization during machine-cp post PrologEpilogInserter if the same |
1693 | // register was assigned for return value ABI lowering with a COPY |
1694 | // instruction. As given below, with the epilog reload, the earlier COPY |
1695 | // appeared to be dead during machine-cp. |
1696 | // ... |
1697 | // v0 in WWM operation, needs the WWM spill at prolog/epilog. |
1698 | // $vgpr0 = V_WRITELANE_B32 $sgpr20, 0, $vgpr0 |
1699 | // ... |
1700 | // Epilog block: |
1701 | // $vgpr0 = COPY $vgpr1 // outgoing value moved to v0 |
1702 | // ... |
1703 | // WWM spill restore to preserve the inactive lanes of v0. |
1704 | // $sgpr4_sgpr5 = S_XOR_SAVEEXEC_B64 -1 |
1705 | // $vgpr0 = BUFFER_LOAD $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0 |
1706 | // $exec = S_MOV_B64 killed $sgpr4_sgpr5 |
1707 | // ... |
1708 | // SI_RETURN implicit $vgpr0 |
1709 | // ... |
1710 | // To fix it, mark the same reg as a tied op for such restore instructions |
1711 | // so that it marks a usage for the preceding COPY. |
1712 | if (!IsStore && MI != MBB.end() && MI->isReturn() && |
1713 | MI->readsRegister(Reg: SubReg, TRI: this)) { |
1714 | MIB.addReg(RegNo: SubReg, flags: RegState::Implicit); |
1715 | MIB->tieOperands(DefIdx: 0, UseIdx: MIB->getNumOperands() - 1); |
1716 | } |
1717 | } |
1718 | |
1719 | if (ScratchOffsetRegDelta != 0) { |
1720 | // Subtract the offset we added to the ScratchOffset register. |
1721 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: SOffset) |
1722 | .addReg(RegNo: SOffset) |
1723 | .addImm(Val: -ScratchOffsetRegDelta); |
1724 | } |
1725 | } |
1726 | |
1727 | void SIRegisterInfo::buildVGPRSpillLoadStore(SGPRSpillBuilder &SB, int Index, |
1728 | int Offset, bool IsLoad, |
1729 | bool IsKill) const { |
1730 | // Load/store VGPR |
1731 | MachineFrameInfo &FrameInfo = SB.MF.getFrameInfo(); |
1732 | assert(FrameInfo.getStackID(Index) != TargetStackID::SGPRSpill); |
1733 | |
1734 | Register FrameReg = |
1735 | FrameInfo.isFixedObjectIndex(ObjectIdx: Index) && hasBasePointer(MF: SB.MF) |
1736 | ? getBaseRegister() |
1737 | : getFrameRegister(MF: SB.MF); |
1738 | |
1739 | Align Alignment = FrameInfo.getObjectAlign(ObjectIdx: Index); |
1740 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF&: SB.MF, FI: Index); |
1741 | MachineMemOperand *MMO = SB.MF.getMachineMemOperand( |
1742 | PtrInfo, F: IsLoad ? MachineMemOperand::MOLoad : MachineMemOperand::MOStore, |
1743 | Size: SB.EltSize, BaseAlignment: Alignment); |
1744 | |
1745 | if (IsLoad) { |
1746 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_LOAD_DWORD_SADDR |
1747 | : AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
1748 | buildSpillLoadStore(MBB&: *SB.MBB, MI: SB.MI, DL: SB.DL, LoadStoreOp: Opc, Index, ValueReg: SB.TmpVGPR, IsKill: false, |
1749 | ScratchOffsetReg: FrameReg, InstOffset: (int64_t)Offset * SB.EltSize, MMO, RS: SB.RS); |
1750 | } else { |
1751 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
1752 | : AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
1753 | buildSpillLoadStore(MBB&: *SB.MBB, MI: SB.MI, DL: SB.DL, LoadStoreOp: Opc, Index, ValueReg: SB.TmpVGPR, IsKill, |
1754 | ScratchOffsetReg: FrameReg, InstOffset: (int64_t)Offset * SB.EltSize, MMO, RS: SB.RS); |
1755 | // This only ever adds one VGPR spill |
1756 | SB.MFI.addToSpilledVGPRs(num: 1); |
1757 | } |
1758 | } |
1759 | |
1760 | bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index, |
1761 | RegScavenger *RS, SlotIndexes *Indexes, |
1762 | LiveIntervals *LIS, bool OnlyToVGPR, |
1763 | bool SpillToPhysVGPRLane) const { |
1764 | SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS); |
1765 | |
1766 | ArrayRef<SpilledReg> VGPRSpills = |
1767 | SpillToPhysVGPRLane ? SB.MFI.getSGPRSpillToPhysicalVGPRLanes(FrameIndex: Index) |
1768 | : SB.MFI.getSGPRSpillToVirtualVGPRLanes(FrameIndex: Index); |
1769 | bool SpillToVGPR = !VGPRSpills.empty(); |
1770 | if (OnlyToVGPR && !SpillToVGPR) |
1771 | return false; |
1772 | |
1773 | assert(SpillToVGPR || (SB.SuperReg != SB.MFI.getStackPtrOffsetReg() && |
1774 | SB.SuperReg != SB.MFI.getFrameOffsetReg())); |
1775 | |
1776 | if (SpillToVGPR) { |
1777 | |
1778 | assert(SB.NumSubRegs == VGPRSpills.size() && |
1779 | "Num of VGPR lanes should be equal to num of SGPRs spilled" ); |
1780 | |
1781 | for (unsigned i = 0, e = SB.NumSubRegs; i < e; ++i) { |
1782 | Register SubReg = |
1783 | SB.NumSubRegs == 1 |
1784 | ? SB.SuperReg |
1785 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
1786 | SpilledReg Spill = VGPRSpills[i]; |
1787 | |
1788 | bool IsFirstSubreg = i == 0; |
1789 | bool IsLastSubreg = i == SB.NumSubRegs - 1; |
1790 | bool UseKill = SB.IsKill && IsLastSubreg; |
1791 | |
1792 | |
1793 | // Mark the "old value of vgpr" input undef only if this is the first sgpr |
1794 | // spill to this specific vgpr in the first basic block. |
1795 | auto MIB = BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, |
1796 | MCID: SB.TII.get(Opcode: AMDGPU::SI_SPILL_S32_TO_VGPR), DestReg: Spill.VGPR) |
1797 | .addReg(RegNo: SubReg, flags: getKillRegState(B: UseKill)) |
1798 | .addImm(Val: Spill.Lane) |
1799 | .addReg(RegNo: Spill.VGPR); |
1800 | if (Indexes) { |
1801 | if (IsFirstSubreg) |
1802 | Indexes->replaceMachineInstrInMaps(MI&: *MI, NewMI&: *MIB); |
1803 | else |
1804 | Indexes->insertMachineInstrInMaps(MI&: *MIB); |
1805 | } |
1806 | |
1807 | if (IsFirstSubreg && SB.NumSubRegs > 1) { |
1808 | // We may be spilling a super-register which is only partially defined, |
1809 | // and need to ensure later spills think the value is defined. |
1810 | MIB.addReg(RegNo: SB.SuperReg, flags: RegState::ImplicitDefine); |
1811 | } |
1812 | |
1813 | if (SB.NumSubRegs > 1 && (IsFirstSubreg || IsLastSubreg)) |
1814 | MIB.addReg(RegNo: SB.SuperReg, flags: getKillRegState(B: UseKill) | RegState::Implicit); |
1815 | |
1816 | // FIXME: Since this spills to another register instead of an actual |
1817 | // frame index, we should delete the frame index when all references to |
1818 | // it are fixed. |
1819 | } |
1820 | } else { |
1821 | SB.prepare(); |
1822 | |
1823 | // SubReg carries the "Kill" flag when SubReg == SB.SuperReg. |
1824 | unsigned SubKillState = getKillRegState(B: (SB.NumSubRegs == 1) && SB.IsKill); |
1825 | |
1826 | // Per VGPR helper data |
1827 | auto PVD = SB.getPerVGPRData(); |
1828 | |
1829 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
1830 | unsigned TmpVGPRFlags = RegState::Undef; |
1831 | |
1832 | // Write sub registers into the VGPR |
1833 | for (unsigned i = Offset * PVD.PerVGPR, |
1834 | e = std::min(a: (Offset + 1) * PVD.PerVGPR, b: SB.NumSubRegs); |
1835 | i < e; ++i) { |
1836 | Register SubReg = |
1837 | SB.NumSubRegs == 1 |
1838 | ? SB.SuperReg |
1839 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
1840 | |
1841 | MachineInstrBuilder WriteLane = |
1842 | BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, |
1843 | MCID: SB.TII.get(Opcode: AMDGPU::SI_SPILL_S32_TO_VGPR), DestReg: SB.TmpVGPR) |
1844 | .addReg(RegNo: SubReg, flags: SubKillState) |
1845 | .addImm(Val: i % PVD.PerVGPR) |
1846 | .addReg(RegNo: SB.TmpVGPR, flags: TmpVGPRFlags); |
1847 | TmpVGPRFlags = 0; |
1848 | |
1849 | if (Indexes) { |
1850 | if (i == 0) |
1851 | Indexes->replaceMachineInstrInMaps(MI&: *MI, NewMI&: *WriteLane); |
1852 | else |
1853 | Indexes->insertMachineInstrInMaps(MI&: *WriteLane); |
1854 | } |
1855 | |
1856 | // There could be undef components of a spilled super register. |
1857 | // TODO: Can we detect this and skip the spill? |
1858 | if (SB.NumSubRegs > 1) { |
1859 | // The last implicit use of the SB.SuperReg carries the "Kill" flag. |
1860 | unsigned SuperKillState = 0; |
1861 | if (i + 1 == SB.NumSubRegs) |
1862 | SuperKillState |= getKillRegState(B: SB.IsKill); |
1863 | WriteLane.addReg(RegNo: SB.SuperReg, flags: RegState::Implicit | SuperKillState); |
1864 | } |
1865 | } |
1866 | |
1867 | // Write out VGPR |
1868 | SB.readWriteTmpVGPR(Offset, /*IsLoad*/ false); |
1869 | } |
1870 | |
1871 | SB.restore(); |
1872 | } |
1873 | |
1874 | MI->eraseFromParent(); |
1875 | SB.MFI.addToSpilledSGPRs(num: SB.NumSubRegs); |
1876 | |
1877 | if (LIS) |
1878 | LIS->removeAllRegUnitsForPhysReg(Reg: SB.SuperReg); |
1879 | |
1880 | return true; |
1881 | } |
1882 | |
1883 | bool SIRegisterInfo::restoreSGPR(MachineBasicBlock::iterator MI, int Index, |
1884 | RegScavenger *RS, SlotIndexes *Indexes, |
1885 | LiveIntervals *LIS, bool OnlyToVGPR, |
1886 | bool SpillToPhysVGPRLane) const { |
1887 | SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS); |
1888 | |
1889 | ArrayRef<SpilledReg> VGPRSpills = |
1890 | SpillToPhysVGPRLane ? SB.MFI.getSGPRSpillToPhysicalVGPRLanes(FrameIndex: Index) |
1891 | : SB.MFI.getSGPRSpillToVirtualVGPRLanes(FrameIndex: Index); |
1892 | bool SpillToVGPR = !VGPRSpills.empty(); |
1893 | if (OnlyToVGPR && !SpillToVGPR) |
1894 | return false; |
1895 | |
1896 | if (SpillToVGPR) { |
1897 | for (unsigned i = 0, e = SB.NumSubRegs; i < e; ++i) { |
1898 | Register SubReg = |
1899 | SB.NumSubRegs == 1 |
1900 | ? SB.SuperReg |
1901 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
1902 | |
1903 | SpilledReg Spill = VGPRSpills[i]; |
1904 | auto MIB = BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, |
1905 | MCID: SB.TII.get(Opcode: AMDGPU::SI_RESTORE_S32_FROM_VGPR), DestReg: SubReg) |
1906 | .addReg(RegNo: Spill.VGPR) |
1907 | .addImm(Val: Spill.Lane); |
1908 | if (SB.NumSubRegs > 1 && i == 0) |
1909 | MIB.addReg(RegNo: SB.SuperReg, flags: RegState::ImplicitDefine); |
1910 | if (Indexes) { |
1911 | if (i == e - 1) |
1912 | Indexes->replaceMachineInstrInMaps(MI&: *MI, NewMI&: *MIB); |
1913 | else |
1914 | Indexes->insertMachineInstrInMaps(MI&: *MIB); |
1915 | } |
1916 | } |
1917 | } else { |
1918 | SB.prepare(); |
1919 | |
1920 | // Per VGPR helper data |
1921 | auto PVD = SB.getPerVGPRData(); |
1922 | |
1923 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
1924 | // Load in VGPR data |
1925 | SB.readWriteTmpVGPR(Offset, /*IsLoad*/ true); |
1926 | |
1927 | // Unpack lanes |
1928 | for (unsigned i = Offset * PVD.PerVGPR, |
1929 | e = std::min(a: (Offset + 1) * PVD.PerVGPR, b: SB.NumSubRegs); |
1930 | i < e; ++i) { |
1931 | Register SubReg = |
1932 | SB.NumSubRegs == 1 |
1933 | ? SB.SuperReg |
1934 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
1935 | |
1936 | bool LastSubReg = (i + 1 == e); |
1937 | auto MIB = BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, |
1938 | MCID: SB.TII.get(Opcode: AMDGPU::SI_RESTORE_S32_FROM_VGPR), DestReg: SubReg) |
1939 | .addReg(RegNo: SB.TmpVGPR, flags: getKillRegState(B: LastSubReg)) |
1940 | .addImm(Val: i); |
1941 | if (SB.NumSubRegs > 1 && i == 0) |
1942 | MIB.addReg(RegNo: SB.SuperReg, flags: RegState::ImplicitDefine); |
1943 | if (Indexes) { |
1944 | if (i == e - 1) |
1945 | Indexes->replaceMachineInstrInMaps(MI&: *MI, NewMI&: *MIB); |
1946 | else |
1947 | Indexes->insertMachineInstrInMaps(MI&: *MIB); |
1948 | } |
1949 | } |
1950 | } |
1951 | |
1952 | SB.restore(); |
1953 | } |
1954 | |
1955 | MI->eraseFromParent(); |
1956 | |
1957 | if (LIS) |
1958 | LIS->removeAllRegUnitsForPhysReg(Reg: SB.SuperReg); |
1959 | |
1960 | return true; |
1961 | } |
1962 | |
1963 | bool SIRegisterInfo::spillEmergencySGPR(MachineBasicBlock::iterator MI, |
1964 | MachineBasicBlock &RestoreMBB, |
1965 | Register SGPR, RegScavenger *RS) const { |
1966 | SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, SGPR, false, 0, |
1967 | RS); |
1968 | SB.prepare(); |
1969 | // Generate the spill of SGPR to SB.TmpVGPR. |
1970 | unsigned SubKillState = getKillRegState(B: (SB.NumSubRegs == 1) && SB.IsKill); |
1971 | auto PVD = SB.getPerVGPRData(); |
1972 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
1973 | unsigned TmpVGPRFlags = RegState::Undef; |
1974 | // Write sub registers into the VGPR |
1975 | for (unsigned i = Offset * PVD.PerVGPR, |
1976 | e = std::min(a: (Offset + 1) * PVD.PerVGPR, b: SB.NumSubRegs); |
1977 | i < e; ++i) { |
1978 | Register SubReg = |
1979 | SB.NumSubRegs == 1 |
1980 | ? SB.SuperReg |
1981 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
1982 | |
1983 | MachineInstrBuilder WriteLane = |
1984 | BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, MCID: SB.TII.get(Opcode: AMDGPU::V_WRITELANE_B32), |
1985 | DestReg: SB.TmpVGPR) |
1986 | .addReg(RegNo: SubReg, flags: SubKillState) |
1987 | .addImm(Val: i % PVD.PerVGPR) |
1988 | .addReg(RegNo: SB.TmpVGPR, flags: TmpVGPRFlags); |
1989 | TmpVGPRFlags = 0; |
1990 | // There could be undef components of a spilled super register. |
1991 | // TODO: Can we detect this and skip the spill? |
1992 | if (SB.NumSubRegs > 1) { |
1993 | // The last implicit use of the SB.SuperReg carries the "Kill" flag. |
1994 | unsigned SuperKillState = 0; |
1995 | if (i + 1 == SB.NumSubRegs) |
1996 | SuperKillState |= getKillRegState(B: SB.IsKill); |
1997 | WriteLane.addReg(RegNo: SB.SuperReg, flags: RegState::Implicit | SuperKillState); |
1998 | } |
1999 | } |
2000 | // Don't need to write VGPR out. |
2001 | } |
2002 | |
2003 | // Restore clobbered registers in the specified restore block. |
2004 | MI = RestoreMBB.end(); |
2005 | SB.setMI(NewMBB: &RestoreMBB, NewMI: MI); |
2006 | // Generate the restore of SGPR from SB.TmpVGPR. |
2007 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
2008 | // Don't need to load VGPR in. |
2009 | // Unpack lanes |
2010 | for (unsigned i = Offset * PVD.PerVGPR, |
2011 | e = std::min(a: (Offset + 1) * PVD.PerVGPR, b: SB.NumSubRegs); |
2012 | i < e; ++i) { |
2013 | Register SubReg = |
2014 | SB.NumSubRegs == 1 |
2015 | ? SB.SuperReg |
2016 | : Register(getSubReg(Reg: SB.SuperReg, Idx: SB.SplitParts[i])); |
2017 | bool LastSubReg = (i + 1 == e); |
2018 | auto MIB = BuildMI(BB&: *SB.MBB, I: MI, MIMD: SB.DL, MCID: SB.TII.get(Opcode: AMDGPU::V_READLANE_B32), |
2019 | DestReg: SubReg) |
2020 | .addReg(RegNo: SB.TmpVGPR, flags: getKillRegState(B: LastSubReg)) |
2021 | .addImm(Val: i); |
2022 | if (SB.NumSubRegs > 1 && i == 0) |
2023 | MIB.addReg(RegNo: SB.SuperReg, flags: RegState::ImplicitDefine); |
2024 | } |
2025 | } |
2026 | SB.restore(); |
2027 | |
2028 | SB.MFI.addToSpilledSGPRs(num: SB.NumSubRegs); |
2029 | return false; |
2030 | } |
2031 | |
2032 | /// Special case of eliminateFrameIndex. Returns true if the SGPR was spilled to |
2033 | /// a VGPR and the stack slot can be safely eliminated when all other users are |
2034 | /// handled. |
2035 | bool SIRegisterInfo::eliminateSGPRToVGPRSpillFrameIndex( |
2036 | MachineBasicBlock::iterator MI, int FI, RegScavenger *RS, |
2037 | SlotIndexes *Indexes, LiveIntervals *LIS, bool SpillToPhysVGPRLane) const { |
2038 | switch (MI->getOpcode()) { |
2039 | case AMDGPU::SI_SPILL_S1024_SAVE: |
2040 | case AMDGPU::SI_SPILL_S512_SAVE: |
2041 | case AMDGPU::SI_SPILL_S384_SAVE: |
2042 | case AMDGPU::SI_SPILL_S352_SAVE: |
2043 | case AMDGPU::SI_SPILL_S320_SAVE: |
2044 | case AMDGPU::SI_SPILL_S288_SAVE: |
2045 | case AMDGPU::SI_SPILL_S256_SAVE: |
2046 | case AMDGPU::SI_SPILL_S224_SAVE: |
2047 | case AMDGPU::SI_SPILL_S192_SAVE: |
2048 | case AMDGPU::SI_SPILL_S160_SAVE: |
2049 | case AMDGPU::SI_SPILL_S128_SAVE: |
2050 | case AMDGPU::SI_SPILL_S96_SAVE: |
2051 | case AMDGPU::SI_SPILL_S64_SAVE: |
2052 | case AMDGPU::SI_SPILL_S32_SAVE: |
2053 | return spillSGPR(MI, Index: FI, RS, Indexes, LIS, OnlyToVGPR: true, SpillToPhysVGPRLane); |
2054 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
2055 | case AMDGPU::SI_SPILL_S512_RESTORE: |
2056 | case AMDGPU::SI_SPILL_S384_RESTORE: |
2057 | case AMDGPU::SI_SPILL_S352_RESTORE: |
2058 | case AMDGPU::SI_SPILL_S320_RESTORE: |
2059 | case AMDGPU::SI_SPILL_S288_RESTORE: |
2060 | case AMDGPU::SI_SPILL_S256_RESTORE: |
2061 | case AMDGPU::SI_SPILL_S224_RESTORE: |
2062 | case AMDGPU::SI_SPILL_S192_RESTORE: |
2063 | case AMDGPU::SI_SPILL_S160_RESTORE: |
2064 | case AMDGPU::SI_SPILL_S128_RESTORE: |
2065 | case AMDGPU::SI_SPILL_S96_RESTORE: |
2066 | case AMDGPU::SI_SPILL_S64_RESTORE: |
2067 | case AMDGPU::SI_SPILL_S32_RESTORE: |
2068 | return restoreSGPR(MI, Index: FI, RS, Indexes, LIS, OnlyToVGPR: true, SpillToPhysVGPRLane); |
2069 | default: |
2070 | llvm_unreachable("not an SGPR spill instruction" ); |
2071 | } |
2072 | } |
2073 | |
2074 | bool SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI, |
2075 | int SPAdj, unsigned FIOperandNum, |
2076 | RegScavenger *RS) const { |
2077 | MachineFunction *MF = MI->getParent()->getParent(); |
2078 | MachineBasicBlock *MBB = MI->getParent(); |
2079 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
2080 | MachineFrameInfo &FrameInfo = MF->getFrameInfo(); |
2081 | const SIInstrInfo *TII = ST.getInstrInfo(); |
2082 | DebugLoc DL = MI->getDebugLoc(); |
2083 | |
2084 | assert(SPAdj == 0 && "unhandled SP adjustment in call sequence?" ); |
2085 | |
2086 | assert(MF->getRegInfo().isReserved(MFI->getScratchRSrcReg()) && |
2087 | "unreserved scratch RSRC register" ); |
2088 | |
2089 | MachineOperand &FIOp = MI->getOperand(i: FIOperandNum); |
2090 | int Index = MI->getOperand(i: FIOperandNum).getIndex(); |
2091 | |
2092 | Register FrameReg = FrameInfo.isFixedObjectIndex(ObjectIdx: Index) && hasBasePointer(MF: *MF) |
2093 | ? getBaseRegister() |
2094 | : getFrameRegister(MF: *MF); |
2095 | |
2096 | switch (MI->getOpcode()) { |
2097 | // SGPR register spill |
2098 | case AMDGPU::SI_SPILL_S1024_SAVE: |
2099 | case AMDGPU::SI_SPILL_S512_SAVE: |
2100 | case AMDGPU::SI_SPILL_S384_SAVE: |
2101 | case AMDGPU::SI_SPILL_S352_SAVE: |
2102 | case AMDGPU::SI_SPILL_S320_SAVE: |
2103 | case AMDGPU::SI_SPILL_S288_SAVE: |
2104 | case AMDGPU::SI_SPILL_S256_SAVE: |
2105 | case AMDGPU::SI_SPILL_S224_SAVE: |
2106 | case AMDGPU::SI_SPILL_S192_SAVE: |
2107 | case AMDGPU::SI_SPILL_S160_SAVE: |
2108 | case AMDGPU::SI_SPILL_S128_SAVE: |
2109 | case AMDGPU::SI_SPILL_S96_SAVE: |
2110 | case AMDGPU::SI_SPILL_S64_SAVE: |
2111 | case AMDGPU::SI_SPILL_S32_SAVE: { |
2112 | return spillSGPR(MI, Index, RS); |
2113 | } |
2114 | |
2115 | // SGPR register restore |
2116 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
2117 | case AMDGPU::SI_SPILL_S512_RESTORE: |
2118 | case AMDGPU::SI_SPILL_S384_RESTORE: |
2119 | case AMDGPU::SI_SPILL_S352_RESTORE: |
2120 | case AMDGPU::SI_SPILL_S320_RESTORE: |
2121 | case AMDGPU::SI_SPILL_S288_RESTORE: |
2122 | case AMDGPU::SI_SPILL_S256_RESTORE: |
2123 | case AMDGPU::SI_SPILL_S224_RESTORE: |
2124 | case AMDGPU::SI_SPILL_S192_RESTORE: |
2125 | case AMDGPU::SI_SPILL_S160_RESTORE: |
2126 | case AMDGPU::SI_SPILL_S128_RESTORE: |
2127 | case AMDGPU::SI_SPILL_S96_RESTORE: |
2128 | case AMDGPU::SI_SPILL_S64_RESTORE: |
2129 | case AMDGPU::SI_SPILL_S32_RESTORE: { |
2130 | return restoreSGPR(MI, Index, RS); |
2131 | } |
2132 | |
2133 | // VGPR register spill |
2134 | case AMDGPU::SI_SPILL_V1024_SAVE: |
2135 | case AMDGPU::SI_SPILL_V512_SAVE: |
2136 | case AMDGPU::SI_SPILL_V384_SAVE: |
2137 | case AMDGPU::SI_SPILL_V352_SAVE: |
2138 | case AMDGPU::SI_SPILL_V320_SAVE: |
2139 | case AMDGPU::SI_SPILL_V288_SAVE: |
2140 | case AMDGPU::SI_SPILL_V256_SAVE: |
2141 | case AMDGPU::SI_SPILL_V224_SAVE: |
2142 | case AMDGPU::SI_SPILL_V192_SAVE: |
2143 | case AMDGPU::SI_SPILL_V160_SAVE: |
2144 | case AMDGPU::SI_SPILL_V128_SAVE: |
2145 | case AMDGPU::SI_SPILL_V96_SAVE: |
2146 | case AMDGPU::SI_SPILL_V64_SAVE: |
2147 | case AMDGPU::SI_SPILL_V32_SAVE: |
2148 | case AMDGPU::SI_SPILL_A1024_SAVE: |
2149 | case AMDGPU::SI_SPILL_A512_SAVE: |
2150 | case AMDGPU::SI_SPILL_A384_SAVE: |
2151 | case AMDGPU::SI_SPILL_A352_SAVE: |
2152 | case AMDGPU::SI_SPILL_A320_SAVE: |
2153 | case AMDGPU::SI_SPILL_A288_SAVE: |
2154 | case AMDGPU::SI_SPILL_A256_SAVE: |
2155 | case AMDGPU::SI_SPILL_A224_SAVE: |
2156 | case AMDGPU::SI_SPILL_A192_SAVE: |
2157 | case AMDGPU::SI_SPILL_A160_SAVE: |
2158 | case AMDGPU::SI_SPILL_A128_SAVE: |
2159 | case AMDGPU::SI_SPILL_A96_SAVE: |
2160 | case AMDGPU::SI_SPILL_A64_SAVE: |
2161 | case AMDGPU::SI_SPILL_A32_SAVE: |
2162 | case AMDGPU::SI_SPILL_AV1024_SAVE: |
2163 | case AMDGPU::SI_SPILL_AV512_SAVE: |
2164 | case AMDGPU::SI_SPILL_AV384_SAVE: |
2165 | case AMDGPU::SI_SPILL_AV352_SAVE: |
2166 | case AMDGPU::SI_SPILL_AV320_SAVE: |
2167 | case AMDGPU::SI_SPILL_AV288_SAVE: |
2168 | case AMDGPU::SI_SPILL_AV256_SAVE: |
2169 | case AMDGPU::SI_SPILL_AV224_SAVE: |
2170 | case AMDGPU::SI_SPILL_AV192_SAVE: |
2171 | case AMDGPU::SI_SPILL_AV160_SAVE: |
2172 | case AMDGPU::SI_SPILL_AV128_SAVE: |
2173 | case AMDGPU::SI_SPILL_AV96_SAVE: |
2174 | case AMDGPU::SI_SPILL_AV64_SAVE: |
2175 | case AMDGPU::SI_SPILL_AV32_SAVE: |
2176 | case AMDGPU::SI_SPILL_WWM_V32_SAVE: |
2177 | case AMDGPU::SI_SPILL_WWM_AV32_SAVE: { |
2178 | const MachineOperand *VData = TII->getNamedOperand(MI&: *MI, |
2179 | OperandName: AMDGPU::OpName::vdata); |
2180 | assert(TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == |
2181 | MFI->getStackPtrOffsetReg()); |
2182 | |
2183 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
2184 | : AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
2185 | auto *MBB = MI->getParent(); |
2186 | bool IsWWMRegSpill = TII->isWWMRegSpillOpcode(Opcode: MI->getOpcode()); |
2187 | if (IsWWMRegSpill) { |
2188 | TII->insertScratchExecCopy(MF&: *MF, MBB&: *MBB, MBBI: MI, DL, Reg: MFI->getSGPRForEXECCopy(), |
2189 | IsSCCLive: RS->isRegUsed(Reg: AMDGPU::SCC)); |
2190 | } |
2191 | buildSpillLoadStore( |
2192 | MBB&: *MBB, MI, DL, LoadStoreOp: Opc, Index, ValueReg: VData->getReg(), IsKill: VData->isKill(), ScratchOffsetReg: FrameReg, |
2193 | InstOffset: TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::offset)->getImm(), |
2194 | MMO: *MI->memoperands_begin(), RS); |
2195 | MFI->addToSpilledVGPRs(num: getNumSubRegsForSpillOp(Op: MI->getOpcode())); |
2196 | if (IsWWMRegSpill) |
2197 | TII->restoreExec(MF&: *MF, MBB&: *MBB, MBBI: MI, DL, Reg: MFI->getSGPRForEXECCopy()); |
2198 | |
2199 | MI->eraseFromParent(); |
2200 | return true; |
2201 | } |
2202 | case AMDGPU::SI_SPILL_V32_RESTORE: |
2203 | case AMDGPU::SI_SPILL_V64_RESTORE: |
2204 | case AMDGPU::SI_SPILL_V96_RESTORE: |
2205 | case AMDGPU::SI_SPILL_V128_RESTORE: |
2206 | case AMDGPU::SI_SPILL_V160_RESTORE: |
2207 | case AMDGPU::SI_SPILL_V192_RESTORE: |
2208 | case AMDGPU::SI_SPILL_V224_RESTORE: |
2209 | case AMDGPU::SI_SPILL_V256_RESTORE: |
2210 | case AMDGPU::SI_SPILL_V288_RESTORE: |
2211 | case AMDGPU::SI_SPILL_V320_RESTORE: |
2212 | case AMDGPU::SI_SPILL_V352_RESTORE: |
2213 | case AMDGPU::SI_SPILL_V384_RESTORE: |
2214 | case AMDGPU::SI_SPILL_V512_RESTORE: |
2215 | case AMDGPU::SI_SPILL_V1024_RESTORE: |
2216 | case AMDGPU::SI_SPILL_A32_RESTORE: |
2217 | case AMDGPU::SI_SPILL_A64_RESTORE: |
2218 | case AMDGPU::SI_SPILL_A96_RESTORE: |
2219 | case AMDGPU::SI_SPILL_A128_RESTORE: |
2220 | case AMDGPU::SI_SPILL_A160_RESTORE: |
2221 | case AMDGPU::SI_SPILL_A192_RESTORE: |
2222 | case AMDGPU::SI_SPILL_A224_RESTORE: |
2223 | case AMDGPU::SI_SPILL_A256_RESTORE: |
2224 | case AMDGPU::SI_SPILL_A288_RESTORE: |
2225 | case AMDGPU::SI_SPILL_A320_RESTORE: |
2226 | case AMDGPU::SI_SPILL_A352_RESTORE: |
2227 | case AMDGPU::SI_SPILL_A384_RESTORE: |
2228 | case AMDGPU::SI_SPILL_A512_RESTORE: |
2229 | case AMDGPU::SI_SPILL_A1024_RESTORE: |
2230 | case AMDGPU::SI_SPILL_AV32_RESTORE: |
2231 | case AMDGPU::SI_SPILL_AV64_RESTORE: |
2232 | case AMDGPU::SI_SPILL_AV96_RESTORE: |
2233 | case AMDGPU::SI_SPILL_AV128_RESTORE: |
2234 | case AMDGPU::SI_SPILL_AV160_RESTORE: |
2235 | case AMDGPU::SI_SPILL_AV192_RESTORE: |
2236 | case AMDGPU::SI_SPILL_AV224_RESTORE: |
2237 | case AMDGPU::SI_SPILL_AV256_RESTORE: |
2238 | case AMDGPU::SI_SPILL_AV288_RESTORE: |
2239 | case AMDGPU::SI_SPILL_AV320_RESTORE: |
2240 | case AMDGPU::SI_SPILL_AV352_RESTORE: |
2241 | case AMDGPU::SI_SPILL_AV384_RESTORE: |
2242 | case AMDGPU::SI_SPILL_AV512_RESTORE: |
2243 | case AMDGPU::SI_SPILL_AV1024_RESTORE: |
2244 | case AMDGPU::SI_SPILL_WWM_V32_RESTORE: |
2245 | case AMDGPU::SI_SPILL_WWM_AV32_RESTORE: { |
2246 | const MachineOperand *VData = TII->getNamedOperand(MI&: *MI, |
2247 | OperandName: AMDGPU::OpName::vdata); |
2248 | assert(TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == |
2249 | MFI->getStackPtrOffsetReg()); |
2250 | |
2251 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_LOAD_DWORD_SADDR |
2252 | : AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
2253 | auto *MBB = MI->getParent(); |
2254 | bool IsWWMRegSpill = TII->isWWMRegSpillOpcode(Opcode: MI->getOpcode()); |
2255 | if (IsWWMRegSpill) { |
2256 | TII->insertScratchExecCopy(MF&: *MF, MBB&: *MBB, MBBI: MI, DL, Reg: MFI->getSGPRForEXECCopy(), |
2257 | IsSCCLive: RS->isRegUsed(Reg: AMDGPU::SCC)); |
2258 | } |
2259 | |
2260 | buildSpillLoadStore( |
2261 | MBB&: *MBB, MI, DL, LoadStoreOp: Opc, Index, ValueReg: VData->getReg(), IsKill: VData->isKill(), ScratchOffsetReg: FrameReg, |
2262 | InstOffset: TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::offset)->getImm(), |
2263 | MMO: *MI->memoperands_begin(), RS); |
2264 | |
2265 | if (IsWWMRegSpill) |
2266 | TII->restoreExec(MF&: *MF, MBB&: *MBB, MBBI: MI, DL, Reg: MFI->getSGPRForEXECCopy()); |
2267 | |
2268 | MI->eraseFromParent(); |
2269 | return true; |
2270 | } |
2271 | |
2272 | default: { |
2273 | // Other access to frame index |
2274 | const DebugLoc &DL = MI->getDebugLoc(); |
2275 | |
2276 | int64_t Offset = FrameInfo.getObjectOffset(ObjectIdx: Index); |
2277 | if (ST.enableFlatScratch()) { |
2278 | if (TII->isFLATScratch(MI: *MI)) { |
2279 | assert((int16_t)FIOperandNum == |
2280 | AMDGPU::getNamedOperandIdx(MI->getOpcode(), |
2281 | AMDGPU::OpName::saddr)); |
2282 | |
2283 | // The offset is always swizzled, just replace it |
2284 | if (FrameReg) |
2285 | FIOp.ChangeToRegister(Reg: FrameReg, isDef: false); |
2286 | |
2287 | MachineOperand *OffsetOp = |
2288 | TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::offset); |
2289 | int64_t NewOffset = Offset + OffsetOp->getImm(); |
2290 | if (TII->isLegalFLATOffset(Offset: NewOffset, AddrSpace: AMDGPUAS::PRIVATE_ADDRESS, |
2291 | FlatVariant: SIInstrFlags::FlatScratch)) { |
2292 | OffsetOp->setImm(NewOffset); |
2293 | if (FrameReg) |
2294 | return false; |
2295 | Offset = 0; |
2296 | } |
2297 | |
2298 | if (!Offset) { |
2299 | unsigned Opc = MI->getOpcode(); |
2300 | int NewOpc = -1; |
2301 | if (AMDGPU::hasNamedOperand(Opcode: Opc, NamedIdx: AMDGPU::OpName::vaddr)) { |
2302 | NewOpc = AMDGPU::getFlatScratchInstSVfromSVS(Opcode: Opc); |
2303 | } else if (ST.hasFlatScratchSTMode()) { |
2304 | // On GFX10 we have ST mode to use no registers for an address. |
2305 | // Otherwise we need to materialize 0 into an SGPR. |
2306 | NewOpc = AMDGPU::getFlatScratchInstSTfromSS(Opcode: Opc); |
2307 | } |
2308 | |
2309 | if (NewOpc != -1) { |
2310 | // removeOperand doesn't fixup tied operand indexes as it goes, so |
2311 | // it asserts. Untie vdst_in for now and retie them afterwards. |
2312 | int VDstIn = AMDGPU::getNamedOperandIdx(Opcode: Opc, |
2313 | NamedIdx: AMDGPU::OpName::vdst_in); |
2314 | bool TiedVDst = VDstIn != -1 && |
2315 | MI->getOperand(i: VDstIn).isReg() && |
2316 | MI->getOperand(i: VDstIn).isTied(); |
2317 | if (TiedVDst) |
2318 | MI->untieRegOperand(OpIdx: VDstIn); |
2319 | |
2320 | MI->removeOperand( |
2321 | OpNo: AMDGPU::getNamedOperandIdx(Opcode: Opc, NamedIdx: AMDGPU::OpName::saddr)); |
2322 | |
2323 | if (TiedVDst) { |
2324 | int NewVDst = |
2325 | AMDGPU::getNamedOperandIdx(Opcode: NewOpc, NamedIdx: AMDGPU::OpName::vdst); |
2326 | int NewVDstIn = |
2327 | AMDGPU::getNamedOperandIdx(Opcode: NewOpc, NamedIdx: AMDGPU::OpName::vdst_in); |
2328 | assert (NewVDst != -1 && NewVDstIn != -1 && "Must be tied!" ); |
2329 | MI->tieOperands(DefIdx: NewVDst, UseIdx: NewVDstIn); |
2330 | } |
2331 | MI->setDesc(TII->get(Opcode: NewOpc)); |
2332 | return false; |
2333 | } |
2334 | } |
2335 | } |
2336 | |
2337 | if (!FrameReg) { |
2338 | FIOp.ChangeToImmediate(ImmVal: Offset); |
2339 | if (TII->isImmOperandLegal(MI: *MI, OpNo: FIOperandNum, MO: FIOp)) |
2340 | return false; |
2341 | } |
2342 | |
2343 | // We need to use register here. Check if we can use an SGPR or need |
2344 | // a VGPR. |
2345 | FIOp.ChangeToRegister(Reg: AMDGPU::M0, isDef: false); |
2346 | bool UseSGPR = TII->isOperandLegal(MI: *MI, OpIdx: FIOperandNum, MO: &FIOp); |
2347 | |
2348 | if (!Offset && FrameReg && UseSGPR) { |
2349 | FIOp.setReg(FrameReg); |
2350 | return false; |
2351 | } |
2352 | |
2353 | const TargetRegisterClass *RC = UseSGPR ? &AMDGPU::SReg_32_XM0RegClass |
2354 | : &AMDGPU::VGPR_32RegClass; |
2355 | |
2356 | Register TmpReg = |
2357 | RS->scavengeRegisterBackwards(RC: *RC, To: MI, RestoreAfter: false, SPAdj: 0, AllowSpill: !UseSGPR); |
2358 | FIOp.setReg(TmpReg); |
2359 | FIOp.setIsKill(); |
2360 | |
2361 | if ((!FrameReg || !Offset) && TmpReg) { |
2362 | unsigned Opc = UseSGPR ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32; |
2363 | auto MIB = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: Opc), DestReg: TmpReg); |
2364 | if (FrameReg) |
2365 | MIB.addReg(RegNo: FrameReg); |
2366 | else |
2367 | MIB.addImm(Val: Offset); |
2368 | |
2369 | return false; |
2370 | } |
2371 | |
2372 | bool NeedSaveSCC = RS->isRegUsed(Reg: AMDGPU::SCC) && |
2373 | !MI->definesRegister(Reg: AMDGPU::SCC, /*TRI=*/nullptr); |
2374 | |
2375 | Register TmpSReg = |
2376 | UseSGPR ? TmpReg |
2377 | : RS->scavengeRegisterBackwards(RC: AMDGPU::SReg_32_XM0RegClass, |
2378 | To: MI, RestoreAfter: false, SPAdj: 0, AllowSpill: !UseSGPR); |
2379 | |
2380 | // TODO: for flat scratch another attempt can be made with a VGPR index |
2381 | // if no SGPRs can be scavenged. |
2382 | if ((!TmpSReg && !FrameReg) || (!TmpReg && !UseSGPR)) |
2383 | report_fatal_error(reason: "Cannot scavenge register in FI elimination!" ); |
2384 | |
2385 | if (!TmpSReg) { |
2386 | // Use frame register and restore it after. |
2387 | TmpSReg = FrameReg; |
2388 | FIOp.setReg(FrameReg); |
2389 | FIOp.setIsKill(false); |
2390 | } |
2391 | |
2392 | if (NeedSaveSCC) { |
2393 | assert(!(Offset & 0x1) && "Flat scratch offset must be aligned!" ); |
2394 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADDC_U32), DestReg: TmpSReg) |
2395 | .addReg(RegNo: FrameReg) |
2396 | .addImm(Val: Offset); |
2397 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_BITCMP1_B32)) |
2398 | .addReg(RegNo: TmpSReg) |
2399 | .addImm(Val: 0); |
2400 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_BITSET0_B32), DestReg: TmpSReg) |
2401 | .addImm(Val: 0) |
2402 | .addReg(RegNo: TmpSReg); |
2403 | } else { |
2404 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: TmpSReg) |
2405 | .addReg(RegNo: FrameReg) |
2406 | .addImm(Val: Offset); |
2407 | } |
2408 | |
2409 | if (!UseSGPR) |
2410 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32), DestReg: TmpReg) |
2411 | .addReg(RegNo: TmpSReg, flags: RegState::Kill); |
2412 | |
2413 | if (TmpSReg == FrameReg) { |
2414 | // Undo frame register modification. |
2415 | if (NeedSaveSCC && |
2416 | !MI->registerDefIsDead(Reg: AMDGPU::SCC, /*TRI=*/nullptr)) { |
2417 | MachineBasicBlock::iterator I = |
2418 | BuildMI(BB&: *MBB, I: std::next(x: MI), MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADDC_U32), |
2419 | DestReg: TmpSReg) |
2420 | .addReg(RegNo: FrameReg) |
2421 | .addImm(Val: -Offset); |
2422 | I = BuildMI(BB&: *MBB, I: std::next(x: I), MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_BITCMP1_B32)) |
2423 | .addReg(RegNo: TmpSReg) |
2424 | .addImm(Val: 0); |
2425 | BuildMI(BB&: *MBB, I: std::next(x: I), MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_BITSET0_B32), |
2426 | DestReg: TmpSReg) |
2427 | .addImm(Val: 0) |
2428 | .addReg(RegNo: TmpSReg); |
2429 | } else { |
2430 | BuildMI(BB&: *MBB, I: std::next(x: MI), MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), |
2431 | DestReg: FrameReg) |
2432 | .addReg(RegNo: FrameReg) |
2433 | .addImm(Val: -Offset); |
2434 | } |
2435 | } |
2436 | |
2437 | return false; |
2438 | } |
2439 | |
2440 | bool IsMUBUF = TII->isMUBUF(MI: *MI); |
2441 | |
2442 | if (!IsMUBUF && !MFI->isBottomOfStack()) { |
2443 | // Convert to a swizzled stack address by scaling by the wave size. |
2444 | // In an entry function/kernel the offset is already swizzled. |
2445 | bool IsSALU = isSGPRClass(RC: TII->getOpRegClass(MI: *MI, OpNo: FIOperandNum)); |
2446 | bool LiveSCC = RS->isRegUsed(Reg: AMDGPU::SCC) && |
2447 | !MI->definesRegister(Reg: AMDGPU::SCC, /*TRI=*/nullptr); |
2448 | const TargetRegisterClass *RC = IsSALU && !LiveSCC |
2449 | ? &AMDGPU::SReg_32RegClass |
2450 | : &AMDGPU::VGPR_32RegClass; |
2451 | bool IsCopy = MI->getOpcode() == AMDGPU::V_MOV_B32_e32 || |
2452 | MI->getOpcode() == AMDGPU::V_MOV_B32_e64; |
2453 | Register ResultReg = |
2454 | IsCopy ? MI->getOperand(i: 0).getReg() |
2455 | : RS->scavengeRegisterBackwards(RC: *RC, To: MI, RestoreAfter: false, SPAdj: 0); |
2456 | |
2457 | int64_t Offset = FrameInfo.getObjectOffset(ObjectIdx: Index); |
2458 | if (Offset == 0) { |
2459 | unsigned OpCode = IsSALU && !LiveSCC ? AMDGPU::S_LSHR_B32 |
2460 | : AMDGPU::V_LSHRREV_B32_e64; |
2461 | auto Shift = BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: OpCode), DestReg: ResultReg); |
2462 | if (OpCode == AMDGPU::V_LSHRREV_B32_e64) |
2463 | // For V_LSHRREV, the operands are reversed (the shift count goes |
2464 | // first). |
2465 | Shift.addImm(Val: ST.getWavefrontSizeLog2()).addReg(RegNo: FrameReg); |
2466 | else |
2467 | Shift.addReg(RegNo: FrameReg).addImm(Val: ST.getWavefrontSizeLog2()); |
2468 | if (IsSALU && !LiveSCC) |
2469 | Shift.getInstr()->getOperand(i: 3).setIsDead(); // Mark SCC as dead. |
2470 | if (IsSALU && LiveSCC) { |
2471 | Register NewDest = RS->scavengeRegisterBackwards( |
2472 | RC: AMDGPU::SReg_32RegClass, To: Shift, RestoreAfter: false, SPAdj: 0); |
2473 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_READFIRSTLANE_B32), |
2474 | DestReg: NewDest) |
2475 | .addReg(RegNo: ResultReg); |
2476 | ResultReg = NewDest; |
2477 | } |
2478 | } else { |
2479 | MachineInstrBuilder MIB; |
2480 | if (!IsSALU) { |
2481 | if ((MIB = TII->getAddNoCarry(MBB&: *MBB, I: MI, DL, DestReg: ResultReg, RS&: *RS)) != |
2482 | nullptr) { |
2483 | // Reuse ResultReg in intermediate step. |
2484 | Register ScaledReg = ResultReg; |
2485 | |
2486 | BuildMI(BB&: *MBB, I&: *MIB, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_LSHRREV_B32_e64), |
2487 | DestReg: ScaledReg) |
2488 | .addImm(Val: ST.getWavefrontSizeLog2()) |
2489 | .addReg(RegNo: FrameReg); |
2490 | |
2491 | const bool IsVOP2 = MIB->getOpcode() == AMDGPU::V_ADD_U32_e32; |
2492 | |
2493 | // TODO: Fold if use instruction is another add of a constant. |
2494 | if (IsVOP2 || AMDGPU::isInlinableLiteral32(Literal: Offset, HasInv2Pi: ST.hasInv2PiInlineImm())) { |
2495 | // FIXME: This can fail |
2496 | MIB.addImm(Val: Offset); |
2497 | MIB.addReg(RegNo: ScaledReg, flags: RegState::Kill); |
2498 | if (!IsVOP2) |
2499 | MIB.addImm(Val: 0); // clamp bit |
2500 | } else { |
2501 | assert(MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 && |
2502 | "Need to reuse carry out register" ); |
2503 | |
2504 | // Use scavenged unused carry out as offset register. |
2505 | Register ConstOffsetReg; |
2506 | if (!isWave32) |
2507 | ConstOffsetReg = getSubReg(Reg: MIB.getReg(Idx: 1), Idx: AMDGPU::sub0); |
2508 | else |
2509 | ConstOffsetReg = MIB.getReg(Idx: 1); |
2510 | |
2511 | BuildMI(BB&: *MBB, I&: *MIB, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: ConstOffsetReg) |
2512 | .addImm(Val: Offset); |
2513 | MIB.addReg(RegNo: ConstOffsetReg, flags: RegState::Kill); |
2514 | MIB.addReg(RegNo: ScaledReg, flags: RegState::Kill); |
2515 | MIB.addImm(Val: 0); // clamp bit |
2516 | } |
2517 | } |
2518 | } |
2519 | if (!MIB || IsSALU) { |
2520 | // We have to produce a carry out, and there isn't a free SGPR pair |
2521 | // for it. We can keep the whole computation on the SALU to avoid |
2522 | // clobbering an additional register at the cost of an extra mov. |
2523 | |
2524 | // We may have 1 free scratch SGPR even though a carry out is |
2525 | // unavailable. Only one additional mov is needed. |
2526 | Register TmpScaledReg = RS->scavengeRegisterBackwards( |
2527 | RC: AMDGPU::SReg_32_XM0RegClass, To: MI, RestoreAfter: false, SPAdj: 0, AllowSpill: false); |
2528 | Register ScaledReg = TmpScaledReg.isValid() ? TmpScaledReg : FrameReg; |
2529 | |
2530 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_LSHR_B32), DestReg: ScaledReg) |
2531 | .addReg(RegNo: FrameReg) |
2532 | .addImm(Val: ST.getWavefrontSizeLog2()); |
2533 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: ScaledReg) |
2534 | .addReg(RegNo: ScaledReg, flags: RegState::Kill) |
2535 | .addImm(Val: Offset); |
2536 | if (!IsSALU) |
2537 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: ResultReg) |
2538 | .addReg(RegNo: ScaledReg, flags: RegState::Kill); |
2539 | else |
2540 | ResultReg = ScaledReg; |
2541 | |
2542 | // If there were truly no free SGPRs, we need to undo everything. |
2543 | if (!TmpScaledReg.isValid()) { |
2544 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_ADD_I32), DestReg: ScaledReg) |
2545 | .addReg(RegNo: ScaledReg, flags: RegState::Kill) |
2546 | .addImm(Val: -Offset); |
2547 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_LSHL_B32), DestReg: ScaledReg) |
2548 | .addReg(RegNo: FrameReg) |
2549 | .addImm(Val: ST.getWavefrontSizeLog2()); |
2550 | } |
2551 | } |
2552 | } |
2553 | |
2554 | // Don't introduce an extra copy if we're just materializing in a mov. |
2555 | if (IsCopy) { |
2556 | MI->eraseFromParent(); |
2557 | return true; |
2558 | } |
2559 | FIOp.ChangeToRegister(Reg: ResultReg, isDef: false, isImp: false, isKill: true); |
2560 | return false; |
2561 | } |
2562 | |
2563 | if (IsMUBUF) { |
2564 | // Disable offen so we don't need a 0 vgpr base. |
2565 | assert(static_cast<int>(FIOperandNum) == |
2566 | AMDGPU::getNamedOperandIdx(MI->getOpcode(), |
2567 | AMDGPU::OpName::vaddr)); |
2568 | |
2569 | auto &SOffset = *TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::soffset); |
2570 | assert((SOffset.isImm() && SOffset.getImm() == 0)); |
2571 | |
2572 | if (FrameReg != AMDGPU::NoRegister) |
2573 | SOffset.ChangeToRegister(Reg: FrameReg, isDef: false); |
2574 | |
2575 | int64_t Offset = FrameInfo.getObjectOffset(ObjectIdx: Index); |
2576 | int64_t OldImm |
2577 | = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::offset)->getImm(); |
2578 | int64_t NewOffset = OldImm + Offset; |
2579 | |
2580 | if (TII->isLegalMUBUFImmOffset(Imm: NewOffset) && |
2581 | buildMUBUFOffsetLoadStore(ST, MFI&: FrameInfo, MI, Index, Offset: NewOffset)) { |
2582 | MI->eraseFromParent(); |
2583 | return true; |
2584 | } |
2585 | } |
2586 | |
2587 | // If the offset is simply too big, don't convert to a scratch wave offset |
2588 | // relative index. |
2589 | |
2590 | FIOp.ChangeToImmediate(ImmVal: Offset); |
2591 | if (!TII->isImmOperandLegal(MI: *MI, OpNo: FIOperandNum, MO: FIOp)) { |
2592 | Register TmpReg = RS->scavengeRegisterBackwards(RC: AMDGPU::VGPR_32RegClass, |
2593 | To: MI, RestoreAfter: false, SPAdj: 0); |
2594 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32), DestReg: TmpReg) |
2595 | .addImm(Val: Offset); |
2596 | FIOp.ChangeToRegister(Reg: TmpReg, isDef: false, isImp: false, isKill: true); |
2597 | } |
2598 | } |
2599 | } |
2600 | return false; |
2601 | } |
2602 | |
2603 | StringRef SIRegisterInfo::getRegAsmName(MCRegister Reg) const { |
2604 | return AMDGPUInstPrinter::getRegisterName(Reg); |
2605 | } |
2606 | |
2607 | unsigned AMDGPU::getRegBitWidth(const TargetRegisterClass &RC) { |
2608 | return getRegBitWidth(RCID: RC.getID()); |
2609 | } |
2610 | |
2611 | static const TargetRegisterClass * |
2612 | getAnyVGPRClassForBitWidth(unsigned BitWidth) { |
2613 | if (BitWidth == 64) |
2614 | return &AMDGPU::VReg_64RegClass; |
2615 | if (BitWidth == 96) |
2616 | return &AMDGPU::VReg_96RegClass; |
2617 | if (BitWidth == 128) |
2618 | return &AMDGPU::VReg_128RegClass; |
2619 | if (BitWidth == 160) |
2620 | return &AMDGPU::VReg_160RegClass; |
2621 | if (BitWidth == 192) |
2622 | return &AMDGPU::VReg_192RegClass; |
2623 | if (BitWidth == 224) |
2624 | return &AMDGPU::VReg_224RegClass; |
2625 | if (BitWidth == 256) |
2626 | return &AMDGPU::VReg_256RegClass; |
2627 | if (BitWidth == 288) |
2628 | return &AMDGPU::VReg_288RegClass; |
2629 | if (BitWidth == 320) |
2630 | return &AMDGPU::VReg_320RegClass; |
2631 | if (BitWidth == 352) |
2632 | return &AMDGPU::VReg_352RegClass; |
2633 | if (BitWidth == 384) |
2634 | return &AMDGPU::VReg_384RegClass; |
2635 | if (BitWidth == 512) |
2636 | return &AMDGPU::VReg_512RegClass; |
2637 | if (BitWidth == 1024) |
2638 | return &AMDGPU::VReg_1024RegClass; |
2639 | |
2640 | return nullptr; |
2641 | } |
2642 | |
2643 | static const TargetRegisterClass * |
2644 | getAlignedVGPRClassForBitWidth(unsigned BitWidth) { |
2645 | if (BitWidth == 64) |
2646 | return &AMDGPU::VReg_64_Align2RegClass; |
2647 | if (BitWidth == 96) |
2648 | return &AMDGPU::VReg_96_Align2RegClass; |
2649 | if (BitWidth == 128) |
2650 | return &AMDGPU::VReg_128_Align2RegClass; |
2651 | if (BitWidth == 160) |
2652 | return &AMDGPU::VReg_160_Align2RegClass; |
2653 | if (BitWidth == 192) |
2654 | return &AMDGPU::VReg_192_Align2RegClass; |
2655 | if (BitWidth == 224) |
2656 | return &AMDGPU::VReg_224_Align2RegClass; |
2657 | if (BitWidth == 256) |
2658 | return &AMDGPU::VReg_256_Align2RegClass; |
2659 | if (BitWidth == 288) |
2660 | return &AMDGPU::VReg_288_Align2RegClass; |
2661 | if (BitWidth == 320) |
2662 | return &AMDGPU::VReg_320_Align2RegClass; |
2663 | if (BitWidth == 352) |
2664 | return &AMDGPU::VReg_352_Align2RegClass; |
2665 | if (BitWidth == 384) |
2666 | return &AMDGPU::VReg_384_Align2RegClass; |
2667 | if (BitWidth == 512) |
2668 | return &AMDGPU::VReg_512_Align2RegClass; |
2669 | if (BitWidth == 1024) |
2670 | return &AMDGPU::VReg_1024_Align2RegClass; |
2671 | |
2672 | return nullptr; |
2673 | } |
2674 | |
2675 | const TargetRegisterClass * |
2676 | SIRegisterInfo::getVGPRClassForBitWidth(unsigned BitWidth) const { |
2677 | if (BitWidth == 1) |
2678 | return &AMDGPU::VReg_1RegClass; |
2679 | if (BitWidth == 16) |
2680 | return &AMDGPU::VGPR_16RegClass; |
2681 | if (BitWidth == 32) |
2682 | return &AMDGPU::VGPR_32RegClass; |
2683 | return ST.needsAlignedVGPRs() ? getAlignedVGPRClassForBitWidth(BitWidth) |
2684 | : getAnyVGPRClassForBitWidth(BitWidth); |
2685 | } |
2686 | |
2687 | static const TargetRegisterClass * |
2688 | getAnyAGPRClassForBitWidth(unsigned BitWidth) { |
2689 | if (BitWidth == 64) |
2690 | return &AMDGPU::AReg_64RegClass; |
2691 | if (BitWidth == 96) |
2692 | return &AMDGPU::AReg_96RegClass; |
2693 | if (BitWidth == 128) |
2694 | return &AMDGPU::AReg_128RegClass; |
2695 | if (BitWidth == 160) |
2696 | return &AMDGPU::AReg_160RegClass; |
2697 | if (BitWidth == 192) |
2698 | return &AMDGPU::AReg_192RegClass; |
2699 | if (BitWidth == 224) |
2700 | return &AMDGPU::AReg_224RegClass; |
2701 | if (BitWidth == 256) |
2702 | return &AMDGPU::AReg_256RegClass; |
2703 | if (BitWidth == 288) |
2704 | return &AMDGPU::AReg_288RegClass; |
2705 | if (BitWidth == 320) |
2706 | return &AMDGPU::AReg_320RegClass; |
2707 | if (BitWidth == 352) |
2708 | return &AMDGPU::AReg_352RegClass; |
2709 | if (BitWidth == 384) |
2710 | return &AMDGPU::AReg_384RegClass; |
2711 | if (BitWidth == 512) |
2712 | return &AMDGPU::AReg_512RegClass; |
2713 | if (BitWidth == 1024) |
2714 | return &AMDGPU::AReg_1024RegClass; |
2715 | |
2716 | return nullptr; |
2717 | } |
2718 | |
2719 | static const TargetRegisterClass * |
2720 | getAlignedAGPRClassForBitWidth(unsigned BitWidth) { |
2721 | if (BitWidth == 64) |
2722 | return &AMDGPU::AReg_64_Align2RegClass; |
2723 | if (BitWidth == 96) |
2724 | return &AMDGPU::AReg_96_Align2RegClass; |
2725 | if (BitWidth == 128) |
2726 | return &AMDGPU::AReg_128_Align2RegClass; |
2727 | if (BitWidth == 160) |
2728 | return &AMDGPU::AReg_160_Align2RegClass; |
2729 | if (BitWidth == 192) |
2730 | return &AMDGPU::AReg_192_Align2RegClass; |
2731 | if (BitWidth == 224) |
2732 | return &AMDGPU::AReg_224_Align2RegClass; |
2733 | if (BitWidth == 256) |
2734 | return &AMDGPU::AReg_256_Align2RegClass; |
2735 | if (BitWidth == 288) |
2736 | return &AMDGPU::AReg_288_Align2RegClass; |
2737 | if (BitWidth == 320) |
2738 | return &AMDGPU::AReg_320_Align2RegClass; |
2739 | if (BitWidth == 352) |
2740 | return &AMDGPU::AReg_352_Align2RegClass; |
2741 | if (BitWidth == 384) |
2742 | return &AMDGPU::AReg_384_Align2RegClass; |
2743 | if (BitWidth == 512) |
2744 | return &AMDGPU::AReg_512_Align2RegClass; |
2745 | if (BitWidth == 1024) |
2746 | return &AMDGPU::AReg_1024_Align2RegClass; |
2747 | |
2748 | return nullptr; |
2749 | } |
2750 | |
2751 | const TargetRegisterClass * |
2752 | SIRegisterInfo::getAGPRClassForBitWidth(unsigned BitWidth) const { |
2753 | if (BitWidth == 16) |
2754 | return &AMDGPU::AGPR_LO16RegClass; |
2755 | if (BitWidth == 32) |
2756 | return &AMDGPU::AGPR_32RegClass; |
2757 | return ST.needsAlignedVGPRs() ? getAlignedAGPRClassForBitWidth(BitWidth) |
2758 | : getAnyAGPRClassForBitWidth(BitWidth); |
2759 | } |
2760 | |
2761 | static const TargetRegisterClass * |
2762 | getAnyVectorSuperClassForBitWidth(unsigned BitWidth) { |
2763 | if (BitWidth == 64) |
2764 | return &AMDGPU::AV_64RegClass; |
2765 | if (BitWidth == 96) |
2766 | return &AMDGPU::AV_96RegClass; |
2767 | if (BitWidth == 128) |
2768 | return &AMDGPU::AV_128RegClass; |
2769 | if (BitWidth == 160) |
2770 | return &AMDGPU::AV_160RegClass; |
2771 | if (BitWidth == 192) |
2772 | return &AMDGPU::AV_192RegClass; |
2773 | if (BitWidth == 224) |
2774 | return &AMDGPU::AV_224RegClass; |
2775 | if (BitWidth == 256) |
2776 | return &AMDGPU::AV_256RegClass; |
2777 | if (BitWidth == 288) |
2778 | return &AMDGPU::AV_288RegClass; |
2779 | if (BitWidth == 320) |
2780 | return &AMDGPU::AV_320RegClass; |
2781 | if (BitWidth == 352) |
2782 | return &AMDGPU::AV_352RegClass; |
2783 | if (BitWidth == 384) |
2784 | return &AMDGPU::AV_384RegClass; |
2785 | if (BitWidth == 512) |
2786 | return &AMDGPU::AV_512RegClass; |
2787 | if (BitWidth == 1024) |
2788 | return &AMDGPU::AV_1024RegClass; |
2789 | |
2790 | return nullptr; |
2791 | } |
2792 | |
2793 | static const TargetRegisterClass * |
2794 | getAlignedVectorSuperClassForBitWidth(unsigned BitWidth) { |
2795 | if (BitWidth == 64) |
2796 | return &AMDGPU::AV_64_Align2RegClass; |
2797 | if (BitWidth == 96) |
2798 | return &AMDGPU::AV_96_Align2RegClass; |
2799 | if (BitWidth == 128) |
2800 | return &AMDGPU::AV_128_Align2RegClass; |
2801 | if (BitWidth == 160) |
2802 | return &AMDGPU::AV_160_Align2RegClass; |
2803 | if (BitWidth == 192) |
2804 | return &AMDGPU::AV_192_Align2RegClass; |
2805 | if (BitWidth == 224) |
2806 | return &AMDGPU::AV_224_Align2RegClass; |
2807 | if (BitWidth == 256) |
2808 | return &AMDGPU::AV_256_Align2RegClass; |
2809 | if (BitWidth == 288) |
2810 | return &AMDGPU::AV_288_Align2RegClass; |
2811 | if (BitWidth == 320) |
2812 | return &AMDGPU::AV_320_Align2RegClass; |
2813 | if (BitWidth == 352) |
2814 | return &AMDGPU::AV_352_Align2RegClass; |
2815 | if (BitWidth == 384) |
2816 | return &AMDGPU::AV_384_Align2RegClass; |
2817 | if (BitWidth == 512) |
2818 | return &AMDGPU::AV_512_Align2RegClass; |
2819 | if (BitWidth == 1024) |
2820 | return &AMDGPU::AV_1024_Align2RegClass; |
2821 | |
2822 | return nullptr; |
2823 | } |
2824 | |
2825 | const TargetRegisterClass * |
2826 | SIRegisterInfo::getVectorSuperClassForBitWidth(unsigned BitWidth) const { |
2827 | if (BitWidth == 32) |
2828 | return &AMDGPU::AV_32RegClass; |
2829 | return ST.needsAlignedVGPRs() |
2830 | ? getAlignedVectorSuperClassForBitWidth(BitWidth) |
2831 | : getAnyVectorSuperClassForBitWidth(BitWidth); |
2832 | } |
2833 | |
2834 | const TargetRegisterClass * |
2835 | SIRegisterInfo::getSGPRClassForBitWidth(unsigned BitWidth) { |
2836 | if (BitWidth == 16) |
2837 | return &AMDGPU::SGPR_LO16RegClass; |
2838 | if (BitWidth == 32) |
2839 | return &AMDGPU::SReg_32RegClass; |
2840 | if (BitWidth == 64) |
2841 | return &AMDGPU::SReg_64RegClass; |
2842 | if (BitWidth == 96) |
2843 | return &AMDGPU::SGPR_96RegClass; |
2844 | if (BitWidth == 128) |
2845 | return &AMDGPU::SGPR_128RegClass; |
2846 | if (BitWidth == 160) |
2847 | return &AMDGPU::SGPR_160RegClass; |
2848 | if (BitWidth == 192) |
2849 | return &AMDGPU::SGPR_192RegClass; |
2850 | if (BitWidth == 224) |
2851 | return &AMDGPU::SGPR_224RegClass; |
2852 | if (BitWidth == 256) |
2853 | return &AMDGPU::SGPR_256RegClass; |
2854 | if (BitWidth == 288) |
2855 | return &AMDGPU::SGPR_288RegClass; |
2856 | if (BitWidth == 320) |
2857 | return &AMDGPU::SGPR_320RegClass; |
2858 | if (BitWidth == 352) |
2859 | return &AMDGPU::SGPR_352RegClass; |
2860 | if (BitWidth == 384) |
2861 | return &AMDGPU::SGPR_384RegClass; |
2862 | if (BitWidth == 512) |
2863 | return &AMDGPU::SGPR_512RegClass; |
2864 | if (BitWidth == 1024) |
2865 | return &AMDGPU::SGPR_1024RegClass; |
2866 | |
2867 | return nullptr; |
2868 | } |
2869 | |
2870 | bool SIRegisterInfo::isSGPRReg(const MachineRegisterInfo &MRI, |
2871 | Register Reg) const { |
2872 | const TargetRegisterClass *RC; |
2873 | if (Reg.isVirtual()) |
2874 | RC = MRI.getRegClass(Reg); |
2875 | else |
2876 | RC = getPhysRegBaseClass(Reg); |
2877 | return RC ? isSGPRClass(RC) : false; |
2878 | } |
2879 | |
2880 | const TargetRegisterClass * |
2881 | SIRegisterInfo::getEquivalentVGPRClass(const TargetRegisterClass *SRC) const { |
2882 | unsigned Size = getRegSizeInBits(RC: *SRC); |
2883 | const TargetRegisterClass *VRC = getVGPRClassForBitWidth(BitWidth: Size); |
2884 | assert(VRC && "Invalid register class size" ); |
2885 | return VRC; |
2886 | } |
2887 | |
2888 | const TargetRegisterClass * |
2889 | SIRegisterInfo::getEquivalentAGPRClass(const TargetRegisterClass *SRC) const { |
2890 | unsigned Size = getRegSizeInBits(RC: *SRC); |
2891 | const TargetRegisterClass *ARC = getAGPRClassForBitWidth(BitWidth: Size); |
2892 | assert(ARC && "Invalid register class size" ); |
2893 | return ARC; |
2894 | } |
2895 | |
2896 | const TargetRegisterClass * |
2897 | SIRegisterInfo::getEquivalentSGPRClass(const TargetRegisterClass *VRC) const { |
2898 | unsigned Size = getRegSizeInBits(RC: *VRC); |
2899 | if (Size == 32) |
2900 | return &AMDGPU::SGPR_32RegClass; |
2901 | const TargetRegisterClass *SRC = getSGPRClassForBitWidth(BitWidth: Size); |
2902 | assert(SRC && "Invalid register class size" ); |
2903 | return SRC; |
2904 | } |
2905 | |
2906 | const TargetRegisterClass * |
2907 | SIRegisterInfo::getCompatibleSubRegClass(const TargetRegisterClass *SuperRC, |
2908 | const TargetRegisterClass *SubRC, |
2909 | unsigned SubIdx) const { |
2910 | // Ensure this subregister index is aligned in the super register. |
2911 | const TargetRegisterClass *MatchRC = |
2912 | getMatchingSuperRegClass(A: SuperRC, B: SubRC, Idx: SubIdx); |
2913 | return MatchRC && MatchRC->hasSubClassEq(RC: SuperRC) ? MatchRC : nullptr; |
2914 | } |
2915 | |
2916 | bool SIRegisterInfo::opCanUseInlineConstant(unsigned OpType) const { |
2917 | if (OpType >= AMDGPU::OPERAND_REG_INLINE_AC_FIRST && |
2918 | OpType <= AMDGPU::OPERAND_REG_INLINE_AC_LAST) |
2919 | return !ST.hasMFMAInlineLiteralBug(); |
2920 | |
2921 | return OpType >= AMDGPU::OPERAND_SRC_FIRST && |
2922 | OpType <= AMDGPU::OPERAND_SRC_LAST; |
2923 | } |
2924 | |
2925 | bool SIRegisterInfo::shouldRewriteCopySrc( |
2926 | const TargetRegisterClass *DefRC, |
2927 | unsigned DefSubReg, |
2928 | const TargetRegisterClass *SrcRC, |
2929 | unsigned SrcSubReg) const { |
2930 | // We want to prefer the smallest register class possible, so we don't want to |
2931 | // stop and rewrite on anything that looks like a subregister |
2932 | // extract. Operations mostly don't care about the super register class, so we |
2933 | // only want to stop on the most basic of copies between the same register |
2934 | // class. |
2935 | // |
2936 | // e.g. if we have something like |
2937 | // %0 = ... |
2938 | // %1 = ... |
2939 | // %2 = REG_SEQUENCE %0, sub0, %1, sub1, %2, sub2 |
2940 | // %3 = COPY %2, sub0 |
2941 | // |
2942 | // We want to look through the COPY to find: |
2943 | // => %3 = COPY %0 |
2944 | |
2945 | // Plain copy. |
2946 | return getCommonSubClass(A: DefRC, B: SrcRC) != nullptr; |
2947 | } |
2948 | |
2949 | bool SIRegisterInfo::opCanUseLiteralConstant(unsigned OpType) const { |
2950 | // TODO: 64-bit operands have extending behavior from 32-bit literal. |
2951 | return OpType >= AMDGPU::OPERAND_REG_IMM_FIRST && |
2952 | OpType <= AMDGPU::OPERAND_REG_IMM_LAST; |
2953 | } |
2954 | |
2955 | /// Returns a lowest register that is not used at any point in the function. |
2956 | /// If all registers are used, then this function will return |
2957 | /// AMDGPU::NoRegister. If \p ReserveHighestRegister = true, then return |
2958 | /// highest unused register. |
2959 | MCRegister SIRegisterInfo::findUnusedRegister( |
2960 | const MachineRegisterInfo &MRI, const TargetRegisterClass *RC, |
2961 | const MachineFunction &MF, bool ReserveHighestRegister) const { |
2962 | if (ReserveHighestRegister) { |
2963 | for (MCRegister Reg : reverse(C: *RC)) |
2964 | if (MRI.isAllocatable(PhysReg: Reg) && !MRI.isPhysRegUsed(PhysReg: Reg)) |
2965 | return Reg; |
2966 | } else { |
2967 | for (MCRegister Reg : *RC) |
2968 | if (MRI.isAllocatable(PhysReg: Reg) && !MRI.isPhysRegUsed(PhysReg: Reg)) |
2969 | return Reg; |
2970 | } |
2971 | return MCRegister(); |
2972 | } |
2973 | |
2974 | bool SIRegisterInfo::isUniformReg(const MachineRegisterInfo &MRI, |
2975 | const RegisterBankInfo &RBI, |
2976 | Register Reg) const { |
2977 | auto *RB = RBI.getRegBank(Reg, MRI, TRI: *MRI.getTargetRegisterInfo()); |
2978 | if (!RB) |
2979 | return false; |
2980 | |
2981 | return !RBI.isDivergentRegBank(RB); |
2982 | } |
2983 | |
2984 | ArrayRef<int16_t> SIRegisterInfo::getRegSplitParts(const TargetRegisterClass *RC, |
2985 | unsigned EltSize) const { |
2986 | const unsigned RegBitWidth = AMDGPU::getRegBitWidth(RC: *RC); |
2987 | assert(RegBitWidth >= 32 && RegBitWidth <= 1024); |
2988 | |
2989 | const unsigned RegDWORDs = RegBitWidth / 32; |
2990 | const unsigned EltDWORDs = EltSize / 4; |
2991 | assert(RegSplitParts.size() + 1 >= EltDWORDs); |
2992 | |
2993 | const std::vector<int16_t> &Parts = RegSplitParts[EltDWORDs - 1]; |
2994 | const unsigned NumParts = RegDWORDs / EltDWORDs; |
2995 | |
2996 | return ArrayRef(Parts.data(), NumParts); |
2997 | } |
2998 | |
2999 | const TargetRegisterClass* |
3000 | SIRegisterInfo::getRegClassForReg(const MachineRegisterInfo &MRI, |
3001 | Register Reg) const { |
3002 | return Reg.isVirtual() ? MRI.getRegClass(Reg) : getPhysRegBaseClass(Reg); |
3003 | } |
3004 | |
3005 | const TargetRegisterClass * |
3006 | SIRegisterInfo::getRegClassForOperandReg(const MachineRegisterInfo &MRI, |
3007 | const MachineOperand &MO) const { |
3008 | const TargetRegisterClass *SrcRC = getRegClassForReg(MRI, Reg: MO.getReg()); |
3009 | return getSubRegisterClass(RC: SrcRC, Idx: MO.getSubReg()); |
3010 | } |
3011 | |
3012 | bool SIRegisterInfo::isVGPR(const MachineRegisterInfo &MRI, |
3013 | Register Reg) const { |
3014 | const TargetRegisterClass *RC = getRegClassForReg(MRI, Reg); |
3015 | // Registers without classes are unaddressable, SGPR-like registers. |
3016 | return RC && isVGPRClass(RC); |
3017 | } |
3018 | |
3019 | bool SIRegisterInfo::isAGPR(const MachineRegisterInfo &MRI, |
3020 | Register Reg) const { |
3021 | const TargetRegisterClass *RC = getRegClassForReg(MRI, Reg); |
3022 | |
3023 | // Registers without classes are unaddressable, SGPR-like registers. |
3024 | return RC && isAGPRClass(RC); |
3025 | } |
3026 | |
3027 | bool SIRegisterInfo::shouldCoalesce(MachineInstr *MI, |
3028 | const TargetRegisterClass *SrcRC, |
3029 | unsigned SubReg, |
3030 | const TargetRegisterClass *DstRC, |
3031 | unsigned DstSubReg, |
3032 | const TargetRegisterClass *NewRC, |
3033 | LiveIntervals &LIS) const { |
3034 | unsigned SrcSize = getRegSizeInBits(RC: *SrcRC); |
3035 | unsigned DstSize = getRegSizeInBits(RC: *DstRC); |
3036 | unsigned NewSize = getRegSizeInBits(RC: *NewRC); |
3037 | |
3038 | // Do not increase size of registers beyond dword, we would need to allocate |
3039 | // adjacent registers and constraint regalloc more than needed. |
3040 | |
3041 | // Always allow dword coalescing. |
3042 | if (SrcSize <= 32 || DstSize <= 32) |
3043 | return true; |
3044 | |
3045 | return NewSize <= DstSize || NewSize <= SrcSize; |
3046 | } |
3047 | |
3048 | unsigned SIRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, |
3049 | MachineFunction &MF) const { |
3050 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
3051 | |
3052 | unsigned Occupancy = ST.getOccupancyWithLocalMemSize(Bytes: MFI->getLDSSize(), |
3053 | MF.getFunction()); |
3054 | switch (RC->getID()) { |
3055 | default: |
3056 | return AMDGPUGenRegisterInfo::getRegPressureLimit(RC, MF); |
3057 | case AMDGPU::VGPR_32RegClassID: |
3058 | return std::min(a: ST.getMaxNumVGPRs(WavesPerEU: Occupancy), b: ST.getMaxNumVGPRs(MF)); |
3059 | case AMDGPU::SGPR_32RegClassID: |
3060 | case AMDGPU::SGPR_LO16RegClassID: |
3061 | return std::min(a: ST.getMaxNumSGPRs(WavesPerEU: Occupancy, Addressable: true), b: ST.getMaxNumSGPRs(MF)); |
3062 | } |
3063 | } |
3064 | |
3065 | unsigned SIRegisterInfo::getRegPressureSetLimit(const MachineFunction &MF, |
3066 | unsigned Idx) const { |
3067 | if (Idx == AMDGPU::RegisterPressureSets::VGPR_32 || |
3068 | Idx == AMDGPU::RegisterPressureSets::AGPR_32) |
3069 | return getRegPressureLimit(RC: &AMDGPU::VGPR_32RegClass, |
3070 | MF&: const_cast<MachineFunction &>(MF)); |
3071 | |
3072 | if (Idx == AMDGPU::RegisterPressureSets::SReg_32) |
3073 | return getRegPressureLimit(RC: &AMDGPU::SGPR_32RegClass, |
3074 | MF&: const_cast<MachineFunction &>(MF)); |
3075 | |
3076 | llvm_unreachable("Unexpected register pressure set!" ); |
3077 | } |
3078 | |
3079 | const int *SIRegisterInfo::getRegUnitPressureSets(unsigned RegUnit) const { |
3080 | static const int Empty[] = { -1 }; |
3081 | |
3082 | if (RegPressureIgnoredUnits[RegUnit]) |
3083 | return Empty; |
3084 | |
3085 | return AMDGPUGenRegisterInfo::getRegUnitPressureSets(RegUnit); |
3086 | } |
3087 | |
3088 | MCRegister SIRegisterInfo::getReturnAddressReg(const MachineFunction &MF) const { |
3089 | // Not a callee saved register. |
3090 | return AMDGPU::SGPR30_SGPR31; |
3091 | } |
3092 | |
3093 | const TargetRegisterClass * |
3094 | SIRegisterInfo::getRegClassForSizeOnBank(unsigned Size, |
3095 | const RegisterBank &RB) const { |
3096 | switch (RB.getID()) { |
3097 | case AMDGPU::VGPRRegBankID: |
3098 | return getVGPRClassForBitWidth( |
3099 | BitWidth: std::max(a: ST.useRealTrue16Insts() ? 16u : 32u, b: Size)); |
3100 | case AMDGPU::VCCRegBankID: |
3101 | assert(Size == 1); |
3102 | return isWave32 ? &AMDGPU::SReg_32_XM0_XEXECRegClass |
3103 | : &AMDGPU::SReg_64_XEXECRegClass; |
3104 | case AMDGPU::SGPRRegBankID: |
3105 | return getSGPRClassForBitWidth(BitWidth: std::max(a: 32u, b: Size)); |
3106 | case AMDGPU::AGPRRegBankID: |
3107 | return getAGPRClassForBitWidth(BitWidth: std::max(a: 32u, b: Size)); |
3108 | default: |
3109 | llvm_unreachable("unknown register bank" ); |
3110 | } |
3111 | } |
3112 | |
3113 | const TargetRegisterClass * |
3114 | SIRegisterInfo::getConstrainedRegClassForOperand(const MachineOperand &MO, |
3115 | const MachineRegisterInfo &MRI) const { |
3116 | const RegClassOrRegBank &RCOrRB = MRI.getRegClassOrRegBank(Reg: MO.getReg()); |
3117 | if (const RegisterBank *RB = RCOrRB.dyn_cast<const RegisterBank*>()) |
3118 | return getRegClassForTypeOnBank(Ty: MRI.getType(Reg: MO.getReg()), Bank: *RB); |
3119 | |
3120 | if (const auto *RC = RCOrRB.dyn_cast<const TargetRegisterClass *>()) |
3121 | return getAllocatableClass(RC); |
3122 | |
3123 | return nullptr; |
3124 | } |
3125 | |
3126 | MCRegister SIRegisterInfo::getVCC() const { |
3127 | return isWave32 ? AMDGPU::VCC_LO : AMDGPU::VCC; |
3128 | } |
3129 | |
3130 | MCRegister SIRegisterInfo::getExec() const { |
3131 | return isWave32 ? AMDGPU::EXEC_LO : AMDGPU::EXEC; |
3132 | } |
3133 | |
3134 | const TargetRegisterClass *SIRegisterInfo::getVGPR64Class() const { |
3135 | // VGPR tuples have an alignment requirement on gfx90a variants. |
3136 | return ST.needsAlignedVGPRs() ? &AMDGPU::VReg_64_Align2RegClass |
3137 | : &AMDGPU::VReg_64RegClass; |
3138 | } |
3139 | |
3140 | const TargetRegisterClass * |
3141 | SIRegisterInfo::getRegClass(unsigned RCID) const { |
3142 | switch ((int)RCID) { |
3143 | case AMDGPU::SReg_1RegClassID: |
3144 | return getBoolRC(); |
3145 | case AMDGPU::SReg_1_XEXECRegClassID: |
3146 | return isWave32 ? &AMDGPU::SReg_32_XM0_XEXECRegClass |
3147 | : &AMDGPU::SReg_64_XEXECRegClass; |
3148 | case -1: |
3149 | return nullptr; |
3150 | default: |
3151 | return AMDGPUGenRegisterInfo::getRegClass(i: RCID); |
3152 | } |
3153 | } |
3154 | |
3155 | // Find reaching register definition |
3156 | MachineInstr *SIRegisterInfo::findReachingDef(Register Reg, unsigned SubReg, |
3157 | MachineInstr &Use, |
3158 | MachineRegisterInfo &MRI, |
3159 | LiveIntervals *LIS) const { |
3160 | auto &MDT = LIS->getDomTree(); |
3161 | SlotIndex UseIdx = LIS->getInstructionIndex(Instr: Use); |
3162 | SlotIndex DefIdx; |
3163 | |
3164 | if (Reg.isVirtual()) { |
3165 | if (!LIS->hasInterval(Reg)) |
3166 | return nullptr; |
3167 | LiveInterval &LI = LIS->getInterval(Reg); |
3168 | LaneBitmask SubLanes = SubReg ? getSubRegIndexLaneMask(SubIdx: SubReg) |
3169 | : MRI.getMaxLaneMaskForVReg(Reg); |
3170 | VNInfo *V = nullptr; |
3171 | if (LI.hasSubRanges()) { |
3172 | for (auto &S : LI.subranges()) { |
3173 | if ((S.LaneMask & SubLanes) == SubLanes) { |
3174 | V = S.getVNInfoAt(Idx: UseIdx); |
3175 | break; |
3176 | } |
3177 | } |
3178 | } else { |
3179 | V = LI.getVNInfoAt(Idx: UseIdx); |
3180 | } |
3181 | if (!V) |
3182 | return nullptr; |
3183 | DefIdx = V->def; |
3184 | } else { |
3185 | // Find last def. |
3186 | for (MCRegUnit Unit : regunits(Reg: Reg.asMCReg())) { |
3187 | LiveRange &LR = LIS->getRegUnit(Unit); |
3188 | if (VNInfo *V = LR.getVNInfoAt(Idx: UseIdx)) { |
3189 | if (!DefIdx.isValid() || |
3190 | MDT.dominates(A: LIS->getInstructionFromIndex(index: DefIdx), |
3191 | B: LIS->getInstructionFromIndex(index: V->def))) |
3192 | DefIdx = V->def; |
3193 | } else { |
3194 | return nullptr; |
3195 | } |
3196 | } |
3197 | } |
3198 | |
3199 | MachineInstr *Def = LIS->getInstructionFromIndex(index: DefIdx); |
3200 | |
3201 | if (!Def || !MDT.dominates(A: Def, B: &Use)) |
3202 | return nullptr; |
3203 | |
3204 | assert(Def->modifiesRegister(Reg, this)); |
3205 | |
3206 | return Def; |
3207 | } |
3208 | |
3209 | MCPhysReg SIRegisterInfo::get32BitRegister(MCPhysReg Reg) const { |
3210 | assert(getRegSizeInBits(*getPhysRegBaseClass(Reg)) <= 32); |
3211 | |
3212 | for (const TargetRegisterClass &RC : { AMDGPU::VGPR_32RegClass, |
3213 | AMDGPU::SReg_32RegClass, |
3214 | AMDGPU::AGPR_32RegClass } ) { |
3215 | if (MCPhysReg Super = getMatchingSuperReg(Reg, SubIdx: AMDGPU::lo16, RC: &RC)) |
3216 | return Super; |
3217 | } |
3218 | if (MCPhysReg Super = getMatchingSuperReg(Reg, SubIdx: AMDGPU::hi16, |
3219 | RC: &AMDGPU::VGPR_32RegClass)) { |
3220 | return Super; |
3221 | } |
3222 | |
3223 | return AMDGPU::NoRegister; |
3224 | } |
3225 | |
3226 | bool SIRegisterInfo::isProperlyAlignedRC(const TargetRegisterClass &RC) const { |
3227 | if (!ST.needsAlignedVGPRs()) |
3228 | return true; |
3229 | |
3230 | if (isVGPRClass(RC: &RC)) |
3231 | return RC.hasSuperClassEq(RC: getVGPRClassForBitWidth(BitWidth: getRegSizeInBits(RC))); |
3232 | if (isAGPRClass(RC: &RC)) |
3233 | return RC.hasSuperClassEq(RC: getAGPRClassForBitWidth(BitWidth: getRegSizeInBits(RC))); |
3234 | if (isVectorSuperClass(RC: &RC)) |
3235 | return RC.hasSuperClassEq( |
3236 | RC: getVectorSuperClassForBitWidth(BitWidth: getRegSizeInBits(RC))); |
3237 | |
3238 | return true; |
3239 | } |
3240 | |
3241 | const TargetRegisterClass * |
3242 | SIRegisterInfo::getProperlyAlignedRC(const TargetRegisterClass *RC) const { |
3243 | if (!RC || !ST.needsAlignedVGPRs()) |
3244 | return RC; |
3245 | |
3246 | unsigned Size = getRegSizeInBits(RC: *RC); |
3247 | if (Size <= 32) |
3248 | return RC; |
3249 | |
3250 | if (isVGPRClass(RC)) |
3251 | return getAlignedVGPRClassForBitWidth(BitWidth: Size); |
3252 | if (isAGPRClass(RC)) |
3253 | return getAlignedAGPRClassForBitWidth(BitWidth: Size); |
3254 | if (isVectorSuperClass(RC)) |
3255 | return getAlignedVectorSuperClassForBitWidth(BitWidth: Size); |
3256 | |
3257 | return RC; |
3258 | } |
3259 | |
3260 | ArrayRef<MCPhysReg> |
3261 | SIRegisterInfo::getAllSGPR128(const MachineFunction &MF) const { |
3262 | return ArrayRef(AMDGPU::SGPR_128RegClass.begin(), ST.getMaxNumSGPRs(MF) / 4); |
3263 | } |
3264 | |
3265 | ArrayRef<MCPhysReg> |
3266 | SIRegisterInfo::getAllSGPR64(const MachineFunction &MF) const { |
3267 | return ArrayRef(AMDGPU::SGPR_64RegClass.begin(), ST.getMaxNumSGPRs(MF) / 2); |
3268 | } |
3269 | |
3270 | ArrayRef<MCPhysReg> |
3271 | SIRegisterInfo::getAllSGPR32(const MachineFunction &MF) const { |
3272 | return ArrayRef(AMDGPU::SGPR_32RegClass.begin(), ST.getMaxNumSGPRs(MF)); |
3273 | } |
3274 | |
3275 | unsigned |
3276 | SIRegisterInfo::getSubRegAlignmentNumBits(const TargetRegisterClass *RC, |
3277 | unsigned SubReg) const { |
3278 | switch (RC->TSFlags & SIRCFlags::RegKindMask) { |
3279 | case SIRCFlags::HasSGPR: |
3280 | return std::min(a: 128u, b: getSubRegIdxSize(Idx: SubReg)); |
3281 | case SIRCFlags::HasAGPR: |
3282 | case SIRCFlags::HasVGPR: |
3283 | case SIRCFlags::HasVGPR | SIRCFlags::HasAGPR: |
3284 | return std::min(a: 32u, b: getSubRegIdxSize(Idx: SubReg)); |
3285 | default: |
3286 | break; |
3287 | } |
3288 | return 0; |
3289 | } |
3290 | |