File: | llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp |
Warning: | line 1085, column 5 Value stored to 'Desc' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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 "SIRegisterInfo.h" |
15 | #include "AMDGPU.h" |
16 | #include "AMDGPURegisterBankInfo.h" |
17 | #include "GCNSubtarget.h" |
18 | #include "MCTargetDesc/AMDGPUInstPrinter.h" |
19 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
20 | #include "SIMachineFunctionInfo.h" |
21 | #include "llvm/CodeGen/LiveIntervals.h" |
22 | #include "llvm/CodeGen/MachineDominators.h" |
23 | #include "llvm/CodeGen/RegisterScavenging.h" |
24 | |
25 | using namespace llvm; |
26 | |
27 | #define GET_REGINFO_TARGET_DESC |
28 | #include "AMDGPUGenRegisterInfo.inc" |
29 | |
30 | static cl::opt<bool> EnableSpillSGPRToVGPR( |
31 | "amdgpu-spill-sgpr-to-vgpr", |
32 | cl::desc("Enable spilling VGPRs to SGPRs"), |
33 | cl::ReallyHidden, |
34 | cl::init(true)); |
35 | |
36 | std::array<std::vector<int16_t>, 16> SIRegisterInfo::RegSplitParts; |
37 | std::array<std::array<uint16_t, 32>, 9> SIRegisterInfo::SubRegFromChannelTable; |
38 | |
39 | // Map numbers of DWORDs to indexes in SubRegFromChannelTable. |
40 | // Valid indexes are shifted 1, such that a 0 mapping means unsupported. |
41 | // e.g. for 8 DWORDs (256-bit), SubRegFromChannelTableWidthMap[8] = 8, |
42 | // meaning index 7 in SubRegFromChannelTable. |
43 | static const std::array<unsigned, 17> SubRegFromChannelTableWidthMap = { |
44 | 0, 1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 9}; |
45 | |
46 | namespace llvm { |
47 | |
48 | // A temporary struct to spill SGPRs. |
49 | // This is mostly to spill SGPRs to memory. Spilling SGPRs into VGPR lanes emits |
50 | // just v_writelane and v_readlane. |
51 | // |
52 | // When spilling to memory, the SGPRs are written into VGPR lanes and the VGPR |
53 | // is saved to scratch (or the other way around for loads). |
54 | // For this, a VGPR is required where the needed lanes can be clobbered. The |
55 | // RegScavenger can provide a VGPR where currently active lanes can be |
56 | // clobbered, but we still need to save inactive lanes. |
57 | // The high-level steps are: |
58 | // - Try to scavenge SGPR(s) to save exec |
59 | // - Try to scavenge VGPR |
60 | // - Save needed, all or inactive lanes of a TmpVGPR |
61 | // - Spill/Restore SGPRs using TmpVGPR |
62 | // - Restore TmpVGPR |
63 | // |
64 | // To save all lanes of TmpVGPR, exec needs to be saved and modified. If we |
65 | // cannot scavenge temporary SGPRs to save exec, we use the following code: |
66 | // buffer_store_dword TmpVGPR ; only if active lanes need to be saved |
67 | // s_not exec, exec |
68 | // buffer_store_dword TmpVGPR ; save inactive lanes |
69 | // s_not exec, exec |
70 | struct SGPRSpillBuilder { |
71 | struct PerVGPRData { |
72 | unsigned PerVGPR; |
73 | unsigned NumVGPRs; |
74 | int64_t VGPRLanes; |
75 | }; |
76 | |
77 | // The SGPR to save |
78 | Register SuperReg; |
79 | MachineBasicBlock::iterator MI; |
80 | ArrayRef<int16_t> SplitParts; |
81 | unsigned NumSubRegs; |
82 | bool IsKill; |
83 | const DebugLoc &DL; |
84 | |
85 | /* When spilling to stack */ |
86 | // The SGPRs are written into this VGPR, which is then written to scratch |
87 | // (or vice versa for loads). |
88 | Register TmpVGPR = AMDGPU::NoRegister; |
89 | // Temporary spill slot to save TmpVGPR to. |
90 | int TmpVGPRIndex = 0; |
91 | // If TmpVGPR is live before the spill or if it is scavenged. |
92 | bool TmpVGPRLive = false; |
93 | // Scavenged SGPR to save EXEC. |
94 | Register SavedExecReg = AMDGPU::NoRegister; |
95 | // Stack index to write the SGPRs to. |
96 | int Index; |
97 | unsigned EltSize = 4; |
98 | |
99 | RegScavenger *RS; |
100 | MachineBasicBlock &MBB; |
101 | MachineFunction &MF; |
102 | SIMachineFunctionInfo &MFI; |
103 | const SIInstrInfo &TII; |
104 | const SIRegisterInfo &TRI; |
105 | bool IsWave32; |
106 | Register ExecReg; |
107 | unsigned MovOpc; |
108 | unsigned NotOpc; |
109 | |
110 | SGPRSpillBuilder(const SIRegisterInfo &TRI, const SIInstrInfo &TII, |
111 | bool IsWave32, MachineBasicBlock::iterator MI, int Index, |
112 | RegScavenger *RS) |
113 | : SuperReg(MI->getOperand(0).getReg()), MI(MI), |
114 | IsKill(MI->getOperand(0).isKill()), DL(MI->getDebugLoc()), Index(Index), |
115 | RS(RS), MBB(*MI->getParent()), MF(*MBB.getParent()), |
116 | MFI(*MF.getInfo<SIMachineFunctionInfo>()), TII(TII), TRI(TRI), |
117 | IsWave32(IsWave32) { |
118 | const TargetRegisterClass *RC = TRI.getPhysRegClass(SuperReg); |
119 | SplitParts = TRI.getRegSplitParts(RC, EltSize); |
120 | NumSubRegs = SplitParts.empty() ? 1 : SplitParts.size(); |
121 | |
122 | if (IsWave32) { |
123 | ExecReg = AMDGPU::EXEC_LO; |
124 | MovOpc = AMDGPU::S_MOV_B32; |
125 | NotOpc = AMDGPU::S_NOT_B32; |
126 | } else { |
127 | ExecReg = AMDGPU::EXEC; |
128 | MovOpc = AMDGPU::S_MOV_B64; |
129 | NotOpc = AMDGPU::S_NOT_B64; |
130 | } |
131 | |
132 | assert(SuperReg != AMDGPU::M0 && "m0 should never spill")(static_cast <bool> (SuperReg != AMDGPU::M0 && "m0 should never spill" ) ? void (0) : __assert_fail ("SuperReg != AMDGPU::M0 && \"m0 should never spill\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 132, __extension__ __PRETTY_FUNCTION__)); |
133 | assert(SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI &&(static_cast <bool> (SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI && SuperReg != AMDGPU::EXEC && "exec should never spill") ? void (0) : __assert_fail ("SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI && SuperReg != AMDGPU::EXEC && \"exec should never spill\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 134, __extension__ __PRETTY_FUNCTION__)) |
134 | SuperReg != AMDGPU::EXEC && "exec should never spill")(static_cast <bool> (SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI && SuperReg != AMDGPU::EXEC && "exec should never spill") ? void (0) : __assert_fail ("SuperReg != AMDGPU::EXEC_LO && SuperReg != AMDGPU::EXEC_HI && SuperReg != AMDGPU::EXEC && \"exec should never spill\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 134, __extension__ __PRETTY_FUNCTION__)); |
135 | } |
136 | |
137 | PerVGPRData getPerVGPRData() { |
138 | PerVGPRData Data; |
139 | Data.PerVGPR = IsWave32 ? 32 : 64; |
140 | Data.NumVGPRs = (NumSubRegs + (Data.PerVGPR - 1)) / Data.PerVGPR; |
141 | Data.VGPRLanes = (1LL << std::min(Data.PerVGPR, NumSubRegs)) - 1LL; |
142 | return Data; |
143 | } |
144 | |
145 | // Tries to scavenge SGPRs to save EXEC and a VGPR. Uses v0 if no VGPR is |
146 | // free. |
147 | // Writes these instructions if an SGPR can be scavenged: |
148 | // s_mov_b64 s[6:7], exec ; Save exec |
149 | // s_mov_b64 exec, 3 ; Wanted lanemask |
150 | // buffer_store_dword v1 ; Write scavenged VGPR to emergency slot |
151 | // |
152 | // Writes these instructions if no SGPR can be scavenged: |
153 | // buffer_store_dword v0 ; Only if no free VGPR was found |
154 | // s_not_b64 exec, exec |
155 | // buffer_store_dword v0 ; Save inactive lanes |
156 | // ; exec stays inverted, it is flipped back in |
157 | // ; restore. |
158 | void prepare() { |
159 | // Scavenged temporary VGPR to use. It must be scavenged once for any number |
160 | // of spilled subregs. |
161 | // FIXME: The liveness analysis is limited and does not tell if a register |
162 | // is in use in lanes that are currently inactive. We can never be sure if |
163 | // a register as actually in use in another lane, so we need to save all |
164 | // used lanes of the chosen VGPR. |
165 | assert(RS && "Cannot spill SGPR to memory without RegScavenger")(static_cast <bool> (RS && "Cannot spill SGPR to memory without RegScavenger" ) ? void (0) : __assert_fail ("RS && \"Cannot spill SGPR to memory without RegScavenger\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 165, __extension__ __PRETTY_FUNCTION__)); |
166 | TmpVGPR = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0, false); |
167 | |
168 | // Reserve temporary stack slot |
169 | TmpVGPRIndex = MFI.getScavengeFI(MF.getFrameInfo(), TRI); |
170 | if (TmpVGPR) { |
171 | // Found a register that is dead in the currently active lanes, we only |
172 | // need to spill inactive lanes. |
173 | TmpVGPRLive = false; |
174 | } else { |
175 | // Pick v0 because it doesn't make a difference. |
176 | TmpVGPR = AMDGPU::VGPR0; |
177 | TmpVGPRLive = true; |
178 | } |
179 | |
180 | // Try to scavenge SGPRs to save exec |
181 | assert(!SavedExecReg && "Exec is already saved, refuse to save again")(static_cast <bool> (!SavedExecReg && "Exec is already saved, refuse to save again" ) ? void (0) : __assert_fail ("!SavedExecReg && \"Exec is already saved, refuse to save again\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 181, __extension__ __PRETTY_FUNCTION__)); |
182 | const TargetRegisterClass &RC = |
183 | IsWave32 ? AMDGPU::SGPR_32RegClass : AMDGPU::SGPR_64RegClass; |
184 | RS->setRegUsed(SuperReg); |
185 | SavedExecReg = RS->scavengeRegister(&RC, MI, 0, false); |
186 | |
187 | int64_t VGPRLanes = getPerVGPRData().VGPRLanes; |
188 | |
189 | if (SavedExecReg) { |
190 | RS->setRegUsed(SavedExecReg); |
191 | // Set exec to needed lanes |
192 | BuildMI(MBB, MI, DL, TII.get(MovOpc), SavedExecReg).addReg(ExecReg); |
193 | auto I = BuildMI(MBB, MI, DL, TII.get(MovOpc), ExecReg).addImm(VGPRLanes); |
194 | if (!TmpVGPRLive) |
195 | I.addReg(TmpVGPR, RegState::ImplicitDefine); |
196 | // Spill needed lanes |
197 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ false); |
198 | } else { |
199 | // Spill active lanes |
200 | if (TmpVGPRLive) |
201 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ false, |
202 | /*IsKill*/ false); |
203 | // Spill inactive lanes |
204 | auto I = BuildMI(MBB, MI, DL, TII.get(NotOpc), ExecReg).addReg(ExecReg); |
205 | if (!TmpVGPRLive) |
206 | I.addReg(TmpVGPR, RegState::ImplicitDefine); |
207 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ false); |
208 | } |
209 | } |
210 | |
211 | // Writes these instructions if an SGPR can be scavenged: |
212 | // buffer_load_dword v1 ; Write scavenged VGPR to emergency slot |
213 | // s_waitcnt vmcnt(0) ; If a free VGPR was found |
214 | // s_mov_b64 exec, s[6:7] ; Save exec |
215 | // |
216 | // Writes these instructions if no SGPR can be scavenged: |
217 | // buffer_load_dword v0 ; Restore inactive lanes |
218 | // s_waitcnt vmcnt(0) ; If a free VGPR was found |
219 | // s_not_b64 exec, exec |
220 | // buffer_load_dword v0 ; Only if no free VGPR was found |
221 | void restore() { |
222 | if (SavedExecReg) { |
223 | // Restore used lanes |
224 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ true, |
225 | /*IsKill*/ false); |
226 | // Restore exec |
227 | auto I = BuildMI(MBB, MI, DL, TII.get(MovOpc), ExecReg) |
228 | .addReg(SavedExecReg, RegState::Kill); |
229 | // Add an implicit use of the load so it is not dead. |
230 | // FIXME This inserts an unnecessary waitcnt |
231 | if (!TmpVGPRLive) { |
232 | I.addReg(TmpVGPR, RegState::ImplicitKill); |
233 | } |
234 | } else { |
235 | // Restore inactive lanes |
236 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ true, |
237 | /*IsKill*/ false); |
238 | auto I = BuildMI(MBB, MI, DL, TII.get(NotOpc), ExecReg).addReg(ExecReg); |
239 | if (!TmpVGPRLive) { |
240 | I.addReg(TmpVGPR, RegState::ImplicitKill); |
241 | } |
242 | // Restore active lanes |
243 | if (TmpVGPRLive) |
244 | TRI.buildVGPRSpillLoadStore(*this, TmpVGPRIndex, 0, /*IsLoad*/ true); |
245 | } |
246 | } |
247 | |
248 | // Write TmpVGPR to memory or read TmpVGPR from memory. |
249 | // Either using a single buffer_load/store if exec is set to the needed mask |
250 | // or using |
251 | // buffer_load |
252 | // s_not exec, exec |
253 | // buffer_load |
254 | // s_not exec, exec |
255 | void readWriteTmpVGPR(unsigned Offset, bool IsLoad) { |
256 | if (SavedExecReg) { |
257 | // Spill needed lanes |
258 | TRI.buildVGPRSpillLoadStore(*this, Index, Offset, IsLoad); |
259 | } else { |
260 | // Spill active lanes |
261 | TRI.buildVGPRSpillLoadStore(*this, Index, Offset, IsLoad, |
262 | /*IsKill*/ false); |
263 | // Spill inactive lanes |
264 | BuildMI(MBB, MI, DL, TII.get(NotOpc), ExecReg).addReg(ExecReg); |
265 | TRI.buildVGPRSpillLoadStore(*this, Index, Offset, IsLoad); |
266 | BuildMI(MBB, MI, DL, TII.get(NotOpc), ExecReg).addReg(ExecReg); |
267 | } |
268 | } |
269 | }; |
270 | |
271 | } // namespace llvm |
272 | |
273 | SIRegisterInfo::SIRegisterInfo(const GCNSubtarget &ST) |
274 | : AMDGPUGenRegisterInfo(AMDGPU::PC_REG, ST.getAMDGPUDwarfFlavour()), ST(ST), |
275 | SpillSGPRToVGPR(EnableSpillSGPRToVGPR), isWave32(ST.isWave32()) { |
276 | |
277 | assert(getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 &&(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)) |
278 | getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) &&(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)) |
279 | (getSubRegIndexLaneMask(AMDGPU::lo16) |(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)) |
280 | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() ==(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)) |
281 | getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() &&(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)) |
282 | "getNumCoveredRegs() will not work with generated subreg masks!")(static_cast <bool> (getSubRegIndexLaneMask(AMDGPU::sub0 ).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU ::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask (AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger () == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && "getNumCoveredRegs() will not work with generated subreg masks!" ) ? void (0) : __assert_fail ("getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() == 3 && getSubRegIndexLaneMask(AMDGPU::sub31).getAsInteger() == (3ULL << 62) && (getSubRegIndexLaneMask(AMDGPU::lo16) | getSubRegIndexLaneMask(AMDGPU::hi16)).getAsInteger() == getSubRegIndexLaneMask(AMDGPU::sub0).getAsInteger() && \"getNumCoveredRegs() will not work with generated subreg masks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 282, __extension__ __PRETTY_FUNCTION__)); |
283 | |
284 | RegPressureIgnoredUnits.resize(getNumRegUnits()); |
285 | RegPressureIgnoredUnits.set( |
286 | *MCRegUnitIterator(MCRegister::from(AMDGPU::M0), this)); |
287 | for (auto Reg : AMDGPU::VGPR_HI16RegClass) |
288 | RegPressureIgnoredUnits.set(*MCRegUnitIterator(Reg, this)); |
289 | |
290 | // HACK: Until this is fully tablegen'd. |
291 | static llvm::once_flag InitializeRegSplitPartsFlag; |
292 | |
293 | static auto InitializeRegSplitPartsOnce = [this]() { |
294 | for (unsigned Idx = 1, E = getNumSubRegIndices() - 1; Idx < E; ++Idx) { |
295 | unsigned Size = getSubRegIdxSize(Idx); |
296 | if (Size & 31) |
297 | continue; |
298 | std::vector<int16_t> &Vec = RegSplitParts[Size / 32 - 1]; |
299 | unsigned Pos = getSubRegIdxOffset(Idx); |
300 | if (Pos % Size) |
301 | continue; |
302 | Pos /= Size; |
303 | if (Vec.empty()) { |
304 | unsigned MaxNumParts = 1024 / Size; // Maximum register is 1024 bits. |
305 | Vec.resize(MaxNumParts); |
306 | } |
307 | Vec[Pos] = Idx; |
308 | } |
309 | }; |
310 | |
311 | static llvm::once_flag InitializeSubRegFromChannelTableFlag; |
312 | |
313 | static auto InitializeSubRegFromChannelTableOnce = [this]() { |
314 | for (auto &Row : SubRegFromChannelTable) |
315 | Row.fill(AMDGPU::NoSubRegister); |
316 | for (uint16_t Idx = 1; Idx < getNumSubRegIndices(); ++Idx) { |
317 | unsigned Width = AMDGPUSubRegIdxRanges[Idx].Size / 32; |
318 | unsigned Offset = AMDGPUSubRegIdxRanges[Idx].Offset / 32; |
319 | assert(Width < SubRegFromChannelTableWidthMap.size())(static_cast <bool> (Width < SubRegFromChannelTableWidthMap .size()) ? void (0) : __assert_fail ("Width < SubRegFromChannelTableWidthMap.size()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 319, __extension__ __PRETTY_FUNCTION__)); |
320 | Width = SubRegFromChannelTableWidthMap[Width]; |
321 | if (Width == 0) |
322 | continue; |
323 | unsigned TableIdx = Width - 1; |
324 | assert(TableIdx < SubRegFromChannelTable.size())(static_cast <bool> (TableIdx < SubRegFromChannelTable .size()) ? void (0) : __assert_fail ("TableIdx < SubRegFromChannelTable.size()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 324, __extension__ __PRETTY_FUNCTION__)); |
325 | assert(Offset < SubRegFromChannelTable[TableIdx].size())(static_cast <bool> (Offset < SubRegFromChannelTable [TableIdx].size()) ? void (0) : __assert_fail ("Offset < SubRegFromChannelTable[TableIdx].size()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 325, __extension__ __PRETTY_FUNCTION__)); |
326 | SubRegFromChannelTable[TableIdx][Offset] = Idx; |
327 | } |
328 | }; |
329 | |
330 | llvm::call_once(InitializeRegSplitPartsFlag, InitializeRegSplitPartsOnce); |
331 | llvm::call_once(InitializeSubRegFromChannelTableFlag, |
332 | InitializeSubRegFromChannelTableOnce); |
333 | } |
334 | |
335 | void SIRegisterInfo::reserveRegisterTuples(BitVector &Reserved, |
336 | MCRegister Reg) const { |
337 | MCRegAliasIterator R(Reg, this, true); |
338 | |
339 | for (; R.isValid(); ++R) |
340 | Reserved.set(*R); |
341 | } |
342 | |
343 | // Forced to be here by one .inc |
344 | const MCPhysReg *SIRegisterInfo::getCalleeSavedRegs( |
345 | const MachineFunction *MF) const { |
346 | CallingConv::ID CC = MF->getFunction().getCallingConv(); |
347 | switch (CC) { |
348 | case CallingConv::C: |
349 | case CallingConv::Fast: |
350 | case CallingConv::Cold: |
351 | case CallingConv::AMDGPU_Gfx: |
352 | return MF->getSubtarget<GCNSubtarget>().hasGFX90AInsts() |
353 | ? CSR_AMDGPU_HighRegs_With_AGPRs_SaveList |
354 | : CSR_AMDGPU_HighRegs_SaveList; |
355 | default: { |
356 | // Dummy to not crash RegisterClassInfo. |
357 | static const MCPhysReg NoCalleeSavedReg = AMDGPU::NoRegister; |
358 | return &NoCalleeSavedReg; |
359 | } |
360 | } |
361 | } |
362 | |
363 | const MCPhysReg * |
364 | SIRegisterInfo::getCalleeSavedRegsViaCopy(const MachineFunction *MF) const { |
365 | return nullptr; |
366 | } |
367 | |
368 | const uint32_t *SIRegisterInfo::getCallPreservedMask(const MachineFunction &MF, |
369 | CallingConv::ID CC) const { |
370 | switch (CC) { |
371 | case CallingConv::C: |
372 | case CallingConv::Fast: |
373 | case CallingConv::Cold: |
374 | case CallingConv::AMDGPU_Gfx: |
375 | return MF.getSubtarget<GCNSubtarget>().hasGFX90AInsts() |
376 | ? CSR_AMDGPU_HighRegs_With_AGPRs_RegMask |
377 | : CSR_AMDGPU_HighRegs_RegMask; |
378 | default: |
379 | return nullptr; |
380 | } |
381 | } |
382 | |
383 | const uint32_t *SIRegisterInfo::getNoPreservedMask() const { |
384 | return CSR_AMDGPU_NoRegs_RegMask; |
385 | } |
386 | |
387 | Register SIRegisterInfo::getFrameRegister(const MachineFunction &MF) const { |
388 | const SIFrameLowering *TFI = |
389 | MF.getSubtarget<GCNSubtarget>().getFrameLowering(); |
390 | const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>(); |
391 | // During ISel lowering we always reserve the stack pointer in entry |
392 | // functions, but never actually want to reference it when accessing our own |
393 | // frame. If we need a frame pointer we use it, but otherwise we can just use |
394 | // an immediate "0" which we represent by returning NoRegister. |
395 | if (FuncInfo->isEntryFunction()) { |
396 | return TFI->hasFP(MF) ? FuncInfo->getFrameOffsetReg() : Register(); |
397 | } |
398 | return TFI->hasFP(MF) ? FuncInfo->getFrameOffsetReg() |
399 | : FuncInfo->getStackPtrOffsetReg(); |
400 | } |
401 | |
402 | bool SIRegisterInfo::hasBasePointer(const MachineFunction &MF) const { |
403 | // When we need stack realignment, we can't reference off of the |
404 | // stack pointer, so we reserve a base pointer. |
405 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
406 | return MFI.getNumFixedObjects() && shouldRealignStack(MF); |
407 | } |
408 | |
409 | Register SIRegisterInfo::getBaseRegister() const { return AMDGPU::SGPR34; } |
410 | |
411 | const uint32_t *SIRegisterInfo::getAllVGPRRegMask() const { |
412 | return CSR_AMDGPU_AllVGPRs_RegMask; |
413 | } |
414 | |
415 | const uint32_t *SIRegisterInfo::getAllAGPRRegMask() const { |
416 | return CSR_AMDGPU_AllAGPRs_RegMask; |
417 | } |
418 | |
419 | const uint32_t *SIRegisterInfo::getAllVectorRegMask() const { |
420 | return CSR_AMDGPU_AllVectorRegs_RegMask; |
421 | } |
422 | |
423 | const uint32_t *SIRegisterInfo::getAllAllocatableSRegMask() const { |
424 | return CSR_AMDGPU_AllAllocatableSRegs_RegMask; |
425 | } |
426 | |
427 | unsigned SIRegisterInfo::getSubRegFromChannel(unsigned Channel, |
428 | unsigned NumRegs) { |
429 | assert(NumRegs < SubRegFromChannelTableWidthMap.size())(static_cast <bool> (NumRegs < SubRegFromChannelTableWidthMap .size()) ? void (0) : __assert_fail ("NumRegs < SubRegFromChannelTableWidthMap.size()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 429, __extension__ __PRETTY_FUNCTION__)); |
430 | unsigned NumRegIndex = SubRegFromChannelTableWidthMap[NumRegs]; |
431 | assert(NumRegIndex && "Not implemented")(static_cast <bool> (NumRegIndex && "Not implemented" ) ? void (0) : __assert_fail ("NumRegIndex && \"Not implemented\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 431, __extension__ __PRETTY_FUNCTION__)); |
432 | assert(Channel < SubRegFromChannelTable[NumRegIndex - 1].size())(static_cast <bool> (Channel < SubRegFromChannelTable [NumRegIndex - 1].size()) ? void (0) : __assert_fail ("Channel < SubRegFromChannelTable[NumRegIndex - 1].size()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 432, __extension__ __PRETTY_FUNCTION__)); |
433 | return SubRegFromChannelTable[NumRegIndex - 1][Channel]; |
434 | } |
435 | |
436 | MCRegister SIRegisterInfo::reservedPrivateSegmentBufferReg( |
437 | const MachineFunction &MF) const { |
438 | unsigned BaseIdx = alignDown(ST.getMaxNumSGPRs(MF), 4) - 4; |
439 | MCRegister BaseReg(AMDGPU::SGPR_32RegClass.getRegister(BaseIdx)); |
440 | return getMatchingSuperReg(BaseReg, AMDGPU::sub0, &AMDGPU::SGPR_128RegClass); |
441 | } |
442 | |
443 | BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const { |
444 | BitVector Reserved(getNumRegs()); |
445 | Reserved.set(AMDGPU::MODE); |
446 | |
447 | // EXEC_LO and EXEC_HI could be allocated and used as regular register, but |
448 | // this seems likely to result in bugs, so I'm marking them as reserved. |
449 | reserveRegisterTuples(Reserved, AMDGPU::EXEC); |
450 | reserveRegisterTuples(Reserved, AMDGPU::FLAT_SCR); |
451 | |
452 | // M0 has to be reserved so that llvm accepts it as a live-in into a block. |
453 | reserveRegisterTuples(Reserved, AMDGPU::M0); |
454 | |
455 | // Reserve src_vccz, src_execz, src_scc. |
456 | reserveRegisterTuples(Reserved, AMDGPU::SRC_VCCZ); |
457 | reserveRegisterTuples(Reserved, AMDGPU::SRC_EXECZ); |
458 | reserveRegisterTuples(Reserved, AMDGPU::SRC_SCC); |
459 | |
460 | // Reserve the memory aperture registers. |
461 | reserveRegisterTuples(Reserved, AMDGPU::SRC_SHARED_BASE); |
462 | reserveRegisterTuples(Reserved, AMDGPU::SRC_SHARED_LIMIT); |
463 | reserveRegisterTuples(Reserved, AMDGPU::SRC_PRIVATE_BASE); |
464 | reserveRegisterTuples(Reserved, AMDGPU::SRC_PRIVATE_LIMIT); |
465 | |
466 | // Reserve src_pops_exiting_wave_id - support is not implemented in Codegen. |
467 | reserveRegisterTuples(Reserved, AMDGPU::SRC_POPS_EXITING_WAVE_ID); |
468 | |
469 | // Reserve xnack_mask registers - support is not implemented in Codegen. |
470 | reserveRegisterTuples(Reserved, AMDGPU::XNACK_MASK); |
471 | |
472 | // Reserve lds_direct register - support is not implemented in Codegen. |
473 | reserveRegisterTuples(Reserved, AMDGPU::LDS_DIRECT); |
474 | |
475 | // Reserve Trap Handler registers - support is not implemented in Codegen. |
476 | reserveRegisterTuples(Reserved, AMDGPU::TBA); |
477 | reserveRegisterTuples(Reserved, AMDGPU::TMA); |
478 | reserveRegisterTuples(Reserved, AMDGPU::TTMP0_TTMP1); |
479 | reserveRegisterTuples(Reserved, AMDGPU::TTMP2_TTMP3); |
480 | reserveRegisterTuples(Reserved, AMDGPU::TTMP4_TTMP5); |
481 | reserveRegisterTuples(Reserved, AMDGPU::TTMP6_TTMP7); |
482 | reserveRegisterTuples(Reserved, AMDGPU::TTMP8_TTMP9); |
483 | reserveRegisterTuples(Reserved, AMDGPU::TTMP10_TTMP11); |
484 | reserveRegisterTuples(Reserved, AMDGPU::TTMP12_TTMP13); |
485 | reserveRegisterTuples(Reserved, AMDGPU::TTMP14_TTMP15); |
486 | |
487 | // Reserve null register - it shall never be allocated |
488 | reserveRegisterTuples(Reserved, AMDGPU::SGPR_NULL); |
489 | |
490 | // Disallow vcc_hi allocation in wave32. It may be allocated but most likely |
491 | // will result in bugs. |
492 | if (isWave32) { |
493 | Reserved.set(AMDGPU::VCC); |
494 | Reserved.set(AMDGPU::VCC_HI); |
495 | } |
496 | |
497 | unsigned MaxNumSGPRs = ST.getMaxNumSGPRs(MF); |
498 | unsigned TotalNumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs(); |
499 | for (unsigned i = MaxNumSGPRs; i < TotalNumSGPRs; ++i) { |
500 | unsigned Reg = AMDGPU::SGPR_32RegClass.getRegister(i); |
501 | reserveRegisterTuples(Reserved, Reg); |
502 | } |
503 | |
504 | unsigned MaxNumVGPRs = ST.getMaxNumVGPRs(MF); |
505 | // TODO: In an entry function without calls and AGPRs used it is possible |
506 | // to use the whole register budget for VGPRs. Even more it shall |
507 | // be possible to estimate maximum AGPR/VGPR pressure and split |
508 | // register file accordingly. |
509 | if (ST.hasGFX90AInsts()) |
510 | MaxNumVGPRs /= 2; |
511 | unsigned TotalNumVGPRs = AMDGPU::VGPR_32RegClass.getNumRegs(); |
512 | for (unsigned i = MaxNumVGPRs; i < TotalNumVGPRs; ++i) { |
513 | unsigned Reg = AMDGPU::VGPR_32RegClass.getRegister(i); |
514 | reserveRegisterTuples(Reserved, Reg); |
515 | Reg = AMDGPU::AGPR_32RegClass.getRegister(i); |
516 | reserveRegisterTuples(Reserved, Reg); |
517 | } |
518 | |
519 | for (auto Reg : AMDGPU::SReg_32RegClass) { |
520 | Reserved.set(getSubReg(Reg, AMDGPU::hi16)); |
521 | Register Low = getSubReg(Reg, AMDGPU::lo16); |
522 | // This is to prevent BB vcc liveness errors. |
523 | if (!AMDGPU::SGPR_LO16RegClass.contains(Low)) |
524 | Reserved.set(Low); |
525 | } |
526 | |
527 | for (auto Reg : AMDGPU::AGPR_32RegClass) { |
528 | Reserved.set(getSubReg(Reg, AMDGPU::hi16)); |
529 | } |
530 | |
531 | // Reserve all the rest AGPRs if there are no instructions to use it. |
532 | if (!ST.hasMAIInsts()) { |
533 | for (unsigned i = 0; i < MaxNumVGPRs; ++i) { |
534 | unsigned Reg = AMDGPU::AGPR_32RegClass.getRegister(i); |
535 | reserveRegisterTuples(Reserved, Reg); |
536 | } |
537 | } |
538 | |
539 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
540 | |
541 | Register ScratchRSrcReg = MFI->getScratchRSrcReg(); |
542 | if (ScratchRSrcReg != AMDGPU::NoRegister) { |
543 | // Reserve 4 SGPRs for the scratch buffer resource descriptor in case we need |
544 | // to spill. |
545 | // TODO: May need to reserve a VGPR if doing LDS spilling. |
546 | reserveRegisterTuples(Reserved, ScratchRSrcReg); |
547 | } |
548 | |
549 | // We have to assume the SP is needed in case there are calls in the function, |
550 | // which is detected after the function is lowered. If we aren't really going |
551 | // to need SP, don't bother reserving it. |
552 | MCRegister StackPtrReg = MFI->getStackPtrOffsetReg(); |
553 | |
554 | if (StackPtrReg) { |
555 | reserveRegisterTuples(Reserved, StackPtrReg); |
556 | assert(!isSubRegister(ScratchRSrcReg, StackPtrReg))(static_cast <bool> (!isSubRegister(ScratchRSrcReg, StackPtrReg )) ? void (0) : __assert_fail ("!isSubRegister(ScratchRSrcReg, StackPtrReg)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 556, __extension__ __PRETTY_FUNCTION__)); |
557 | } |
558 | |
559 | MCRegister FrameReg = MFI->getFrameOffsetReg(); |
560 | if (FrameReg) { |
561 | reserveRegisterTuples(Reserved, FrameReg); |
562 | assert(!isSubRegister(ScratchRSrcReg, FrameReg))(static_cast <bool> (!isSubRegister(ScratchRSrcReg, FrameReg )) ? void (0) : __assert_fail ("!isSubRegister(ScratchRSrcReg, FrameReg)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 562, __extension__ __PRETTY_FUNCTION__)); |
563 | } |
564 | |
565 | if (hasBasePointer(MF)) { |
566 | MCRegister BasePtrReg = getBaseRegister(); |
567 | reserveRegisterTuples(Reserved, BasePtrReg); |
568 | assert(!isSubRegister(ScratchRSrcReg, BasePtrReg))(static_cast <bool> (!isSubRegister(ScratchRSrcReg, BasePtrReg )) ? void (0) : __assert_fail ("!isSubRegister(ScratchRSrcReg, BasePtrReg)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 568, __extension__ __PRETTY_FUNCTION__)); |
569 | } |
570 | |
571 | for (auto Reg : MFI->WWMReservedRegs) { |
572 | reserveRegisterTuples(Reserved, Reg.first); |
573 | } |
574 | |
575 | // Reserve VGPRs used for SGPR spilling. |
576 | // Note we treat freezeReservedRegs unusually because we run register |
577 | // allocation in two phases. It's OK to re-freeze with new registers for the |
578 | // second run. |
579 | #if 0 |
580 | for (auto &SpilledFI : MFI->sgpr_spill_vgprs()) { |
581 | for (auto &SpilledVGPR : SpilledFI.second) |
582 | reserveRegisterTuples(Reserved, SpilledVGPR.VGPR); |
583 | } |
584 | #endif |
585 | |
586 | // FIXME: Stop using reserved registers for this. |
587 | for (MCPhysReg Reg : MFI->getAGPRSpillVGPRs()) |
588 | reserveRegisterTuples(Reserved, Reg); |
589 | |
590 | for (MCPhysReg Reg : MFI->getVGPRSpillAGPRs()) |
591 | reserveRegisterTuples(Reserved, Reg); |
592 | |
593 | for (auto SSpill : MFI->getSGPRSpillVGPRs()) |
594 | reserveRegisterTuples(Reserved, SSpill.VGPR); |
595 | |
596 | return Reserved; |
597 | } |
598 | |
599 | bool SIRegisterInfo::shouldRealignStack(const MachineFunction &MF) const { |
600 | const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>(); |
601 | // On entry, the base address is 0, so it can't possibly need any more |
602 | // alignment. |
603 | |
604 | // FIXME: Should be able to specify the entry frame alignment per calling |
605 | // convention instead. |
606 | if (Info->isEntryFunction()) |
607 | return false; |
608 | |
609 | return TargetRegisterInfo::shouldRealignStack(MF); |
610 | } |
611 | |
612 | bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const { |
613 | const SIMachineFunctionInfo *Info = Fn.getInfo<SIMachineFunctionInfo>(); |
614 | if (Info->isEntryFunction()) { |
615 | const MachineFrameInfo &MFI = Fn.getFrameInfo(); |
616 | return MFI.hasStackObjects() || MFI.hasCalls(); |
617 | } |
618 | |
619 | // May need scavenger for dealing with callee saved registers. |
620 | return true; |
621 | } |
622 | |
623 | bool SIRegisterInfo::requiresFrameIndexScavenging( |
624 | const MachineFunction &MF) const { |
625 | // Do not use frame virtual registers. They used to be used for SGPRs, but |
626 | // once we reach PrologEpilogInserter, we can no longer spill SGPRs. If the |
627 | // scavenger fails, we can increment/decrement the necessary SGPRs to avoid a |
628 | // spill. |
629 | return false; |
630 | } |
631 | |
632 | bool SIRegisterInfo::requiresFrameIndexReplacementScavenging( |
633 | const MachineFunction &MF) const { |
634 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
635 | return MFI.hasStackObjects(); |
636 | } |
637 | |
638 | bool SIRegisterInfo::requiresVirtualBaseRegisters( |
639 | const MachineFunction &) const { |
640 | // There are no special dedicated stack or frame pointers. |
641 | return true; |
642 | } |
643 | |
644 | int64_t SIRegisterInfo::getScratchInstrOffset(const MachineInstr *MI) const { |
645 | assert(SIInstrInfo::isMUBUF(*MI) || SIInstrInfo::isFLATScratch(*MI))(static_cast <bool> (SIInstrInfo::isMUBUF(*MI) || SIInstrInfo ::isFLATScratch(*MI)) ? void (0) : __assert_fail ("SIInstrInfo::isMUBUF(*MI) || SIInstrInfo::isFLATScratch(*MI)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 645, __extension__ __PRETTY_FUNCTION__)); |
646 | |
647 | int OffIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(), |
648 | AMDGPU::OpName::offset); |
649 | return MI->getOperand(OffIdx).getImm(); |
650 | } |
651 | |
652 | int64_t SIRegisterInfo::getFrameIndexInstrOffset(const MachineInstr *MI, |
653 | int Idx) const { |
654 | if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isFLATScratch(*MI)) |
655 | return 0; |
656 | |
657 | assert((Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(),(static_cast <bool> ((Idx == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU ::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr ))) && "Should never see frame index on non-address operand" ) ? void (0) : __assert_fail ("(Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr))) && \"Should never see frame index on non-address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 661, __extension__ __PRETTY_FUNCTION__)) |
658 | AMDGPU::OpName::vaddr) ||(static_cast <bool> ((Idx == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU ::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr ))) && "Should never see frame index on non-address operand" ) ? void (0) : __assert_fail ("(Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr))) && \"Should never see frame index on non-address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 661, __extension__ __PRETTY_FUNCTION__)) |
659 | (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(),(static_cast <bool> ((Idx == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU ::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr ))) && "Should never see frame index on non-address operand" ) ? void (0) : __assert_fail ("(Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr))) && \"Should never see frame index on non-address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 661, __extension__ __PRETTY_FUNCTION__)) |
660 | AMDGPU::OpName::saddr))) &&(static_cast <bool> ((Idx == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU ::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr ))) && "Should never see frame index on non-address operand" ) ? void (0) : __assert_fail ("(Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr))) && \"Should never see frame index on non-address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 661, __extension__ __PRETTY_FUNCTION__)) |
661 | "Should never see frame index on non-address operand")(static_cast <bool> ((Idx == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU ::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr ))) && "Should never see frame index on non-address operand" ) ? void (0) : __assert_fail ("(Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr) || (Idx == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr))) && \"Should never see frame index on non-address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 661, __extension__ __PRETTY_FUNCTION__)); |
662 | |
663 | return getScratchInstrOffset(MI); |
664 | } |
665 | |
666 | bool SIRegisterInfo::needsFrameBaseReg(MachineInstr *MI, int64_t Offset) const { |
667 | if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isFLATScratch(*MI)) |
668 | return false; |
669 | |
670 | int64_t FullOffset = Offset + getScratchInstrOffset(MI); |
671 | |
672 | if (SIInstrInfo::isMUBUF(*MI)) |
673 | return !SIInstrInfo::isLegalMUBUFImmOffset(FullOffset); |
674 | |
675 | const SIInstrInfo *TII = ST.getInstrInfo(); |
676 | return !TII->isLegalFLATOffset(FullOffset, AMDGPUAS::PRIVATE_ADDRESS, |
677 | SIInstrFlags::FlatScratch); |
678 | } |
679 | |
680 | Register SIRegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, |
681 | int FrameIdx, |
682 | int64_t Offset) const { |
683 | MachineBasicBlock::iterator Ins = MBB->begin(); |
684 | DebugLoc DL; // Defaults to "unknown" |
685 | |
686 | if (Ins != MBB->end()) |
687 | DL = Ins->getDebugLoc(); |
688 | |
689 | MachineFunction *MF = MBB->getParent(); |
690 | const SIInstrInfo *TII = ST.getInstrInfo(); |
691 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
692 | unsigned MovOpc = ST.enableFlatScratch() ? AMDGPU::S_MOV_B32 |
693 | : AMDGPU::V_MOV_B32_e32; |
694 | |
695 | Register BaseReg = MRI.createVirtualRegister( |
696 | ST.enableFlatScratch() ? &AMDGPU::SReg_32_XEXEC_HIRegClass |
697 | : &AMDGPU::VGPR_32RegClass); |
698 | |
699 | if (Offset == 0) { |
700 | BuildMI(*MBB, Ins, DL, TII->get(MovOpc), BaseReg) |
701 | .addFrameIndex(FrameIdx); |
702 | return BaseReg; |
703 | } |
704 | |
705 | Register OffsetReg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass); |
706 | |
707 | Register FIReg = MRI.createVirtualRegister( |
708 | ST.enableFlatScratch() ? &AMDGPU::SReg_32_XM0RegClass |
709 | : &AMDGPU::VGPR_32RegClass); |
710 | |
711 | BuildMI(*MBB, Ins, DL, TII->get(AMDGPU::S_MOV_B32), OffsetReg) |
712 | .addImm(Offset); |
713 | BuildMI(*MBB, Ins, DL, TII->get(MovOpc), FIReg) |
714 | .addFrameIndex(FrameIdx); |
715 | |
716 | if (ST.enableFlatScratch() ) { |
717 | BuildMI(*MBB, Ins, DL, TII->get(AMDGPU::S_ADD_I32), BaseReg) |
718 | .addReg(OffsetReg, RegState::Kill) |
719 | .addReg(FIReg); |
720 | return BaseReg; |
721 | } |
722 | |
723 | TII->getAddNoCarry(*MBB, Ins, DL, BaseReg) |
724 | .addReg(OffsetReg, RegState::Kill) |
725 | .addReg(FIReg) |
726 | .addImm(0); // clamp bit |
727 | |
728 | return BaseReg; |
729 | } |
730 | |
731 | void SIRegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, |
732 | int64_t Offset) const { |
733 | const SIInstrInfo *TII = ST.getInstrInfo(); |
734 | bool IsFlat = TII->isFLATScratch(MI); |
735 | |
736 | #ifndef NDEBUG |
737 | // FIXME: Is it possible to be storing a frame index to itself? |
738 | bool SeenFI = false; |
739 | for (const MachineOperand &MO: MI.operands()) { |
740 | if (MO.isFI()) { |
741 | if (SeenFI) |
742 | llvm_unreachable("should not see multiple frame indices")::llvm::llvm_unreachable_internal("should not see multiple frame indices" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 742); |
743 | |
744 | SeenFI = true; |
745 | } |
746 | } |
747 | #endif |
748 | |
749 | MachineOperand *FIOp = |
750 | TII->getNamedOperand(MI, IsFlat ? AMDGPU::OpName::saddr |
751 | : AMDGPU::OpName::vaddr); |
752 | |
753 | MachineOperand *OffsetOp = TII->getNamedOperand(MI, AMDGPU::OpName::offset); |
754 | int64_t NewOffset = OffsetOp->getImm() + Offset; |
755 | |
756 | assert(FIOp && FIOp->isFI() && "frame index must be address operand")(static_cast <bool> (FIOp && FIOp->isFI() && "frame index must be address operand") ? void (0) : __assert_fail ("FIOp && FIOp->isFI() && \"frame index must be address operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 756, __extension__ __PRETTY_FUNCTION__)); |
757 | assert(TII->isMUBUF(MI) || TII->isFLATScratch(MI))(static_cast <bool> (TII->isMUBUF(MI) || TII->isFLATScratch (MI)) ? void (0) : __assert_fail ("TII->isMUBUF(MI) || TII->isFLATScratch(MI)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 757, __extension__ __PRETTY_FUNCTION__)); |
758 | |
759 | if (IsFlat) { |
760 | assert(TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS,(static_cast <bool> (TII->isLegalFLATOffset(NewOffset , AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && "offset should be legal") ? void (0) : __assert_fail ("TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && \"offset should be legal\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 762, __extension__ __PRETTY_FUNCTION__)) |
761 | SIInstrFlags::FlatScratch) &&(static_cast <bool> (TII->isLegalFLATOffset(NewOffset , AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && "offset should be legal") ? void (0) : __assert_fail ("TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && \"offset should be legal\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 762, __extension__ __PRETTY_FUNCTION__)) |
762 | "offset should be legal")(static_cast <bool> (TII->isLegalFLATOffset(NewOffset , AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && "offset should be legal") ? void (0) : __assert_fail ("TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, SIInstrFlags::FlatScratch) && \"offset should be legal\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 762, __extension__ __PRETTY_FUNCTION__)); |
763 | FIOp->ChangeToRegister(BaseReg, false); |
764 | OffsetOp->setImm(NewOffset); |
765 | return; |
766 | } |
767 | |
768 | #ifndef NDEBUG |
769 | MachineOperand *SOffset = TII->getNamedOperand(MI, AMDGPU::OpName::soffset); |
770 | assert(SOffset->isImm() && SOffset->getImm() == 0)(static_cast <bool> (SOffset->isImm() && SOffset ->getImm() == 0) ? void (0) : __assert_fail ("SOffset->isImm() && SOffset->getImm() == 0" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 770, __extension__ __PRETTY_FUNCTION__)); |
771 | #endif |
772 | |
773 | assert(SIInstrInfo::isLegalMUBUFImmOffset(NewOffset) &&(static_cast <bool> (SIInstrInfo::isLegalMUBUFImmOffset (NewOffset) && "offset should be legal") ? void (0) : __assert_fail ("SIInstrInfo::isLegalMUBUFImmOffset(NewOffset) && \"offset should be legal\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 774, __extension__ __PRETTY_FUNCTION__)) |
774 | "offset should be legal")(static_cast <bool> (SIInstrInfo::isLegalMUBUFImmOffset (NewOffset) && "offset should be legal") ? void (0) : __assert_fail ("SIInstrInfo::isLegalMUBUFImmOffset(NewOffset) && \"offset should be legal\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 774, __extension__ __PRETTY_FUNCTION__)); |
775 | |
776 | FIOp->ChangeToRegister(BaseReg, false); |
777 | OffsetOp->setImm(NewOffset); |
778 | } |
779 | |
780 | bool SIRegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, |
781 | Register BaseReg, |
782 | int64_t Offset) const { |
783 | if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isFLATScratch(*MI)) |
784 | return false; |
785 | |
786 | int64_t NewOffset = Offset + getScratchInstrOffset(MI); |
787 | |
788 | if (SIInstrInfo::isMUBUF(*MI)) |
789 | return SIInstrInfo::isLegalMUBUFImmOffset(NewOffset); |
790 | |
791 | const SIInstrInfo *TII = ST.getInstrInfo(); |
792 | return TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, |
793 | SIInstrFlags::FlatScratch); |
794 | } |
795 | |
796 | const TargetRegisterClass *SIRegisterInfo::getPointerRegClass( |
797 | const MachineFunction &MF, unsigned Kind) const { |
798 | // This is inaccurate. It depends on the instruction and address space. The |
799 | // only place where we should hit this is for dealing with frame indexes / |
800 | // private accesses, so this is correct in that case. |
801 | return &AMDGPU::VGPR_32RegClass; |
802 | } |
803 | |
804 | const TargetRegisterClass * |
805 | SIRegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { |
806 | if (isAGPRClass(RC) && !ST.hasGFX90AInsts()) |
807 | return getEquivalentVGPRClass(RC); |
808 | |
809 | return RC; |
810 | } |
811 | |
812 | static unsigned getNumSubRegsForSpillOp(unsigned Op) { |
813 | |
814 | switch (Op) { |
815 | case AMDGPU::SI_SPILL_S1024_SAVE: |
816 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
817 | case AMDGPU::SI_SPILL_V1024_SAVE: |
818 | case AMDGPU::SI_SPILL_V1024_RESTORE: |
819 | case AMDGPU::SI_SPILL_A1024_SAVE: |
820 | case AMDGPU::SI_SPILL_A1024_RESTORE: |
821 | return 32; |
822 | case AMDGPU::SI_SPILL_S512_SAVE: |
823 | case AMDGPU::SI_SPILL_S512_RESTORE: |
824 | case AMDGPU::SI_SPILL_V512_SAVE: |
825 | case AMDGPU::SI_SPILL_V512_RESTORE: |
826 | case AMDGPU::SI_SPILL_A512_SAVE: |
827 | case AMDGPU::SI_SPILL_A512_RESTORE: |
828 | return 16; |
829 | case AMDGPU::SI_SPILL_S256_SAVE: |
830 | case AMDGPU::SI_SPILL_S256_RESTORE: |
831 | case AMDGPU::SI_SPILL_V256_SAVE: |
832 | case AMDGPU::SI_SPILL_V256_RESTORE: |
833 | case AMDGPU::SI_SPILL_A256_SAVE: |
834 | case AMDGPU::SI_SPILL_A256_RESTORE: |
835 | return 8; |
836 | case AMDGPU::SI_SPILL_S224_SAVE: |
837 | case AMDGPU::SI_SPILL_S224_RESTORE: |
838 | case AMDGPU::SI_SPILL_V224_SAVE: |
839 | case AMDGPU::SI_SPILL_V224_RESTORE: |
840 | case AMDGPU::SI_SPILL_A224_SAVE: |
841 | case AMDGPU::SI_SPILL_A224_RESTORE: |
842 | return 7; |
843 | case AMDGPU::SI_SPILL_S192_SAVE: |
844 | case AMDGPU::SI_SPILL_S192_RESTORE: |
845 | case AMDGPU::SI_SPILL_V192_SAVE: |
846 | case AMDGPU::SI_SPILL_V192_RESTORE: |
847 | case AMDGPU::SI_SPILL_A192_SAVE: |
848 | case AMDGPU::SI_SPILL_A192_RESTORE: |
849 | return 6; |
850 | case AMDGPU::SI_SPILL_S160_SAVE: |
851 | case AMDGPU::SI_SPILL_S160_RESTORE: |
852 | case AMDGPU::SI_SPILL_V160_SAVE: |
853 | case AMDGPU::SI_SPILL_V160_RESTORE: |
854 | case AMDGPU::SI_SPILL_A160_SAVE: |
855 | case AMDGPU::SI_SPILL_A160_RESTORE: |
856 | return 5; |
857 | case AMDGPU::SI_SPILL_S128_SAVE: |
858 | case AMDGPU::SI_SPILL_S128_RESTORE: |
859 | case AMDGPU::SI_SPILL_V128_SAVE: |
860 | case AMDGPU::SI_SPILL_V128_RESTORE: |
861 | case AMDGPU::SI_SPILL_A128_SAVE: |
862 | case AMDGPU::SI_SPILL_A128_RESTORE: |
863 | return 4; |
864 | case AMDGPU::SI_SPILL_S96_SAVE: |
865 | case AMDGPU::SI_SPILL_S96_RESTORE: |
866 | case AMDGPU::SI_SPILL_V96_SAVE: |
867 | case AMDGPU::SI_SPILL_V96_RESTORE: |
868 | case AMDGPU::SI_SPILL_A96_SAVE: |
869 | case AMDGPU::SI_SPILL_A96_RESTORE: |
870 | return 3; |
871 | case AMDGPU::SI_SPILL_S64_SAVE: |
872 | case AMDGPU::SI_SPILL_S64_RESTORE: |
873 | case AMDGPU::SI_SPILL_V64_SAVE: |
874 | case AMDGPU::SI_SPILL_V64_RESTORE: |
875 | case AMDGPU::SI_SPILL_A64_SAVE: |
876 | case AMDGPU::SI_SPILL_A64_RESTORE: |
877 | return 2; |
878 | case AMDGPU::SI_SPILL_S32_SAVE: |
879 | case AMDGPU::SI_SPILL_S32_RESTORE: |
880 | case AMDGPU::SI_SPILL_V32_SAVE: |
881 | case AMDGPU::SI_SPILL_V32_RESTORE: |
882 | case AMDGPU::SI_SPILL_A32_SAVE: |
883 | case AMDGPU::SI_SPILL_A32_RESTORE: |
884 | return 1; |
885 | default: llvm_unreachable("Invalid spill opcode")::llvm::llvm_unreachable_internal("Invalid spill opcode", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 885); |
886 | } |
887 | } |
888 | |
889 | static int getOffsetMUBUFStore(unsigned Opc) { |
890 | switch (Opc) { |
891 | case AMDGPU::BUFFER_STORE_DWORD_OFFEN: |
892 | return AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
893 | case AMDGPU::BUFFER_STORE_BYTE_OFFEN: |
894 | return AMDGPU::BUFFER_STORE_BYTE_OFFSET; |
895 | case AMDGPU::BUFFER_STORE_SHORT_OFFEN: |
896 | return AMDGPU::BUFFER_STORE_SHORT_OFFSET; |
897 | case AMDGPU::BUFFER_STORE_DWORDX2_OFFEN: |
898 | return AMDGPU::BUFFER_STORE_DWORDX2_OFFSET; |
899 | case AMDGPU::BUFFER_STORE_DWORDX4_OFFEN: |
900 | return AMDGPU::BUFFER_STORE_DWORDX4_OFFSET; |
901 | case AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFEN: |
902 | return AMDGPU::BUFFER_STORE_SHORT_D16_HI_OFFSET; |
903 | case AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFEN: |
904 | return AMDGPU::BUFFER_STORE_BYTE_D16_HI_OFFSET; |
905 | default: |
906 | return -1; |
907 | } |
908 | } |
909 | |
910 | static int getOffsetMUBUFLoad(unsigned Opc) { |
911 | switch (Opc) { |
912 | case AMDGPU::BUFFER_LOAD_DWORD_OFFEN: |
913 | return AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
914 | case AMDGPU::BUFFER_LOAD_UBYTE_OFFEN: |
915 | return AMDGPU::BUFFER_LOAD_UBYTE_OFFSET; |
916 | case AMDGPU::BUFFER_LOAD_SBYTE_OFFEN: |
917 | return AMDGPU::BUFFER_LOAD_SBYTE_OFFSET; |
918 | case AMDGPU::BUFFER_LOAD_USHORT_OFFEN: |
919 | return AMDGPU::BUFFER_LOAD_USHORT_OFFSET; |
920 | case AMDGPU::BUFFER_LOAD_SSHORT_OFFEN: |
921 | return AMDGPU::BUFFER_LOAD_SSHORT_OFFSET; |
922 | case AMDGPU::BUFFER_LOAD_DWORDX2_OFFEN: |
923 | return AMDGPU::BUFFER_LOAD_DWORDX2_OFFSET; |
924 | case AMDGPU::BUFFER_LOAD_DWORDX4_OFFEN: |
925 | return AMDGPU::BUFFER_LOAD_DWORDX4_OFFSET; |
926 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFEN: |
927 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_OFFSET; |
928 | case AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFEN: |
929 | return AMDGPU::BUFFER_LOAD_UBYTE_D16_HI_OFFSET; |
930 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFEN: |
931 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_OFFSET; |
932 | case AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFEN: |
933 | return AMDGPU::BUFFER_LOAD_SBYTE_D16_HI_OFFSET; |
934 | case AMDGPU::BUFFER_LOAD_SHORT_D16_OFFEN: |
935 | return AMDGPU::BUFFER_LOAD_SHORT_D16_OFFSET; |
936 | case AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFEN: |
937 | return AMDGPU::BUFFER_LOAD_SHORT_D16_HI_OFFSET; |
938 | default: |
939 | return -1; |
940 | } |
941 | } |
942 | |
943 | static MachineInstrBuilder spillVGPRtoAGPR(const GCNSubtarget &ST, |
944 | MachineBasicBlock &MBB, |
945 | MachineBasicBlock::iterator MI, |
946 | int Index, unsigned Lane, |
947 | unsigned ValueReg, bool IsKill) { |
948 | MachineFunction *MF = MBB.getParent(); |
949 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
950 | const SIInstrInfo *TII = ST.getInstrInfo(); |
951 | |
952 | MCPhysReg Reg = MFI->getVGPRToAGPRSpill(Index, Lane); |
953 | |
954 | if (Reg == AMDGPU::NoRegister) |
955 | return MachineInstrBuilder(); |
956 | |
957 | bool IsStore = MI->mayStore(); |
958 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
959 | auto *TRI = static_cast<const SIRegisterInfo*>(MRI.getTargetRegisterInfo()); |
960 | |
961 | unsigned Dst = IsStore ? Reg : ValueReg; |
962 | unsigned Src = IsStore ? ValueReg : Reg; |
963 | unsigned Opc = (IsStore ^ TRI->isVGPR(MRI, Reg)) ? AMDGPU::V_ACCVGPR_WRITE_B32_e64 |
964 | : AMDGPU::V_ACCVGPR_READ_B32_e64; |
965 | |
966 | auto MIB = BuildMI(MBB, MI, MI->getDebugLoc(), TII->get(Opc), Dst) |
967 | .addReg(Src, getKillRegState(IsKill)); |
968 | MIB->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
969 | return MIB; |
970 | } |
971 | |
972 | // This differs from buildSpillLoadStore by only scavenging a VGPR. It does not |
973 | // need to handle the case where an SGPR may need to be spilled while spilling. |
974 | static bool buildMUBUFOffsetLoadStore(const GCNSubtarget &ST, |
975 | MachineFrameInfo &MFI, |
976 | MachineBasicBlock::iterator MI, |
977 | int Index, |
978 | int64_t Offset) { |
979 | const SIInstrInfo *TII = ST.getInstrInfo(); |
980 | MachineBasicBlock *MBB = MI->getParent(); |
981 | const DebugLoc &DL = MI->getDebugLoc(); |
982 | bool IsStore = MI->mayStore(); |
983 | |
984 | unsigned Opc = MI->getOpcode(); |
985 | int LoadStoreOp = IsStore ? |
986 | getOffsetMUBUFStore(Opc) : getOffsetMUBUFLoad(Opc); |
987 | if (LoadStoreOp == -1) |
988 | return false; |
989 | |
990 | const MachineOperand *Reg = TII->getNamedOperand(*MI, AMDGPU::OpName::vdata); |
991 | if (spillVGPRtoAGPR(ST, *MBB, MI, Index, 0, Reg->getReg(), false).getInstr()) |
992 | return true; |
993 | |
994 | MachineInstrBuilder NewMI = |
995 | BuildMI(*MBB, MI, DL, TII->get(LoadStoreOp)) |
996 | .add(*Reg) |
997 | .add(*TII->getNamedOperand(*MI, AMDGPU::OpName::srsrc)) |
998 | .add(*TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)) |
999 | .addImm(Offset) |
1000 | .addImm(0) // cpol |
1001 | .addImm(0) // tfe |
1002 | .addImm(0) // swz |
1003 | .cloneMemRefs(*MI); |
1004 | |
1005 | const MachineOperand *VDataIn = TII->getNamedOperand(*MI, |
1006 | AMDGPU::OpName::vdata_in); |
1007 | if (VDataIn) |
1008 | NewMI.add(*VDataIn); |
1009 | return true; |
1010 | } |
1011 | |
1012 | static unsigned getFlatScratchSpillOpcode(const SIInstrInfo *TII, |
1013 | unsigned LoadStoreOp, |
1014 | unsigned EltSize) { |
1015 | bool IsStore = TII->get(LoadStoreOp).mayStore(); |
1016 | bool UseST = |
1017 | AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::vaddr) < 0 && |
1018 | AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::saddr) < 0; |
1019 | |
1020 | switch (EltSize) { |
1021 | case 4: |
1022 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
1023 | : AMDGPU::SCRATCH_LOAD_DWORD_SADDR; |
1024 | break; |
1025 | case 8: |
1026 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX2_SADDR |
1027 | : AMDGPU::SCRATCH_LOAD_DWORDX2_SADDR; |
1028 | break; |
1029 | case 12: |
1030 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX3_SADDR |
1031 | : AMDGPU::SCRATCH_LOAD_DWORDX3_SADDR; |
1032 | break; |
1033 | case 16: |
1034 | LoadStoreOp = IsStore ? AMDGPU::SCRATCH_STORE_DWORDX4_SADDR |
1035 | : AMDGPU::SCRATCH_LOAD_DWORDX4_SADDR; |
1036 | break; |
1037 | default: |
1038 | llvm_unreachable("Unexpected spill load/store size!")::llvm::llvm_unreachable_internal("Unexpected spill load/store size!" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1038); |
1039 | } |
1040 | |
1041 | if (UseST) |
1042 | LoadStoreOp = AMDGPU::getFlatScratchInstSTfromSS(LoadStoreOp); |
1043 | |
1044 | return LoadStoreOp; |
1045 | } |
1046 | |
1047 | void SIRegisterInfo::buildSpillLoadStore( |
1048 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
1049 | unsigned LoadStoreOp, int Index, Register ValueReg, bool IsKill, |
1050 | MCRegister ScratchOffsetReg, int64_t InstOffset, MachineMemOperand *MMO, |
1051 | RegScavenger *RS, LivePhysRegs *LiveRegs) const { |
1052 | assert((!RS || !LiveRegs) && "Only RS or LiveRegs can be set but not both")(static_cast <bool> ((!RS || !LiveRegs) && "Only RS or LiveRegs can be set but not both" ) ? void (0) : __assert_fail ("(!RS || !LiveRegs) && \"Only RS or LiveRegs can be set but not both\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1052, __extension__ __PRETTY_FUNCTION__)); |
1053 | |
1054 | MachineFunction *MF = MBB.getParent(); |
1055 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1056 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
1057 | const SIMachineFunctionInfo *FuncInfo = MF->getInfo<SIMachineFunctionInfo>(); |
1058 | |
1059 | const MCInstrDesc *Desc = &TII->get(LoadStoreOp); |
1060 | const DebugLoc &DL = MI != MBB.end() ? MI->getDebugLoc() : DebugLoc(); |
1061 | bool IsStore = Desc->mayStore(); |
1062 | bool IsFlat = TII->isFLATScratch(LoadStoreOp); |
1063 | |
1064 | bool Scavenged = false; |
1065 | MCRegister SOffset = ScratchOffsetReg; |
1066 | |
1067 | const TargetRegisterClass *RC = getRegClassForReg(MF->getRegInfo(), ValueReg); |
1068 | // On gfx90a+ AGPR is a regular VGPR acceptable for loads and stores. |
1069 | const bool IsAGPR = !ST.hasGFX90AInsts() && hasAGPRs(RC); |
1070 | const unsigned RegWidth = AMDGPU::getRegBitWidth(RC->getID()) / 8; |
1071 | |
1072 | // Always use 4 byte operations for AGPRs because we need to scavenge |
1073 | // a temporary VGPR. |
1074 | unsigned EltSize = (IsFlat && !IsAGPR) ? std::min(RegWidth, 16u) : 4u; |
1075 | unsigned NumSubRegs = RegWidth / EltSize; |
1076 | unsigned Size = NumSubRegs * EltSize; |
1077 | unsigned RemSize = RegWidth - Size; |
1078 | unsigned NumRemSubRegs = RemSize ? 1 : 0; |
1079 | int64_t Offset = InstOffset + MFI.getObjectOffset(Index); |
1080 | int64_t MaxOffset = Offset + Size + RemSize - EltSize; |
1081 | int64_t ScratchOffsetRegDelta = 0; |
1082 | |
1083 | if (IsFlat && EltSize > 4) { |
1084 | LoadStoreOp = getFlatScratchSpillOpcode(TII, LoadStoreOp, EltSize); |
1085 | Desc = &TII->get(LoadStoreOp); |
Value stored to 'Desc' is never read | |
1086 | } |
1087 | |
1088 | Align Alignment = MFI.getObjectAlign(Index); |
1089 | const MachinePointerInfo &BasePtrInfo = MMO->getPointerInfo(); |
1090 | |
1091 | assert((IsFlat || ((Offset % EltSize) == 0)) &&(static_cast <bool> ((IsFlat || ((Offset % EltSize) == 0 )) && "unexpected VGPR spill offset") ? void (0) : __assert_fail ("(IsFlat || ((Offset % EltSize) == 0)) && \"unexpected VGPR spill offset\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1092, __extension__ __PRETTY_FUNCTION__)) |
1092 | "unexpected VGPR spill offset")(static_cast <bool> ((IsFlat || ((Offset % EltSize) == 0 )) && "unexpected VGPR spill offset") ? void (0) : __assert_fail ("(IsFlat || ((Offset % EltSize) == 0)) && \"unexpected VGPR spill offset\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1092, __extension__ __PRETTY_FUNCTION__)); |
1093 | |
1094 | bool IsOffsetLegal = |
1095 | IsFlat ? TII->isLegalFLATOffset(MaxOffset, AMDGPUAS::PRIVATE_ADDRESS, |
1096 | SIInstrFlags::FlatScratch) |
1097 | : SIInstrInfo::isLegalMUBUFImmOffset(MaxOffset); |
1098 | if (!IsOffsetLegal || (IsFlat && !SOffset && !ST.hasFlatScratchSTMode())) { |
1099 | SOffset = MCRegister(); |
1100 | |
1101 | // We currently only support spilling VGPRs to EltSize boundaries, meaning |
1102 | // we can simplify the adjustment of Offset here to just scale with |
1103 | // WavefrontSize. |
1104 | if (!IsFlat) |
1105 | Offset *= ST.getWavefrontSize(); |
1106 | |
1107 | // We don't have access to the register scavenger if this function is called |
1108 | // during PEI::scavengeFrameVirtualRegs() so use LiveRegs in this case. |
1109 | if (RS) { |
1110 | SOffset = RS->scavengeRegister(&AMDGPU::SGPR_32RegClass, MI, 0, false); |
1111 | } else if (LiveRegs) { |
1112 | for (MCRegister Reg : AMDGPU::SGPR_32RegClass) { |
1113 | if (LiveRegs->available(MF->getRegInfo(), Reg)) { |
1114 | SOffset = Reg; |
1115 | break; |
1116 | } |
1117 | } |
1118 | } |
1119 | |
1120 | if (!SOffset) { |
1121 | // There are no free SGPRs, and since we are in the process of spilling |
1122 | // VGPRs too. Since we need a VGPR in order to spill SGPRs (this is true |
1123 | // on SI/CI and on VI it is true until we implement spilling using scalar |
1124 | // stores), we have no way to free up an SGPR. Our solution here is to |
1125 | // add the offset directly to the ScratchOffset or StackPtrOffset |
1126 | // register, and then subtract the offset after the spill to return the |
1127 | // register to it's original value. |
1128 | if (!ScratchOffsetReg) |
1129 | ScratchOffsetReg = FuncInfo->getStackPtrOffsetReg(); |
1130 | SOffset = ScratchOffsetReg; |
1131 | ScratchOffsetRegDelta = Offset; |
1132 | } else { |
1133 | Scavenged = true; |
1134 | } |
1135 | |
1136 | if (!SOffset) |
1137 | report_fatal_error("could not scavenge SGPR to spill in entry function"); |
1138 | |
1139 | if (ScratchOffsetReg == AMDGPU::NoRegister) { |
1140 | BuildMI(MBB, MI, DL, TII->get(AMDGPU::S_MOV_B32), SOffset).addImm(Offset); |
1141 | } else { |
1142 | BuildMI(MBB, MI, DL, TII->get(AMDGPU::S_ADD_I32), SOffset) |
1143 | .addReg(ScratchOffsetReg) |
1144 | .addImm(Offset); |
1145 | } |
1146 | |
1147 | Offset = 0; |
1148 | } |
1149 | |
1150 | if (IsFlat && SOffset == AMDGPU::NoRegister) { |
1151 | assert(AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::vaddr) < 0(static_cast <bool> (AMDGPU::getNamedOperandIdx(LoadStoreOp , AMDGPU::OpName::vaddr) < 0 && "Unexpected vaddr for flat scratch with a FI operand" ) ? void (0) : __assert_fail ("AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::vaddr) < 0 && \"Unexpected vaddr for flat scratch with a FI operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1152, __extension__ __PRETTY_FUNCTION__)) |
1152 | && "Unexpected vaddr for flat scratch with a FI operand")(static_cast <bool> (AMDGPU::getNamedOperandIdx(LoadStoreOp , AMDGPU::OpName::vaddr) < 0 && "Unexpected vaddr for flat scratch with a FI operand" ) ? void (0) : __assert_fail ("AMDGPU::getNamedOperandIdx(LoadStoreOp, AMDGPU::OpName::vaddr) < 0 && \"Unexpected vaddr for flat scratch with a FI operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1152, __extension__ __PRETTY_FUNCTION__)); |
1153 | |
1154 | assert(ST.hasFlatScratchSTMode())(static_cast <bool> (ST.hasFlatScratchSTMode()) ? void ( 0) : __assert_fail ("ST.hasFlatScratchSTMode()", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1154, __extension__ __PRETTY_FUNCTION__)); |
1155 | LoadStoreOp = AMDGPU::getFlatScratchInstSTfromSS(LoadStoreOp); |
1156 | Desc = &TII->get(LoadStoreOp); |
1157 | } |
1158 | |
1159 | Register TmpReg; |
1160 | |
1161 | for (unsigned i = 0, e = NumSubRegs + NumRemSubRegs, RegOffset = 0; i != e; |
1162 | ++i, RegOffset += EltSize) { |
1163 | if (i == NumSubRegs) { |
1164 | EltSize = RemSize; |
1165 | LoadStoreOp = getFlatScratchSpillOpcode(TII, LoadStoreOp, EltSize); |
1166 | } |
1167 | Desc = &TII->get(LoadStoreOp); |
1168 | |
1169 | unsigned NumRegs = EltSize / 4; |
1170 | Register SubReg = e == 1 |
1171 | ? ValueReg |
1172 | : Register(getSubReg(ValueReg, |
1173 | getSubRegFromChannel(RegOffset / 4, NumRegs))); |
1174 | |
1175 | unsigned SOffsetRegState = 0; |
1176 | unsigned SrcDstRegState = getDefRegState(!IsStore); |
1177 | if (i + 1 == e) { |
1178 | SOffsetRegState |= getKillRegState(Scavenged); |
1179 | // The last implicit use carries the "Kill" flag. |
1180 | SrcDstRegState |= getKillRegState(IsKill); |
1181 | } |
1182 | |
1183 | // Make sure the whole register is defined if there are undef components by |
1184 | // adding an implicit def of the super-reg on the first instruction. |
1185 | bool NeedSuperRegDef = e > 1 && IsStore && i == 0; |
1186 | bool NeedSuperRegImpOperand = e > 1; |
1187 | |
1188 | // Remaining element size to spill into memory after some parts of it |
1189 | // spilled into either AGPRs or VGPRs. |
1190 | unsigned RemEltSize = EltSize; |
1191 | |
1192 | // AGPRs to spill VGPRs and vice versa are allocated in a reverse order, |
1193 | // starting from the last lane. In case if a register cannot be completely |
1194 | // spilled into another register that will ensure its alignment does not |
1195 | // change. For targets with VGPR alignment requirement this is important |
1196 | // in case of flat scratch usage as we might get a scratch_load or |
1197 | // scratch_store of an unaligned register otherwise. |
1198 | for (int LaneS = (RegOffset + EltSize) / 4 - 1, Lane = LaneS, |
1199 | LaneE = RegOffset / 4; |
1200 | Lane >= LaneE; --Lane) { |
1201 | bool IsSubReg = e > 1 || EltSize > 4; |
1202 | Register Sub = IsSubReg |
1203 | ? Register(getSubReg(ValueReg, getSubRegFromChannel(Lane))) |
1204 | : ValueReg; |
1205 | auto MIB = spillVGPRtoAGPR(ST, MBB, MI, Index, Lane, Sub, IsKill); |
1206 | if (!MIB.getInstr()) |
1207 | break; |
1208 | if (NeedSuperRegDef || (IsSubReg && IsStore && Lane == LaneS && !i)) { |
1209 | MIB.addReg(ValueReg, RegState::ImplicitDefine); |
1210 | NeedSuperRegDef = false; |
1211 | } |
1212 | if (IsSubReg || NeedSuperRegImpOperand) { |
1213 | NeedSuperRegImpOperand = true; |
1214 | unsigned State = SrcDstRegState; |
1215 | if (Lane != LaneE) |
1216 | State &= ~RegState::Kill; |
1217 | MIB.addReg(ValueReg, RegState::Implicit | State); |
1218 | } |
1219 | RemEltSize -= 4; |
1220 | } |
1221 | |
1222 | if (!RemEltSize) // Fully spilled into AGPRs. |
1223 | continue; |
1224 | |
1225 | if (RemEltSize != EltSize) { // Partially spilled to AGPRs |
1226 | assert(IsFlat && EltSize > 4)(static_cast <bool> (IsFlat && EltSize > 4) ? void (0) : __assert_fail ("IsFlat && EltSize > 4" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1226, __extension__ __PRETTY_FUNCTION__)); |
1227 | |
1228 | unsigned NumRegs = RemEltSize / 4; |
1229 | SubReg = Register(getSubReg(ValueReg, |
1230 | getSubRegFromChannel(RegOffset / 4, NumRegs))); |
1231 | unsigned Opc = getFlatScratchSpillOpcode(TII, LoadStoreOp, RemEltSize); |
1232 | Desc = &TII->get(Opc); |
1233 | } |
1234 | |
1235 | unsigned FinalReg = SubReg; |
1236 | |
1237 | if (IsAGPR) { |
1238 | assert(EltSize == 4)(static_cast <bool> (EltSize == 4) ? void (0) : __assert_fail ("EltSize == 4", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1238, __extension__ __PRETTY_FUNCTION__)); |
1239 | |
1240 | if (!TmpReg) { |
1241 | assert(RS && "Needs to have RegScavenger to spill an AGPR!")(static_cast <bool> (RS && "Needs to have RegScavenger to spill an AGPR!" ) ? void (0) : __assert_fail ("RS && \"Needs to have RegScavenger to spill an AGPR!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1241, __extension__ __PRETTY_FUNCTION__)); |
1242 | // FIXME: change to scavengeRegisterBackwards() |
1243 | TmpReg = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0); |
1244 | RS->setRegUsed(TmpReg); |
1245 | } |
1246 | if (IsStore) { |
1247 | auto AccRead = BuildMI(MBB, MI, DL, |
1248 | TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64), TmpReg) |
1249 | .addReg(SubReg, getKillRegState(IsKill)); |
1250 | if (NeedSuperRegDef) |
1251 | AccRead.addReg(ValueReg, RegState::ImplicitDefine); |
1252 | AccRead->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1253 | } |
1254 | SubReg = TmpReg; |
1255 | } |
1256 | |
1257 | MachinePointerInfo PInfo = BasePtrInfo.getWithOffset(RegOffset); |
1258 | MachineMemOperand *NewMMO = |
1259 | MF->getMachineMemOperand(PInfo, MMO->getFlags(), RemEltSize, |
1260 | commonAlignment(Alignment, RegOffset)); |
1261 | |
1262 | auto MIB = |
1263 | BuildMI(MBB, MI, DL, *Desc) |
1264 | .addReg(SubReg, getDefRegState(!IsStore) | getKillRegState(IsKill)); |
1265 | if (!IsFlat) |
1266 | MIB.addReg(FuncInfo->getScratchRSrcReg()); |
1267 | |
1268 | if (SOffset == AMDGPU::NoRegister) { |
1269 | if (!IsFlat) |
1270 | MIB.addImm(0); |
1271 | } else { |
1272 | MIB.addReg(SOffset, SOffsetRegState); |
1273 | } |
1274 | MIB.addImm(Offset + RegOffset) |
1275 | .addImm(0); // cpol |
1276 | if (!IsFlat) |
1277 | MIB.addImm(0) // tfe |
1278 | .addImm(0); // swz |
1279 | MIB.addMemOperand(NewMMO); |
1280 | |
1281 | if (!IsAGPR && NeedSuperRegDef) |
1282 | MIB.addReg(ValueReg, RegState::ImplicitDefine); |
1283 | |
1284 | if (!IsStore && TmpReg != AMDGPU::NoRegister) { |
1285 | MIB = BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), |
1286 | FinalReg) |
1287 | .addReg(TmpReg, RegState::Kill); |
1288 | MIB->setAsmPrinterFlag(MachineInstr::ReloadReuse); |
1289 | } |
1290 | |
1291 | if (NeedSuperRegImpOperand) |
1292 | MIB.addReg(ValueReg, RegState::Implicit | SrcDstRegState); |
1293 | } |
1294 | |
1295 | if (ScratchOffsetRegDelta != 0) { |
1296 | // Subtract the offset we added to the ScratchOffset register. |
1297 | BuildMI(MBB, MI, DL, TII->get(AMDGPU::S_ADD_I32), SOffset) |
1298 | .addReg(SOffset) |
1299 | .addImm(-ScratchOffsetRegDelta); |
1300 | } |
1301 | } |
1302 | |
1303 | void SIRegisterInfo::buildVGPRSpillLoadStore(SGPRSpillBuilder &SB, int Index, |
1304 | int Offset, bool IsLoad, |
1305 | bool IsKill) const { |
1306 | // Load/store VGPR |
1307 | MachineFrameInfo &FrameInfo = SB.MF.getFrameInfo(); |
1308 | assert(FrameInfo.getStackID(Index) != TargetStackID::SGPRSpill)(static_cast <bool> (FrameInfo.getStackID(Index) != TargetStackID ::SGPRSpill) ? void (0) : __assert_fail ("FrameInfo.getStackID(Index) != TargetStackID::SGPRSpill" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1308, __extension__ __PRETTY_FUNCTION__)); |
1309 | |
1310 | Register FrameReg = |
1311 | FrameInfo.isFixedObjectIndex(Index) && hasBasePointer(SB.MF) |
1312 | ? getBaseRegister() |
1313 | : getFrameRegister(SB.MF); |
1314 | |
1315 | Align Alignment = FrameInfo.getObjectAlign(Index); |
1316 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SB.MF, Index); |
1317 | MachineMemOperand *MMO = SB.MF.getMachineMemOperand( |
1318 | PtrInfo, IsLoad ? MachineMemOperand::MOLoad : MachineMemOperand::MOStore, |
1319 | SB.EltSize, Alignment); |
1320 | |
1321 | if (IsLoad) { |
1322 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_LOAD_DWORD_SADDR |
1323 | : AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
1324 | buildSpillLoadStore(SB.MBB, SB.MI, Opc, Index, SB.TmpVGPR, false, FrameReg, |
1325 | Offset * SB.EltSize, MMO, SB.RS); |
1326 | } else { |
1327 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
1328 | : AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
1329 | buildSpillLoadStore(SB.MBB, SB.MI, Opc, Index, SB.TmpVGPR, IsKill, FrameReg, |
1330 | Offset * SB.EltSize, MMO, SB.RS); |
1331 | // This only ever adds one VGPR spill |
1332 | SB.MFI.addToSpilledVGPRs(1); |
1333 | } |
1334 | } |
1335 | |
1336 | bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, |
1337 | int Index, |
1338 | RegScavenger *RS, |
1339 | LiveIntervals *LIS, |
1340 | bool OnlyToVGPR) const { |
1341 | SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS); |
1342 | |
1343 | ArrayRef<SIMachineFunctionInfo::SpilledReg> VGPRSpills = |
1344 | SB.MFI.getSGPRToVGPRSpills(Index); |
1345 | bool SpillToVGPR = !VGPRSpills.empty(); |
1346 | if (OnlyToVGPR && !SpillToVGPR) |
1347 | return false; |
1348 | |
1349 | assert(SpillToVGPR || (SB.SuperReg != SB.MFI.getStackPtrOffsetReg() &&(static_cast <bool> (SpillToVGPR || (SB.SuperReg != SB. MFI.getStackPtrOffsetReg() && SB.SuperReg != SB.MFI.getFrameOffsetReg ())) ? void (0) : __assert_fail ("SpillToVGPR || (SB.SuperReg != SB.MFI.getStackPtrOffsetReg() && SB.SuperReg != SB.MFI.getFrameOffsetReg())" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1350, __extension__ __PRETTY_FUNCTION__)) |
1350 | SB.SuperReg != SB.MFI.getFrameOffsetReg()))(static_cast <bool> (SpillToVGPR || (SB.SuperReg != SB. MFI.getStackPtrOffsetReg() && SB.SuperReg != SB.MFI.getFrameOffsetReg ())) ? void (0) : __assert_fail ("SpillToVGPR || (SB.SuperReg != SB.MFI.getStackPtrOffsetReg() && SB.SuperReg != SB.MFI.getFrameOffsetReg())" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1350, __extension__ __PRETTY_FUNCTION__)); |
1351 | |
1352 | if (SpillToVGPR) { |
1353 | for (unsigned i = 0, e = SB.NumSubRegs; i < e; ++i) { |
1354 | Register SubReg = |
1355 | SB.NumSubRegs == 1 |
1356 | ? SB.SuperReg |
1357 | : Register(getSubReg(SB.SuperReg, SB.SplitParts[i])); |
1358 | SIMachineFunctionInfo::SpilledReg Spill = VGPRSpills[i]; |
1359 | |
1360 | bool UseKill = SB.IsKill && i == SB.NumSubRegs - 1; |
1361 | |
1362 | // Mark the "old value of vgpr" input undef only if this is the first sgpr |
1363 | // spill to this specific vgpr in the first basic block. |
1364 | auto MIB = BuildMI(SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_WRITELANE_B32), |
1365 | Spill.VGPR) |
1366 | .addReg(SubReg, getKillRegState(UseKill)) |
1367 | .addImm(Spill.Lane) |
1368 | .addReg(Spill.VGPR); |
1369 | if (LIS) { |
1370 | if (i == 0) |
1371 | LIS->ReplaceMachineInstrInMaps(*MI, *MIB); |
1372 | else |
1373 | LIS->InsertMachineInstrInMaps(*MIB); |
1374 | } |
1375 | |
1376 | if (i == 0 && SB.NumSubRegs > 1) { |
1377 | // We may be spilling a super-register which is only partially defined, |
1378 | // and need to ensure later spills think the value is defined. |
1379 | MIB.addReg(SB.SuperReg, RegState::ImplicitDefine); |
1380 | } |
1381 | |
1382 | if (SB.NumSubRegs > 1) |
1383 | MIB.addReg(SB.SuperReg, getKillRegState(UseKill) | RegState::Implicit); |
1384 | |
1385 | // FIXME: Since this spills to another register instead of an actual |
1386 | // frame index, we should delete the frame index when all references to |
1387 | // it are fixed. |
1388 | } |
1389 | } else { |
1390 | SB.prepare(); |
1391 | |
1392 | // SubReg carries the "Kill" flag when SubReg == SB.SuperReg. |
1393 | unsigned SubKillState = getKillRegState((SB.NumSubRegs == 1) && SB.IsKill); |
1394 | |
1395 | // Per VGPR helper data |
1396 | auto PVD = SB.getPerVGPRData(); |
1397 | |
1398 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
1399 | unsigned TmpVGPRFlags = RegState::Undef; |
1400 | |
1401 | // Write sub registers into the VGPR |
1402 | for (unsigned i = Offset * PVD.PerVGPR, |
1403 | e = std::min((Offset + 1) * PVD.PerVGPR, SB.NumSubRegs); |
1404 | i < e; ++i) { |
1405 | Register SubReg = |
1406 | SB.NumSubRegs == 1 |
1407 | ? SB.SuperReg |
1408 | : Register(getSubReg(SB.SuperReg, SB.SplitParts[i])); |
1409 | |
1410 | MachineInstrBuilder WriteLane = |
1411 | BuildMI(SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_WRITELANE_B32), |
1412 | SB.TmpVGPR) |
1413 | .addReg(SubReg, SubKillState) |
1414 | .addImm(i % PVD.PerVGPR) |
1415 | .addReg(SB.TmpVGPR, TmpVGPRFlags); |
1416 | TmpVGPRFlags = 0; |
1417 | |
1418 | if (LIS) { |
1419 | if (i == 0) |
1420 | LIS->ReplaceMachineInstrInMaps(*MI, *WriteLane); |
1421 | else |
1422 | LIS->InsertMachineInstrInMaps(*WriteLane); |
1423 | } |
1424 | |
1425 | // There could be undef components of a spilled super register. |
1426 | // TODO: Can we detect this and skip the spill? |
1427 | if (SB.NumSubRegs > 1) { |
1428 | // The last implicit use of the SB.SuperReg carries the "Kill" flag. |
1429 | unsigned SuperKillState = 0; |
1430 | if (i + 1 == SB.NumSubRegs) |
1431 | SuperKillState |= getKillRegState(SB.IsKill); |
1432 | WriteLane.addReg(SB.SuperReg, RegState::Implicit | SuperKillState); |
1433 | } |
1434 | } |
1435 | |
1436 | // Write out VGPR |
1437 | SB.readWriteTmpVGPR(Offset, /*IsLoad*/ false); |
1438 | } |
1439 | |
1440 | SB.restore(); |
1441 | } |
1442 | |
1443 | MI->eraseFromParent(); |
1444 | SB.MFI.addToSpilledSGPRs(SB.NumSubRegs); |
1445 | |
1446 | if (LIS) |
1447 | LIS->removeAllRegUnitsForPhysReg(SB.SuperReg); |
1448 | |
1449 | return true; |
1450 | } |
1451 | |
1452 | bool SIRegisterInfo::restoreSGPR(MachineBasicBlock::iterator MI, |
1453 | int Index, |
1454 | RegScavenger *RS, |
1455 | LiveIntervals *LIS, |
1456 | bool OnlyToVGPR) const { |
1457 | SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS); |
1458 | |
1459 | ArrayRef<SIMachineFunctionInfo::SpilledReg> VGPRSpills = |
1460 | SB.MFI.getSGPRToVGPRSpills(Index); |
1461 | bool SpillToVGPR = !VGPRSpills.empty(); |
1462 | if (OnlyToVGPR && !SpillToVGPR) |
1463 | return false; |
1464 | |
1465 | if (SpillToVGPR) { |
1466 | for (unsigned i = 0, e = SB.NumSubRegs; i < e; ++i) { |
1467 | Register SubReg = |
1468 | SB.NumSubRegs == 1 |
1469 | ? SB.SuperReg |
1470 | : Register(getSubReg(SB.SuperReg, SB.SplitParts[i])); |
1471 | |
1472 | SIMachineFunctionInfo::SpilledReg Spill = VGPRSpills[i]; |
1473 | auto MIB = |
1474 | BuildMI(SB.MBB, MI, SB.DL, SB.TII.get(AMDGPU::V_READLANE_B32), SubReg) |
1475 | .addReg(Spill.VGPR) |
1476 | .addImm(Spill.Lane); |
1477 | if (SB.NumSubRegs > 1 && i == 0) |
1478 | MIB.addReg(SB.SuperReg, RegState::ImplicitDefine); |
1479 | if (LIS) { |
1480 | if (i == e - 1) |
1481 | LIS->ReplaceMachineInstrInMaps(*MI, *MIB); |
1482 | else |
1483 | LIS->InsertMachineInstrInMaps(*MIB); |
1484 | } |
1485 | |
1486 | } |
1487 | } else { |
1488 | SB.prepare(); |
1489 | |
1490 | // Per VGPR helper data |
1491 | auto PVD = SB.getPerVGPRData(); |
1492 | |
1493 | for (unsigned Offset = 0; Offset < PVD.NumVGPRs; ++Offset) { |
1494 | // Load in VGPR data |
1495 | SB.readWriteTmpVGPR(Offset, /*IsLoad*/ true); |
1496 | |
1497 | // Unpack lanes |
1498 | for (unsigned i = Offset * PVD.PerVGPR, |
1499 | e = std::min((Offset + 1) * PVD.PerVGPR, SB.NumSubRegs); |
1500 | i < e; ++i) { |
1501 | Register SubReg = |
1502 | SB.NumSubRegs == 1 |
1503 | ? SB.SuperReg |
1504 | : Register(getSubReg(SB.SuperReg, SB.SplitParts[i])); |
1505 | |
1506 | bool LastSubReg = (i + 1 == e); |
1507 | auto MIB = BuildMI(SB.MBB, MI, SB.DL, |
1508 | SB.TII.get(AMDGPU::V_READLANE_B32), SubReg) |
1509 | .addReg(SB.TmpVGPR, getKillRegState(LastSubReg)) |
1510 | .addImm(i); |
1511 | if (SB.NumSubRegs > 1 && i == 0) |
1512 | MIB.addReg(SB.SuperReg, RegState::ImplicitDefine); |
1513 | if (LIS) { |
1514 | if (i == e - 1) |
1515 | LIS->ReplaceMachineInstrInMaps(*MI, *MIB); |
1516 | else |
1517 | LIS->InsertMachineInstrInMaps(*MIB); |
1518 | } |
1519 | } |
1520 | } |
1521 | |
1522 | SB.restore(); |
1523 | } |
1524 | |
1525 | MI->eraseFromParent(); |
1526 | |
1527 | if (LIS) |
1528 | LIS->removeAllRegUnitsForPhysReg(SB.SuperReg); |
1529 | |
1530 | return true; |
1531 | } |
1532 | |
1533 | /// Special case of eliminateFrameIndex. Returns true if the SGPR was spilled to |
1534 | /// a VGPR and the stack slot can be safely eliminated when all other users are |
1535 | /// handled. |
1536 | bool SIRegisterInfo::eliminateSGPRToVGPRSpillFrameIndex( |
1537 | MachineBasicBlock::iterator MI, |
1538 | int FI, |
1539 | RegScavenger *RS, |
1540 | LiveIntervals *LIS) const { |
1541 | switch (MI->getOpcode()) { |
1542 | case AMDGPU::SI_SPILL_S1024_SAVE: |
1543 | case AMDGPU::SI_SPILL_S512_SAVE: |
1544 | case AMDGPU::SI_SPILL_S256_SAVE: |
1545 | case AMDGPU::SI_SPILL_S224_SAVE: |
1546 | case AMDGPU::SI_SPILL_S192_SAVE: |
1547 | case AMDGPU::SI_SPILL_S160_SAVE: |
1548 | case AMDGPU::SI_SPILL_S128_SAVE: |
1549 | case AMDGPU::SI_SPILL_S96_SAVE: |
1550 | case AMDGPU::SI_SPILL_S64_SAVE: |
1551 | case AMDGPU::SI_SPILL_S32_SAVE: |
1552 | return spillSGPR(MI, FI, RS, LIS, true); |
1553 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
1554 | case AMDGPU::SI_SPILL_S512_RESTORE: |
1555 | case AMDGPU::SI_SPILL_S256_RESTORE: |
1556 | case AMDGPU::SI_SPILL_S224_RESTORE: |
1557 | case AMDGPU::SI_SPILL_S192_RESTORE: |
1558 | case AMDGPU::SI_SPILL_S160_RESTORE: |
1559 | case AMDGPU::SI_SPILL_S128_RESTORE: |
1560 | case AMDGPU::SI_SPILL_S96_RESTORE: |
1561 | case AMDGPU::SI_SPILL_S64_RESTORE: |
1562 | case AMDGPU::SI_SPILL_S32_RESTORE: |
1563 | return restoreSGPR(MI, FI, RS, LIS, true); |
1564 | default: |
1565 | llvm_unreachable("not an SGPR spill instruction")::llvm::llvm_unreachable_internal("not an SGPR spill instruction" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1565); |
1566 | } |
1567 | } |
1568 | |
1569 | void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI, |
1570 | int SPAdj, unsigned FIOperandNum, |
1571 | RegScavenger *RS) const { |
1572 | MachineFunction *MF = MI->getParent()->getParent(); |
1573 | MachineBasicBlock *MBB = MI->getParent(); |
1574 | SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); |
1575 | MachineFrameInfo &FrameInfo = MF->getFrameInfo(); |
1576 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1577 | DebugLoc DL = MI->getDebugLoc(); |
1578 | |
1579 | assert(SPAdj == 0 && "unhandled SP adjustment in call sequence?")(static_cast <bool> (SPAdj == 0 && "unhandled SP adjustment in call sequence?" ) ? void (0) : __assert_fail ("SPAdj == 0 && \"unhandled SP adjustment in call sequence?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1579, __extension__ __PRETTY_FUNCTION__)); |
1580 | |
1581 | MachineOperand &FIOp = MI->getOperand(FIOperandNum); |
1582 | int Index = MI->getOperand(FIOperandNum).getIndex(); |
1583 | |
1584 | Register FrameReg = FrameInfo.isFixedObjectIndex(Index) && hasBasePointer(*MF) |
1585 | ? getBaseRegister() |
1586 | : getFrameRegister(*MF); |
1587 | |
1588 | switch (MI->getOpcode()) { |
1589 | // SGPR register spill |
1590 | case AMDGPU::SI_SPILL_S1024_SAVE: |
1591 | case AMDGPU::SI_SPILL_S512_SAVE: |
1592 | case AMDGPU::SI_SPILL_S256_SAVE: |
1593 | case AMDGPU::SI_SPILL_S224_SAVE: |
1594 | case AMDGPU::SI_SPILL_S192_SAVE: |
1595 | case AMDGPU::SI_SPILL_S160_SAVE: |
1596 | case AMDGPU::SI_SPILL_S128_SAVE: |
1597 | case AMDGPU::SI_SPILL_S96_SAVE: |
1598 | case AMDGPU::SI_SPILL_S64_SAVE: |
1599 | case AMDGPU::SI_SPILL_S32_SAVE: { |
1600 | spillSGPR(MI, Index, RS); |
1601 | break; |
1602 | } |
1603 | |
1604 | // SGPR register restore |
1605 | case AMDGPU::SI_SPILL_S1024_RESTORE: |
1606 | case AMDGPU::SI_SPILL_S512_RESTORE: |
1607 | case AMDGPU::SI_SPILL_S256_RESTORE: |
1608 | case AMDGPU::SI_SPILL_S224_RESTORE: |
1609 | case AMDGPU::SI_SPILL_S192_RESTORE: |
1610 | case AMDGPU::SI_SPILL_S160_RESTORE: |
1611 | case AMDGPU::SI_SPILL_S128_RESTORE: |
1612 | case AMDGPU::SI_SPILL_S96_RESTORE: |
1613 | case AMDGPU::SI_SPILL_S64_RESTORE: |
1614 | case AMDGPU::SI_SPILL_S32_RESTORE: { |
1615 | restoreSGPR(MI, Index, RS); |
1616 | break; |
1617 | } |
1618 | |
1619 | // VGPR register spill |
1620 | case AMDGPU::SI_SPILL_V1024_SAVE: |
1621 | case AMDGPU::SI_SPILL_V512_SAVE: |
1622 | case AMDGPU::SI_SPILL_V256_SAVE: |
1623 | case AMDGPU::SI_SPILL_V224_SAVE: |
1624 | case AMDGPU::SI_SPILL_V192_SAVE: |
1625 | case AMDGPU::SI_SPILL_V160_SAVE: |
1626 | case AMDGPU::SI_SPILL_V128_SAVE: |
1627 | case AMDGPU::SI_SPILL_V96_SAVE: |
1628 | case AMDGPU::SI_SPILL_V64_SAVE: |
1629 | case AMDGPU::SI_SPILL_V32_SAVE: |
1630 | case AMDGPU::SI_SPILL_A1024_SAVE: |
1631 | case AMDGPU::SI_SPILL_A512_SAVE: |
1632 | case AMDGPU::SI_SPILL_A256_SAVE: |
1633 | case AMDGPU::SI_SPILL_A224_SAVE: |
1634 | case AMDGPU::SI_SPILL_A192_SAVE: |
1635 | case AMDGPU::SI_SPILL_A160_SAVE: |
1636 | case AMDGPU::SI_SPILL_A128_SAVE: |
1637 | case AMDGPU::SI_SPILL_A96_SAVE: |
1638 | case AMDGPU::SI_SPILL_A64_SAVE: |
1639 | case AMDGPU::SI_SPILL_A32_SAVE: { |
1640 | const MachineOperand *VData = TII->getNamedOperand(*MI, |
1641 | AMDGPU::OpName::vdata); |
1642 | assert(TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() ==(static_cast <bool> (TII->getNamedOperand(*MI, AMDGPU ::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg ()) ? void (0) : __assert_fail ("TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1643, __extension__ __PRETTY_FUNCTION__)) |
1643 | MFI->getStackPtrOffsetReg())(static_cast <bool> (TII->getNamedOperand(*MI, AMDGPU ::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg ()) ? void (0) : __assert_fail ("TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1643, __extension__ __PRETTY_FUNCTION__)); |
1644 | |
1645 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_STORE_DWORD_SADDR |
1646 | : AMDGPU::BUFFER_STORE_DWORD_OFFSET; |
1647 | auto *MBB = MI->getParent(); |
1648 | buildSpillLoadStore( |
1649 | *MBB, MI, Opc, Index, VData->getReg(), VData->isKill(), FrameReg, |
1650 | TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm(), |
1651 | *MI->memoperands_begin(), RS); |
1652 | MFI->addToSpilledVGPRs(getNumSubRegsForSpillOp(MI->getOpcode())); |
1653 | MI->eraseFromParent(); |
1654 | break; |
1655 | } |
1656 | case AMDGPU::SI_SPILL_V32_RESTORE: |
1657 | case AMDGPU::SI_SPILL_V64_RESTORE: |
1658 | case AMDGPU::SI_SPILL_V96_RESTORE: |
1659 | case AMDGPU::SI_SPILL_V128_RESTORE: |
1660 | case AMDGPU::SI_SPILL_V160_RESTORE: |
1661 | case AMDGPU::SI_SPILL_V192_RESTORE: |
1662 | case AMDGPU::SI_SPILL_V224_RESTORE: |
1663 | case AMDGPU::SI_SPILL_V256_RESTORE: |
1664 | case AMDGPU::SI_SPILL_V512_RESTORE: |
1665 | case AMDGPU::SI_SPILL_V1024_RESTORE: |
1666 | case AMDGPU::SI_SPILL_A32_RESTORE: |
1667 | case AMDGPU::SI_SPILL_A64_RESTORE: |
1668 | case AMDGPU::SI_SPILL_A96_RESTORE: |
1669 | case AMDGPU::SI_SPILL_A128_RESTORE: |
1670 | case AMDGPU::SI_SPILL_A160_RESTORE: |
1671 | case AMDGPU::SI_SPILL_A192_RESTORE: |
1672 | case AMDGPU::SI_SPILL_A224_RESTORE: |
1673 | case AMDGPU::SI_SPILL_A256_RESTORE: |
1674 | case AMDGPU::SI_SPILL_A512_RESTORE: |
1675 | case AMDGPU::SI_SPILL_A1024_RESTORE: { |
1676 | const MachineOperand *VData = TII->getNamedOperand(*MI, |
1677 | AMDGPU::OpName::vdata); |
1678 | assert(TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() ==(static_cast <bool> (TII->getNamedOperand(*MI, AMDGPU ::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg ()) ? void (0) : __assert_fail ("TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1679, __extension__ __PRETTY_FUNCTION__)) |
1679 | MFI->getStackPtrOffsetReg())(static_cast <bool> (TII->getNamedOperand(*MI, AMDGPU ::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg ()) ? void (0) : __assert_fail ("TII->getNamedOperand(*MI, AMDGPU::OpName::soffset)->getReg() == MFI->getStackPtrOffsetReg()" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1679, __extension__ __PRETTY_FUNCTION__)); |
1680 | |
1681 | unsigned Opc = ST.enableFlatScratch() ? AMDGPU::SCRATCH_LOAD_DWORD_SADDR |
1682 | : AMDGPU::BUFFER_LOAD_DWORD_OFFSET; |
1683 | auto *MBB = MI->getParent(); |
1684 | buildSpillLoadStore( |
1685 | *MBB, MI, Opc, Index, VData->getReg(), VData->isKill(), FrameReg, |
1686 | TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm(), |
1687 | *MI->memoperands_begin(), RS); |
1688 | MI->eraseFromParent(); |
1689 | break; |
1690 | } |
1691 | |
1692 | default: { |
1693 | // Other access to frame index |
1694 | const DebugLoc &DL = MI->getDebugLoc(); |
1695 | |
1696 | int64_t Offset = FrameInfo.getObjectOffset(Index); |
1697 | if (ST.enableFlatScratch()) { |
1698 | if (TII->isFLATScratch(*MI)) { |
1699 | assert((int16_t)FIOperandNum ==(static_cast <bool> ((int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::saddr)) ? void (0) : __assert_fail ("(int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1701, __extension__ __PRETTY_FUNCTION__)) |
1700 | AMDGPU::getNamedOperandIdx(MI->getOpcode(),(static_cast <bool> ((int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::saddr)) ? void (0) : __assert_fail ("(int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1701, __extension__ __PRETTY_FUNCTION__)) |
1701 | AMDGPU::OpName::saddr))(static_cast <bool> ((int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx (MI->getOpcode(), AMDGPU::OpName::saddr)) ? void (0) : __assert_fail ("(int16_t)FIOperandNum == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::saddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1701, __extension__ __PRETTY_FUNCTION__)); |
1702 | |
1703 | // The offset is always swizzled, just replace it |
1704 | if (FrameReg) |
1705 | FIOp.ChangeToRegister(FrameReg, false); |
1706 | |
1707 | if (!Offset) |
1708 | return; |
1709 | |
1710 | MachineOperand *OffsetOp = |
1711 | TII->getNamedOperand(*MI, AMDGPU::OpName::offset); |
1712 | int64_t NewOffset = Offset + OffsetOp->getImm(); |
1713 | if (TII->isLegalFLATOffset(NewOffset, AMDGPUAS::PRIVATE_ADDRESS, |
1714 | SIInstrFlags::FlatScratch)) { |
1715 | OffsetOp->setImm(NewOffset); |
1716 | if (FrameReg) |
1717 | return; |
1718 | Offset = 0; |
1719 | } |
1720 | |
1721 | assert(!TII->getNamedOperand(*MI, AMDGPU::OpName::vaddr) &&(static_cast <bool> (!TII->getNamedOperand(*MI, AMDGPU ::OpName::vaddr) && "Unexpected vaddr for flat scratch with a FI operand" ) ? void (0) : __assert_fail ("!TII->getNamedOperand(*MI, AMDGPU::OpName::vaddr) && \"Unexpected vaddr for flat scratch with a FI operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1722, __extension__ __PRETTY_FUNCTION__)) |
1722 | "Unexpected vaddr for flat scratch with a FI operand")(static_cast <bool> (!TII->getNamedOperand(*MI, AMDGPU ::OpName::vaddr) && "Unexpected vaddr for flat scratch with a FI operand" ) ? void (0) : __assert_fail ("!TII->getNamedOperand(*MI, AMDGPU::OpName::vaddr) && \"Unexpected vaddr for flat scratch with a FI operand\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1722, __extension__ __PRETTY_FUNCTION__)); |
1723 | |
1724 | // On GFX10 we have ST mode to use no registers for an address. |
1725 | // Otherwise we need to materialize 0 into an SGPR. |
1726 | if (!Offset && ST.hasFlatScratchSTMode()) { |
1727 | unsigned Opc = MI->getOpcode(); |
1728 | unsigned NewOpc = AMDGPU::getFlatScratchInstSTfromSS(Opc); |
1729 | MI->RemoveOperand( |
1730 | AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::saddr)); |
1731 | MI->setDesc(TII->get(NewOpc)); |
1732 | return; |
1733 | } |
1734 | } |
1735 | |
1736 | if (!FrameReg) { |
1737 | FIOp.ChangeToImmediate(Offset); |
1738 | if (TII->isImmOperandLegal(*MI, FIOperandNum, FIOp)) |
1739 | return; |
1740 | } |
1741 | |
1742 | // We need to use register here. Check if we can use an SGPR or need |
1743 | // a VGPR. |
1744 | FIOp.ChangeToRegister(AMDGPU::M0, false); |
1745 | bool UseSGPR = TII->isOperandLegal(*MI, FIOperandNum, &FIOp); |
1746 | |
1747 | if (!Offset && FrameReg && UseSGPR) { |
1748 | FIOp.setReg(FrameReg); |
1749 | return; |
1750 | } |
1751 | |
1752 | const TargetRegisterClass *RC = UseSGPR ? &AMDGPU::SReg_32_XM0RegClass |
1753 | : &AMDGPU::VGPR_32RegClass; |
1754 | |
1755 | Register TmpReg = RS->scavengeRegister(RC, MI, 0, !UseSGPR); |
1756 | FIOp.setReg(TmpReg); |
1757 | FIOp.setIsKill(true); |
1758 | |
1759 | if ((!FrameReg || !Offset) && TmpReg) { |
1760 | unsigned Opc = UseSGPR ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32; |
1761 | auto MIB = BuildMI(*MBB, MI, DL, TII->get(Opc), TmpReg); |
1762 | if (FrameReg) |
1763 | MIB.addReg(FrameReg); |
1764 | else |
1765 | MIB.addImm(Offset); |
1766 | |
1767 | return; |
1768 | } |
1769 | |
1770 | Register TmpSReg = |
1771 | UseSGPR ? TmpReg |
1772 | : RS->scavengeRegister(&AMDGPU::SReg_32_XM0RegClass, MI, 0, |
1773 | !UseSGPR); |
1774 | |
1775 | // TODO: for flat scratch another attempt can be made with a VGPR index |
1776 | // if no SGPRs can be scavenged. |
1777 | if ((!TmpSReg && !FrameReg) || (!TmpReg && !UseSGPR)) |
1778 | report_fatal_error("Cannot scavenge register in FI elimination!"); |
1779 | |
1780 | if (!TmpSReg) { |
1781 | // Use frame register and restore it after. |
1782 | TmpSReg = FrameReg; |
1783 | FIOp.setReg(FrameReg); |
1784 | FIOp.setIsKill(false); |
1785 | } |
1786 | |
1787 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_ADD_I32), TmpSReg) |
1788 | .addReg(FrameReg) |
1789 | .addImm(Offset); |
1790 | |
1791 | if (!UseSGPR) |
1792 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_MOV_B32_e32), TmpReg) |
1793 | .addReg(TmpSReg, RegState::Kill); |
1794 | |
1795 | if (TmpSReg == FrameReg) { |
1796 | // Undo frame register modification. |
1797 | BuildMI(*MBB, std::next(MI), DL, TII->get(AMDGPU::S_ADD_I32), |
1798 | FrameReg) |
1799 | .addReg(FrameReg) |
1800 | .addImm(-Offset); |
1801 | } |
1802 | |
1803 | return; |
1804 | } |
1805 | |
1806 | bool IsMUBUF = TII->isMUBUF(*MI); |
1807 | |
1808 | if (!IsMUBUF && !MFI->isEntryFunction()) { |
1809 | // Convert to a swizzled stack address by scaling by the wave size. |
1810 | // |
1811 | // In an entry function/kernel the offset is already swizzled. |
1812 | |
1813 | bool IsCopy = MI->getOpcode() == AMDGPU::V_MOV_B32_e32; |
1814 | Register ResultReg = |
1815 | IsCopy ? MI->getOperand(0).getReg() |
1816 | : RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0); |
1817 | |
1818 | int64_t Offset = FrameInfo.getObjectOffset(Index); |
1819 | if (Offset == 0) { |
1820 | // XXX - This never happens because of emergency scavenging slot at 0? |
1821 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_LSHRREV_B32_e64), ResultReg) |
1822 | .addImm(ST.getWavefrontSizeLog2()) |
1823 | .addReg(FrameReg); |
1824 | } else { |
1825 | if (auto MIB = TII->getAddNoCarry(*MBB, MI, DL, ResultReg, *RS)) { |
1826 | // Reuse ResultReg in intermediate step. |
1827 | Register ScaledReg = ResultReg; |
1828 | |
1829 | BuildMI(*MBB, *MIB, DL, TII->get(AMDGPU::V_LSHRREV_B32_e64), |
1830 | ScaledReg) |
1831 | .addImm(ST.getWavefrontSizeLog2()) |
1832 | .addReg(FrameReg); |
1833 | |
1834 | const bool IsVOP2 = MIB->getOpcode() == AMDGPU::V_ADD_U32_e32; |
1835 | |
1836 | // TODO: Fold if use instruction is another add of a constant. |
1837 | if (IsVOP2 || AMDGPU::isInlinableLiteral32(Offset, ST.hasInv2PiInlineImm())) { |
1838 | // FIXME: This can fail |
1839 | MIB.addImm(Offset); |
1840 | MIB.addReg(ScaledReg, RegState::Kill); |
1841 | if (!IsVOP2) |
1842 | MIB.addImm(0); // clamp bit |
1843 | } else { |
1844 | assert(MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 &&(static_cast <bool> (MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 && "Need to reuse carry out register") ? void (0) : __assert_fail ("MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 && \"Need to reuse carry out register\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1845, __extension__ __PRETTY_FUNCTION__)) |
1845 | "Need to reuse carry out register")(static_cast <bool> (MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 && "Need to reuse carry out register") ? void (0) : __assert_fail ("MIB->getOpcode() == AMDGPU::V_ADD_CO_U32_e64 && \"Need to reuse carry out register\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1845, __extension__ __PRETTY_FUNCTION__)); |
1846 | |
1847 | // Use scavenged unused carry out as offset register. |
1848 | Register ConstOffsetReg; |
1849 | if (!isWave32) |
1850 | ConstOffsetReg = getSubReg(MIB.getReg(1), AMDGPU::sub0); |
1851 | else |
1852 | ConstOffsetReg = MIB.getReg(1); |
1853 | |
1854 | BuildMI(*MBB, *MIB, DL, TII->get(AMDGPU::S_MOV_B32), ConstOffsetReg) |
1855 | .addImm(Offset); |
1856 | MIB.addReg(ConstOffsetReg, RegState::Kill); |
1857 | MIB.addReg(ScaledReg, RegState::Kill); |
1858 | MIB.addImm(0); // clamp bit |
1859 | } |
1860 | } else { |
1861 | // We have to produce a carry out, and there isn't a free SGPR pair |
1862 | // for it. We can keep the whole computation on the SALU to avoid |
1863 | // clobbering an additional register at the cost of an extra mov. |
1864 | |
1865 | // We may have 1 free scratch SGPR even though a carry out is |
1866 | // unavailable. Only one additional mov is needed. |
1867 | Register TmpScaledReg = |
1868 | RS->scavengeRegister(&AMDGPU::SReg_32_XM0RegClass, MI, 0, false); |
1869 | Register ScaledReg = TmpScaledReg.isValid() ? TmpScaledReg : FrameReg; |
1870 | |
1871 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_LSHR_B32), ScaledReg) |
1872 | .addReg(FrameReg) |
1873 | .addImm(ST.getWavefrontSizeLog2()); |
1874 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_ADD_I32), ScaledReg) |
1875 | .addReg(ScaledReg, RegState::Kill) |
1876 | .addImm(Offset); |
1877 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::COPY), ResultReg) |
1878 | .addReg(ScaledReg, RegState::Kill); |
1879 | |
1880 | // If there were truly no free SGPRs, we need to undo everything. |
1881 | if (!TmpScaledReg.isValid()) { |
1882 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_ADD_I32), ScaledReg) |
1883 | .addReg(ScaledReg, RegState::Kill) |
1884 | .addImm(-Offset); |
1885 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_LSHL_B32), ScaledReg) |
1886 | .addReg(FrameReg) |
1887 | .addImm(ST.getWavefrontSizeLog2()); |
1888 | } |
1889 | } |
1890 | } |
1891 | |
1892 | // Don't introduce an extra copy if we're just materializing in a mov. |
1893 | if (IsCopy) |
1894 | MI->eraseFromParent(); |
1895 | else |
1896 | FIOp.ChangeToRegister(ResultReg, false, false, true); |
1897 | return; |
1898 | } |
1899 | |
1900 | if (IsMUBUF) { |
1901 | // Disable offen so we don't need a 0 vgpr base. |
1902 | assert(static_cast<int>(FIOperandNum) ==(static_cast <bool> (static_cast<int>(FIOperandNum ) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName ::vaddr)) ? void (0) : __assert_fail ("static_cast<int>(FIOperandNum) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1904, __extension__ __PRETTY_FUNCTION__)) |
1903 | AMDGPU::getNamedOperandIdx(MI->getOpcode(),(static_cast <bool> (static_cast<int>(FIOperandNum ) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName ::vaddr)) ? void (0) : __assert_fail ("static_cast<int>(FIOperandNum) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1904, __extension__ __PRETTY_FUNCTION__)) |
1904 | AMDGPU::OpName::vaddr))(static_cast <bool> (static_cast<int>(FIOperandNum ) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName ::vaddr)) ? void (0) : __assert_fail ("static_cast<int>(FIOperandNum) == AMDGPU::getNamedOperandIdx(MI->getOpcode(), AMDGPU::OpName::vaddr)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1904, __extension__ __PRETTY_FUNCTION__)); |
1905 | |
1906 | auto &SOffset = *TII->getNamedOperand(*MI, AMDGPU::OpName::soffset); |
1907 | assert((SOffset.isImm() && SOffset.getImm() == 0))(static_cast <bool> ((SOffset.isImm() && SOffset .getImm() == 0)) ? void (0) : __assert_fail ("(SOffset.isImm() && SOffset.getImm() == 0)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 1907, __extension__ __PRETTY_FUNCTION__)); |
1908 | |
1909 | if (FrameReg != AMDGPU::NoRegister) |
1910 | SOffset.ChangeToRegister(FrameReg, false); |
1911 | |
1912 | int64_t Offset = FrameInfo.getObjectOffset(Index); |
1913 | int64_t OldImm |
1914 | = TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm(); |
1915 | int64_t NewOffset = OldImm + Offset; |
1916 | |
1917 | if (SIInstrInfo::isLegalMUBUFImmOffset(NewOffset) && |
1918 | buildMUBUFOffsetLoadStore(ST, FrameInfo, MI, Index, NewOffset)) { |
1919 | MI->eraseFromParent(); |
1920 | return; |
1921 | } |
1922 | } |
1923 | |
1924 | // If the offset is simply too big, don't convert to a scratch wave offset |
1925 | // relative index. |
1926 | |
1927 | FIOp.ChangeToImmediate(Offset); |
1928 | if (!TII->isImmOperandLegal(*MI, FIOperandNum, FIOp)) { |
1929 | Register TmpReg = RS->scavengeRegister(&AMDGPU::VGPR_32RegClass, MI, 0); |
1930 | BuildMI(*MBB, MI, DL, TII->get(AMDGPU::V_MOV_B32_e32), TmpReg) |
1931 | .addImm(Offset); |
1932 | FIOp.ChangeToRegister(TmpReg, false, false, true); |
1933 | } |
1934 | } |
1935 | } |
1936 | } |
1937 | |
1938 | StringRef SIRegisterInfo::getRegAsmName(MCRegister Reg) const { |
1939 | return AMDGPUInstPrinter::getRegisterName(Reg); |
1940 | } |
1941 | |
1942 | static const TargetRegisterClass * |
1943 | getAnyVGPRClassForBitWidth(unsigned BitWidth) { |
1944 | if (BitWidth <= 64) |
1945 | return &AMDGPU::VReg_64RegClass; |
1946 | if (BitWidth <= 96) |
1947 | return &AMDGPU::VReg_96RegClass; |
1948 | if (BitWidth <= 128) |
1949 | return &AMDGPU::VReg_128RegClass; |
1950 | if (BitWidth <= 160) |
1951 | return &AMDGPU::VReg_160RegClass; |
1952 | if (BitWidth <= 192) |
1953 | return &AMDGPU::VReg_192RegClass; |
1954 | if (BitWidth <= 224) |
1955 | return &AMDGPU::VReg_224RegClass; |
1956 | if (BitWidth <= 256) |
1957 | return &AMDGPU::VReg_256RegClass; |
1958 | if (BitWidth <= 512) |
1959 | return &AMDGPU::VReg_512RegClass; |
1960 | if (BitWidth <= 1024) |
1961 | return &AMDGPU::VReg_1024RegClass; |
1962 | |
1963 | return nullptr; |
1964 | } |
1965 | |
1966 | static const TargetRegisterClass * |
1967 | getAlignedVGPRClassForBitWidth(unsigned BitWidth) { |
1968 | if (BitWidth <= 64) |
1969 | return &AMDGPU::VReg_64_Align2RegClass; |
1970 | if (BitWidth <= 96) |
1971 | return &AMDGPU::VReg_96_Align2RegClass; |
1972 | if (BitWidth <= 128) |
1973 | return &AMDGPU::VReg_128_Align2RegClass; |
1974 | if (BitWidth <= 160) |
1975 | return &AMDGPU::VReg_160_Align2RegClass; |
1976 | if (BitWidth <= 192) |
1977 | return &AMDGPU::VReg_192_Align2RegClass; |
1978 | if (BitWidth <= 224) |
1979 | return &AMDGPU::VReg_224_Align2RegClass; |
1980 | if (BitWidth <= 256) |
1981 | return &AMDGPU::VReg_256_Align2RegClass; |
1982 | if (BitWidth <= 512) |
1983 | return &AMDGPU::VReg_512_Align2RegClass; |
1984 | if (BitWidth <= 1024) |
1985 | return &AMDGPU::VReg_1024_Align2RegClass; |
1986 | |
1987 | return nullptr; |
1988 | } |
1989 | |
1990 | const TargetRegisterClass * |
1991 | SIRegisterInfo::getVGPRClassForBitWidth(unsigned BitWidth) const { |
1992 | if (BitWidth == 1) |
1993 | return &AMDGPU::VReg_1RegClass; |
1994 | if (BitWidth <= 16) |
1995 | return &AMDGPU::VGPR_LO16RegClass; |
1996 | if (BitWidth <= 32) |
1997 | return &AMDGPU::VGPR_32RegClass; |
1998 | return ST.needsAlignedVGPRs() ? getAlignedVGPRClassForBitWidth(BitWidth) |
1999 | : getAnyVGPRClassForBitWidth(BitWidth); |
2000 | } |
2001 | |
2002 | static const TargetRegisterClass * |
2003 | getAnyAGPRClassForBitWidth(unsigned BitWidth) { |
2004 | if (BitWidth <= 64) |
2005 | return &AMDGPU::AReg_64RegClass; |
2006 | if (BitWidth <= 96) |
2007 | return &AMDGPU::AReg_96RegClass; |
2008 | if (BitWidth <= 128) |
2009 | return &AMDGPU::AReg_128RegClass; |
2010 | if (BitWidth <= 160) |
2011 | return &AMDGPU::AReg_160RegClass; |
2012 | if (BitWidth <= 192) |
2013 | return &AMDGPU::AReg_192RegClass; |
2014 | if (BitWidth <= 224) |
2015 | return &AMDGPU::AReg_224RegClass; |
2016 | if (BitWidth <= 256) |
2017 | return &AMDGPU::AReg_256RegClass; |
2018 | if (BitWidth <= 512) |
2019 | return &AMDGPU::AReg_512RegClass; |
2020 | if (BitWidth <= 1024) |
2021 | return &AMDGPU::AReg_1024RegClass; |
2022 | |
2023 | return nullptr; |
2024 | } |
2025 | |
2026 | static const TargetRegisterClass * |
2027 | getAlignedAGPRClassForBitWidth(unsigned BitWidth) { |
2028 | if (BitWidth <= 64) |
2029 | return &AMDGPU::AReg_64_Align2RegClass; |
2030 | if (BitWidth <= 96) |
2031 | return &AMDGPU::AReg_96_Align2RegClass; |
2032 | if (BitWidth <= 128) |
2033 | return &AMDGPU::AReg_128_Align2RegClass; |
2034 | if (BitWidth <= 160) |
2035 | return &AMDGPU::AReg_160_Align2RegClass; |
2036 | if (BitWidth <= 192) |
2037 | return &AMDGPU::AReg_192_Align2RegClass; |
2038 | if (BitWidth <= 224) |
2039 | return &AMDGPU::AReg_224_Align2RegClass; |
2040 | if (BitWidth <= 256) |
2041 | return &AMDGPU::AReg_256_Align2RegClass; |
2042 | if (BitWidth <= 512) |
2043 | return &AMDGPU::AReg_512_Align2RegClass; |
2044 | if (BitWidth <= 1024) |
2045 | return &AMDGPU::AReg_1024_Align2RegClass; |
2046 | |
2047 | return nullptr; |
2048 | } |
2049 | |
2050 | const TargetRegisterClass * |
2051 | SIRegisterInfo::getAGPRClassForBitWidth(unsigned BitWidth) const { |
2052 | if (BitWidth <= 16) |
2053 | return &AMDGPU::AGPR_LO16RegClass; |
2054 | if (BitWidth <= 32) |
2055 | return &AMDGPU::AGPR_32RegClass; |
2056 | return ST.needsAlignedVGPRs() ? getAlignedAGPRClassForBitWidth(BitWidth) |
2057 | : getAnyAGPRClassForBitWidth(BitWidth); |
2058 | } |
2059 | |
2060 | const TargetRegisterClass * |
2061 | SIRegisterInfo::getSGPRClassForBitWidth(unsigned BitWidth) { |
2062 | if (BitWidth <= 16) |
2063 | return &AMDGPU::SGPR_LO16RegClass; |
2064 | if (BitWidth <= 32) |
2065 | return &AMDGPU::SReg_32RegClass; |
2066 | if (BitWidth <= 64) |
2067 | return &AMDGPU::SReg_64RegClass; |
2068 | if (BitWidth <= 96) |
2069 | return &AMDGPU::SGPR_96RegClass; |
2070 | if (BitWidth <= 128) |
2071 | return &AMDGPU::SGPR_128RegClass; |
2072 | if (BitWidth <= 160) |
2073 | return &AMDGPU::SGPR_160RegClass; |
2074 | if (BitWidth <= 192) |
2075 | return &AMDGPU::SGPR_192RegClass; |
2076 | if (BitWidth <= 224) |
2077 | return &AMDGPU::SGPR_224RegClass; |
2078 | if (BitWidth <= 256) |
2079 | return &AMDGPU::SGPR_256RegClass; |
2080 | if (BitWidth <= 512) |
2081 | return &AMDGPU::SGPR_512RegClass; |
2082 | if (BitWidth <= 1024) |
2083 | return &AMDGPU::SGPR_1024RegClass; |
2084 | |
2085 | return nullptr; |
2086 | } |
2087 | |
2088 | // FIXME: This is very slow. It might be worth creating a map from physreg to |
2089 | // register class. |
2090 | const TargetRegisterClass * |
2091 | SIRegisterInfo::getPhysRegClass(MCRegister Reg) const { |
2092 | static const TargetRegisterClass *const BaseClasses[] = { |
2093 | &AMDGPU::VGPR_LO16RegClass, |
2094 | &AMDGPU::VGPR_HI16RegClass, |
2095 | &AMDGPU::SReg_LO16RegClass, |
2096 | &AMDGPU::AGPR_LO16RegClass, |
2097 | &AMDGPU::VGPR_32RegClass, |
2098 | &AMDGPU::SReg_32RegClass, |
2099 | &AMDGPU::AGPR_32RegClass, |
2100 | &AMDGPU::AGPR_32RegClass, |
2101 | &AMDGPU::VReg_64_Align2RegClass, |
2102 | &AMDGPU::VReg_64RegClass, |
2103 | &AMDGPU::SReg_64RegClass, |
2104 | &AMDGPU::AReg_64_Align2RegClass, |
2105 | &AMDGPU::AReg_64RegClass, |
2106 | &AMDGPU::VReg_96_Align2RegClass, |
2107 | &AMDGPU::VReg_96RegClass, |
2108 | &AMDGPU::SReg_96RegClass, |
2109 | &AMDGPU::AReg_96_Align2RegClass, |
2110 | &AMDGPU::AReg_96RegClass, |
2111 | &AMDGPU::VReg_128_Align2RegClass, |
2112 | &AMDGPU::VReg_128RegClass, |
2113 | &AMDGPU::SReg_128RegClass, |
2114 | &AMDGPU::AReg_128_Align2RegClass, |
2115 | &AMDGPU::AReg_128RegClass, |
2116 | &AMDGPU::VReg_160_Align2RegClass, |
2117 | &AMDGPU::VReg_160RegClass, |
2118 | &AMDGPU::SReg_160RegClass, |
2119 | &AMDGPU::AReg_160_Align2RegClass, |
2120 | &AMDGPU::AReg_160RegClass, |
2121 | &AMDGPU::VReg_192_Align2RegClass, |
2122 | &AMDGPU::VReg_192RegClass, |
2123 | &AMDGPU::SReg_192RegClass, |
2124 | &AMDGPU::AReg_192_Align2RegClass, |
2125 | &AMDGPU::AReg_192RegClass, |
2126 | &AMDGPU::VReg_224_Align2RegClass, |
2127 | &AMDGPU::VReg_224RegClass, |
2128 | &AMDGPU::SReg_224RegClass, |
2129 | &AMDGPU::AReg_224_Align2RegClass, |
2130 | &AMDGPU::AReg_224RegClass, |
2131 | &AMDGPU::VReg_256_Align2RegClass, |
2132 | &AMDGPU::VReg_256RegClass, |
2133 | &AMDGPU::SReg_256RegClass, |
2134 | &AMDGPU::AReg_256_Align2RegClass, |
2135 | &AMDGPU::AReg_256RegClass, |
2136 | &AMDGPU::VReg_512_Align2RegClass, |
2137 | &AMDGPU::VReg_512RegClass, |
2138 | &AMDGPU::SReg_512RegClass, |
2139 | &AMDGPU::AReg_512_Align2RegClass, |
2140 | &AMDGPU::AReg_512RegClass, |
2141 | &AMDGPU::SReg_1024RegClass, |
2142 | &AMDGPU::VReg_1024_Align2RegClass, |
2143 | &AMDGPU::VReg_1024RegClass, |
2144 | &AMDGPU::AReg_1024_Align2RegClass, |
2145 | &AMDGPU::AReg_1024RegClass, |
2146 | &AMDGPU::SCC_CLASSRegClass, |
2147 | &AMDGPU::Pseudo_SReg_32RegClass, |
2148 | &AMDGPU::Pseudo_SReg_128RegClass, |
2149 | }; |
2150 | |
2151 | for (const TargetRegisterClass *BaseClass : BaseClasses) { |
2152 | if (BaseClass->contains(Reg)) { |
2153 | return BaseClass; |
2154 | } |
2155 | } |
2156 | return nullptr; |
2157 | } |
2158 | |
2159 | bool SIRegisterInfo::isSGPRReg(const MachineRegisterInfo &MRI, |
2160 | Register Reg) const { |
2161 | const TargetRegisterClass *RC; |
2162 | if (Reg.isVirtual()) |
2163 | RC = MRI.getRegClass(Reg); |
2164 | else |
2165 | RC = getPhysRegClass(Reg); |
2166 | return isSGPRClass(RC); |
2167 | } |
2168 | |
2169 | // TODO: It might be helpful to have some target specific flags in |
2170 | // TargetRegisterClass to mark which classes are VGPRs to make this trivial. |
2171 | bool SIRegisterInfo::hasVGPRs(const TargetRegisterClass *RC) const { |
2172 | unsigned Size = getRegSizeInBits(*RC); |
2173 | if (Size == 16) { |
2174 | return getCommonSubClass(&AMDGPU::VGPR_LO16RegClass, RC) != nullptr || |
2175 | getCommonSubClass(&AMDGPU::VGPR_HI16RegClass, RC) != nullptr; |
2176 | } |
2177 | const TargetRegisterClass *VRC = getVGPRClassForBitWidth(Size); |
2178 | if (!VRC) { |
2179 | assert(Size < 32 && "Invalid register class size")(static_cast <bool> (Size < 32 && "Invalid register class size" ) ? void (0) : __assert_fail ("Size < 32 && \"Invalid register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2179, __extension__ __PRETTY_FUNCTION__)); |
2180 | return false; |
2181 | } |
2182 | return getCommonSubClass(VRC, RC) != nullptr; |
2183 | } |
2184 | |
2185 | bool SIRegisterInfo::hasAGPRs(const TargetRegisterClass *RC) const { |
2186 | unsigned Size = getRegSizeInBits(*RC); |
2187 | if (Size < 16) |
2188 | return false; |
2189 | const TargetRegisterClass *ARC = getAGPRClassForBitWidth(Size); |
2190 | if (!ARC) { |
2191 | assert(getVGPRClassForBitWidth(Size) && "Invalid register class size")(static_cast <bool> (getVGPRClassForBitWidth(Size) && "Invalid register class size") ? void (0) : __assert_fail ("getVGPRClassForBitWidth(Size) && \"Invalid register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2191, __extension__ __PRETTY_FUNCTION__)); |
2192 | return false; |
2193 | } |
2194 | return getCommonSubClass(ARC, RC) != nullptr; |
2195 | } |
2196 | |
2197 | const TargetRegisterClass * |
2198 | SIRegisterInfo::getEquivalentVGPRClass(const TargetRegisterClass *SRC) const { |
2199 | unsigned Size = getRegSizeInBits(*SRC); |
2200 | const TargetRegisterClass *VRC = getVGPRClassForBitWidth(Size); |
2201 | assert(VRC && "Invalid register class size")(static_cast <bool> (VRC && "Invalid register class size" ) ? void (0) : __assert_fail ("VRC && \"Invalid register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2201, __extension__ __PRETTY_FUNCTION__)); |
2202 | return VRC; |
2203 | } |
2204 | |
2205 | const TargetRegisterClass * |
2206 | SIRegisterInfo::getEquivalentAGPRClass(const TargetRegisterClass *SRC) const { |
2207 | unsigned Size = getRegSizeInBits(*SRC); |
2208 | const TargetRegisterClass *ARC = getAGPRClassForBitWidth(Size); |
2209 | assert(ARC && "Invalid register class size")(static_cast <bool> (ARC && "Invalid register class size" ) ? void (0) : __assert_fail ("ARC && \"Invalid register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2209, __extension__ __PRETTY_FUNCTION__)); |
2210 | return ARC; |
2211 | } |
2212 | |
2213 | const TargetRegisterClass * |
2214 | SIRegisterInfo::getEquivalentSGPRClass(const TargetRegisterClass *VRC) const { |
2215 | unsigned Size = getRegSizeInBits(*VRC); |
2216 | if (Size == 32) |
2217 | return &AMDGPU::SGPR_32RegClass; |
2218 | const TargetRegisterClass *SRC = getSGPRClassForBitWidth(Size); |
2219 | assert(SRC && "Invalid register class size")(static_cast <bool> (SRC && "Invalid register class size" ) ? void (0) : __assert_fail ("SRC && \"Invalid register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2219, __extension__ __PRETTY_FUNCTION__)); |
2220 | return SRC; |
2221 | } |
2222 | |
2223 | const TargetRegisterClass *SIRegisterInfo::getSubRegClass( |
2224 | const TargetRegisterClass *RC, unsigned SubIdx) const { |
2225 | if (SubIdx == AMDGPU::NoSubRegister) |
2226 | return RC; |
2227 | |
2228 | // We can assume that each lane corresponds to one 32-bit register. |
2229 | unsigned Size = getNumChannelsFromSubReg(SubIdx) * 32; |
2230 | if (isSGPRClass(RC)) { |
2231 | if (Size == 32) |
2232 | RC = &AMDGPU::SGPR_32RegClass; |
2233 | else |
2234 | RC = getSGPRClassForBitWidth(Size); |
2235 | } else if (hasAGPRs(RC)) { |
2236 | RC = getAGPRClassForBitWidth(Size); |
2237 | } else { |
2238 | RC = getVGPRClassForBitWidth(Size); |
2239 | } |
2240 | assert(RC && "Invalid sub-register class size")(static_cast <bool> (RC && "Invalid sub-register class size" ) ? void (0) : __assert_fail ("RC && \"Invalid sub-register class size\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2240, __extension__ __PRETTY_FUNCTION__)); |
2241 | return RC; |
2242 | } |
2243 | |
2244 | const TargetRegisterClass * |
2245 | SIRegisterInfo::getCompatibleSubRegClass(const TargetRegisterClass *SuperRC, |
2246 | const TargetRegisterClass *SubRC, |
2247 | unsigned SubIdx) const { |
2248 | // Ensure this subregister index is aligned in the super register. |
2249 | const TargetRegisterClass *MatchRC = |
2250 | getMatchingSuperRegClass(SuperRC, SubRC, SubIdx); |
2251 | return MatchRC && MatchRC->hasSubClassEq(SuperRC) ? MatchRC : nullptr; |
2252 | } |
2253 | |
2254 | bool SIRegisterInfo::opCanUseInlineConstant(unsigned OpType) const { |
2255 | if (OpType >= AMDGPU::OPERAND_REG_INLINE_AC_FIRST && |
2256 | OpType <= AMDGPU::OPERAND_REG_INLINE_AC_LAST) |
2257 | return !ST.hasMFMAInlineLiteralBug(); |
2258 | |
2259 | return OpType >= AMDGPU::OPERAND_SRC_FIRST && |
2260 | OpType <= AMDGPU::OPERAND_SRC_LAST; |
2261 | } |
2262 | |
2263 | bool SIRegisterInfo::shouldRewriteCopySrc( |
2264 | const TargetRegisterClass *DefRC, |
2265 | unsigned DefSubReg, |
2266 | const TargetRegisterClass *SrcRC, |
2267 | unsigned SrcSubReg) const { |
2268 | // We want to prefer the smallest register class possible, so we don't want to |
2269 | // stop and rewrite on anything that looks like a subregister |
2270 | // extract. Operations mostly don't care about the super register class, so we |
2271 | // only want to stop on the most basic of copies between the same register |
2272 | // class. |
2273 | // |
2274 | // e.g. if we have something like |
2275 | // %0 = ... |
2276 | // %1 = ... |
2277 | // %2 = REG_SEQUENCE %0, sub0, %1, sub1, %2, sub2 |
2278 | // %3 = COPY %2, sub0 |
2279 | // |
2280 | // We want to look through the COPY to find: |
2281 | // => %3 = COPY %0 |
2282 | |
2283 | // Plain copy. |
2284 | return getCommonSubClass(DefRC, SrcRC) != nullptr; |
2285 | } |
2286 | |
2287 | bool SIRegisterInfo::opCanUseLiteralConstant(unsigned OpType) const { |
2288 | // TODO: 64-bit operands have extending behavior from 32-bit literal. |
2289 | return OpType >= AMDGPU::OPERAND_REG_IMM_FIRST && |
2290 | OpType <= AMDGPU::OPERAND_REG_IMM_LAST; |
2291 | } |
2292 | |
2293 | /// Returns a lowest register that is not used at any point in the function. |
2294 | /// If all registers are used, then this function will return |
2295 | /// AMDGPU::NoRegister. If \p ReserveHighestVGPR = true, then return |
2296 | /// highest unused register. |
2297 | MCRegister SIRegisterInfo::findUnusedRegister(const MachineRegisterInfo &MRI, |
2298 | const TargetRegisterClass *RC, |
2299 | const MachineFunction &MF, |
2300 | bool ReserveHighestVGPR) const { |
2301 | if (ReserveHighestVGPR) { |
2302 | for (MCRegister Reg : reverse(*RC)) |
2303 | if (MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg)) |
2304 | return Reg; |
2305 | } else { |
2306 | for (MCRegister Reg : *RC) |
2307 | if (MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg)) |
2308 | return Reg; |
2309 | } |
2310 | return MCRegister(); |
2311 | } |
2312 | |
2313 | ArrayRef<int16_t> SIRegisterInfo::getRegSplitParts(const TargetRegisterClass *RC, |
2314 | unsigned EltSize) const { |
2315 | const unsigned RegBitWidth = AMDGPU::getRegBitWidth(*RC->MC); |
2316 | assert(RegBitWidth >= 32 && RegBitWidth <= 1024)(static_cast <bool> (RegBitWidth >= 32 && RegBitWidth <= 1024) ? void (0) : __assert_fail ("RegBitWidth >= 32 && RegBitWidth <= 1024" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2316, __extension__ __PRETTY_FUNCTION__)); |
2317 | |
2318 | const unsigned RegDWORDs = RegBitWidth / 32; |
2319 | const unsigned EltDWORDs = EltSize / 4; |
2320 | assert(RegSplitParts.size() + 1 >= EltDWORDs)(static_cast <bool> (RegSplitParts.size() + 1 >= EltDWORDs ) ? void (0) : __assert_fail ("RegSplitParts.size() + 1 >= EltDWORDs" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2320, __extension__ __PRETTY_FUNCTION__)); |
2321 | |
2322 | const std::vector<int16_t> &Parts = RegSplitParts[EltDWORDs - 1]; |
2323 | const unsigned NumParts = RegDWORDs / EltDWORDs; |
2324 | |
2325 | return makeArrayRef(Parts.data(), NumParts); |
2326 | } |
2327 | |
2328 | const TargetRegisterClass* |
2329 | SIRegisterInfo::getRegClassForReg(const MachineRegisterInfo &MRI, |
2330 | Register Reg) const { |
2331 | return Reg.isVirtual() ? MRI.getRegClass(Reg) : getPhysRegClass(Reg); |
2332 | } |
2333 | |
2334 | bool SIRegisterInfo::isVGPR(const MachineRegisterInfo &MRI, |
2335 | Register Reg) const { |
2336 | const TargetRegisterClass *RC = getRegClassForReg(MRI, Reg); |
2337 | // Registers without classes are unaddressable, SGPR-like registers. |
2338 | return RC && hasVGPRs(RC); |
2339 | } |
2340 | |
2341 | bool SIRegisterInfo::isAGPR(const MachineRegisterInfo &MRI, |
2342 | Register Reg) const { |
2343 | const TargetRegisterClass *RC = getRegClassForReg(MRI, Reg); |
2344 | |
2345 | // Registers without classes are unaddressable, SGPR-like registers. |
2346 | return RC && hasAGPRs(RC); |
2347 | } |
2348 | |
2349 | bool SIRegisterInfo::shouldCoalesce(MachineInstr *MI, |
2350 | const TargetRegisterClass *SrcRC, |
2351 | unsigned SubReg, |
2352 | const TargetRegisterClass *DstRC, |
2353 | unsigned DstSubReg, |
2354 | const TargetRegisterClass *NewRC, |
2355 | LiveIntervals &LIS) const { |
2356 | unsigned SrcSize = getRegSizeInBits(*SrcRC); |
2357 | unsigned DstSize = getRegSizeInBits(*DstRC); |
2358 | unsigned NewSize = getRegSizeInBits(*NewRC); |
2359 | |
2360 | // Do not increase size of registers beyond dword, we would need to allocate |
2361 | // adjacent registers and constraint regalloc more than needed. |
2362 | |
2363 | // Always allow dword coalescing. |
2364 | if (SrcSize <= 32 || DstSize <= 32) |
2365 | return true; |
2366 | |
2367 | return NewSize <= DstSize || NewSize <= SrcSize; |
2368 | } |
2369 | |
2370 | unsigned SIRegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, |
2371 | MachineFunction &MF) const { |
2372 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
2373 | |
2374 | unsigned Occupancy = ST.getOccupancyWithLocalMemSize(MFI->getLDSSize(), |
2375 | MF.getFunction()); |
2376 | switch (RC->getID()) { |
2377 | default: |
2378 | return AMDGPUGenRegisterInfo::getRegPressureLimit(RC, MF); |
2379 | case AMDGPU::VGPR_32RegClassID: |
2380 | case AMDGPU::VGPR_LO16RegClassID: |
2381 | case AMDGPU::VGPR_HI16RegClassID: |
2382 | return std::min(ST.getMaxNumVGPRs(Occupancy), ST.getMaxNumVGPRs(MF)); |
2383 | case AMDGPU::SGPR_32RegClassID: |
2384 | case AMDGPU::SGPR_LO16RegClassID: |
2385 | return std::min(ST.getMaxNumSGPRs(Occupancy, true), ST.getMaxNumSGPRs(MF)); |
2386 | } |
2387 | } |
2388 | |
2389 | unsigned SIRegisterInfo::getRegPressureSetLimit(const MachineFunction &MF, |
2390 | unsigned Idx) const { |
2391 | if (Idx == AMDGPU::RegisterPressureSets::VGPR_32 || |
2392 | Idx == AMDGPU::RegisterPressureSets::AGPR_32) |
2393 | return getRegPressureLimit(&AMDGPU::VGPR_32RegClass, |
2394 | const_cast<MachineFunction &>(MF)); |
2395 | |
2396 | if (Idx == AMDGPU::RegisterPressureSets::SReg_32) |
2397 | return getRegPressureLimit(&AMDGPU::SGPR_32RegClass, |
2398 | const_cast<MachineFunction &>(MF)); |
2399 | |
2400 | llvm_unreachable("Unexpected register pressure set!")::llvm::llvm_unreachable_internal("Unexpected register pressure set!" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2400); |
2401 | } |
2402 | |
2403 | const int *SIRegisterInfo::getRegUnitPressureSets(unsigned RegUnit) const { |
2404 | static const int Empty[] = { -1 }; |
2405 | |
2406 | if (RegPressureIgnoredUnits[RegUnit]) |
2407 | return Empty; |
2408 | |
2409 | return AMDGPUGenRegisterInfo::getRegUnitPressureSets(RegUnit); |
2410 | } |
2411 | |
2412 | MCRegister SIRegisterInfo::getReturnAddressReg(const MachineFunction &MF) const { |
2413 | // Not a callee saved register. |
2414 | return AMDGPU::SGPR30_SGPR31; |
2415 | } |
2416 | |
2417 | const TargetRegisterClass * |
2418 | SIRegisterInfo::getRegClassForSizeOnBank(unsigned Size, |
2419 | const RegisterBank &RB, |
2420 | const MachineRegisterInfo &MRI) const { |
2421 | switch (RB.getID()) { |
2422 | case AMDGPU::VGPRRegBankID: |
2423 | return getVGPRClassForBitWidth(std::max(32u, Size)); |
2424 | case AMDGPU::VCCRegBankID: |
2425 | assert(Size == 1)(static_cast <bool> (Size == 1) ? void (0) : __assert_fail ("Size == 1", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2425, __extension__ __PRETTY_FUNCTION__)); |
2426 | return isWave32 ? &AMDGPU::SReg_32_XM0_XEXECRegClass |
2427 | : &AMDGPU::SReg_64_XEXECRegClass; |
2428 | case AMDGPU::SGPRRegBankID: |
2429 | return getSGPRClassForBitWidth(std::max(32u, Size)); |
2430 | case AMDGPU::AGPRRegBankID: |
2431 | return getAGPRClassForBitWidth(std::max(32u, Size)); |
2432 | default: |
2433 | llvm_unreachable("unknown register bank")::llvm::llvm_unreachable_internal("unknown register bank", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2433); |
2434 | } |
2435 | } |
2436 | |
2437 | const TargetRegisterClass * |
2438 | SIRegisterInfo::getConstrainedRegClassForOperand(const MachineOperand &MO, |
2439 | const MachineRegisterInfo &MRI) const { |
2440 | const RegClassOrRegBank &RCOrRB = MRI.getRegClassOrRegBank(MO.getReg()); |
2441 | if (const RegisterBank *RB = RCOrRB.dyn_cast<const RegisterBank*>()) |
2442 | return getRegClassForTypeOnBank(MRI.getType(MO.getReg()), *RB, MRI); |
2443 | |
2444 | const TargetRegisterClass *RC = RCOrRB.get<const TargetRegisterClass*>(); |
2445 | return getAllocatableClass(RC); |
2446 | } |
2447 | |
2448 | MCRegister SIRegisterInfo::getVCC() const { |
2449 | return isWave32 ? AMDGPU::VCC_LO : AMDGPU::VCC; |
2450 | } |
2451 | |
2452 | const TargetRegisterClass *SIRegisterInfo::getVGPR64Class() const { |
2453 | // VGPR tuples have an alignment requirement on gfx90a variants. |
2454 | return ST.needsAlignedVGPRs() ? &AMDGPU::VReg_64_Align2RegClass |
2455 | : &AMDGPU::VReg_64RegClass; |
2456 | } |
2457 | |
2458 | const TargetRegisterClass * |
2459 | SIRegisterInfo::getRegClass(unsigned RCID) const { |
2460 | switch ((int)RCID) { |
2461 | case AMDGPU::SReg_1RegClassID: |
2462 | return getBoolRC(); |
2463 | case AMDGPU::SReg_1_XEXECRegClassID: |
2464 | return isWave32 ? &AMDGPU::SReg_32_XM0_XEXECRegClass |
2465 | : &AMDGPU::SReg_64_XEXECRegClass; |
2466 | case -1: |
2467 | return nullptr; |
2468 | default: |
2469 | return AMDGPUGenRegisterInfo::getRegClass(RCID); |
2470 | } |
2471 | } |
2472 | |
2473 | // Find reaching register definition |
2474 | MachineInstr *SIRegisterInfo::findReachingDef(Register Reg, unsigned SubReg, |
2475 | MachineInstr &Use, |
2476 | MachineRegisterInfo &MRI, |
2477 | LiveIntervals *LIS) const { |
2478 | auto &MDT = LIS->getAnalysis<MachineDominatorTree>(); |
2479 | SlotIndex UseIdx = LIS->getInstructionIndex(Use); |
2480 | SlotIndex DefIdx; |
2481 | |
2482 | if (Reg.isVirtual()) { |
2483 | if (!LIS->hasInterval(Reg)) |
2484 | return nullptr; |
2485 | LiveInterval &LI = LIS->getInterval(Reg); |
2486 | LaneBitmask SubLanes = SubReg ? getSubRegIndexLaneMask(SubReg) |
2487 | : MRI.getMaxLaneMaskForVReg(Reg); |
2488 | VNInfo *V = nullptr; |
2489 | if (LI.hasSubRanges()) { |
2490 | for (auto &S : LI.subranges()) { |
2491 | if ((S.LaneMask & SubLanes) == SubLanes) { |
2492 | V = S.getVNInfoAt(UseIdx); |
2493 | break; |
2494 | } |
2495 | } |
2496 | } else { |
2497 | V = LI.getVNInfoAt(UseIdx); |
2498 | } |
2499 | if (!V) |
2500 | return nullptr; |
2501 | DefIdx = V->def; |
2502 | } else { |
2503 | // Find last def. |
2504 | for (MCRegUnitIterator Units(Reg.asMCReg(), this); Units.isValid(); |
2505 | ++Units) { |
2506 | LiveRange &LR = LIS->getRegUnit(*Units); |
2507 | if (VNInfo *V = LR.getVNInfoAt(UseIdx)) { |
2508 | if (!DefIdx.isValid() || |
2509 | MDT.dominates(LIS->getInstructionFromIndex(DefIdx), |
2510 | LIS->getInstructionFromIndex(V->def))) |
2511 | DefIdx = V->def; |
2512 | } else { |
2513 | return nullptr; |
2514 | } |
2515 | } |
2516 | } |
2517 | |
2518 | MachineInstr *Def = LIS->getInstructionFromIndex(DefIdx); |
2519 | |
2520 | if (!Def || !MDT.dominates(Def, &Use)) |
2521 | return nullptr; |
2522 | |
2523 | assert(Def->modifiesRegister(Reg, this))(static_cast <bool> (Def->modifiesRegister(Reg, this )) ? void (0) : __assert_fail ("Def->modifiesRegister(Reg, this)" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2523, __extension__ __PRETTY_FUNCTION__)); |
2524 | |
2525 | return Def; |
2526 | } |
2527 | |
2528 | MCPhysReg SIRegisterInfo::get32BitRegister(MCPhysReg Reg) const { |
2529 | assert(getRegSizeInBits(*getPhysRegClass(Reg)) <= 32)(static_cast <bool> (getRegSizeInBits(*getPhysRegClass( Reg)) <= 32) ? void (0) : __assert_fail ("getRegSizeInBits(*getPhysRegClass(Reg)) <= 32" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp" , 2529, __extension__ __PRETTY_FUNCTION__)); |
2530 | |
2531 | for (const TargetRegisterClass &RC : { AMDGPU::VGPR_32RegClass, |
2532 | AMDGPU::SReg_32RegClass, |
2533 | AMDGPU::AGPR_32RegClass } ) { |
2534 | if (MCPhysReg Super = getMatchingSuperReg(Reg, AMDGPU::lo16, &RC)) |
2535 | return Super; |
2536 | } |
2537 | if (MCPhysReg Super = getMatchingSuperReg(Reg, AMDGPU::hi16, |
2538 | &AMDGPU::VGPR_32RegClass)) { |
2539 | return Super; |
2540 | } |
2541 | |
2542 | return AMDGPU::NoRegister; |
2543 | } |
2544 | |
2545 | bool SIRegisterInfo::isProperlyAlignedRC(const TargetRegisterClass &RC) const { |
2546 | if (!ST.needsAlignedVGPRs()) |
2547 | return true; |
2548 | |
2549 | if (hasVGPRs(&RC)) |
2550 | return RC.hasSuperClassEq(getVGPRClassForBitWidth(getRegSizeInBits(RC))); |
2551 | if (hasAGPRs(&RC)) |
2552 | return RC.hasSuperClassEq(getAGPRClassForBitWidth(getRegSizeInBits(RC))); |
2553 | |
2554 | return true; |
2555 | } |
2556 | |
2557 | bool SIRegisterInfo::isConstantPhysReg(MCRegister PhysReg) const { |
2558 | switch (PhysReg) { |
2559 | case AMDGPU::SGPR_NULL: |
2560 | case AMDGPU::SRC_SHARED_BASE: |
2561 | case AMDGPU::SRC_PRIVATE_BASE: |
2562 | case AMDGPU::SRC_SHARED_LIMIT: |
2563 | case AMDGPU::SRC_PRIVATE_LIMIT: |
2564 | return true; |
2565 | default: |
2566 | return false; |
2567 | } |
2568 | } |
2569 | |
2570 | ArrayRef<MCPhysReg> |
2571 | SIRegisterInfo::getAllSGPR128(const MachineFunction &MF) const { |
2572 | return makeArrayRef(AMDGPU::SGPR_128RegClass.begin(), |
2573 | ST.getMaxNumSGPRs(MF) / 4); |
2574 | } |
2575 | |
2576 | ArrayRef<MCPhysReg> |
2577 | SIRegisterInfo::getAllSGPR64(const MachineFunction &MF) const { |
2578 | return makeArrayRef(AMDGPU::SGPR_64RegClass.begin(), |
2579 | ST.getMaxNumSGPRs(MF) / 2); |
2580 | } |
2581 | |
2582 | ArrayRef<MCPhysReg> |
2583 | SIRegisterInfo::getAllSGPR32(const MachineFunction &MF) const { |
2584 | return makeArrayRef(AMDGPU::SGPR_32RegClass.begin(), ST.getMaxNumSGPRs(MF)); |
2585 | } |