1 | //===-- SIFormMemoryClauses.cpp -------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file This pass extends the live ranges of registers used as pointers in |
10 | /// sequences of adjacent SMEM and VMEM instructions if XNACK is enabled. A |
11 | /// load that would overwrite a pointer would require breaking the soft clause. |
12 | /// Artificially extend the live ranges of the pointer operands by adding |
13 | /// implicit-def early-clobber operands throughout the soft clause. |
14 | /// |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #include "AMDGPU.h" |
18 | #include "GCNRegPressure.h" |
19 | #include "SIMachineFunctionInfo.h" |
20 | #include "llvm/InitializePasses.h" |
21 | |
22 | using namespace llvm; |
23 | |
24 | #define DEBUG_TYPE "si-form-memory-clauses" |
25 | |
26 | // Clauses longer then 15 instructions would overflow one of the counters |
27 | // and stall. They can stall even earlier if there are outstanding counters. |
28 | static cl::opt<unsigned> |
29 | MaxClause("amdgpu-max-memory-clause" , cl::Hidden, cl::init(Val: 15), |
30 | cl::desc("Maximum length of a memory clause, instructions" )); |
31 | |
32 | namespace { |
33 | |
34 | class SIFormMemoryClauses : public MachineFunctionPass { |
35 | using RegUse = DenseMap<unsigned, std::pair<unsigned, LaneBitmask>>; |
36 | |
37 | public: |
38 | static char ID; |
39 | |
40 | public: |
41 | SIFormMemoryClauses() : MachineFunctionPass(ID) { |
42 | initializeSIFormMemoryClausesPass(*PassRegistry::getPassRegistry()); |
43 | } |
44 | |
45 | bool runOnMachineFunction(MachineFunction &MF) override; |
46 | |
47 | StringRef getPassName() const override { |
48 | return "SI Form memory clauses" ; |
49 | } |
50 | |
51 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
52 | AU.addRequired<LiveIntervalsWrapperPass>(); |
53 | AU.setPreservesAll(); |
54 | MachineFunctionPass::getAnalysisUsage(AU); |
55 | } |
56 | |
57 | MachineFunctionProperties getClearedProperties() const override { |
58 | return MachineFunctionProperties().set( |
59 | MachineFunctionProperties::Property::IsSSA); |
60 | } |
61 | |
62 | private: |
63 | bool canBundle(const MachineInstr &MI, const RegUse &Defs, |
64 | const RegUse &Uses) const; |
65 | bool checkPressure(const MachineInstr &MI, GCNDownwardRPTracker &RPT); |
66 | void collectRegUses(const MachineInstr &MI, RegUse &Defs, RegUse &Uses) const; |
67 | bool processRegUses(const MachineInstr &MI, RegUse &Defs, RegUse &Uses, |
68 | GCNDownwardRPTracker &RPT); |
69 | |
70 | const GCNSubtarget *ST; |
71 | const SIRegisterInfo *TRI; |
72 | const MachineRegisterInfo *MRI; |
73 | SIMachineFunctionInfo *MFI; |
74 | |
75 | unsigned LastRecordedOccupancy; |
76 | unsigned MaxVGPRs; |
77 | unsigned MaxSGPRs; |
78 | }; |
79 | |
80 | } // End anonymous namespace. |
81 | |
82 | INITIALIZE_PASS_BEGIN(SIFormMemoryClauses, DEBUG_TYPE, |
83 | "SI Form memory clauses" , false, false) |
84 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
85 | INITIALIZE_PASS_END(SIFormMemoryClauses, DEBUG_TYPE, |
86 | "SI Form memory clauses" , false, false) |
87 | |
88 | |
89 | char SIFormMemoryClauses::ID = 0; |
90 | |
91 | char &llvm::SIFormMemoryClausesID = SIFormMemoryClauses::ID; |
92 | |
93 | FunctionPass *llvm::createSIFormMemoryClausesPass() { |
94 | return new SIFormMemoryClauses(); |
95 | } |
96 | |
97 | static bool isVMEMClauseInst(const MachineInstr &MI) { |
98 | return SIInstrInfo::isFLAT(MI) || SIInstrInfo::isVMEM(MI); |
99 | } |
100 | |
101 | static bool isSMEMClauseInst(const MachineInstr &MI) { |
102 | return SIInstrInfo::isSMRD(MI); |
103 | } |
104 | |
105 | // There no sense to create store clauses, they do not define anything, |
106 | // thus there is nothing to set early-clobber. |
107 | static bool isValidClauseInst(const MachineInstr &MI, bool IsVMEMClause) { |
108 | assert(!MI.isDebugInstr() && "debug instructions should not reach here" ); |
109 | if (MI.isBundled()) |
110 | return false; |
111 | if (!MI.mayLoad() || MI.mayStore()) |
112 | return false; |
113 | if (SIInstrInfo::isAtomic(MI)) |
114 | return false; |
115 | if (IsVMEMClause && !isVMEMClauseInst(MI)) |
116 | return false; |
117 | if (!IsVMEMClause && !isSMEMClauseInst(MI)) |
118 | return false; |
119 | // If this is a load instruction where the result has been coalesced with an operand, then we cannot clause it. |
120 | for (const MachineOperand &ResMO : MI.defs()) { |
121 | Register ResReg = ResMO.getReg(); |
122 | for (const MachineOperand &MO : MI.all_uses()) { |
123 | if (MO.getReg() == ResReg) |
124 | return false; |
125 | } |
126 | break; // Only check the first def. |
127 | } |
128 | return true; |
129 | } |
130 | |
131 | static unsigned getMopState(const MachineOperand &MO) { |
132 | unsigned S = 0; |
133 | if (MO.isImplicit()) |
134 | S |= RegState::Implicit; |
135 | if (MO.isDead()) |
136 | S |= RegState::Dead; |
137 | if (MO.isUndef()) |
138 | S |= RegState::Undef; |
139 | if (MO.isKill()) |
140 | S |= RegState::Kill; |
141 | if (MO.isEarlyClobber()) |
142 | S |= RegState::EarlyClobber; |
143 | if (MO.getReg().isPhysical() && MO.isRenamable()) |
144 | S |= RegState::Renamable; |
145 | return S; |
146 | } |
147 | |
148 | // Returns false if there is a use of a def already in the map. |
149 | // In this case we must break the clause. |
150 | bool SIFormMemoryClauses::canBundle(const MachineInstr &MI, const RegUse &Defs, |
151 | const RegUse &Uses) const { |
152 | // Check interference with defs. |
153 | for (const MachineOperand &MO : MI.operands()) { |
154 | // TODO: Prologue/Epilogue Insertion pass does not process bundled |
155 | // instructions. |
156 | if (MO.isFI()) |
157 | return false; |
158 | |
159 | if (!MO.isReg()) |
160 | continue; |
161 | |
162 | Register Reg = MO.getReg(); |
163 | |
164 | // If it is tied we will need to write same register as we read. |
165 | if (MO.isTied()) |
166 | return false; |
167 | |
168 | const RegUse &Map = MO.isDef() ? Uses : Defs; |
169 | auto Conflict = Map.find(Val: Reg); |
170 | if (Conflict == Map.end()) |
171 | continue; |
172 | |
173 | if (Reg.isPhysical()) |
174 | return false; |
175 | |
176 | LaneBitmask Mask = TRI->getSubRegIndexLaneMask(SubIdx: MO.getSubReg()); |
177 | if ((Conflict->second.second & Mask).any()) |
178 | return false; |
179 | } |
180 | |
181 | return true; |
182 | } |
183 | |
184 | // Since all defs in the clause are early clobber we can run out of registers. |
185 | // Function returns false if pressure would hit the limit if instruction is |
186 | // bundled into a memory clause. |
187 | bool SIFormMemoryClauses::checkPressure(const MachineInstr &MI, |
188 | GCNDownwardRPTracker &RPT) { |
189 | // NB: skip advanceBeforeNext() call. Since all defs will be marked |
190 | // early-clobber they will all stay alive at least to the end of the |
191 | // clause. Therefor we should not decrease pressure even if load |
192 | // pointer becomes dead and could otherwise be reused for destination. |
193 | RPT.advanceToNext(); |
194 | GCNRegPressure MaxPressure = RPT.moveMaxPressure(); |
195 | unsigned Occupancy = MaxPressure.getOccupancy(ST: *ST); |
196 | |
197 | // Don't push over half the register budget. We don't want to introduce |
198 | // spilling just to form a soft clause. |
199 | // |
200 | // FIXME: This pressure check is fundamentally broken. First, this is checking |
201 | // the global pressure, not the pressure at this specific point in the |
202 | // program. Second, it's not accounting for the increased liveness of the use |
203 | // operands due to the early clobber we will introduce. Third, the pressure |
204 | // tracking does not account for the alignment requirements for SGPRs, or the |
205 | // fragmentation of registers the allocator will need to satisfy. |
206 | if (Occupancy >= MFI->getMinAllowedOccupancy() && |
207 | MaxPressure.getVGPRNum(UnifiedVGPRFile: ST->hasGFX90AInsts()) <= MaxVGPRs / 2 && |
208 | MaxPressure.getSGPRNum() <= MaxSGPRs / 2) { |
209 | LastRecordedOccupancy = Occupancy; |
210 | return true; |
211 | } |
212 | return false; |
213 | } |
214 | |
215 | // Collect register defs and uses along with their lane masks and states. |
216 | void SIFormMemoryClauses::collectRegUses(const MachineInstr &MI, |
217 | RegUse &Defs, RegUse &Uses) const { |
218 | for (const MachineOperand &MO : MI.operands()) { |
219 | if (!MO.isReg()) |
220 | continue; |
221 | Register Reg = MO.getReg(); |
222 | if (!Reg) |
223 | continue; |
224 | |
225 | LaneBitmask Mask = Reg.isVirtual() |
226 | ? TRI->getSubRegIndexLaneMask(SubIdx: MO.getSubReg()) |
227 | : LaneBitmask::getAll(); |
228 | RegUse &Map = MO.isDef() ? Defs : Uses; |
229 | |
230 | auto Loc = Map.find(Val: Reg); |
231 | unsigned State = getMopState(MO); |
232 | if (Loc == Map.end()) { |
233 | Map[Reg] = std::pair(State, Mask); |
234 | } else { |
235 | Loc->second.first |= State; |
236 | Loc->second.second |= Mask; |
237 | } |
238 | } |
239 | } |
240 | |
241 | // Check register def/use conflicts, occupancy limits and collect def/use maps. |
242 | // Return true if instruction can be bundled with previous. If it cannot |
243 | // def/use maps are not updated. |
244 | bool SIFormMemoryClauses::processRegUses(const MachineInstr &MI, |
245 | RegUse &Defs, RegUse &Uses, |
246 | GCNDownwardRPTracker &RPT) { |
247 | if (!canBundle(MI, Defs, Uses)) |
248 | return false; |
249 | |
250 | if (!checkPressure(MI, RPT)) |
251 | return false; |
252 | |
253 | collectRegUses(MI, Defs, Uses); |
254 | return true; |
255 | } |
256 | |
257 | bool SIFormMemoryClauses::runOnMachineFunction(MachineFunction &MF) { |
258 | if (skipFunction(F: MF.getFunction())) |
259 | return false; |
260 | |
261 | ST = &MF.getSubtarget<GCNSubtarget>(); |
262 | if (!ST->isXNACKEnabled()) |
263 | return false; |
264 | |
265 | const SIInstrInfo *TII = ST->getInstrInfo(); |
266 | TRI = ST->getRegisterInfo(); |
267 | MRI = &MF.getRegInfo(); |
268 | MFI = MF.getInfo<SIMachineFunctionInfo>(); |
269 | LiveIntervals *LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
270 | SlotIndexes *Ind = LIS->getSlotIndexes(); |
271 | bool Changed = false; |
272 | |
273 | MaxVGPRs = TRI->getAllocatableSet(MF, RC: &AMDGPU::VGPR_32RegClass).count(); |
274 | MaxSGPRs = TRI->getAllocatableSet(MF, RC: &AMDGPU::SGPR_32RegClass).count(); |
275 | unsigned FuncMaxClause = MF.getFunction().getFnAttributeAsParsedInteger( |
276 | Kind: "amdgpu-max-memory-clause" , Default: MaxClause); |
277 | |
278 | for (MachineBasicBlock &MBB : MF) { |
279 | GCNDownwardRPTracker RPT(*LIS); |
280 | MachineBasicBlock::instr_iterator Next; |
281 | for (auto I = MBB.instr_begin(), E = MBB.instr_end(); I != E; I = Next) { |
282 | MachineInstr &MI = *I; |
283 | Next = std::next(x: I); |
284 | |
285 | if (MI.isMetaInstruction()) |
286 | continue; |
287 | |
288 | bool IsVMEM = isVMEMClauseInst(MI); |
289 | |
290 | if (!isValidClauseInst(MI, IsVMEMClause: IsVMEM)) |
291 | continue; |
292 | |
293 | if (!RPT.getNext().isValid()) |
294 | RPT.reset(MI); |
295 | else { // Advance the state to the current MI. |
296 | RPT.advance(End: MachineBasicBlock::const_iterator(MI)); |
297 | RPT.advanceBeforeNext(); |
298 | } |
299 | |
300 | const GCNRPTracker::LiveRegSet LiveRegsCopy(RPT.getLiveRegs()); |
301 | RegUse Defs, Uses; |
302 | if (!processRegUses(MI, Defs, Uses, RPT)) { |
303 | RPT.reset(MI, LiveRegs: &LiveRegsCopy); |
304 | continue; |
305 | } |
306 | |
307 | MachineBasicBlock::iterator LastClauseInst = Next; |
308 | unsigned Length = 1; |
309 | for ( ; Next != E && Length < FuncMaxClause; ++Next) { |
310 | // Debug instructions should not change the kill insertion. |
311 | if (Next->isMetaInstruction()) |
312 | continue; |
313 | |
314 | if (!isValidClauseInst(MI: *Next, IsVMEMClause: IsVMEM)) |
315 | break; |
316 | |
317 | // A load from pointer which was loaded inside the same bundle is an |
318 | // impossible clause because we will need to write and read the same |
319 | // register inside. In this case processRegUses will return false. |
320 | if (!processRegUses(MI: *Next, Defs, Uses, RPT)) |
321 | break; |
322 | |
323 | LastClauseInst = Next; |
324 | ++Length; |
325 | } |
326 | if (Length < 2) { |
327 | RPT.reset(MI, LiveRegs: &LiveRegsCopy); |
328 | continue; |
329 | } |
330 | |
331 | Changed = true; |
332 | MFI->limitOccupancy(Limit: LastRecordedOccupancy); |
333 | |
334 | assert(!LastClauseInst->isMetaInstruction()); |
335 | |
336 | SlotIndex ClauseLiveInIdx = LIS->getInstructionIndex(Instr: MI); |
337 | SlotIndex ClauseLiveOutIdx = |
338 | LIS->getInstructionIndex(Instr: *LastClauseInst).getNextIndex(); |
339 | |
340 | // Track the last inserted kill. |
341 | MachineInstrBuilder Kill; |
342 | |
343 | // Insert one kill per register, with operands covering all necessary |
344 | // subregisters. |
345 | for (auto &&R : Uses) { |
346 | Register Reg = R.first; |
347 | if (Reg.isPhysical()) |
348 | continue; |
349 | |
350 | // Collect the register operands we should extend the live ranges of. |
351 | SmallVector<std::tuple<unsigned, unsigned>> KillOps; |
352 | const LiveInterval &LI = LIS->getInterval(Reg: R.first); |
353 | |
354 | if (!LI.hasSubRanges()) { |
355 | if (!LI.liveAt(index: ClauseLiveOutIdx)) { |
356 | KillOps.emplace_back(Args: R.second.first | RegState::Kill, |
357 | Args: AMDGPU::NoSubRegister); |
358 | } |
359 | } else { |
360 | LaneBitmask KilledMask; |
361 | for (const LiveInterval::SubRange &SR : LI.subranges()) { |
362 | if (SR.liveAt(index: ClauseLiveInIdx) && !SR.liveAt(index: ClauseLiveOutIdx)) |
363 | KilledMask |= SR.LaneMask; |
364 | } |
365 | |
366 | if (KilledMask.none()) |
367 | continue; |
368 | |
369 | SmallVector<unsigned> KilledIndexes; |
370 | bool Success = TRI->getCoveringSubRegIndexes( |
371 | MRI: *MRI, RC: MRI->getRegClass(Reg), LaneMask: KilledMask, Indexes&: KilledIndexes); |
372 | (void)Success; |
373 | assert(Success && "Failed to find subregister mask to cover lanes" ); |
374 | for (unsigned SubReg : KilledIndexes) { |
375 | KillOps.emplace_back(Args: R.second.first | RegState::Kill, Args&: SubReg); |
376 | } |
377 | } |
378 | |
379 | if (KillOps.empty()) |
380 | continue; |
381 | |
382 | // We only want to extend the live ranges of used registers. If they |
383 | // already have existing uses beyond the bundle, we don't need the kill. |
384 | // |
385 | // It's possible all of the use registers were already live past the |
386 | // bundle. |
387 | Kill = BuildMI(BB&: *MI.getParent(), I: std::next(x: LastClauseInst), |
388 | MIMD: DebugLoc(), MCID: TII->get(Opcode: AMDGPU::KILL)); |
389 | for (auto &Op : KillOps) |
390 | Kill.addUse(RegNo: Reg, Flags: std::get<0>(t&: Op), SubReg: std::get<1>(t&: Op)); |
391 | Ind->insertMachineInstrInMaps(MI&: *Kill); |
392 | } |
393 | |
394 | // Restore the state after processing the end of the bundle. |
395 | RPT.reset(MI, LiveRegs: &LiveRegsCopy); |
396 | |
397 | if (!Kill) |
398 | continue; |
399 | |
400 | for (auto &&R : Defs) { |
401 | Register Reg = R.first; |
402 | Uses.erase(Val: Reg); |
403 | if (Reg.isPhysical()) |
404 | continue; |
405 | LIS->removeInterval(Reg); |
406 | LIS->createAndComputeVirtRegInterval(Reg); |
407 | } |
408 | |
409 | for (auto &&R : Uses) { |
410 | Register Reg = R.first; |
411 | if (Reg.isPhysical()) |
412 | continue; |
413 | LIS->removeInterval(Reg); |
414 | LIS->createAndComputeVirtRegInterval(Reg); |
415 | } |
416 | } |
417 | } |
418 | |
419 | return Changed; |
420 | } |
421 | |