1 | //===-- GCNHazardRecognizers.cpp - GCN Hazard Recognizer Impls ------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements hazard recognizers for scheduling on GCN processors. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "GCNHazardRecognizer.h" |
14 | #include "GCNSubtarget.h" |
15 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
16 | #include "SIMachineFunctionInfo.h" |
17 | #include "llvm/CodeGen/MachineFrameInfo.h" |
18 | #include "llvm/CodeGen/MachineFunction.h" |
19 | #include "llvm/CodeGen/ScheduleDAG.h" |
20 | #include "llvm/TargetParser/TargetParser.h" |
21 | |
22 | using namespace llvm; |
23 | |
24 | namespace { |
25 | |
26 | struct MFMAPaddingRatioParser : public cl::parser<unsigned> { |
27 | MFMAPaddingRatioParser(cl::Option &O) : cl::parser<unsigned>(O) {} |
28 | |
29 | bool parse(cl::Option &O, StringRef ArgName, StringRef Arg, unsigned &Value) { |
30 | if (Arg.getAsInteger(Radix: 0, Result&: Value)) |
31 | return O.error(Message: "'" + Arg + "' value invalid for uint argument!" ); |
32 | |
33 | if (Value > 100) |
34 | return O.error(Message: "'" + Arg + "' value must be in the range [0, 100]!" ); |
35 | |
36 | return false; |
37 | } |
38 | }; |
39 | |
40 | } // end anonymous namespace |
41 | |
42 | static cl::opt<unsigned, false, MFMAPaddingRatioParser> |
43 | MFMAPaddingRatio("amdgpu-mfma-padding-ratio" , cl::init(Val: 0), cl::Hidden, |
44 | cl::desc("Fill a percentage of the latency between " |
45 | "neighboring MFMA with s_nops." )); |
46 | |
47 | // This is intended for debugging purposes only. |
48 | static cl::opt<unsigned> |
49 | NopPadding("amdgpu-snop-padding" , cl::init(Val: 0), cl::Hidden, |
50 | cl::desc("Insert a s_nop x before every instruction" )); |
51 | |
52 | //===----------------------------------------------------------------------===// |
53 | // Hazard Recognizer Implementation |
54 | //===----------------------------------------------------------------------===// |
55 | |
56 | static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF, |
57 | const GCNSubtarget &ST); |
58 | |
59 | GCNHazardRecognizer::GCNHazardRecognizer(const MachineFunction &MF) |
60 | : IsHazardRecognizerMode(false), CurrCycleInstr(nullptr), MF(MF), |
61 | ST(MF.getSubtarget<GCNSubtarget>()), TII(*ST.getInstrInfo()), |
62 | TRI(TII.getRegisterInfo()), TSchedModel(TII.getSchedModel()), |
63 | ClauseUses(TRI.getNumRegUnits()), ClauseDefs(TRI.getNumRegUnits()) { |
64 | MaxLookAhead = MF.getRegInfo().isPhysRegUsed(PhysReg: AMDGPU::AGPR0) ? 19 : 5; |
65 | RunLdsBranchVmemWARHazardFixup = shouldRunLdsBranchVmemWARHazardFixup(MF, ST); |
66 | } |
67 | |
68 | void GCNHazardRecognizer::Reset() { |
69 | EmittedInstrs.clear(); |
70 | } |
71 | |
72 | void GCNHazardRecognizer::EmitInstruction(SUnit *SU) { |
73 | EmitInstruction(MI: SU->getInstr()); |
74 | } |
75 | |
76 | void GCNHazardRecognizer::EmitInstruction(MachineInstr *MI) { |
77 | CurrCycleInstr = MI; |
78 | } |
79 | |
80 | static bool isDivFMas(unsigned Opcode) { |
81 | return Opcode == AMDGPU::V_DIV_FMAS_F32_e64 || Opcode == AMDGPU::V_DIV_FMAS_F64_e64; |
82 | } |
83 | |
84 | static bool isSGetReg(unsigned Opcode) { |
85 | return Opcode == AMDGPU::S_GETREG_B32; |
86 | } |
87 | |
88 | static bool isSSetReg(unsigned Opcode) { |
89 | switch (Opcode) { |
90 | case AMDGPU::S_SETREG_B32: |
91 | case AMDGPU::S_SETREG_B32_mode: |
92 | case AMDGPU::S_SETREG_IMM32_B32: |
93 | case AMDGPU::S_SETREG_IMM32_B32_mode: |
94 | return true; |
95 | } |
96 | return false; |
97 | } |
98 | |
99 | static bool isRWLane(unsigned Opcode) { |
100 | return Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32; |
101 | } |
102 | |
103 | static bool isRFE(unsigned Opcode) { |
104 | return Opcode == AMDGPU::S_RFE_B64; |
105 | } |
106 | |
107 | static bool isSMovRel(unsigned Opcode) { |
108 | switch (Opcode) { |
109 | case AMDGPU::S_MOVRELS_B32: |
110 | case AMDGPU::S_MOVRELS_B64: |
111 | case AMDGPU::S_MOVRELD_B32: |
112 | case AMDGPU::S_MOVRELD_B64: |
113 | return true; |
114 | default: |
115 | return false; |
116 | } |
117 | } |
118 | |
119 | static bool isSendMsgTraceDataOrGDS(const SIInstrInfo &TII, |
120 | const MachineInstr &MI) { |
121 | if (TII.isAlwaysGDS(Opcode: MI.getOpcode())) |
122 | return true; |
123 | |
124 | switch (MI.getOpcode()) { |
125 | case AMDGPU::S_SENDMSG: |
126 | case AMDGPU::S_SENDMSGHALT: |
127 | case AMDGPU::S_TTRACEDATA: |
128 | return true; |
129 | // These DS opcodes don't support GDS. |
130 | case AMDGPU::DS_NOP: |
131 | case AMDGPU::DS_PERMUTE_B32: |
132 | case AMDGPU::DS_BPERMUTE_B32: |
133 | return false; |
134 | default: |
135 | if (TII.isDS(Opcode: MI.getOpcode())) { |
136 | int GDS = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), |
137 | Name: AMDGPU::OpName::gds); |
138 | if (MI.getOperand(i: GDS).getImm()) |
139 | return true; |
140 | } |
141 | return false; |
142 | } |
143 | } |
144 | |
145 | static bool isPermlane(const MachineInstr &MI) { |
146 | unsigned Opcode = MI.getOpcode(); |
147 | return Opcode == AMDGPU::V_PERMLANE16_B32_e64 || |
148 | Opcode == AMDGPU::V_PERMLANE64_B32 || |
149 | Opcode == AMDGPU::V_PERMLANEX16_B32_e64 || |
150 | Opcode == AMDGPU::V_PERMLANE16_VAR_B32_e64 || |
151 | Opcode == AMDGPU::V_PERMLANEX16_VAR_B32_e64 || |
152 | Opcode == AMDGPU::V_PERMLANE16_SWAP_B32_e32 || |
153 | Opcode == AMDGPU::V_PERMLANE16_SWAP_B32_e64 || |
154 | Opcode == AMDGPU::V_PERMLANE32_SWAP_B32_e32 || |
155 | Opcode == AMDGPU::V_PERMLANE32_SWAP_B32_e64; |
156 | } |
157 | |
158 | static bool isLdsDma(const MachineInstr &MI) { |
159 | return SIInstrInfo::isVALU(MI) && |
160 | (SIInstrInfo::isMUBUF(MI) || SIInstrInfo::isFLAT(MI)); |
161 | } |
162 | |
163 | static unsigned getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) { |
164 | const MachineOperand *RegOp = TII->getNamedOperand(MI: RegInstr, |
165 | OperandName: AMDGPU::OpName::simm16); |
166 | return std::get<0>(t: AMDGPU::Hwreg::HwregEncoding::decode(Encoded: RegOp->getImm())); |
167 | } |
168 | |
169 | ScheduleHazardRecognizer::HazardType |
170 | GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) { |
171 | MachineInstr *MI = SU->getInstr(); |
172 | // If we are not in "HazardRecognizerMode" and therefore not being run from |
173 | // the scheduler, track possible stalls from hazards but don't insert noops. |
174 | auto HazardType = IsHazardRecognizerMode ? NoopHazard : Hazard; |
175 | |
176 | if (MI->isBundle()) |
177 | return NoHazard; |
178 | |
179 | if (SIInstrInfo::isSMRD(MI: *MI) && checkSMRDHazards(SMRD: MI) > 0) |
180 | return HazardType; |
181 | |
182 | if (ST.hasNSAtoVMEMBug() && checkNSAtoVMEMHazard(MI) > 0) |
183 | return HazardType; |
184 | |
185 | if (checkFPAtomicToDenormModeHazard(MI) > 0) |
186 | return HazardType; |
187 | |
188 | if (ST.hasNoDataDepHazard()) |
189 | return NoHazard; |
190 | |
191 | if (SIInstrInfo::isVMEM(MI: *MI) && checkVMEMHazards(VMEM: MI) > 0) |
192 | return HazardType; |
193 | |
194 | if (SIInstrInfo::isVALU(MI: *MI) && checkVALUHazards(VALU: MI) > 0) |
195 | return HazardType; |
196 | |
197 | if (SIInstrInfo::isDPP(MI: *MI) && checkDPPHazards(DPP: MI) > 0) |
198 | return HazardType; |
199 | |
200 | if (isDivFMas(Opcode: MI->getOpcode()) && checkDivFMasHazards(DivFMas: MI) > 0) |
201 | return HazardType; |
202 | |
203 | if (isRWLane(Opcode: MI->getOpcode()) && checkRWLaneHazards(RWLane: MI) > 0) |
204 | return HazardType; |
205 | |
206 | if ((SIInstrInfo::isVALU(MI: *MI) || SIInstrInfo::isVMEM(MI: *MI) || |
207 | SIInstrInfo::isDS(MI: *MI) || SIInstrInfo::isEXP(MI: *MI)) && |
208 | checkMAIVALUHazards(MI) > 0) |
209 | return HazardType; |
210 | |
211 | if (isSGetReg(Opcode: MI->getOpcode()) && checkGetRegHazards(GetRegInstr: MI) > 0) |
212 | return HazardType; |
213 | |
214 | if (isSSetReg(Opcode: MI->getOpcode()) && checkSetRegHazards(SetRegInstr: MI) > 0) |
215 | return HazardType; |
216 | |
217 | if (isRFE(Opcode: MI->getOpcode()) && checkRFEHazards(RFE: MI) > 0) |
218 | return HazardType; |
219 | |
220 | if (((ST.hasReadM0MovRelInterpHazard() && |
221 | (TII.isVINTRP(MI: *MI) || isSMovRel(Opcode: MI->getOpcode()) || |
222 | MI->getOpcode() == AMDGPU::DS_WRITE_ADDTID_B32 || |
223 | MI->getOpcode() == AMDGPU::DS_READ_ADDTID_B32)) || |
224 | (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, MI: *MI)) || |
225 | (ST.hasReadM0LdsDmaHazard() && isLdsDma(MI: *MI)) || |
226 | (ST.hasReadM0LdsDirectHazard() && |
227 | MI->readsRegister(Reg: AMDGPU::LDS_DIRECT, /*TRI=*/nullptr))) && |
228 | checkReadM0Hazards(SMovRel: MI) > 0) |
229 | return HazardType; |
230 | |
231 | if (SIInstrInfo::isMAI(MI: *MI) && checkMAIHazards(MI) > 0) |
232 | return HazardType; |
233 | |
234 | if ((SIInstrInfo::isVMEM(MI: *MI) || SIInstrInfo::isDS(MI: *MI)) && |
235 | checkMAILdStHazards(MI) > 0) |
236 | return HazardType; |
237 | |
238 | if (MI->isInlineAsm() && checkInlineAsmHazards(IA: MI) > 0) |
239 | return HazardType; |
240 | |
241 | return NoHazard; |
242 | } |
243 | |
244 | static void insertNoopsInBundle(MachineInstr *MI, const SIInstrInfo &TII, |
245 | unsigned Quantity) { |
246 | while (Quantity > 0) { |
247 | unsigned Arg = std::min(a: Quantity, b: 8u); |
248 | Quantity -= Arg; |
249 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), MCID: TII.get(Opcode: AMDGPU::S_NOP)) |
250 | .addImm(Val: Arg - 1); |
251 | } |
252 | } |
253 | |
254 | unsigned |
255 | GCNHazardRecognizer::getMFMAPipelineWaitStates(const MachineInstr &MI) const { |
256 | const MCSchedClassDesc *SC = TSchedModel.resolveSchedClass(MI: &MI); |
257 | assert(TSchedModel.getWriteProcResBegin(SC) != |
258 | TSchedModel.getWriteProcResEnd(SC)); |
259 | return TSchedModel.getWriteProcResBegin(SC)->ReleaseAtCycle; |
260 | } |
261 | |
262 | void GCNHazardRecognizer::processBundle() { |
263 | MachineBasicBlock::instr_iterator MI = std::next(x: CurrCycleInstr->getIterator()); |
264 | MachineBasicBlock::instr_iterator E = CurrCycleInstr->getParent()->instr_end(); |
265 | // Check bundled MachineInstr's for hazards. |
266 | for (; MI != E && MI->isInsideBundle(); ++MI) { |
267 | CurrCycleInstr = &*MI; |
268 | unsigned WaitStates = PreEmitNoopsCommon(CurrCycleInstr); |
269 | |
270 | if (IsHazardRecognizerMode) { |
271 | fixHazards(MI: CurrCycleInstr); |
272 | |
273 | insertNoopsInBundle(MI: CurrCycleInstr, TII, Quantity: WaitStates); |
274 | } |
275 | |
276 | // It’s unnecessary to track more than MaxLookAhead instructions. Since we |
277 | // include the bundled MI directly after, only add a maximum of |
278 | // (MaxLookAhead - 1) noops to EmittedInstrs. |
279 | for (unsigned i = 0, e = std::min(a: WaitStates, b: MaxLookAhead - 1); i < e; ++i) |
280 | EmittedInstrs.push_front(x: nullptr); |
281 | |
282 | EmittedInstrs.push_front(x: CurrCycleInstr); |
283 | EmittedInstrs.resize(new_size: MaxLookAhead); |
284 | } |
285 | CurrCycleInstr = nullptr; |
286 | } |
287 | |
288 | void GCNHazardRecognizer::runOnInstruction(MachineInstr *MI) { |
289 | assert(IsHazardRecognizerMode); |
290 | |
291 | unsigned NumPreNoops = PreEmitNoops(MI); |
292 | EmitNoops(Quantity: NumPreNoops); |
293 | if (MI->isInsideBundle()) |
294 | insertNoopsInBundle(MI, TII, Quantity: NumPreNoops); |
295 | else |
296 | TII.insertNoops(MBB&: *MI->getParent(), MI: MachineBasicBlock::iterator(MI), |
297 | Quantity: NumPreNoops); |
298 | EmitInstruction(MI); |
299 | AdvanceCycle(); |
300 | } |
301 | |
302 | unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) { |
303 | IsHazardRecognizerMode = true; |
304 | CurrCycleInstr = MI; |
305 | unsigned W = PreEmitNoopsCommon(MI); |
306 | fixHazards(MI); |
307 | CurrCycleInstr = nullptr; |
308 | return std::max(a: W, b: NopPadding.getValue()); |
309 | } |
310 | |
311 | unsigned GCNHazardRecognizer::PreEmitNoopsCommon(MachineInstr *MI) { |
312 | if (MI->isBundle()) |
313 | return 0; |
314 | |
315 | int WaitStates = 0; |
316 | |
317 | if (SIInstrInfo::isSMRD(MI: *MI)) |
318 | return std::max(a: WaitStates, b: checkSMRDHazards(SMRD: MI)); |
319 | |
320 | if (ST.hasNSAtoVMEMBug()) |
321 | WaitStates = std::max(a: WaitStates, b: checkNSAtoVMEMHazard(MI)); |
322 | |
323 | WaitStates = std::max(a: WaitStates, b: checkFPAtomicToDenormModeHazard(MI)); |
324 | |
325 | if (ST.hasNoDataDepHazard()) |
326 | return WaitStates; |
327 | |
328 | if (SIInstrInfo::isVMEM(MI: *MI)) |
329 | WaitStates = std::max(a: WaitStates, b: checkVMEMHazards(VMEM: MI)); |
330 | |
331 | if (SIInstrInfo::isVALU(MI: *MI)) |
332 | WaitStates = std::max(a: WaitStates, b: checkVALUHazards(VALU: MI)); |
333 | |
334 | if (SIInstrInfo::isDPP(MI: *MI)) |
335 | WaitStates = std::max(a: WaitStates, b: checkDPPHazards(DPP: MI)); |
336 | |
337 | if (isDivFMas(Opcode: MI->getOpcode())) |
338 | WaitStates = std::max(a: WaitStates, b: checkDivFMasHazards(DivFMas: MI)); |
339 | |
340 | if (isRWLane(Opcode: MI->getOpcode())) |
341 | WaitStates = std::max(a: WaitStates, b: checkRWLaneHazards(RWLane: MI)); |
342 | |
343 | if ((SIInstrInfo::isVALU(MI: *MI) || SIInstrInfo::isVMEM(MI: *MI) || |
344 | SIInstrInfo::isDS(MI: *MI) || SIInstrInfo::isEXP(MI: *MI)) && |
345 | checkMAIVALUHazards(MI) > 0) |
346 | WaitStates = std::max(a: WaitStates, b: checkMAIVALUHazards(MI)); |
347 | |
348 | if (MI->isInlineAsm()) |
349 | return std::max(a: WaitStates, b: checkInlineAsmHazards(IA: MI)); |
350 | |
351 | if (isSGetReg(Opcode: MI->getOpcode())) |
352 | return std::max(a: WaitStates, b: checkGetRegHazards(GetRegInstr: MI)); |
353 | |
354 | if (isSSetReg(Opcode: MI->getOpcode())) |
355 | return std::max(a: WaitStates, b: checkSetRegHazards(SetRegInstr: MI)); |
356 | |
357 | if (isRFE(Opcode: MI->getOpcode())) |
358 | return std::max(a: WaitStates, b: checkRFEHazards(RFE: MI)); |
359 | |
360 | if ((ST.hasReadM0MovRelInterpHazard() && |
361 | (TII.isVINTRP(MI: *MI) || isSMovRel(Opcode: MI->getOpcode()) || |
362 | MI->getOpcode() == AMDGPU::DS_WRITE_ADDTID_B32 || |
363 | MI->getOpcode() == AMDGPU::DS_READ_ADDTID_B32)) || |
364 | (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, MI: *MI)) || |
365 | (ST.hasReadM0LdsDmaHazard() && isLdsDma(MI: *MI)) || |
366 | (ST.hasReadM0LdsDirectHazard() && |
367 | MI->readsRegister(Reg: AMDGPU::LDS_DIRECT, /*TRI=*/nullptr))) |
368 | return std::max(a: WaitStates, b: checkReadM0Hazards(SMovRel: MI)); |
369 | |
370 | if (SIInstrInfo::isMAI(MI: *MI)) |
371 | return std::max(a: WaitStates, b: checkMAIHazards(MI)); |
372 | |
373 | if (SIInstrInfo::isVMEM(MI: *MI) || SIInstrInfo::isDS(MI: *MI)) |
374 | return std::max(a: WaitStates, b: checkMAILdStHazards(MI)); |
375 | |
376 | if (ST.hasGFX950Insts() && isPermlane(MI: *MI)) |
377 | return std::max(a: WaitStates, b: checkPermlaneHazards(MI)); |
378 | |
379 | return WaitStates; |
380 | } |
381 | |
382 | void GCNHazardRecognizer::EmitNoop() { |
383 | EmittedInstrs.push_front(x: nullptr); |
384 | } |
385 | |
386 | void GCNHazardRecognizer::AdvanceCycle() { |
387 | // When the scheduler detects a stall, it will call AdvanceCycle() without |
388 | // emitting any instructions. |
389 | if (!CurrCycleInstr) { |
390 | EmittedInstrs.push_front(x: nullptr); |
391 | return; |
392 | } |
393 | |
394 | if (CurrCycleInstr->isBundle()) { |
395 | processBundle(); |
396 | return; |
397 | } |
398 | |
399 | unsigned NumWaitStates = TII.getNumWaitStates(MI: *CurrCycleInstr); |
400 | if (!NumWaitStates) { |
401 | CurrCycleInstr = nullptr; |
402 | return; |
403 | } |
404 | |
405 | // Keep track of emitted instructions |
406 | EmittedInstrs.push_front(x: CurrCycleInstr); |
407 | |
408 | // Add a nullptr for each additional wait state after the first. Make sure |
409 | // not to add more than getMaxLookAhead() items to the list, since we |
410 | // truncate the list to that size right after this loop. |
411 | for (unsigned i = 1, e = std::min(a: NumWaitStates, b: getMaxLookAhead()); |
412 | i < e; ++i) { |
413 | EmittedInstrs.push_front(x: nullptr); |
414 | } |
415 | |
416 | // getMaxLookahead() is the largest number of wait states we will ever need |
417 | // to insert, so there is no point in keeping track of more than that many |
418 | // wait states. |
419 | EmittedInstrs.resize(new_size: getMaxLookAhead()); |
420 | |
421 | CurrCycleInstr = nullptr; |
422 | } |
423 | |
424 | void GCNHazardRecognizer::RecedeCycle() { |
425 | assert(!IsHazardRecognizerMode && |
426 | "Bottom-up scheduling shouldn't run in hazard recognizer mode" ); |
427 | } |
428 | |
429 | //===----------------------------------------------------------------------===// |
430 | // Helper Functions |
431 | //===----------------------------------------------------------------------===// |
432 | |
433 | using HazardFnResult = enum { HazardFound, HazardExpired, NoHazardFound }; |
434 | |
435 | using IsExpiredFn = function_ref<bool(const MachineInstr &, int WaitStates)>; |
436 | using GetNumWaitStatesFn = function_ref<unsigned int(const MachineInstr &)>; |
437 | |
438 | // Search for a hazard in a block and its predecessors. |
439 | template <typename StateT> |
440 | static bool |
441 | hasHazard(StateT State, |
442 | function_ref<HazardFnResult(StateT &, const MachineInstr &)> IsHazard, |
443 | function_ref<void(StateT &, const MachineInstr &)> UpdateState, |
444 | const MachineBasicBlock *MBB, |
445 | MachineBasicBlock::const_reverse_instr_iterator I, |
446 | DenseSet<const MachineBasicBlock *> &Visited) { |
447 | for (auto E = MBB->instr_rend(); I != E; ++I) { |
448 | // No need to look at parent BUNDLE instructions. |
449 | if (I->isBundle()) |
450 | continue; |
451 | |
452 | switch (IsHazard(State, *I)) { |
453 | case HazardFound: |
454 | return true; |
455 | case HazardExpired: |
456 | return false; |
457 | default: |
458 | // Continue search |
459 | break; |
460 | } |
461 | |
462 | if (I->isInlineAsm() || I->isMetaInstruction()) |
463 | continue; |
464 | |
465 | UpdateState(State, *I); |
466 | } |
467 | |
468 | for (MachineBasicBlock *Pred : MBB->predecessors()) { |
469 | if (!Visited.insert(V: Pred).second) |
470 | continue; |
471 | |
472 | if (hasHazard(State, IsHazard, UpdateState, Pred, Pred->instr_rbegin(), |
473 | Visited)) |
474 | return true; |
475 | } |
476 | |
477 | return false; |
478 | } |
479 | |
480 | // Returns a minimum wait states since \p I walking all predecessors. |
481 | // Only scans until \p IsExpired does not return true. |
482 | // Can only be run in a hazard recognizer mode. |
483 | static int getWaitStatesSince( |
484 | GCNHazardRecognizer::IsHazardFn IsHazard, const MachineBasicBlock *MBB, |
485 | MachineBasicBlock::const_reverse_instr_iterator I, int WaitStates, |
486 | IsExpiredFn IsExpired, DenseSet<const MachineBasicBlock *> &Visited, |
487 | GetNumWaitStatesFn GetNumWaitStates = SIInstrInfo::getNumWaitStates) { |
488 | for (auto E = MBB->instr_rend(); I != E; ++I) { |
489 | // Don't add WaitStates for parent BUNDLE instructions. |
490 | if (I->isBundle()) |
491 | continue; |
492 | |
493 | if (IsHazard(*I)) |
494 | return WaitStates; |
495 | |
496 | if (I->isInlineAsm()) |
497 | continue; |
498 | |
499 | WaitStates += GetNumWaitStates(*I); |
500 | |
501 | if (IsExpired(*I, WaitStates)) |
502 | return std::numeric_limits<int>::max(); |
503 | } |
504 | |
505 | int MinWaitStates = std::numeric_limits<int>::max(); |
506 | for (MachineBasicBlock *Pred : MBB->predecessors()) { |
507 | if (!Visited.insert(V: Pred).second) |
508 | continue; |
509 | |
510 | int W = getWaitStatesSince(IsHazard, MBB: Pred, I: Pred->instr_rbegin(), WaitStates, |
511 | IsExpired, Visited, GetNumWaitStates); |
512 | |
513 | MinWaitStates = std::min(a: MinWaitStates, b: W); |
514 | } |
515 | |
516 | return MinWaitStates; |
517 | } |
518 | |
519 | static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard, |
520 | const MachineInstr *MI, IsExpiredFn IsExpired) { |
521 | DenseSet<const MachineBasicBlock *> Visited; |
522 | return getWaitStatesSince(IsHazard, MBB: MI->getParent(), |
523 | I: std::next(x: MI->getReverseIterator()), |
524 | WaitStates: 0, IsExpired, Visited); |
525 | } |
526 | |
527 | int GCNHazardRecognizer::getWaitStatesSince(IsHazardFn IsHazard, int Limit) { |
528 | if (IsHazardRecognizerMode) { |
529 | auto IsExpiredFn = [Limit](const MachineInstr &, int WaitStates) { |
530 | return WaitStates >= Limit; |
531 | }; |
532 | return ::getWaitStatesSince(IsHazard, MI: CurrCycleInstr, IsExpired: IsExpiredFn); |
533 | } |
534 | |
535 | int WaitStates = 0; |
536 | for (MachineInstr *MI : EmittedInstrs) { |
537 | if (MI) { |
538 | if (IsHazard(*MI)) |
539 | return WaitStates; |
540 | |
541 | if (MI->isInlineAsm()) |
542 | continue; |
543 | } |
544 | ++WaitStates; |
545 | |
546 | if (WaitStates >= Limit) |
547 | break; |
548 | } |
549 | return std::numeric_limits<int>::max(); |
550 | } |
551 | |
552 | int GCNHazardRecognizer::getWaitStatesSinceDef(unsigned Reg, |
553 | IsHazardFn IsHazardDef, |
554 | int Limit) { |
555 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
556 | |
557 | auto IsHazardFn = [IsHazardDef, TRI, Reg](const MachineInstr &MI) { |
558 | return IsHazardDef(MI) && MI.modifiesRegister(Reg, TRI); |
559 | }; |
560 | |
561 | return getWaitStatesSince(IsHazard: IsHazardFn, Limit); |
562 | } |
563 | |
564 | int GCNHazardRecognizer::getWaitStatesSinceSetReg(IsHazardFn IsHazard, |
565 | int Limit) { |
566 | auto IsHazardFn = [IsHazard](const MachineInstr &MI) { |
567 | return isSSetReg(Opcode: MI.getOpcode()) && IsHazard(MI); |
568 | }; |
569 | |
570 | return getWaitStatesSince(IsHazard: IsHazardFn, Limit); |
571 | } |
572 | |
573 | //===----------------------------------------------------------------------===// |
574 | // No-op Hazard Detection |
575 | //===----------------------------------------------------------------------===// |
576 | |
577 | static void addRegUnits(const SIRegisterInfo &TRI, BitVector &BV, |
578 | MCRegister Reg) { |
579 | for (MCRegUnit Unit : TRI.regunits(Reg)) |
580 | BV.set(Unit); |
581 | } |
582 | |
583 | static void addRegsToSet(const SIRegisterInfo &TRI, |
584 | iterator_range<MachineInstr::const_mop_iterator> Ops, |
585 | BitVector &DefSet, BitVector &UseSet) { |
586 | for (const MachineOperand &Op : Ops) { |
587 | if (Op.isReg()) |
588 | addRegUnits(TRI, BV&: Op.isDef() ? DefSet : UseSet, Reg: Op.getReg().asMCReg()); |
589 | } |
590 | } |
591 | |
592 | void GCNHazardRecognizer::addClauseInst(const MachineInstr &MI) { |
593 | addRegsToSet(TRI, Ops: MI.operands(), DefSet&: ClauseDefs, UseSet&: ClauseUses); |
594 | } |
595 | |
596 | static bool breaksSMEMSoftClause(MachineInstr *MI) { |
597 | return !SIInstrInfo::isSMRD(MI: *MI); |
598 | } |
599 | |
600 | static bool breaksVMEMSoftClause(MachineInstr *MI) { |
601 | return !SIInstrInfo::isVMEM(MI: *MI); |
602 | } |
603 | |
604 | int GCNHazardRecognizer::checkSoftClauseHazards(MachineInstr *MEM) { |
605 | // SMEM soft clause are only present on VI+, and only matter if xnack is |
606 | // enabled. |
607 | if (!ST.isXNACKEnabled()) |
608 | return 0; |
609 | |
610 | bool IsSMRD = TII.isSMRD(MI: *MEM); |
611 | |
612 | resetClause(); |
613 | |
614 | // A soft-clause is any group of consecutive SMEM instructions. The |
615 | // instructions in this group may return out of order and/or may be |
616 | // replayed (i.e. the same instruction issued more than once). |
617 | // |
618 | // In order to handle these situations correctly we need to make sure that |
619 | // when a clause has more than one instruction, no instruction in the clause |
620 | // writes to a register that is read by another instruction in the clause |
621 | // (including itself). If we encounter this situation, we need to break the |
622 | // clause by inserting a non SMEM instruction. |
623 | |
624 | for (MachineInstr *MI : EmittedInstrs) { |
625 | // When we hit a non-SMEM instruction then we have passed the start of the |
626 | // clause and we can stop. |
627 | if (!MI) |
628 | break; |
629 | |
630 | if (IsSMRD ? breaksSMEMSoftClause(MI) : breaksVMEMSoftClause(MI)) |
631 | break; |
632 | |
633 | addClauseInst(MI: *MI); |
634 | } |
635 | |
636 | if (ClauseDefs.none()) |
637 | return 0; |
638 | |
639 | // We need to make sure not to put loads and stores in the same clause if they |
640 | // use the same address. For now, just start a new clause whenever we see a |
641 | // store. |
642 | if (MEM->mayStore()) |
643 | return 1; |
644 | |
645 | addClauseInst(MI: *MEM); |
646 | |
647 | // If the set of defs and uses intersect then we cannot add this instruction |
648 | // to the clause, so we have a hazard. |
649 | return ClauseDefs.anyCommon(RHS: ClauseUses) ? 1 : 0; |
650 | } |
651 | |
652 | int GCNHazardRecognizer::checkSMRDHazards(MachineInstr *SMRD) { |
653 | int WaitStatesNeeded = 0; |
654 | |
655 | WaitStatesNeeded = checkSoftClauseHazards(MEM: SMRD); |
656 | |
657 | // This SMRD hazard only affects SI. |
658 | if (!ST.hasSMRDReadVALUDefHazard()) |
659 | return WaitStatesNeeded; |
660 | |
661 | // A read of an SGPR by SMRD instruction requires 4 wait states when the |
662 | // SGPR was written by a VALU instruction. |
663 | int SmrdSgprWaitStates = 4; |
664 | auto IsHazardDefFn = [this](const MachineInstr &MI) { |
665 | return TII.isVALU(MI); |
666 | }; |
667 | auto IsBufferHazardDefFn = [this](const MachineInstr &MI) { |
668 | return TII.isSALU(MI); |
669 | }; |
670 | |
671 | bool IsBufferSMRD = TII.isBufferSMRD(MI: *SMRD); |
672 | |
673 | for (const MachineOperand &Use : SMRD->uses()) { |
674 | if (!Use.isReg()) |
675 | continue; |
676 | int WaitStatesNeededForUse = |
677 | SmrdSgprWaitStates - getWaitStatesSinceDef(Reg: Use.getReg(), IsHazardDef: IsHazardDefFn, |
678 | Limit: SmrdSgprWaitStates); |
679 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
680 | |
681 | // This fixes what appears to be undocumented hardware behavior in SI where |
682 | // s_mov writing a descriptor and s_buffer_load_dword reading the descriptor |
683 | // needs some number of nops in between. We don't know how many we need, but |
684 | // let's use 4. This wasn't discovered before probably because the only |
685 | // case when this happens is when we expand a 64-bit pointer into a full |
686 | // descriptor and use s_buffer_load_dword instead of s_load_dword, which was |
687 | // probably never encountered in the closed-source land. |
688 | if (IsBufferSMRD) { |
689 | int WaitStatesNeededForUse = |
690 | SmrdSgprWaitStates - getWaitStatesSinceDef(Reg: Use.getReg(), |
691 | IsHazardDef: IsBufferHazardDefFn, |
692 | Limit: SmrdSgprWaitStates); |
693 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
694 | } |
695 | } |
696 | |
697 | return WaitStatesNeeded; |
698 | } |
699 | |
700 | int GCNHazardRecognizer::checkVMEMHazards(MachineInstr* VMEM) { |
701 | if (!ST.hasVMEMReadSGPRVALUDefHazard()) |
702 | return 0; |
703 | |
704 | int WaitStatesNeeded = checkSoftClauseHazards(MEM: VMEM); |
705 | |
706 | // A read of an SGPR by a VMEM instruction requires 5 wait states when the |
707 | // SGPR was written by a VALU Instruction. |
708 | const int VmemSgprWaitStates = 5; |
709 | auto IsHazardDefFn = [this](const MachineInstr &MI) { |
710 | return TII.isVALU(MI); |
711 | }; |
712 | for (const MachineOperand &Use : VMEM->uses()) { |
713 | if (!Use.isReg() || TRI.isVectorRegister(MRI: MF.getRegInfo(), Reg: Use.getReg())) |
714 | continue; |
715 | |
716 | int WaitStatesNeededForUse = |
717 | VmemSgprWaitStates - getWaitStatesSinceDef(Reg: Use.getReg(), IsHazardDef: IsHazardDefFn, |
718 | Limit: VmemSgprWaitStates); |
719 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
720 | } |
721 | return WaitStatesNeeded; |
722 | } |
723 | |
724 | int GCNHazardRecognizer::checkDPPHazards(MachineInstr *DPP) { |
725 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
726 | const SIInstrInfo *TII = ST.getInstrInfo(); |
727 | |
728 | // Check for DPP VGPR read after VALU VGPR write and EXEC write. |
729 | int DppVgprWaitStates = 2; |
730 | int DppExecWaitStates = 5; |
731 | int WaitStatesNeeded = 0; |
732 | auto IsHazardDefFn = [TII](const MachineInstr &MI) { |
733 | return TII->isVALU(MI); |
734 | }; |
735 | |
736 | for (const MachineOperand &Use : DPP->uses()) { |
737 | if (!Use.isReg() || !TRI->isVGPR(MRI: MF.getRegInfo(), Reg: Use.getReg())) |
738 | continue; |
739 | int WaitStatesNeededForUse = |
740 | DppVgprWaitStates - getWaitStatesSinceDef( |
741 | Reg: Use.getReg(), |
742 | IsHazardDef: [](const MachineInstr &) { return true; }, |
743 | Limit: DppVgprWaitStates); |
744 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
745 | } |
746 | |
747 | WaitStatesNeeded = std::max( |
748 | a: WaitStatesNeeded, |
749 | b: DppExecWaitStates - getWaitStatesSinceDef(Reg: AMDGPU::EXEC, IsHazardDef: IsHazardDefFn, |
750 | Limit: DppExecWaitStates)); |
751 | |
752 | return WaitStatesNeeded; |
753 | } |
754 | |
755 | int GCNHazardRecognizer::checkDivFMasHazards(MachineInstr *DivFMas) { |
756 | const SIInstrInfo *TII = ST.getInstrInfo(); |
757 | |
758 | // v_div_fmas requires 4 wait states after a write to vcc from a VALU |
759 | // instruction. |
760 | const int DivFMasWaitStates = 4; |
761 | auto IsHazardDefFn = [TII](const MachineInstr &MI) { |
762 | return TII->isVALU(MI); |
763 | }; |
764 | int WaitStatesNeeded = getWaitStatesSinceDef(Reg: AMDGPU::VCC, IsHazardDef: IsHazardDefFn, |
765 | Limit: DivFMasWaitStates); |
766 | |
767 | return DivFMasWaitStates - WaitStatesNeeded; |
768 | } |
769 | |
770 | int GCNHazardRecognizer::checkGetRegHazards(MachineInstr *GetRegInstr) { |
771 | const SIInstrInfo *TII = ST.getInstrInfo(); |
772 | unsigned GetRegHWReg = getHWReg(TII, RegInstr: *GetRegInstr); |
773 | |
774 | const int GetRegWaitStates = 2; |
775 | auto IsHazardFn = [TII, GetRegHWReg](const MachineInstr &MI) { |
776 | return GetRegHWReg == getHWReg(TII, RegInstr: MI); |
777 | }; |
778 | int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazard: IsHazardFn, Limit: GetRegWaitStates); |
779 | |
780 | return GetRegWaitStates - WaitStatesNeeded; |
781 | } |
782 | |
783 | int GCNHazardRecognizer::checkSetRegHazards(MachineInstr *SetRegInstr) { |
784 | const SIInstrInfo *TII = ST.getInstrInfo(); |
785 | unsigned HWReg = getHWReg(TII, RegInstr: *SetRegInstr); |
786 | |
787 | const int SetRegWaitStates = ST.getSetRegWaitStates(); |
788 | auto IsHazardFn = [TII, HWReg](const MachineInstr &MI) { |
789 | return HWReg == getHWReg(TII, RegInstr: MI); |
790 | }; |
791 | int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazard: IsHazardFn, Limit: SetRegWaitStates); |
792 | return SetRegWaitStates - WaitStatesNeeded; |
793 | } |
794 | |
795 | int GCNHazardRecognizer::createsVALUHazard(const MachineInstr &MI) { |
796 | if (!MI.mayStore()) |
797 | return -1; |
798 | |
799 | const SIInstrInfo *TII = ST.getInstrInfo(); |
800 | unsigned Opcode = MI.getOpcode(); |
801 | const MCInstrDesc &Desc = MI.getDesc(); |
802 | |
803 | int VDataIdx = AMDGPU::getNamedOperandIdx(Opcode, Name: AMDGPU::OpName::vdata); |
804 | int VDataRCID = -1; |
805 | if (VDataIdx != -1) |
806 | VDataRCID = Desc.operands()[VDataIdx].RegClass; |
807 | |
808 | if (TII->isMUBUF(MI) || TII->isMTBUF(MI)) { |
809 | // There is no hazard if the instruction does not use vector regs |
810 | // (like wbinvl1) |
811 | if (VDataIdx == -1) |
812 | return -1; |
813 | // For MUBUF/MTBUF instructions this hazard only exists if the |
814 | // instruction is not using a register in the soffset field. |
815 | const MachineOperand *SOffset = |
816 | TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::soffset); |
817 | // If we have no soffset operand, then assume this field has been |
818 | // hardcoded to zero. |
819 | if (AMDGPU::getRegBitWidth(RCID: VDataRCID) > 64 && |
820 | (!SOffset || !SOffset->isReg())) |
821 | return VDataIdx; |
822 | } |
823 | |
824 | // MIMG instructions create a hazard if they don't use a 256-bit T# and |
825 | // the store size is greater than 8 bytes and they have more than two bits |
826 | // of their dmask set. |
827 | // All our MIMG definitions use a 256-bit T#, so we can skip checking for them. |
828 | if (TII->isMIMG(MI)) { |
829 | int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opcode, Name: AMDGPU::OpName::srsrc); |
830 | assert(SRsrcIdx != -1 && |
831 | AMDGPU::getRegBitWidth(Desc.operands()[SRsrcIdx].RegClass) == 256); |
832 | (void)SRsrcIdx; |
833 | } |
834 | |
835 | if (TII->isFLAT(MI)) { |
836 | // There is no hazard if the instruction does not use vector regs |
837 | if (VDataIdx == -1) |
838 | return -1; |
839 | |
840 | if (AMDGPU::getRegBitWidth(RCID: VDataRCID) > 64) |
841 | return VDataIdx; |
842 | } |
843 | |
844 | return -1; |
845 | } |
846 | |
847 | int |
848 | GCNHazardRecognizer::checkVALUHazardsHelper(const MachineOperand &Def, |
849 | const MachineRegisterInfo &MRI) { |
850 | // Helper to check for the hazard where VMEM instructions that store more than |
851 | // 8 bytes can have there store data over written by the next instruction. |
852 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
853 | |
854 | const int VALUWaitStates = ST.hasGFX940Insts() ? 2 : 1; |
855 | int WaitStatesNeeded = 0; |
856 | |
857 | if (!TRI->isVectorRegister(MRI, Reg: Def.getReg())) |
858 | return WaitStatesNeeded; |
859 | Register Reg = Def.getReg(); |
860 | auto IsHazardFn = [this, Reg, TRI](const MachineInstr &MI) { |
861 | int DataIdx = createsVALUHazard(MI); |
862 | return DataIdx >= 0 && |
863 | TRI->regsOverlap(RegA: MI.getOperand(i: DataIdx).getReg(), RegB: Reg); |
864 | }; |
865 | |
866 | int WaitStatesNeededForDef = |
867 | VALUWaitStates - getWaitStatesSince(IsHazard: IsHazardFn, Limit: VALUWaitStates); |
868 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
869 | |
870 | return WaitStatesNeeded; |
871 | } |
872 | |
873 | /// Dest sel forwarding issue occurs if additional logic is needed to swizzle / |
874 | /// pack the computed value into correct bit position of the dest register. This |
875 | /// occurs if we have SDWA with dst_sel != DWORD or if we have op_sel with |
876 | /// dst_sel that is not aligned to the register. This function analayzes the \p |
877 | /// MI and \returns an operand with dst forwarding issue, or nullptr if |
878 | /// none exists. |
879 | static const MachineOperand * |
880 | getDstSelForwardingOperand(const MachineInstr &MI, const GCNSubtarget &ST) { |
881 | if (!SIInstrInfo::isVALU(MI)) |
882 | return nullptr; |
883 | |
884 | const SIInstrInfo *TII = ST.getInstrInfo(); |
885 | |
886 | unsigned Opcode = MI.getOpcode(); |
887 | |
888 | // There are three different types of instructions |
889 | // which produce forwarded dest: 1. SDWA with dst_sel != DWORD, 2. VOP3 |
890 | // which write hi bits (e.g. op_sel[3] == 1), and 3. FP8DstSelInst |
891 | // (instructions with dest byte sel, e.g. CVT_SR_BF8_F32) and |
892 | // op_sel[3:2] |
893 | // != 0 |
894 | if (SIInstrInfo::isSDWA(MI)) { |
895 | // Type 1: SDWA with dst_sel != DWORD |
896 | if (auto *DstSel = TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::dst_sel)) |
897 | if (DstSel->getImm() != AMDGPU::SDWA::DWORD) |
898 | return TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdst); |
899 | } |
900 | |
901 | AMDGPU::FPType IsFP4OrFP8ConvOpc = AMDGPU::getFPDstSelType(Opc: Opcode); |
902 | if (AMDGPU::hasNamedOperand(Opcode, NamedIdx: AMDGPU::OpName::op_sel)) { |
903 | // Type 2: VOP3 which write the hi bits |
904 | if (TII->getNamedImmOperand(MI, OperandName: AMDGPU::OpName::src0_modifiers) & |
905 | SISrcMods::DST_OP_SEL) |
906 | return TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdst); |
907 | |
908 | // Type 3: FP8DstSelInst with op_sel[3:2] != 0) |
909 | if (IsFP4OrFP8ConvOpc == AMDGPU::FPType::FP8 && |
910 | (TII->getNamedImmOperand(MI, OperandName: AMDGPU::OpName::src2_modifiers) & |
911 | SISrcMods::OP_SEL_0)) |
912 | return TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdst); |
913 | } |
914 | |
915 | // Special case: nop is required for all the opsel values for fp4 sr variant |
916 | // cvt scale instructions |
917 | if (IsFP4OrFP8ConvOpc == AMDGPU::FPType::FP4) |
918 | return TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdst); |
919 | |
920 | return nullptr; |
921 | } |
922 | |
923 | /// Checks whether the provided \p MI "consumes" the operand with a Dest sel |
924 | /// fowarding issue \p Dst . We may "consume" the Dst via a standard explicit |
925 | /// RAW, or through irregular ways (e.g implicit RAW, certain types of WAW) |
926 | static bool consumesDstSelForwardingOperand(const MachineInstr *VALU, |
927 | const MachineOperand *Dst, |
928 | const SIRegisterInfo *TRI) { |
929 | // We must consider implicit reads of the VALU. SDWA with dst_sel and |
930 | // UNUSED_PRESERVE will implicitly read the result from forwarded dest, |
931 | // and we must account for that hazard. |
932 | // We also must account for WAW hazards. In particular, WAW with dest |
933 | // preserve semantics (e.g. VOP3 with op_sel, VOP2 && |
934 | // !zeroesHigh16BitsOfDest) will read the forwarded dest for parity |
935 | // check for ECC. Without accounting for this hazard, the ECC will be |
936 | // wrong. |
937 | // TODO: limit to RAW (including implicit reads) + problematic WAW (i.e. |
938 | // complete zeroesHigh16BitsOfDest) |
939 | for (auto &Operand : VALU->operands()) { |
940 | if (Operand.isReg() && TRI->regsOverlap(RegA: Dst->getReg(), RegB: Operand.getReg())) { |
941 | return true; |
942 | } |
943 | } |
944 | return false; |
945 | } |
946 | |
947 | int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) { |
948 | int WaitStatesNeeded = 0; |
949 | |
950 | if (ST.hasTransForwardingHazard() && !SIInstrInfo::isTRANS(MI: *VALU)) { |
951 | const int TransDefWaitstates = 1; |
952 | |
953 | auto IsTransDefFn = [this, VALU](const MachineInstr &MI) { |
954 | if (!SIInstrInfo::isTRANS(MI)) |
955 | return false; |
956 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
957 | const SIInstrInfo *TII = ST.getInstrInfo(); |
958 | Register Def = TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::vdst)->getReg(); |
959 | |
960 | for (const MachineOperand &Use : VALU->explicit_uses()) { |
961 | if (Use.isReg() && TRI->regsOverlap(RegA: Def, RegB: Use.getReg())) |
962 | return true; |
963 | } |
964 | |
965 | return false; |
966 | }; |
967 | |
968 | int WaitStatesNeededForDef = |
969 | TransDefWaitstates - |
970 | getWaitStatesSince(IsHazard: IsTransDefFn, Limit: TransDefWaitstates); |
971 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
972 | } |
973 | |
974 | if (ST.hasDstSelForwardingHazard() || ST.hasCvtScaleForwardingHazard()) { |
975 | const int Shift16DefWaitstates = 1; |
976 | |
977 | auto IsShift16BitDefFn = [this, VALU](const MachineInstr &ProducerMI) { |
978 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
979 | const MachineOperand *ForwardedDst = |
980 | getDstSelForwardingOperand(MI: ProducerMI, ST); |
981 | if (ForwardedDst) { |
982 | return consumesDstSelForwardingOperand(VALU, Dst: ForwardedDst, TRI); |
983 | } |
984 | |
985 | if (ProducerMI.isInlineAsm()) { |
986 | // Assume inline asm has dst forwarding hazard |
987 | for (auto &Def : ProducerMI.all_defs()) { |
988 | if (consumesDstSelForwardingOperand(VALU, Dst: &Def, TRI)) |
989 | return true; |
990 | } |
991 | } |
992 | |
993 | return false; |
994 | }; |
995 | |
996 | int WaitStatesNeededForDef = |
997 | Shift16DefWaitstates - |
998 | getWaitStatesSince(IsHazard: IsShift16BitDefFn, Limit: Shift16DefWaitstates); |
999 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1000 | } |
1001 | |
1002 | if (ST.hasVDecCoExecHazard()) { |
1003 | const int VALUWriteSGPRVALUReadWaitstates = 2; |
1004 | const int VALUWriteEXECRWLane = 4; |
1005 | const int VALUWriteVGPRReadlaneRead = 1; |
1006 | |
1007 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1008 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1009 | Register UseReg; |
1010 | auto IsVALUDefSGPRFn = [&UseReg, TRI](const MachineInstr &MI) { |
1011 | if (!SIInstrInfo::isVALU(MI)) |
1012 | return false; |
1013 | return MI.modifiesRegister(Reg: UseReg, TRI); |
1014 | }; |
1015 | |
1016 | for (const MachineOperand &Use : VALU->explicit_uses()) { |
1017 | if (!Use.isReg()) |
1018 | continue; |
1019 | |
1020 | UseReg = Use.getReg(); |
1021 | if (TRI->isSGPRReg(MRI, Reg: UseReg)) { |
1022 | int WaitStatesNeededForDef = |
1023 | VALUWriteSGPRVALUReadWaitstates - |
1024 | getWaitStatesSince(IsHazard: IsVALUDefSGPRFn, |
1025 | Limit: VALUWriteSGPRVALUReadWaitstates); |
1026 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1027 | } |
1028 | } |
1029 | |
1030 | if (VALU->readsRegister(Reg: AMDGPU::VCC, TRI)) { |
1031 | UseReg = AMDGPU::VCC; |
1032 | int WaitStatesNeededForDef = |
1033 | VALUWriteSGPRVALUReadWaitstates - |
1034 | getWaitStatesSince(IsHazard: IsVALUDefSGPRFn, Limit: VALUWriteSGPRVALUReadWaitstates); |
1035 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1036 | } |
1037 | |
1038 | switch (VALU->getOpcode()) { |
1039 | case AMDGPU::V_READLANE_B32: |
1040 | case AMDGPU::V_READFIRSTLANE_B32: { |
1041 | MachineOperand *Src = TII.getNamedOperand(MI&: *VALU, OperandName: AMDGPU::OpName::src0); |
1042 | UseReg = Src->getReg(); |
1043 | int WaitStatesNeededForDef = |
1044 | VALUWriteVGPRReadlaneRead - |
1045 | getWaitStatesSince(IsHazard: IsVALUDefSGPRFn, Limit: VALUWriteVGPRReadlaneRead); |
1046 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1047 | } |
1048 | [[fallthrough]]; |
1049 | case AMDGPU::V_WRITELANE_B32: { |
1050 | UseReg = AMDGPU::EXEC; |
1051 | int WaitStatesNeededForDef = |
1052 | VALUWriteEXECRWLane - |
1053 | getWaitStatesSince(IsHazard: IsVALUDefSGPRFn, Limit: VALUWriteEXECRWLane); |
1054 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1055 | break; |
1056 | } |
1057 | default: |
1058 | break; |
1059 | } |
1060 | } |
1061 | |
1062 | // This checks for the hazard where VMEM instructions that store more than |
1063 | // 8 bytes can have there store data over written by the next instruction. |
1064 | if (!ST.has12DWordStoreHazard()) |
1065 | return WaitStatesNeeded; |
1066 | |
1067 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1068 | |
1069 | for (const MachineOperand &Def : VALU->defs()) { |
1070 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: checkVALUHazardsHelper(Def, MRI)); |
1071 | } |
1072 | |
1073 | return WaitStatesNeeded; |
1074 | } |
1075 | |
1076 | int GCNHazardRecognizer::checkInlineAsmHazards(MachineInstr *IA) { |
1077 | // This checks for hazards associated with inline asm statements. |
1078 | // Since inline asms can contain just about anything, we use this |
1079 | // to call/leverage other check*Hazard routines. Note that |
1080 | // this function doesn't attempt to address all possible inline asm |
1081 | // hazards (good luck), but is a collection of what has been |
1082 | // problematic thus far. |
1083 | |
1084 | // see checkVALUHazards() |
1085 | if (!ST.has12DWordStoreHazard() && !ST.hasDstSelForwardingHazard() && |
1086 | !ST.hasCvtScaleForwardingHazard()) |
1087 | return 0; |
1088 | |
1089 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1090 | int WaitStatesNeeded = 0; |
1091 | |
1092 | for (const MachineOperand &Op : |
1093 | llvm::drop_begin(RangeOrContainer: IA->operands(), N: InlineAsm::MIOp_FirstOperand)) { |
1094 | if (Op.isReg() && Op.isDef()) { |
1095 | if (!TRI.isVectorRegister(MRI, Reg: Op.getReg())) |
1096 | continue; |
1097 | |
1098 | if (ST.has12DWordStoreHazard()) { |
1099 | WaitStatesNeeded = |
1100 | std::max(a: WaitStatesNeeded, b: checkVALUHazardsHelper(Def: Op, MRI)); |
1101 | } |
1102 | } |
1103 | } |
1104 | |
1105 | if (ST.hasDstSelForwardingHazard()) { |
1106 | const int Shift16DefWaitstates = 1; |
1107 | |
1108 | auto IsShift16BitDefFn = [this, &IA](const MachineInstr &ProducerMI) { |
1109 | const MachineOperand *Dst = getDstSelForwardingOperand(MI: ProducerMI, ST); |
1110 | // Assume inline asm reads the dst |
1111 | if (Dst) |
1112 | return IA->modifiesRegister(Reg: Dst->getReg(), TRI: &TRI) || |
1113 | IA->readsRegister(Reg: Dst->getReg(), TRI: &TRI); |
1114 | |
1115 | if (ProducerMI.isInlineAsm()) { |
1116 | // If MI is inline asm, assume it has dst forwarding hazard |
1117 | for (auto &Def : ProducerMI.all_defs()) { |
1118 | if (IA->modifiesRegister(Reg: Def.getReg(), TRI: &TRI) || |
1119 | IA->readsRegister(Reg: Def.getReg(), TRI: &TRI)) { |
1120 | return true; |
1121 | } |
1122 | } |
1123 | } |
1124 | |
1125 | return false; |
1126 | }; |
1127 | |
1128 | int WaitStatesNeededForDef = |
1129 | Shift16DefWaitstates - |
1130 | getWaitStatesSince(IsHazard: IsShift16BitDefFn, Limit: Shift16DefWaitstates); |
1131 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForDef); |
1132 | } |
1133 | |
1134 | return WaitStatesNeeded; |
1135 | } |
1136 | |
1137 | int GCNHazardRecognizer::checkRWLaneHazards(MachineInstr *RWLane) { |
1138 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1139 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1140 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1141 | |
1142 | const MachineOperand *LaneSelectOp = |
1143 | TII->getNamedOperand(MI&: *RWLane, OperandName: AMDGPU::OpName::src1); |
1144 | |
1145 | if (!LaneSelectOp->isReg() || !TRI->isSGPRReg(MRI, Reg: LaneSelectOp->getReg())) |
1146 | return 0; |
1147 | |
1148 | Register LaneSelectReg = LaneSelectOp->getReg(); |
1149 | auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVALU(MI); }; |
1150 | |
1151 | const int RWLaneWaitStates = 4; |
1152 | int WaitStatesSince = getWaitStatesSinceDef(Reg: LaneSelectReg, IsHazardDef: IsHazardFn, |
1153 | Limit: RWLaneWaitStates); |
1154 | return RWLaneWaitStates - WaitStatesSince; |
1155 | } |
1156 | |
1157 | int GCNHazardRecognizer::checkRFEHazards(MachineInstr *RFE) { |
1158 | if (!ST.hasRFEHazards()) |
1159 | return 0; |
1160 | |
1161 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1162 | |
1163 | const int RFEWaitStates = 1; |
1164 | |
1165 | auto IsHazardFn = [TII](const MachineInstr &MI) { |
1166 | return getHWReg(TII, RegInstr: MI) == AMDGPU::Hwreg::ID_TRAPSTS; |
1167 | }; |
1168 | int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazard: IsHazardFn, Limit: RFEWaitStates); |
1169 | return RFEWaitStates - WaitStatesNeeded; |
1170 | } |
1171 | |
1172 | int GCNHazardRecognizer::checkReadM0Hazards(MachineInstr *MI) { |
1173 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1174 | const int ReadM0WaitStates = 1; |
1175 | auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isSALU(MI); }; |
1176 | return ReadM0WaitStates - |
1177 | getWaitStatesSinceDef(Reg: AMDGPU::M0, IsHazardDef: IsHazardFn, Limit: ReadM0WaitStates); |
1178 | } |
1179 | |
1180 | void GCNHazardRecognizer::fixHazards(MachineInstr *MI) { |
1181 | fixVMEMtoScalarWriteHazards(MI); |
1182 | fixVcmpxPermlaneHazards(MI); |
1183 | fixSMEMtoVectorWriteHazards(MI); |
1184 | fixVcmpxExecWARHazard(MI); |
1185 | fixLdsBranchVmemWARHazard(MI); |
1186 | if (ST.hasLdsDirect()) { |
1187 | fixLdsDirectVALUHazard(MI); |
1188 | fixLdsDirectVMEMHazard(MI); |
1189 | } |
1190 | fixVALUPartialForwardingHazard(MI); |
1191 | fixVALUTransUseHazard(MI); |
1192 | fixWMMAHazards(MI); |
1193 | fixShift64HighRegBug(MI); |
1194 | fixVALUMaskWriteHazard(MI); |
1195 | fixRequiredExportPriority(MI); |
1196 | } |
1197 | |
1198 | static bool isVCmpXWritesExec(const SIInstrInfo &TII, const SIRegisterInfo &TRI, |
1199 | const MachineInstr &MI) { |
1200 | return (TII.isVOPC(MI) || |
1201 | (MI.isCompare() && (TII.isVOP3(MI) || TII.isSDWA(MI)))) && |
1202 | MI.modifiesRegister(Reg: AMDGPU::EXEC, TRI: &TRI); |
1203 | } |
1204 | |
1205 | bool GCNHazardRecognizer::fixVcmpxPermlaneHazards(MachineInstr *MI) { |
1206 | if (!ST.hasVcmpxPermlaneHazard() || !isPermlane(MI: *MI)) |
1207 | return false; |
1208 | |
1209 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1210 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1211 | auto IsHazardFn = [TII, TRI](const MachineInstr &MI) { |
1212 | return isVCmpXWritesExec(TII: *TII, TRI: *TRI, MI); |
1213 | }; |
1214 | |
1215 | auto IsExpiredFn = [](const MachineInstr &MI, int) { |
1216 | unsigned Opc = MI.getOpcode(); |
1217 | return SIInstrInfo::isVALU(MI) && Opc != AMDGPU::V_NOP_e32 && |
1218 | Opc != AMDGPU::V_NOP_e64 && Opc != AMDGPU::V_NOP_sdwa; |
1219 | }; |
1220 | |
1221 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1222 | std::numeric_limits<int>::max()) |
1223 | return false; |
1224 | |
1225 | // V_NOP will be discarded by SQ. |
1226 | // Use V_MOV_B32 v?, v?. Register must be alive so use src0 of V_PERMLANE* |
1227 | // which is always a VGPR and available. |
1228 | auto *Src0 = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src0); |
1229 | Register Reg = Src0->getReg(); |
1230 | bool IsUndef = Src0->isUndef(); |
1231 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1232 | MCID: TII->get(Opcode: AMDGPU::V_MOV_B32_e32)) |
1233 | .addReg(RegNo: Reg, flags: RegState::Define | (IsUndef ? RegState::Dead : 0)) |
1234 | .addReg(RegNo: Reg, flags: IsUndef ? RegState::Undef : RegState::Kill); |
1235 | |
1236 | return true; |
1237 | } |
1238 | |
1239 | bool GCNHazardRecognizer::fixVMEMtoScalarWriteHazards(MachineInstr *MI) { |
1240 | if (!ST.hasVMEMtoScalarWriteHazard()) |
1241 | return false; |
1242 | assert(!ST.hasExtendedWaitCounts()); |
1243 | |
1244 | if (!SIInstrInfo::isSALU(MI: *MI) && !SIInstrInfo::isSMRD(MI: *MI)) |
1245 | return false; |
1246 | |
1247 | if (MI->getNumDefs() == 0) |
1248 | return false; |
1249 | |
1250 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1251 | |
1252 | auto IsHazardFn = [TRI, MI](const MachineInstr &I) { |
1253 | if (!SIInstrInfo::isVMEM(MI: I) && !SIInstrInfo::isDS(MI: I)) |
1254 | return false; |
1255 | |
1256 | for (const MachineOperand &Def : MI->defs()) { |
1257 | const MachineOperand *Op = |
1258 | I.findRegisterUseOperand(Reg: Def.getReg(), TRI, isKill: false); |
1259 | if (!Op) |
1260 | continue; |
1261 | return true; |
1262 | } |
1263 | return false; |
1264 | }; |
1265 | |
1266 | auto IsExpiredFn = [](const MachineInstr &MI, int) { |
1267 | return SIInstrInfo::isVALU(MI) || |
1268 | (MI.getOpcode() == AMDGPU::S_WAITCNT && |
1269 | !MI.getOperand(i: 0).getImm()) || |
1270 | (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
1271 | AMDGPU::DepCtr::decodeFieldVmVsrc(Encoded: MI.getOperand(i: 0).getImm()) == 0); |
1272 | }; |
1273 | |
1274 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1275 | std::numeric_limits<int>::max()) |
1276 | return false; |
1277 | |
1278 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1279 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1280 | MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
1281 | .addImm(Val: AMDGPU::DepCtr::encodeFieldVmVsrc(VmVsrc: 0)); |
1282 | return true; |
1283 | } |
1284 | |
1285 | bool GCNHazardRecognizer::fixSMEMtoVectorWriteHazards(MachineInstr *MI) { |
1286 | if (!ST.hasSMEMtoVectorWriteHazard()) |
1287 | return false; |
1288 | assert(!ST.hasExtendedWaitCounts()); |
1289 | |
1290 | if (!SIInstrInfo::isVALU(MI: *MI)) |
1291 | return false; |
1292 | |
1293 | AMDGPU::OpName SDSTName; |
1294 | switch (MI->getOpcode()) { |
1295 | case AMDGPU::V_READLANE_B32: |
1296 | case AMDGPU::V_READFIRSTLANE_B32: |
1297 | SDSTName = AMDGPU::OpName::vdst; |
1298 | break; |
1299 | default: |
1300 | SDSTName = AMDGPU::OpName::sdst; |
1301 | break; |
1302 | } |
1303 | |
1304 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1305 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1306 | const AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(GPU: ST.getCPU()); |
1307 | const MachineOperand *SDST = TII->getNamedOperand(MI&: *MI, OperandName: SDSTName); |
1308 | if (!SDST) { |
1309 | for (const auto &MO : MI->implicit_operands()) { |
1310 | if (MO.isDef() && TRI->isSGPRClass(RC: TRI->getPhysRegBaseClass(Reg: MO.getReg()))) { |
1311 | SDST = &MO; |
1312 | break; |
1313 | } |
1314 | } |
1315 | } |
1316 | |
1317 | if (!SDST) |
1318 | return false; |
1319 | |
1320 | const Register SDSTReg = SDST->getReg(); |
1321 | auto IsHazardFn = [SDSTReg, TRI](const MachineInstr &I) { |
1322 | return SIInstrInfo::isSMRD(MI: I) && I.readsRegister(Reg: SDSTReg, TRI); |
1323 | }; |
1324 | |
1325 | auto IsExpiredFn = [TII, IV](const MachineInstr &MI, int) { |
1326 | if (TII->isSALU(MI)) { |
1327 | switch (MI.getOpcode()) { |
1328 | case AMDGPU::S_SETVSKIP: |
1329 | case AMDGPU::S_VERSION: |
1330 | case AMDGPU::S_WAITCNT_VSCNT: |
1331 | case AMDGPU::S_WAITCNT_VMCNT: |
1332 | case AMDGPU::S_WAITCNT_EXPCNT: |
1333 | // These instructions cannot not mitigate the hazard. |
1334 | return false; |
1335 | case AMDGPU::S_WAITCNT_LGKMCNT: |
1336 | // Reducing lgkmcnt count to 0 always mitigates the hazard. |
1337 | return (MI.getOperand(i: 1).getImm() == 0) && |
1338 | (MI.getOperand(i: 0).getReg() == AMDGPU::SGPR_NULL); |
1339 | case AMDGPU::S_WAITCNT: { |
1340 | const int64_t Imm = MI.getOperand(i: 0).getImm(); |
1341 | AMDGPU::Waitcnt Decoded = AMDGPU::decodeWaitcnt(Version: IV, Encoded: Imm); |
1342 | // DsCnt corresponds to LGKMCnt here. |
1343 | return (Decoded.DsCnt == 0); |
1344 | } |
1345 | default: |
1346 | // SOPP instructions cannot mitigate the hazard. |
1347 | if (TII->isSOPP(MI)) |
1348 | return false; |
1349 | // At this point the SALU can be assumed to mitigate the hazard |
1350 | // because either: |
1351 | // (a) it is independent of the at risk SMEM (breaking chain), |
1352 | // or |
1353 | // (b) it is dependent on the SMEM, in which case an appropriate |
1354 | // s_waitcnt lgkmcnt _must_ exist between it and the at risk |
1355 | // SMEM instruction. |
1356 | return true; |
1357 | } |
1358 | } |
1359 | return false; |
1360 | }; |
1361 | |
1362 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1363 | std::numeric_limits<int>::max()) |
1364 | return false; |
1365 | |
1366 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1367 | MCID: TII->get(Opcode: AMDGPU::S_MOV_B32), DestReg: AMDGPU::SGPR_NULL) |
1368 | .addImm(Val: 0); |
1369 | return true; |
1370 | } |
1371 | |
1372 | bool GCNHazardRecognizer::fixVcmpxExecWARHazard(MachineInstr *MI) { |
1373 | if (!ST.hasVcmpxExecWARHazard()) |
1374 | return false; |
1375 | assert(!ST.hasExtendedWaitCounts()); |
1376 | |
1377 | if (!SIInstrInfo::isVALU(MI: *MI)) |
1378 | return false; |
1379 | |
1380 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1381 | if (!MI->modifiesRegister(Reg: AMDGPU::EXEC, TRI)) |
1382 | return false; |
1383 | |
1384 | auto IsHazardFn = [TRI](const MachineInstr &I) { |
1385 | if (SIInstrInfo::isVALU(MI: I)) |
1386 | return false; |
1387 | return I.readsRegister(Reg: AMDGPU::EXEC, TRI); |
1388 | }; |
1389 | |
1390 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1391 | auto IsExpiredFn = [TII, TRI](const MachineInstr &MI, int) { |
1392 | if (SIInstrInfo::isVALU(MI)) { |
1393 | if (TII->getNamedOperand(MI, OperandName: AMDGPU::OpName::sdst)) |
1394 | return true; |
1395 | for (auto MO : MI.implicit_operands()) |
1396 | if (MO.isDef() && TRI->isSGPRClass(RC: TRI->getPhysRegBaseClass(Reg: MO.getReg()))) |
1397 | return true; |
1398 | } |
1399 | if (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
1400 | AMDGPU::DepCtr::decodeFieldSaSdst(Encoded: MI.getOperand(i: 0).getImm()) == 0) |
1401 | return true; |
1402 | return false; |
1403 | }; |
1404 | |
1405 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1406 | std::numeric_limits<int>::max()) |
1407 | return false; |
1408 | |
1409 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1410 | MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
1411 | .addImm(Val: AMDGPU::DepCtr::encodeFieldSaSdst(SaSdst: 0)); |
1412 | return true; |
1413 | } |
1414 | |
1415 | static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF, |
1416 | const GCNSubtarget &ST) { |
1417 | if (!ST.hasLdsBranchVmemWARHazard()) |
1418 | return false; |
1419 | |
1420 | // Check if the necessary condition for the hazard is met: both LDS and VMEM |
1421 | // instructions need to appear in the same function. |
1422 | bool HasLds = false; |
1423 | bool HasVmem = false; |
1424 | for (auto &MBB : MF) { |
1425 | for (auto &MI : MBB) { |
1426 | HasLds |= SIInstrInfo::isDS(MI); |
1427 | HasVmem |= (SIInstrInfo::isVMEM(MI) && !SIInstrInfo::isFLAT(MI)) || |
1428 | SIInstrInfo::isSegmentSpecificFLAT(MI); |
1429 | if (HasLds && HasVmem) |
1430 | return true; |
1431 | } |
1432 | } |
1433 | return false; |
1434 | } |
1435 | |
1436 | static bool isStoreCountWaitZero(const MachineInstr &I) { |
1437 | return I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT && |
1438 | I.getOperand(i: 0).getReg() == AMDGPU::SGPR_NULL && |
1439 | !I.getOperand(i: 1).getImm(); |
1440 | } |
1441 | |
1442 | bool GCNHazardRecognizer::fixLdsBranchVmemWARHazard(MachineInstr *MI) { |
1443 | if (!RunLdsBranchVmemWARHazardFixup) |
1444 | return false; |
1445 | |
1446 | assert(ST.hasLdsBranchVmemWARHazard()); |
1447 | assert(!ST.hasExtendedWaitCounts()); |
1448 | |
1449 | auto IsHazardInst = [](const MachineInstr &MI) { |
1450 | if (SIInstrInfo::isDS(MI)) |
1451 | return 1; |
1452 | if ((SIInstrInfo::isVMEM(MI) && !SIInstrInfo::isFLAT(MI)) || |
1453 | SIInstrInfo::isSegmentSpecificFLAT(MI)) |
1454 | return 2; |
1455 | return 0; |
1456 | }; |
1457 | |
1458 | auto InstType = IsHazardInst(*MI); |
1459 | if (!InstType) |
1460 | return false; |
1461 | |
1462 | auto IsExpiredFn = [&IsHazardInst](const MachineInstr &I, int) { |
1463 | return IsHazardInst(I) || isStoreCountWaitZero(I); |
1464 | }; |
1465 | |
1466 | auto IsHazardFn = [InstType, &IsHazardInst](const MachineInstr &I) { |
1467 | if (!I.isBranch()) |
1468 | return false; |
1469 | |
1470 | auto IsHazardFn = [InstType, IsHazardInst](const MachineInstr &I) { |
1471 | auto InstType2 = IsHazardInst(I); |
1472 | return InstType2 && InstType != InstType2; |
1473 | }; |
1474 | |
1475 | auto IsExpiredFn = [InstType, &IsHazardInst](const MachineInstr &I, int) { |
1476 | auto InstType2 = IsHazardInst(I); |
1477 | if (InstType == InstType2) |
1478 | return true; |
1479 | |
1480 | return isStoreCountWaitZero(I); |
1481 | }; |
1482 | |
1483 | return ::getWaitStatesSince(IsHazard: IsHazardFn, MI: &I, IsExpired: IsExpiredFn) != |
1484 | std::numeric_limits<int>::max(); |
1485 | }; |
1486 | |
1487 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1488 | std::numeric_limits<int>::max()) |
1489 | return false; |
1490 | |
1491 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1492 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1493 | MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_VSCNT)) |
1494 | .addReg(RegNo: AMDGPU::SGPR_NULL, flags: RegState::Undef) |
1495 | .addImm(Val: 0); |
1496 | |
1497 | return true; |
1498 | } |
1499 | |
1500 | bool GCNHazardRecognizer::fixLdsDirectVALUHazard(MachineInstr *MI) { |
1501 | if (!SIInstrInfo::isLDSDIR(MI: *MI)) |
1502 | return false; |
1503 | |
1504 | const int NoHazardWaitStates = 15; |
1505 | const MachineOperand *VDST = TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::vdst); |
1506 | const Register VDSTReg = VDST->getReg(); |
1507 | |
1508 | bool VisitedTrans = false; |
1509 | auto IsHazardFn = [this, VDSTReg, &VisitedTrans](const MachineInstr &I) { |
1510 | if (!SIInstrInfo::isVALU(MI: I)) |
1511 | return false; |
1512 | VisitedTrans = VisitedTrans || SIInstrInfo::isTRANS(MI: I); |
1513 | // Cover both WAR and WAW |
1514 | return I.readsRegister(Reg: VDSTReg, TRI: &TRI) || I.modifiesRegister(Reg: VDSTReg, TRI: &TRI); |
1515 | }; |
1516 | auto IsExpiredFn = [&](const MachineInstr &I, int WaitStates) { |
1517 | if (WaitStates >= NoHazardWaitStates) |
1518 | return true; |
1519 | // Instructions which cause va_vdst==0 expire hazard |
1520 | return SIInstrInfo::isVMEM(MI: I) || SIInstrInfo::isDS(MI: I) || |
1521 | SIInstrInfo::isEXP(MI: I); |
1522 | }; |
1523 | auto GetWaitStatesFn = [](const MachineInstr &MI) { |
1524 | return SIInstrInfo::isVALU(MI) ? 1 : 0; |
1525 | }; |
1526 | |
1527 | DenseSet<const MachineBasicBlock *> Visited; |
1528 | auto Count = ::getWaitStatesSince(IsHazard: IsHazardFn, MBB: MI->getParent(), |
1529 | I: std::next(x: MI->getReverseIterator()), WaitStates: 0, |
1530 | IsExpired: IsExpiredFn, Visited, GetNumWaitStates: GetWaitStatesFn); |
1531 | |
1532 | // Transcendentals can execute in parallel to other VALUs. |
1533 | // This makes va_vdst count unusable with a mixture of VALU and TRANS. |
1534 | if (VisitedTrans) |
1535 | Count = 0; |
1536 | |
1537 | MachineOperand *WaitVdstOp = |
1538 | TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::waitvdst); |
1539 | WaitVdstOp->setImm(std::min(a: Count, b: NoHazardWaitStates)); |
1540 | |
1541 | return true; |
1542 | } |
1543 | |
1544 | bool GCNHazardRecognizer::fixLdsDirectVMEMHazard(MachineInstr *MI) { |
1545 | if (!SIInstrInfo::isLDSDIR(MI: *MI)) |
1546 | return false; |
1547 | |
1548 | const MachineOperand *VDST = TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::vdst); |
1549 | const Register VDSTReg = VDST->getReg(); |
1550 | |
1551 | auto IsHazardFn = [this, VDSTReg](const MachineInstr &I) { |
1552 | if (!SIInstrInfo::isVMEM(MI: I) && !SIInstrInfo::isDS(MI: I)) |
1553 | return false; |
1554 | return I.readsRegister(Reg: VDSTReg, TRI: &TRI) || I.modifiesRegister(Reg: VDSTReg, TRI: &TRI); |
1555 | }; |
1556 | bool LdsdirCanWait = ST.hasLdsWaitVMSRC(); |
1557 | // TODO: On GFX12 the hazard should expire on S_WAIT_LOADCNT/SAMPLECNT/BVHCNT |
1558 | // according to the type of VMEM instruction. |
1559 | auto IsExpiredFn = [this, LdsdirCanWait](const MachineInstr &I, int) { |
1560 | return SIInstrInfo::isVALU(MI: I) || SIInstrInfo::isEXP(MI: I) || |
1561 | (I.getOpcode() == AMDGPU::S_WAITCNT && !I.getOperand(i: 0).getImm()) || |
1562 | (I.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
1563 | AMDGPU::DepCtr::decodeFieldVmVsrc(Encoded: I.getOperand(i: 0).getImm()) == 0) || |
1564 | (LdsdirCanWait && SIInstrInfo::isLDSDIR(MI: I) && |
1565 | !TII.getNamedOperand(MI: I, OperandName: AMDGPU::OpName::waitvsrc)->getImm()); |
1566 | }; |
1567 | |
1568 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1569 | std::numeric_limits<int>::max()) |
1570 | return false; |
1571 | |
1572 | if (LdsdirCanWait) { |
1573 | TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::waitvsrc)->setImm(0); |
1574 | } else { |
1575 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1576 | MCID: TII.get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
1577 | .addImm(Val: AMDGPU::DepCtr::encodeFieldVmVsrc(VmVsrc: 0)); |
1578 | } |
1579 | |
1580 | return true; |
1581 | } |
1582 | |
1583 | bool GCNHazardRecognizer::fixVALUPartialForwardingHazard(MachineInstr *MI) { |
1584 | if (!ST.hasVALUPartialForwardingHazard()) |
1585 | return false; |
1586 | assert(!ST.hasExtendedWaitCounts()); |
1587 | |
1588 | if (!ST.isWave64() || !SIInstrInfo::isVALU(MI: *MI)) |
1589 | return false; |
1590 | |
1591 | SmallSetVector<Register, 4> SrcVGPRs; |
1592 | |
1593 | for (const MachineOperand &Use : MI->explicit_uses()) { |
1594 | if (Use.isReg() && TRI.isVGPR(MRI: MF.getRegInfo(), Reg: Use.getReg())) |
1595 | SrcVGPRs.insert(X: Use.getReg()); |
1596 | } |
1597 | |
1598 | // Only applies with >= 2 unique VGPR sources |
1599 | if (SrcVGPRs.size() <= 1) |
1600 | return false; |
1601 | |
1602 | // Look for the following pattern: |
1603 | // Va <- VALU [PreExecPos] |
1604 | // intv1 |
1605 | // Exec <- SALU [ExecPos] |
1606 | // intv2 |
1607 | // Vb <- VALU [PostExecPos] |
1608 | // intv3 |
1609 | // MI Va, Vb (WaitState = 0) |
1610 | // |
1611 | // Where: |
1612 | // intv1 + intv2 <= 2 VALUs |
1613 | // intv3 <= 4 VALUs |
1614 | // |
1615 | // If found, insert an appropriate S_WAITCNT_DEPCTR before MI. |
1616 | |
1617 | const int Intv1plus2MaxVALUs = 2; |
1618 | const int Intv3MaxVALUs = 4; |
1619 | const int IntvMaxVALUs = 6; |
1620 | const int NoHazardVALUWaitStates = IntvMaxVALUs + 2; |
1621 | |
1622 | struct StateType { |
1623 | SmallDenseMap<Register, int, 4> DefPos; |
1624 | int ExecPos = std::numeric_limits<int>::max(); |
1625 | int VALUs = 0; |
1626 | }; |
1627 | |
1628 | StateType State; |
1629 | |
1630 | // This overloads expiry testing with all the hazard detection |
1631 | auto IsHazardFn = [&, this](StateType &State, const MachineInstr &I) { |
1632 | // Too many VALU states have passed |
1633 | if (State.VALUs > NoHazardVALUWaitStates) |
1634 | return HazardExpired; |
1635 | |
1636 | // Instructions which cause va_vdst==0 expire hazard |
1637 | if (SIInstrInfo::isVMEM(MI: I) || SIInstrInfo::isDS(MI: I) || |
1638 | SIInstrInfo::isEXP(MI: I) || |
1639 | (I.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
1640 | AMDGPU::DepCtr::decodeFieldVaVdst(Encoded: I.getOperand(i: 0).getImm()) == 0)) |
1641 | return HazardExpired; |
1642 | |
1643 | // Track registers writes |
1644 | bool Changed = false; |
1645 | if (SIInstrInfo::isVALU(MI: I)) { |
1646 | for (Register Src : SrcVGPRs) { |
1647 | if (!State.DefPos.count(Val: Src) && I.modifiesRegister(Reg: Src, TRI: &TRI)) { |
1648 | State.DefPos[Src] = State.VALUs; |
1649 | Changed = true; |
1650 | } |
1651 | } |
1652 | } else if (SIInstrInfo::isSALU(MI: I)) { |
1653 | if (State.ExecPos == std::numeric_limits<int>::max()) { |
1654 | if (!State.DefPos.empty() && I.modifiesRegister(Reg: AMDGPU::EXEC, TRI: &TRI)) { |
1655 | State.ExecPos = State.VALUs; |
1656 | Changed = true; |
1657 | } |
1658 | } |
1659 | } |
1660 | |
1661 | // Early expiration: too many VALUs in intv3 |
1662 | if (State.VALUs > Intv3MaxVALUs && State.DefPos.empty()) |
1663 | return HazardExpired; |
1664 | |
1665 | // Only evaluate state if something changed |
1666 | if (!Changed) |
1667 | return NoHazardFound; |
1668 | |
1669 | // Determine positions of VALUs pre/post exec change |
1670 | if (State.ExecPos == std::numeric_limits<int>::max()) |
1671 | return NoHazardFound; |
1672 | |
1673 | int PreExecPos = std::numeric_limits<int>::max(); |
1674 | int PostExecPos = std::numeric_limits<int>::max(); |
1675 | |
1676 | for (auto Entry : State.DefPos) { |
1677 | int DefVALUs = Entry.second; |
1678 | if (DefVALUs != std::numeric_limits<int>::max()) { |
1679 | if (DefVALUs >= State.ExecPos) |
1680 | PreExecPos = std::min(a: PreExecPos, b: DefVALUs); |
1681 | else |
1682 | PostExecPos = std::min(a: PostExecPos, b: DefVALUs); |
1683 | } |
1684 | } |
1685 | |
1686 | // Need a VALUs post exec change |
1687 | if (PostExecPos == std::numeric_limits<int>::max()) |
1688 | return NoHazardFound; |
1689 | |
1690 | // Too many VALUs in intv3? |
1691 | int Intv3VALUs = PostExecPos; |
1692 | if (Intv3VALUs > Intv3MaxVALUs) |
1693 | return HazardExpired; |
1694 | |
1695 | // Too many VALUs in intv2? |
1696 | int Intv2VALUs = (State.ExecPos - PostExecPos) - 1; |
1697 | if (Intv2VALUs > Intv1plus2MaxVALUs) |
1698 | return HazardExpired; |
1699 | |
1700 | // Need a VALUs pre exec change |
1701 | if (PreExecPos == std::numeric_limits<int>::max()) |
1702 | return NoHazardFound; |
1703 | |
1704 | // Too many VALUs in intv1? |
1705 | int Intv1VALUs = PreExecPos - State.ExecPos; |
1706 | if (Intv1VALUs > Intv1plus2MaxVALUs) |
1707 | return HazardExpired; |
1708 | |
1709 | // Too many VALUs in intv1 + intv2 |
1710 | if (Intv1VALUs + Intv2VALUs > Intv1plus2MaxVALUs) |
1711 | return HazardExpired; |
1712 | |
1713 | return HazardFound; |
1714 | }; |
1715 | auto UpdateStateFn = [](StateType &State, const MachineInstr &MI) { |
1716 | if (SIInstrInfo::isVALU(MI)) |
1717 | State.VALUs += 1; |
1718 | }; |
1719 | |
1720 | DenseSet<const MachineBasicBlock *> Visited; |
1721 | if (!hasHazard<StateType>(State, IsHazard: IsHazardFn, UpdateState: UpdateStateFn, MBB: MI->getParent(), |
1722 | I: std::next(x: MI->getReverseIterator()), Visited)) |
1723 | return false; |
1724 | |
1725 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1726 | MCID: TII.get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
1727 | .addImm(Val: 0x0fff); |
1728 | |
1729 | return true; |
1730 | } |
1731 | |
1732 | bool GCNHazardRecognizer::fixVALUTransUseHazard(MachineInstr *MI) { |
1733 | if (!ST.hasVALUTransUseHazard()) |
1734 | return false; |
1735 | assert(!ST.hasExtendedWaitCounts()); |
1736 | |
1737 | if (!SIInstrInfo::isVALU(MI: *MI)) |
1738 | return false; |
1739 | |
1740 | SmallSet<Register, 4> SrcVGPRs; |
1741 | |
1742 | for (const MachineOperand &Use : MI->explicit_uses()) { |
1743 | if (Use.isReg() && TRI.isVGPR(MRI: MF.getRegInfo(), Reg: Use.getReg())) |
1744 | SrcVGPRs.insert(V: Use.getReg()); |
1745 | } |
1746 | |
1747 | // Look for the following pattern: |
1748 | // Va <- TRANS VALU |
1749 | // intv |
1750 | // MI Va (WaitState = 0) |
1751 | // |
1752 | // Where: |
1753 | // intv <= 5 VALUs / 1 TRANS |
1754 | // |
1755 | // If found, insert an appropriate S_WAITCNT_DEPCTR before MI. |
1756 | |
1757 | const int IntvMaxVALUs = 5; |
1758 | const int IntvMaxTRANS = 1; |
1759 | |
1760 | struct StateType { |
1761 | int VALUs = 0; |
1762 | int TRANS = 0; |
1763 | }; |
1764 | |
1765 | StateType State; |
1766 | |
1767 | // This overloads expiry testing with all the hazard detection |
1768 | auto IsHazardFn = [&, this](StateType &State, const MachineInstr &I) { |
1769 | // Too many VALU states have passed |
1770 | if (State.VALUs > IntvMaxVALUs || State.TRANS > IntvMaxTRANS) |
1771 | return HazardExpired; |
1772 | |
1773 | // Instructions which cause va_vdst==0 expire hazard |
1774 | if (SIInstrInfo::isVMEM(MI: I) || SIInstrInfo::isDS(MI: I) || |
1775 | SIInstrInfo::isEXP(MI: I) || |
1776 | (I.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
1777 | I.getOperand(i: 0).getImm() == 0x0fff)) |
1778 | return HazardExpired; |
1779 | |
1780 | // Track registers writes |
1781 | if (SIInstrInfo::isTRANS(MI: I)) { |
1782 | for (Register Src : SrcVGPRs) { |
1783 | if (I.modifiesRegister(Reg: Src, TRI: &TRI)) { |
1784 | return HazardFound; |
1785 | } |
1786 | } |
1787 | } |
1788 | |
1789 | return NoHazardFound; |
1790 | }; |
1791 | auto UpdateStateFn = [](StateType &State, const MachineInstr &MI) { |
1792 | if (SIInstrInfo::isVALU(MI)) |
1793 | State.VALUs += 1; |
1794 | if (SIInstrInfo::isTRANS(MI)) |
1795 | State.TRANS += 1; |
1796 | }; |
1797 | |
1798 | DenseSet<const MachineBasicBlock *> Visited; |
1799 | if (!hasHazard<StateType>(State, IsHazard: IsHazardFn, UpdateState: UpdateStateFn, MBB: MI->getParent(), |
1800 | I: std::next(x: MI->getReverseIterator()), Visited)) |
1801 | return false; |
1802 | |
1803 | // Hazard is observed - insert a wait on va_dst counter to ensure hazard is |
1804 | // avoided. |
1805 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), |
1806 | MCID: TII.get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
1807 | .addImm(Val: AMDGPU::DepCtr::encodeFieldVaVdst(VaVdst: 0)); |
1808 | |
1809 | return true; |
1810 | } |
1811 | |
1812 | bool GCNHazardRecognizer::fixWMMAHazards(MachineInstr *MI) { |
1813 | if (!SIInstrInfo::isWMMA(MI: *MI) && !SIInstrInfo::isSWMMAC(MI: *MI)) |
1814 | return false; |
1815 | |
1816 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1817 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
1818 | |
1819 | auto IsHazardFn = [MI, TII, TRI, this](const MachineInstr &I) { |
1820 | if (!SIInstrInfo::isWMMA(MI: I) && !SIInstrInfo::isSWMMAC(MI: I)) |
1821 | return false; |
1822 | |
1823 | // Src0(matrix A) or Src1(matrix B) of the current wmma instruction overlaps |
1824 | // with the dest(matrix D) of the previous wmma. |
1825 | const Register CurSrc0Reg = |
1826 | TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src0)->getReg(); |
1827 | const Register CurSrc1Reg = |
1828 | TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src1)->getReg(); |
1829 | |
1830 | const Register PrevDstReg = |
1831 | TII->getNamedOperand(MI: I, OperandName: AMDGPU::OpName::vdst)->getReg(); |
1832 | |
1833 | if (TRI->regsOverlap(RegA: PrevDstReg, RegB: CurSrc0Reg) || |
1834 | TRI->regsOverlap(RegA: PrevDstReg, RegB: CurSrc1Reg)) { |
1835 | return true; |
1836 | } |
1837 | |
1838 | // GFX12+ allows overlap of matrix C with PrevDstReg (hardware will stall) |
1839 | // but Index can't overlap with PrevDstReg. |
1840 | if (AMDGPU::isGFX12Plus(STI: ST)) { |
1841 | if (SIInstrInfo::isSWMMAC(MI: *MI)) { |
1842 | const Register CurIndex = |
1843 | TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src2)->getReg(); |
1844 | if (TRI->regsOverlap(RegA: PrevDstReg, RegB: CurIndex)) |
1845 | return true; |
1846 | } |
1847 | return false; |
1848 | } |
1849 | |
1850 | return false; |
1851 | }; |
1852 | |
1853 | auto IsExpiredFn = [](const MachineInstr &I, int) { |
1854 | return SIInstrInfo::isVALU(MI: I); |
1855 | }; |
1856 | |
1857 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
1858 | std::numeric_limits<int>::max()) |
1859 | return false; |
1860 | |
1861 | BuildMI(BB&: *MI->getParent(), I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: AMDGPU::V_NOP_e32)); |
1862 | |
1863 | return true; |
1864 | } |
1865 | |
1866 | bool GCNHazardRecognizer::fixShift64HighRegBug(MachineInstr *MI) { |
1867 | if (!ST.hasShift64HighRegBug()) |
1868 | return false; |
1869 | assert(!ST.hasExtendedWaitCounts()); |
1870 | |
1871 | switch (MI->getOpcode()) { |
1872 | default: |
1873 | return false; |
1874 | case AMDGPU::V_LSHLREV_B64_e64: |
1875 | case AMDGPU::V_LSHRREV_B64_e64: |
1876 | case AMDGPU::V_ASHRREV_I64_e64: |
1877 | break; |
1878 | } |
1879 | |
1880 | MachineOperand *Amt = TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src0); |
1881 | if (!Amt->isReg()) |
1882 | return false; |
1883 | |
1884 | Register AmtReg = Amt->getReg(); |
1885 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1886 | // Check if this is a last VGPR in the allocation block. |
1887 | if (!TRI.isVGPR(MRI, Reg: AmtReg) || ((AmtReg - AMDGPU::VGPR0) & 7) != 7) |
1888 | return false; |
1889 | |
1890 | if (AmtReg != AMDGPU::VGPR255 && MRI.isPhysRegUsed(PhysReg: AmtReg + 1)) |
1891 | return false; |
1892 | |
1893 | MachineOperand *Src1 = TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src1); |
1894 | bool OverlappedSrc = Src1->isReg() && TRI.regsOverlap(RegA: Src1->getReg(), RegB: AmtReg); |
1895 | bool OverlappedDst = MI->modifiesRegister(Reg: AmtReg, TRI: &TRI); |
1896 | bool Overlapped = OverlappedSrc || OverlappedDst; |
1897 | |
1898 | assert(!OverlappedDst || !OverlappedSrc || |
1899 | Src1->getReg() == MI->getOperand(0).getReg()); |
1900 | assert(ST.needsAlignedVGPRs()); |
1901 | static_assert(AMDGPU::VGPR0 + 1 == AMDGPU::VGPR1); |
1902 | |
1903 | Register NewReg; |
1904 | for (MCRegister Reg : Overlapped ? AMDGPU::VReg_64_Align2RegClass |
1905 | : AMDGPU::VGPR_32RegClass) { |
1906 | if (!MI->modifiesRegister(Reg, TRI: &TRI) && !MI->readsRegister(Reg, TRI: &TRI)) { |
1907 | NewReg = Reg; |
1908 | break; |
1909 | } |
1910 | } |
1911 | |
1912 | Register NewAmt = Overlapped ? (Register)TRI.getSubReg(Reg: NewReg, Idx: AMDGPU::sub1) |
1913 | : NewReg; |
1914 | Register NewAmtLo; |
1915 | |
1916 | if (Overlapped) |
1917 | NewAmtLo = TRI.getSubReg(Reg: NewReg, Idx: AMDGPU::sub0); |
1918 | |
1919 | DebugLoc DL = MI->getDebugLoc(); |
1920 | MachineBasicBlock *MBB = MI->getParent(); |
1921 | // Insert a full wait count because found register might be pending a wait. |
1922 | BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_WAITCNT)) |
1923 | .addImm(Val: 0); |
1924 | |
1925 | // Insert V_SWAP_B32 instruction(s) and run hazard recognizer on them. |
1926 | if (Overlapped) |
1927 | runOnInstruction( |
1928 | MI: BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::V_SWAP_B32), DestReg: NewAmtLo) |
1929 | .addDef(RegNo: AmtReg - 1) |
1930 | .addReg(RegNo: AmtReg - 1, flags: RegState::Undef) |
1931 | .addReg(RegNo: NewAmtLo, flags: RegState::Undef)); |
1932 | runOnInstruction(MI: BuildMI(BB&: *MBB, I: MI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::V_SWAP_B32), DestReg: NewAmt) |
1933 | .addDef(RegNo: AmtReg) |
1934 | .addReg(RegNo: AmtReg, flags: RegState::Undef) |
1935 | .addReg(RegNo: NewAmt, flags: RegState::Undef)); |
1936 | |
1937 | // Instructions emitted after the current instruction will be processed by the |
1938 | // parent loop of the hazard recognizer in a natural way. |
1939 | BuildMI(BB&: *MBB, I: std::next(x: MI->getIterator()), MIMD: DL, MCID: TII.get(Opcode: AMDGPU::V_SWAP_B32), |
1940 | DestReg: AmtReg) |
1941 | .addDef(RegNo: NewAmt) |
1942 | .addReg(RegNo: NewAmt) |
1943 | .addReg(RegNo: AmtReg); |
1944 | if (Overlapped) |
1945 | BuildMI(BB&: *MBB, I: std::next(x: MI->getIterator()), MIMD: DL, MCID: TII.get(Opcode: AMDGPU::V_SWAP_B32), |
1946 | DestReg: AmtReg - 1) |
1947 | .addDef(RegNo: NewAmtLo) |
1948 | .addReg(RegNo: NewAmtLo) |
1949 | .addReg(RegNo: AmtReg - 1); |
1950 | |
1951 | // Re-running hazard recognizer on the modified instruction is not necessary, |
1952 | // inserted V_SWAP_B32 has already both read and write new registers so |
1953 | // hazards related to these register has already been handled. |
1954 | Amt->setReg(NewAmt); |
1955 | Amt->setIsKill(false); |
1956 | // We do not update liveness, so verifier may see it as undef. |
1957 | Amt->setIsUndef(); |
1958 | if (OverlappedDst) |
1959 | MI->getOperand(i: 0).setReg(NewReg); |
1960 | if (OverlappedSrc) { |
1961 | Src1->setReg(NewReg); |
1962 | Src1->setIsKill(false); |
1963 | Src1->setIsUndef(); |
1964 | } |
1965 | |
1966 | return true; |
1967 | } |
1968 | |
1969 | int GCNHazardRecognizer::checkNSAtoVMEMHazard(MachineInstr *MI) { |
1970 | int NSAtoVMEMWaitStates = 1; |
1971 | |
1972 | if (!ST.hasNSAtoVMEMBug()) |
1973 | return 0; |
1974 | |
1975 | if (!SIInstrInfo::isMUBUF(MI: *MI) && !SIInstrInfo::isMTBUF(MI: *MI)) |
1976 | return 0; |
1977 | |
1978 | const SIInstrInfo *TII = ST.getInstrInfo(); |
1979 | const auto *Offset = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::offset); |
1980 | if (!Offset || (Offset->getImm() & 6) == 0) |
1981 | return 0; |
1982 | |
1983 | auto IsHazardFn = [TII](const MachineInstr &I) { |
1984 | if (!SIInstrInfo::isMIMG(MI: I)) |
1985 | return false; |
1986 | const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Opc: I.getOpcode()); |
1987 | return Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA && |
1988 | TII->getInstSizeInBytes(MI: I) >= 16; |
1989 | }; |
1990 | |
1991 | return NSAtoVMEMWaitStates - getWaitStatesSince(IsHazard: IsHazardFn, Limit: 1); |
1992 | } |
1993 | |
1994 | int GCNHazardRecognizer::checkFPAtomicToDenormModeHazard(MachineInstr *MI) { |
1995 | int FPAtomicToDenormModeWaitStates = 3; |
1996 | |
1997 | if (!ST.hasFPAtomicToDenormModeHazard()) |
1998 | return 0; |
1999 | assert(!ST.hasExtendedWaitCounts()); |
2000 | |
2001 | if (MI->getOpcode() != AMDGPU::S_DENORM_MODE) |
2002 | return 0; |
2003 | |
2004 | auto IsHazardFn = [](const MachineInstr &I) { |
2005 | if (!SIInstrInfo::isVMEM(MI: I)) |
2006 | return false; |
2007 | return SIInstrInfo::isFPAtomic(MI: I); |
2008 | }; |
2009 | |
2010 | auto IsExpiredFn = [](const MachineInstr &MI, int WaitStates) { |
2011 | if (WaitStates >= 3 || SIInstrInfo::isVALU(MI)) |
2012 | return true; |
2013 | |
2014 | switch (MI.getOpcode()) { |
2015 | case AMDGPU::S_WAITCNT: |
2016 | case AMDGPU::S_WAITCNT_VSCNT: |
2017 | case AMDGPU::S_WAITCNT_VMCNT: |
2018 | case AMDGPU::S_WAITCNT_EXPCNT: |
2019 | case AMDGPU::S_WAITCNT_LGKMCNT: |
2020 | case AMDGPU::S_WAIT_IDLE: |
2021 | return true; |
2022 | default: |
2023 | break; |
2024 | } |
2025 | |
2026 | return false; |
2027 | }; |
2028 | |
2029 | return FPAtomicToDenormModeWaitStates - |
2030 | ::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn); |
2031 | } |
2032 | |
2033 | int GCNHazardRecognizer::checkMAIHazards(MachineInstr *MI) { |
2034 | assert(SIInstrInfo::isMAI(*MI)); |
2035 | |
2036 | return ST.hasGFX90AInsts() ? checkMAIHazards90A(MI) : checkMAIHazards908(MI); |
2037 | } |
2038 | |
2039 | int GCNHazardRecognizer::checkMFMAPadding(MachineInstr *MI) { |
2040 | // Early exit if no padding is requested. |
2041 | if (MFMAPaddingRatio == 0) |
2042 | return 0; |
2043 | |
2044 | const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); |
2045 | if (!SIInstrInfo::isMFMA(MI: *MI) || MFI->getOccupancy() < 2) |
2046 | return 0; |
2047 | |
2048 | int NeighborMFMALatency = 0; |
2049 | auto IsNeighboringMFMA = [&NeighborMFMALatency, |
2050 | this](const MachineInstr &MI) { |
2051 | if (!SIInstrInfo::isMFMA(MI)) |
2052 | return false; |
2053 | |
2054 | NeighborMFMALatency = this->getMFMAPipelineWaitStates(MI); |
2055 | return true; |
2056 | }; |
2057 | |
2058 | const int MaxMFMAPipelineWaitStates = 16; |
2059 | int WaitStatesSinceNeighborMFMA = |
2060 | getWaitStatesSince(IsHazard: IsNeighboringMFMA, Limit: MaxMFMAPipelineWaitStates); |
2061 | |
2062 | int NeighborMFMAPaddingNeeded = |
2063 | (NeighborMFMALatency * MFMAPaddingRatio / 100) - |
2064 | WaitStatesSinceNeighborMFMA; |
2065 | |
2066 | return std::max(a: 0, b: NeighborMFMAPaddingNeeded); |
2067 | } |
2068 | |
2069 | int GCNHazardRecognizer::checkMAIHazards908(MachineInstr *MI) { |
2070 | int WaitStatesNeeded = 0; |
2071 | unsigned Opc = MI->getOpcode(); |
2072 | |
2073 | auto IsVALUFn = [](const MachineInstr &MI) { |
2074 | return SIInstrInfo::isVALU(MI) || MI.isInlineAsm(); |
2075 | }; |
2076 | |
2077 | if (Opc != AMDGPU::V_ACCVGPR_READ_B32_e64) { // MFMA or v_accvgpr_write |
2078 | const int LegacyVALUWritesVGPRWaitStates = 2; |
2079 | const int VALUWritesExecWaitStates = 4; |
2080 | const int MaxWaitStates = 4; |
2081 | |
2082 | int WaitStatesNeededForUse = VALUWritesExecWaitStates - |
2083 | getWaitStatesSinceDef(Reg: AMDGPU::EXEC, IsHazardDef: IsVALUFn, Limit: MaxWaitStates); |
2084 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2085 | |
2086 | if (WaitStatesNeeded < MaxWaitStates) { |
2087 | for (const MachineOperand &Use : MI->explicit_uses()) { |
2088 | const int MaxWaitStates = 2; |
2089 | |
2090 | if (!Use.isReg() || !TRI.isVGPR(MRI: MF.getRegInfo(), Reg: Use.getReg())) |
2091 | continue; |
2092 | |
2093 | int WaitStatesNeededForUse = LegacyVALUWritesVGPRWaitStates - |
2094 | getWaitStatesSinceDef(Reg: Use.getReg(), IsHazardDef: IsVALUFn, Limit: MaxWaitStates); |
2095 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2096 | |
2097 | if (WaitStatesNeeded == MaxWaitStates) |
2098 | break; |
2099 | } |
2100 | } |
2101 | } |
2102 | |
2103 | for (const MachineOperand &Op : MI->explicit_operands()) { |
2104 | if (!Op.isReg() || !TRI.isAGPR(MRI: MF.getRegInfo(), Reg: Op.getReg())) |
2105 | continue; |
2106 | |
2107 | if (Op.isDef() && Opc != AMDGPU::V_ACCVGPR_WRITE_B32_e64) |
2108 | continue; |
2109 | |
2110 | const int MFMAWritesAGPROverlappedSrcABWaitStates = 4; |
2111 | const int MFMAWritesAGPROverlappedSrcCWaitStates = 2; |
2112 | const int MFMA4x4WritesAGPRAccVgprReadWaitStates = 4; |
2113 | const int MFMA16x16WritesAGPRAccVgprReadWaitStates = 10; |
2114 | const int MFMA32x32WritesAGPRAccVgprReadWaitStates = 18; |
2115 | const int MFMA4x4WritesAGPRAccVgprWriteWaitStates = 1; |
2116 | const int MFMA16x16WritesAGPRAccVgprWriteWaitStates = 7; |
2117 | const int MFMA32x32WritesAGPRAccVgprWriteWaitStates = 15; |
2118 | const int MaxWaitStates = 18; |
2119 | Register Reg = Op.getReg(); |
2120 | unsigned HazardDefLatency = 0; |
2121 | |
2122 | auto IsOverlappedMFMAFn = [Reg, &HazardDefLatency, |
2123 | this](const MachineInstr &MI) { |
2124 | if (!SIInstrInfo::isMFMA(MI)) |
2125 | return false; |
2126 | Register DstReg = MI.getOperand(i: 0).getReg(); |
2127 | if (DstReg == Reg) |
2128 | return false; |
2129 | HazardDefLatency = |
2130 | std::max(a: HazardDefLatency, b: TSchedModel.computeInstrLatency(MI: &MI)); |
2131 | return TRI.regsOverlap(RegA: DstReg, RegB: Reg); |
2132 | }; |
2133 | |
2134 | int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsHazardDef: IsOverlappedMFMAFn, |
2135 | Limit: MaxWaitStates); |
2136 | int NeedWaitStates = MFMAWritesAGPROverlappedSrcABWaitStates; |
2137 | int SrcCIdx = AMDGPU::getNamedOperandIdx(Opcode: Opc, Name: AMDGPU::OpName::src2); |
2138 | int OpNo = Op.getOperandNo(); |
2139 | if (OpNo == SrcCIdx) { |
2140 | NeedWaitStates = MFMAWritesAGPROverlappedSrcCWaitStates; |
2141 | } else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) { |
2142 | switch (HazardDefLatency) { |
2143 | case 2: NeedWaitStates = MFMA4x4WritesAGPRAccVgprReadWaitStates; |
2144 | break; |
2145 | case 8: NeedWaitStates = MFMA16x16WritesAGPRAccVgprReadWaitStates; |
2146 | break; |
2147 | case 16: [[fallthrough]]; |
2148 | default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprReadWaitStates; |
2149 | break; |
2150 | } |
2151 | } else if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) { |
2152 | switch (HazardDefLatency) { |
2153 | case 2: NeedWaitStates = MFMA4x4WritesAGPRAccVgprWriteWaitStates; |
2154 | break; |
2155 | case 8: NeedWaitStates = MFMA16x16WritesAGPRAccVgprWriteWaitStates; |
2156 | break; |
2157 | case 16: [[fallthrough]]; |
2158 | default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprWriteWaitStates; |
2159 | break; |
2160 | } |
2161 | } |
2162 | |
2163 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; |
2164 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2165 | |
2166 | if (WaitStatesNeeded == MaxWaitStates) |
2167 | return WaitStatesNeeded; // Early exit. |
2168 | |
2169 | auto IsAccVgprWriteFn = [Reg, this](const MachineInstr &MI) { |
2170 | if (MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64) |
2171 | return false; |
2172 | Register DstReg = MI.getOperand(i: 0).getReg(); |
2173 | return TRI.regsOverlap(RegA: Reg, RegB: DstReg); |
2174 | }; |
2175 | |
2176 | const int AccVGPRWriteMFMAReadSrcCWaitStates = 1; |
2177 | const int AccVGPRWriteMFMAReadSrcABWaitStates = 3; |
2178 | const int AccVGPRWriteAccVgprReadWaitStates = 3; |
2179 | NeedWaitStates = AccVGPRWriteMFMAReadSrcABWaitStates; |
2180 | if (OpNo == SrcCIdx) |
2181 | NeedWaitStates = AccVGPRWriteMFMAReadSrcCWaitStates; |
2182 | else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) |
2183 | NeedWaitStates = AccVGPRWriteAccVgprReadWaitStates; |
2184 | |
2185 | WaitStatesNeededForUse = NeedWaitStates - |
2186 | getWaitStatesSinceDef(Reg, IsHazardDef: IsAccVgprWriteFn, Limit: MaxWaitStates); |
2187 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2188 | |
2189 | if (WaitStatesNeeded == MaxWaitStates) |
2190 | return WaitStatesNeeded; // Early exit. |
2191 | } |
2192 | |
2193 | if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) { |
2194 | const int MFMA4x4ReadSrcCAccVgprWriteWaitStates = 0; |
2195 | const int MFMA16x16ReadSrcCAccVgprWriteWaitStates = 5; |
2196 | const int MFMA32x32ReadSrcCAccVgprWriteWaitStates = 13; |
2197 | const int MaxWaitStates = 13; |
2198 | Register DstReg = MI->getOperand(i: 0).getReg(); |
2199 | unsigned HazardDefLatency = 0; |
2200 | |
2201 | auto IsSrcCMFMAFn = [DstReg, &HazardDefLatency, |
2202 | this](const MachineInstr &MI) { |
2203 | if (!SIInstrInfo::isMFMA(MI)) |
2204 | return false; |
2205 | Register Reg = TII.getNamedOperand(MI, OperandName: AMDGPU::OpName::src2)->getReg(); |
2206 | HazardDefLatency = |
2207 | std::max(a: HazardDefLatency, b: TSchedModel.computeInstrLatency(MI: &MI)); |
2208 | return TRI.regsOverlap(RegA: Reg, RegB: DstReg); |
2209 | }; |
2210 | |
2211 | int WaitStatesSince = getWaitStatesSince(IsHazard: IsSrcCMFMAFn, Limit: MaxWaitStates); |
2212 | int NeedWaitStates; |
2213 | switch (HazardDefLatency) { |
2214 | case 2: NeedWaitStates = MFMA4x4ReadSrcCAccVgprWriteWaitStates; |
2215 | break; |
2216 | case 8: NeedWaitStates = MFMA16x16ReadSrcCAccVgprWriteWaitStates; |
2217 | break; |
2218 | case 16: [[fallthrough]]; |
2219 | default: NeedWaitStates = MFMA32x32ReadSrcCAccVgprWriteWaitStates; |
2220 | break; |
2221 | } |
2222 | |
2223 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSince; |
2224 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2225 | } |
2226 | |
2227 | // Pad neighboring MFMA with noops for better inter-wave performance. |
2228 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: checkMFMAPadding(MI)); |
2229 | |
2230 | return WaitStatesNeeded; |
2231 | } |
2232 | |
2233 | static int |
2234 | GFX940_XDL_N_PassWritesVGPROverlappedXDLOrSMFMASrcCWaitStates(int NumPasses, |
2235 | bool IsGFX950) { |
2236 | // xdl def cycles | gfx940 | gfx950 |
2237 | // 2 pass | 3 4 |
2238 | // 4 pass | 5 6 |
2239 | // 8 pass | 9 10 |
2240 | // 16 pass | 17 18 |
2241 | return NumPasses + 1 + IsGFX950; |
2242 | } |
2243 | |
2244 | static int |
2245 | GFX940_XDL_N_PassWritesVGPROverlappedSGEMMDGEMMSrcCWaitStates(int NumPasses, |
2246 | bool IsGFX950) { |
2247 | // xdl def cycles | gfx940 | gfx950 |
2248 | // 2 pass | 3 3 |
2249 | // 4 pass | 5 6 |
2250 | // 8 pass | 9 10 |
2251 | // 16 pass | 17 18 |
2252 | return NumPasses + 1 + (NumPasses != 2 && IsGFX950); |
2253 | } |
2254 | |
2255 | static int |
2256 | GFX940_SMFMA_N_PassWritesVGPROverlappedSMFMASrcCWaitStates(int NumPasses) { |
2257 | // 2 pass -> 2 |
2258 | // 4 pass -> 4 |
2259 | // 8 pass -> 8 |
2260 | // 16 pass -> 16 |
2261 | return NumPasses; |
2262 | } |
2263 | |
2264 | static int |
2265 | GFX940_SMFMA_N_PassWritesVGPROverlappedSrcABWaitStates(int NumPasses) { |
2266 | // 2 pass -> 4 |
2267 | // 4 pass -> 6 |
2268 | // 8 pass -> 10 |
2269 | // 16 pass -> 18 |
2270 | return NumPasses + 2; |
2271 | } |
2272 | |
2273 | static int GFX940_XDL_N_PassWritesVGPROverlappedSrcABWaitStates(int NumPasses, |
2274 | bool IsGFX950) { |
2275 | // xdl def cycles | gfx942 | gfx950 |
2276 | // 2 pass | 5 5 |
2277 | // 4 pass | 7 8 |
2278 | // 8 pass | 11 12 |
2279 | // 16 pass | 19 20 |
2280 | return NumPasses + 3 + (NumPasses != 2 && IsGFX950); |
2281 | } |
2282 | |
2283 | int GCNHazardRecognizer::checkMAIHazards90A(MachineInstr *MI) { |
2284 | int WaitStatesNeeded = 0; |
2285 | unsigned Opc = MI->getOpcode(); |
2286 | |
2287 | auto IsLegacyVALUFn = [](const MachineInstr &MI) { |
2288 | return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMFMA(MI); |
2289 | }; |
2290 | |
2291 | auto IsLegacyVALUNotDotFn = [](const MachineInstr &MI) { |
2292 | return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMFMA(MI) && |
2293 | !SIInstrInfo::isDOT(MI); |
2294 | }; |
2295 | |
2296 | if (!SIInstrInfo::isMFMA(MI: *MI)) |
2297 | return WaitStatesNeeded; |
2298 | |
2299 | const int VALUWritesExecWaitStates = 4; |
2300 | int WaitStatesNeededForUse = VALUWritesExecWaitStates - |
2301 | getWaitStatesSinceDef(Reg: AMDGPU::EXEC, IsHazardDef: IsLegacyVALUFn, |
2302 | Limit: VALUWritesExecWaitStates); |
2303 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2304 | |
2305 | int SrcCIdx = AMDGPU::getNamedOperandIdx(Opcode: Opc, Name: AMDGPU::OpName::src2); |
2306 | |
2307 | // Loop for both DGEMM and S/HGEMM 2nd instruction. |
2308 | for (const MachineOperand &Use : MI->explicit_uses()) { |
2309 | const int LegacyVALUNotDotWritesVGPRWaitStates = 2; |
2310 | const int SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates = 2; |
2311 | const int SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates = 8; |
2312 | const int SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates = 16; |
2313 | const int SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates = 3; |
2314 | const int SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates = 9; |
2315 | const int SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates = 17; |
2316 | const int DMFMA16x16WritesVGPROverlappedSrcCWaitStates = 9; |
2317 | const int GFX950_DMFMA16x16WritesVGPROverlappedSrcCWaitStates = 17; |
2318 | const int DMFMA4x4WritesVGPROverlappedSrcCWaitStates = 4; |
2319 | const int SMFMA4x4WritesVGPROverlappedSrcABWaitStates = 5; |
2320 | const int SMFMA16x16WritesVGPROverlappedSrcABWaitStates = 11; |
2321 | const int SMFMA32x32WritesVGPROverlappedSrcABWaitStates = 19; |
2322 | const int DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates = 6; |
2323 | const int DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates = 11; |
2324 | const int GFX950_DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates = 19; |
2325 | const int DMFMA4x4WritesVGPRFullSrcCWaitStates = 4; |
2326 | const int GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates = 2; |
2327 | const int MaxWaitStates = 19; |
2328 | |
2329 | if (!Use.isReg()) |
2330 | continue; |
2331 | Register Reg = Use.getReg(); |
2332 | bool FullReg; |
2333 | const MachineInstr *MI1; |
2334 | |
2335 | auto IsOverlappedMFMAFn = [Reg, &FullReg, &MI1, |
2336 | this](const MachineInstr &MI) { |
2337 | if (!SIInstrInfo::isMFMA(MI)) |
2338 | return false; |
2339 | Register DstReg = MI.getOperand(i: 0).getReg(); |
2340 | FullReg = (DstReg == Reg); |
2341 | MI1 = &MI; |
2342 | return TRI.regsOverlap(RegA: DstReg, RegB: Reg); |
2343 | }; |
2344 | |
2345 | WaitStatesNeededForUse = LegacyVALUNotDotWritesVGPRWaitStates - |
2346 | getWaitStatesSinceDef(Reg, IsHazardDef: IsLegacyVALUNotDotFn, Limit: MaxWaitStates); |
2347 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2348 | |
2349 | int NumWaitStates = |
2350 | getWaitStatesSinceDef(Reg, IsHazardDef: IsOverlappedMFMAFn, Limit: MaxWaitStates); |
2351 | if (NumWaitStates == std::numeric_limits<int>::max()) |
2352 | continue; |
2353 | |
2354 | int OpNo = Use.getOperandNo(); |
2355 | unsigned Opc1 = MI1->getOpcode(); |
2356 | int NeedWaitStates = 0; |
2357 | if (OpNo == SrcCIdx) { |
2358 | if (!SIInstrInfo::isDGEMM(Opcode: Opc) && |
2359 | (!ST.hasGFX940Insts() && SIInstrInfo::isDGEMM(Opcode: Opc1))) { |
2360 | NeedWaitStates = 0; |
2361 | } else if (FullReg) { |
2362 | if ((Opc == AMDGPU::V_MFMA_F64_4X4X4F64_e64 || |
2363 | Opc == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64) && |
2364 | (Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_e64 || |
2365 | Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64)) |
2366 | NeedWaitStates = DMFMA4x4WritesVGPRFullSrcCWaitStates; |
2367 | else if (ST.hasGFX940Insts() && |
2368 | TSchedModel.computeInstrLatency(MI: MI1) == 2) |
2369 | NeedWaitStates = GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates; |
2370 | } else { |
2371 | switch (Opc1) { |
2372 | case AMDGPU::V_MFMA_F64_16X16X4F64_e64: |
2373 | case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64: |
2374 | case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64: |
2375 | case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64: |
2376 | if (!TII.isXDL(MI: *MI)) |
2377 | NeedWaitStates = |
2378 | ST.hasGFX950Insts() |
2379 | ? GFX950_DMFMA16x16WritesVGPROverlappedSrcCWaitStates |
2380 | : DMFMA16x16WritesVGPROverlappedSrcCWaitStates; |
2381 | break; |
2382 | case AMDGPU::V_MFMA_F64_4X4X4F64_e64: |
2383 | case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64: |
2384 | if (!TII.isXDL(MI: *MI)) |
2385 | NeedWaitStates = DMFMA4x4WritesVGPROverlappedSrcCWaitStates; |
2386 | break; |
2387 | default: |
2388 | int NumPasses = TSchedModel.computeInstrLatency(MI: MI1); |
2389 | if (ST.hasGFX940Insts()) { |
2390 | if (TII.isXDL(MI: *MI) && !TII.isXDL(MI: *MI1)) |
2391 | break; |
2392 | |
2393 | NeedWaitStates = |
2394 | TII.isXDL(MI: *MI1) |
2395 | ? (TII.isXDL(MI: *MI) |
2396 | ? GFX940_XDL_N_PassWritesVGPROverlappedXDLOrSMFMASrcCWaitStates( |
2397 | NumPasses, IsGFX950: ST.hasGFX950Insts()) |
2398 | : GFX940_XDL_N_PassWritesVGPROverlappedSGEMMDGEMMSrcCWaitStates( |
2399 | NumPasses, IsGFX950: ST.hasGFX950Insts())) |
2400 | : GFX940_SMFMA_N_PassWritesVGPROverlappedSMFMASrcCWaitStates( |
2401 | NumPasses); |
2402 | break; |
2403 | } |
2404 | |
2405 | switch (NumPasses) { |
2406 | case 2: |
2407 | NeedWaitStates = |
2408 | SIInstrInfo::isDGEMM(Opcode: Opc) |
2409 | ? SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates |
2410 | : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates; |
2411 | break; |
2412 | case 8: |
2413 | NeedWaitStates = |
2414 | SIInstrInfo::isDGEMM(Opcode: Opc) |
2415 | ? SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates |
2416 | : SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates; |
2417 | break; |
2418 | case 16: |
2419 | NeedWaitStates = |
2420 | SIInstrInfo::isDGEMM(Opcode: Opc) |
2421 | ? SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates |
2422 | : SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates; |
2423 | break; |
2424 | default: |
2425 | llvm_unreachable("unexpected number of passes" ); |
2426 | } |
2427 | } |
2428 | } |
2429 | } else { |
2430 | switch (Opc1) { |
2431 | case AMDGPU::V_MFMA_F64_16X16X4F64_e64: |
2432 | case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64: |
2433 | case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64: |
2434 | case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64: |
2435 | NeedWaitStates = |
2436 | ST.hasGFX950Insts() |
2437 | ? GFX950_DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates |
2438 | : DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates; |
2439 | break; |
2440 | case AMDGPU::V_MFMA_F64_4X4X4F64_e64: |
2441 | case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64: |
2442 | NeedWaitStates = DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates; |
2443 | break; |
2444 | default: |
2445 | int NumPasses = TSchedModel.computeInstrLatency(MI: MI1); |
2446 | |
2447 | if (ST.hasGFX940Insts()) { |
2448 | NeedWaitStates = |
2449 | TII.isXDL(MI: *MI1) |
2450 | ? GFX940_XDL_N_PassWritesVGPROverlappedSrcABWaitStates( |
2451 | NumPasses, IsGFX950: ST.hasGFX950Insts()) |
2452 | : GFX940_SMFMA_N_PassWritesVGPROverlappedSrcABWaitStates( |
2453 | NumPasses); |
2454 | break; |
2455 | } |
2456 | |
2457 | switch (NumPasses) { |
2458 | case 2: |
2459 | NeedWaitStates = SMFMA4x4WritesVGPROverlappedSrcABWaitStates; |
2460 | break; |
2461 | case 4: |
2462 | llvm_unreachable("unexpected number of passes for mfma" ); |
2463 | case 8: |
2464 | NeedWaitStates = SMFMA16x16WritesVGPROverlappedSrcABWaitStates; |
2465 | break; |
2466 | case 16: |
2467 | default: |
2468 | NeedWaitStates = SMFMA32x32WritesVGPROverlappedSrcABWaitStates; |
2469 | } |
2470 | } |
2471 | } |
2472 | if (WaitStatesNeeded >= NeedWaitStates) |
2473 | continue; |
2474 | |
2475 | WaitStatesNeededForUse = NeedWaitStates - NumWaitStates; |
2476 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2477 | |
2478 | if (WaitStatesNeeded == MaxWaitStates) |
2479 | break; |
2480 | } |
2481 | |
2482 | // Pad neighboring MFMA with noops for better inter-wave performance. |
2483 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: checkMFMAPadding(MI)); |
2484 | |
2485 | return WaitStatesNeeded; |
2486 | } |
2487 | |
2488 | int GCNHazardRecognizer::checkMAILdStHazards(MachineInstr *MI) { |
2489 | // On gfx90a+ relevant hazards are checked in checkMAIVALUHazards() |
2490 | if (!ST.hasMAIInsts() || ST.hasGFX90AInsts()) |
2491 | return 0; |
2492 | |
2493 | int WaitStatesNeeded = 0; |
2494 | |
2495 | auto IsAccVgprReadFn = [](const MachineInstr &MI) { |
2496 | return MI.getOpcode() == AMDGPU::V_ACCVGPR_READ_B32_e64; |
2497 | }; |
2498 | |
2499 | for (const MachineOperand &Op : MI->explicit_uses()) { |
2500 | if (!Op.isReg() || !TRI.isVGPR(MRI: MF.getRegInfo(), Reg: Op.getReg())) |
2501 | continue; |
2502 | |
2503 | Register Reg = Op.getReg(); |
2504 | |
2505 | const int AccVgprReadLdStWaitStates = 2; |
2506 | const int VALUWriteAccVgprRdWrLdStDepVALUWaitStates = 1; |
2507 | const int MaxWaitStates = 2; |
2508 | |
2509 | int WaitStatesNeededForUse = AccVgprReadLdStWaitStates - |
2510 | getWaitStatesSinceDef(Reg, IsHazardDef: IsAccVgprReadFn, Limit: MaxWaitStates); |
2511 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2512 | |
2513 | if (WaitStatesNeeded == MaxWaitStates) |
2514 | return WaitStatesNeeded; // Early exit. |
2515 | |
2516 | auto IsVALUAccVgprRdWrCheckFn = [Reg, this](const MachineInstr &MI) { |
2517 | if (MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64 && |
2518 | MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64) |
2519 | return false; |
2520 | auto IsVALUFn = [](const MachineInstr &MI) { |
2521 | return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMAI(MI); |
2522 | }; |
2523 | return getWaitStatesSinceDef(Reg, IsHazardDef: IsVALUFn, Limit: 2 /*MaxWaitStates*/) < |
2524 | std::numeric_limits<int>::max(); |
2525 | }; |
2526 | |
2527 | WaitStatesNeededForUse = VALUWriteAccVgprRdWrLdStDepVALUWaitStates - |
2528 | getWaitStatesSince(IsHazard: IsVALUAccVgprRdWrCheckFn, Limit: MaxWaitStates); |
2529 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2530 | } |
2531 | |
2532 | return WaitStatesNeeded; |
2533 | } |
2534 | |
2535 | int GCNHazardRecognizer::checkPermlaneHazards(MachineInstr *MI) { |
2536 | assert(!ST.hasVcmpxPermlaneHazard() && |
2537 | "this is a different vcmpx+permlane hazard" ); |
2538 | const SIRegisterInfo *TRI = ST.getRegisterInfo(); |
2539 | const SIInstrInfo *TII = ST.getInstrInfo(); |
2540 | |
2541 | auto IsVCmpXWritesExecFn = [TII, TRI](const MachineInstr &MI) { |
2542 | return isVCmpXWritesExec(TII: *TII, TRI: *TRI, MI); |
2543 | }; |
2544 | |
2545 | auto IsVALUFn = [](const MachineInstr &MI) { |
2546 | return SIInstrInfo::isVALU(MI); |
2547 | }; |
2548 | |
2549 | const int VCmpXWritesExecWaitStates = 4; |
2550 | const int VALUWritesVDstWaitStates = 2; |
2551 | int WaitStatesNeeded = 0; |
2552 | |
2553 | for (const MachineOperand &Op : MI->explicit_uses()) { |
2554 | if (!Op.isReg() || !TRI->isVGPR(MRI: MF.getRegInfo(), Reg: Op.getReg())) |
2555 | continue; |
2556 | Register Reg = Op.getReg(); |
2557 | |
2558 | int WaitStatesSinceDef = |
2559 | VALUWritesVDstWaitStates - |
2560 | getWaitStatesSinceDef(Reg, IsHazardDef: IsVALUFn, |
2561 | /*MaxWaitStates=*/Limit: VALUWritesVDstWaitStates); |
2562 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesSinceDef); |
2563 | if (WaitStatesNeeded >= VALUWritesVDstWaitStates) |
2564 | break; |
2565 | } |
2566 | |
2567 | int VCmpXHazardWaits = |
2568 | VCmpXWritesExecWaitStates - |
2569 | getWaitStatesSince(IsHazard: IsVCmpXWritesExecFn, Limit: VCmpXWritesExecWaitStates); |
2570 | |
2571 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: VCmpXHazardWaits); |
2572 | return WaitStatesNeeded; |
2573 | } |
2574 | |
2575 | static int GFX940_SMFMA_N_PassWriteVgprVALUWawWaitStates(int NumPasses) { |
2576 | // 2 pass -> 4 |
2577 | // 4 pass -> 6 |
2578 | // 8 pass -> 10 |
2579 | // 16 pass -> 18 |
2580 | return NumPasses + 2; |
2581 | } |
2582 | |
2583 | static int GFX940_XDL_N_PassWriteVgprVALUWawWaitStates(int NumPasses, |
2584 | bool IsGFX950) { |
2585 | // xdl def cycles | gfx942 | gfx950 |
2586 | // 2 pass | 5 5 |
2587 | // 4 pass | 7 8 |
2588 | // 8 pass | 11 12 |
2589 | // 16 pass | 19 20 |
2590 | return NumPasses + 3 + (NumPasses != 2 && IsGFX950); |
2591 | } |
2592 | |
2593 | static int GFX940_XDL_N_PassWriteVgprVALUMemExpReadWaitStates(int NumPasses, |
2594 | bool IsGFX950) { |
2595 | // xdl def cycles | gfx942 | gfx950 |
2596 | // 2 pass | 5 5 |
2597 | // 4 pass | 7 8 |
2598 | // 8 pass | 11 12 |
2599 | // 16 pass | 19 20 |
2600 | return NumPasses + 3 + (NumPasses != 2 && IsGFX950); |
2601 | } |
2602 | |
2603 | static int GFX940_SMFMA_N_PassWriteVgprVALUMemExpReadWaitStates(int NumPasses) { |
2604 | // 2 pass -> 4 |
2605 | // 4 pass -> 6 |
2606 | // 8 pass -> 10 |
2607 | // 16 pass -> 18 |
2608 | return NumPasses + 2; |
2609 | } |
2610 | |
2611 | int GCNHazardRecognizer::checkMAIVALUHazards(MachineInstr *MI) { |
2612 | if (!ST.hasGFX90AInsts()) |
2613 | return 0; |
2614 | |
2615 | auto IsDGEMMFn = [](const MachineInstr &MI) -> bool { |
2616 | return SIInstrInfo::isDGEMM(Opcode: MI.getOpcode()); |
2617 | }; |
2618 | |
2619 | // This is checked in checkMAIHazards90A() |
2620 | if (SIInstrInfo::isMFMA(MI: *MI)) |
2621 | return 0; |
2622 | |
2623 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
2624 | |
2625 | int WaitStatesNeeded = 0; |
2626 | |
2627 | bool IsMem = SIInstrInfo::isVMEM(MI: *MI) || SIInstrInfo::isDS(MI: *MI); |
2628 | bool IsMemOrExport = IsMem || SIInstrInfo::isEXP(MI: *MI); |
2629 | bool IsVALU = SIInstrInfo::isVALU(MI: *MI); |
2630 | |
2631 | const MachineInstr *MFMA = nullptr; |
2632 | unsigned Reg; |
2633 | auto IsMFMAWriteFn = [&Reg, &MFMA, this](const MachineInstr &MI) { |
2634 | if (!SIInstrInfo::isMFMA(MI) || |
2635 | !TRI.regsOverlap(RegA: MI.getOperand(i: 0).getReg(), RegB: Reg)) |
2636 | return false; |
2637 | MFMA = &MI; |
2638 | return true; |
2639 | }; |
2640 | |
2641 | const MachineInstr *DOT = nullptr; |
2642 | auto IsDotWriteFn = [&Reg, &DOT, this](const MachineInstr &MI) { |
2643 | if (!SIInstrInfo::isDOT(MI) || |
2644 | !TRI.regsOverlap(RegA: MI.getOperand(i: 0).getReg(), RegB: Reg)) |
2645 | return false; |
2646 | DOT = &MI; |
2647 | return true; |
2648 | }; |
2649 | |
2650 | bool DGEMMAfterVALUWrite = false; |
2651 | auto IsDGEMMHazard = [&DGEMMAfterVALUWrite, this](const MachineInstr &MI) { |
2652 | // Found DGEMM on reverse traversal to def. |
2653 | if (SIInstrInfo::isDGEMM(Opcode: MI.getOpcode())) |
2654 | DGEMMAfterVALUWrite = true; |
2655 | |
2656 | // Only hazard if register is defined by a VALU and a DGEMM is found after |
2657 | // after the def. |
2658 | if (!TII.isVALU(MI) || !DGEMMAfterVALUWrite) |
2659 | return false; |
2660 | |
2661 | return true; |
2662 | }; |
2663 | |
2664 | int SrcCIdx = AMDGPU::getNamedOperandIdx(Opcode: MI->getOpcode(), |
2665 | Name: AMDGPU::OpName::src2); |
2666 | |
2667 | if (IsMemOrExport || IsVALU) { |
2668 | const int SMFMA4x4WriteVgprVALUMemExpReadWaitStates = 5; |
2669 | const int SMFMA16x16WriteVgprVALUMemExpReadWaitStates = 11; |
2670 | const int SMFMA32x32WriteVgprVALUMemExpReadWaitStates = 19; |
2671 | const int DMFMA4x4WriteVgprMemExpReadWaitStates = 9; |
2672 | const int DMFMA16x16WriteVgprMemExpReadWaitStates = 18; |
2673 | const int DMFMA4x4WriteVgprVALUReadWaitStates = 6; |
2674 | const int DMFMA16x16WriteVgprVALUReadWaitStates = 11; |
2675 | const int GFX950_DMFMA16x16WriteVgprVALUReadWaitStates = 19; |
2676 | const int DotWriteSameDotReadSrcAB = 3; |
2677 | const int DotWriteDifferentVALURead = 3; |
2678 | const int DMFMABetweenVALUWriteVMEMRead = 2; |
2679 | const int MaxWaitStates = 19; |
2680 | |
2681 | for (const MachineOperand &Use : MI->explicit_uses()) { |
2682 | if (!Use.isReg()) |
2683 | continue; |
2684 | Reg = Use.getReg(); |
2685 | |
2686 | DOT = nullptr; |
2687 | int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsHazardDef: IsDotWriteFn, |
2688 | Limit: MaxWaitStates); |
2689 | if (DOT) { |
2690 | int NeedWaitStates = 0; |
2691 | if (DOT->getOpcode() == MI->getOpcode()) { |
2692 | if (&Use - &MI->getOperand(i: 0) != SrcCIdx) |
2693 | NeedWaitStates = DotWriteSameDotReadSrcAB; |
2694 | } else { |
2695 | NeedWaitStates = DotWriteDifferentVALURead; |
2696 | } |
2697 | |
2698 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; |
2699 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2700 | } |
2701 | |
2702 | // Workaround for HW data hazard bug observed only in GFX90A. When there |
2703 | // is a DGEMM instruction in-between a VALU and a VMEM instruction it |
2704 | // causes the SQ to incorrectly not insert two wait states between the two |
2705 | // instructions needed to avoid data hazard. |
2706 | if (IsMem && ST.hasGFX90AInsts() && !ST.hasGFX940Insts()) { |
2707 | DGEMMAfterVALUWrite = false; |
2708 | if (TRI.isVectorRegister(MRI, Reg)) { |
2709 | int WaitStatesNeededForUse = |
2710 | DMFMABetweenVALUWriteVMEMRead - |
2711 | getWaitStatesSinceDef(Reg, IsHazardDef: IsDGEMMHazard, |
2712 | Limit: DMFMABetweenVALUWriteVMEMRead); |
2713 | |
2714 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2715 | } |
2716 | } |
2717 | |
2718 | MFMA = nullptr; |
2719 | WaitStatesSinceDef = |
2720 | getWaitStatesSinceDef(Reg, IsHazardDef: IsMFMAWriteFn, Limit: MaxWaitStates); |
2721 | if (!MFMA) |
2722 | continue; |
2723 | |
2724 | unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MI: MFMA); |
2725 | int NumPasses = HazardDefLatency; |
2726 | int NeedWaitStates = MaxWaitStates; |
2727 | |
2728 | if (SIInstrInfo::isDGEMM(Opcode: MFMA->getOpcode())) { |
2729 | switch (HazardDefLatency) { |
2730 | case 4: |
2731 | NeedWaitStates = IsMemOrExport ? DMFMA4x4WriteVgprMemExpReadWaitStates |
2732 | : DMFMA4x4WriteVgprVALUReadWaitStates; |
2733 | break; |
2734 | case 8: |
2735 | case 16: |
2736 | NeedWaitStates = |
2737 | IsMemOrExport |
2738 | ? DMFMA16x16WriteVgprMemExpReadWaitStates |
2739 | : (ST.hasGFX950Insts() |
2740 | ? GFX950_DMFMA16x16WriteVgprVALUReadWaitStates |
2741 | : DMFMA16x16WriteVgprVALUReadWaitStates); |
2742 | break; |
2743 | default: |
2744 | llvm_unreachable("unexpected dgemm" ); |
2745 | } |
2746 | } else if (ST.hasGFX940Insts()) { |
2747 | NeedWaitStates = |
2748 | TII.isXDL(MI: *MFMA) |
2749 | ? GFX940_XDL_N_PassWriteVgprVALUMemExpReadWaitStates( |
2750 | NumPasses, IsGFX950: ST.hasGFX950Insts()) |
2751 | : GFX940_SMFMA_N_PassWriteVgprVALUMemExpReadWaitStates( |
2752 | NumPasses); |
2753 | } else { |
2754 | switch (HazardDefLatency) { |
2755 | case 2: |
2756 | NeedWaitStates = SMFMA4x4WriteVgprVALUMemExpReadWaitStates; |
2757 | break; |
2758 | case 8: |
2759 | NeedWaitStates = SMFMA16x16WriteVgprVALUMemExpReadWaitStates; |
2760 | break; |
2761 | case 16: |
2762 | NeedWaitStates = SMFMA32x32WriteVgprVALUMemExpReadWaitStates; |
2763 | break; |
2764 | default: |
2765 | llvm_unreachable("unexpected number of passes for mfma" ); |
2766 | } |
2767 | } |
2768 | |
2769 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; |
2770 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2771 | |
2772 | if (WaitStatesNeeded == MaxWaitStates) |
2773 | break; |
2774 | } |
2775 | } |
2776 | |
2777 | unsigned Opc = MI->getOpcode(); |
2778 | const int DMFMAToFMA64WaitStates = 2; |
2779 | if ((Opc == AMDGPU::V_FMA_F64_e64 || |
2780 | Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64 || |
2781 | Opc == AMDGPU::V_FMAC_F64_dpp) && |
2782 | WaitStatesNeeded < DMFMAToFMA64WaitStates) { |
2783 | int WaitStatesNeededForUse = DMFMAToFMA64WaitStates - |
2784 | getWaitStatesSince(IsHazard: IsDGEMMFn, Limit: DMFMAToFMA64WaitStates); |
2785 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2786 | } |
2787 | |
2788 | if (!IsVALU && !IsMemOrExport) |
2789 | return WaitStatesNeeded; |
2790 | |
2791 | for (const MachineOperand &Def : MI->defs()) { |
2792 | const int SMFMA4x4WriteVgprVALUWawWaitStates = 5; |
2793 | const int SMFMA16x16WriteVgprVALUWawWaitStates = 11; |
2794 | const int SMFMA32x32WriteVgprVALUWawWaitStates = 19; |
2795 | const int SMFMA4x4ReadVgprVALUWarWaitStates = 1; |
2796 | const int GFX940_XDL4PassReadVgprVALUWarWaitStates = 3; |
2797 | const int SMFMA16x16ReadVgprVALUWarWaitStates = 7; |
2798 | const int SMFMA32x32ReadVgprVALUWarWaitStates = 15; |
2799 | const int DMFMA4x4WriteVgprVALUWriteWaitStates = 6; |
2800 | const int DMFMA16x16WriteVgprVALUWriteWaitStates = 11; |
2801 | const int DotWriteDifferentVALUWrite = 3; |
2802 | const int MaxWaitStates = 19; |
2803 | const int MaxWarWaitStates = 15; |
2804 | |
2805 | Reg = Def.getReg(); |
2806 | |
2807 | DOT = nullptr; |
2808 | int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsHazardDef: IsDotWriteFn, |
2809 | Limit: MaxWaitStates); |
2810 | if (DOT && DOT->getOpcode() != MI->getOpcode()) |
2811 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: DotWriteDifferentVALUWrite - |
2812 | WaitStatesSinceDef); |
2813 | |
2814 | MFMA = nullptr; |
2815 | WaitStatesSinceDef = |
2816 | getWaitStatesSinceDef(Reg, IsHazardDef: IsMFMAWriteFn, Limit: MaxWaitStates); |
2817 | if (MFMA) { |
2818 | int NeedWaitStates = MaxWaitStates; |
2819 | int NumPasses = TSchedModel.computeInstrLatency(MI: MFMA); |
2820 | |
2821 | if (SIInstrInfo::isDGEMM(Opcode: MFMA->getOpcode())) { |
2822 | switch (NumPasses) { |
2823 | case 4: |
2824 | NeedWaitStates = DMFMA4x4WriteVgprVALUWriteWaitStates; |
2825 | break; |
2826 | case 8: |
2827 | case 16: |
2828 | NeedWaitStates = DMFMA16x16WriteVgprVALUWriteWaitStates; |
2829 | break; |
2830 | default: |
2831 | llvm_unreachable("unexpected number of cycles for dgemm" ); |
2832 | } |
2833 | } else if (ST.hasGFX940Insts()) { |
2834 | NeedWaitStates = |
2835 | TII.isXDL(MI: *MFMA) |
2836 | ? GFX940_XDL_N_PassWriteVgprVALUWawWaitStates( |
2837 | NumPasses, IsGFX950: ST.hasGFX950Insts()) |
2838 | : GFX940_SMFMA_N_PassWriteVgprVALUWawWaitStates(NumPasses); |
2839 | } else { |
2840 | switch (NumPasses) { |
2841 | case 2: |
2842 | NeedWaitStates = SMFMA4x4WriteVgprVALUWawWaitStates; |
2843 | break; |
2844 | case 8: |
2845 | NeedWaitStates = SMFMA16x16WriteVgprVALUWawWaitStates; |
2846 | break; |
2847 | case 16: |
2848 | NeedWaitStates = SMFMA32x32WriteVgprVALUWawWaitStates; |
2849 | break; |
2850 | default: |
2851 | llvm_unreachable("Unexpected number of passes for mfma" ); |
2852 | } |
2853 | } |
2854 | |
2855 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; |
2856 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2857 | |
2858 | if (WaitStatesNeeded == MaxWaitStates) |
2859 | break; |
2860 | } |
2861 | |
2862 | auto IsSMFMAReadAsCFn = [&Reg, &MFMA, this](const MachineInstr &MI) { |
2863 | if (!SIInstrInfo::isMFMA(MI) || SIInstrInfo::isDGEMM(Opcode: MI.getOpcode()) || |
2864 | !MI.readsRegister(Reg, TRI: &TRI)) |
2865 | return false; |
2866 | |
2867 | if (ST.hasGFX940Insts() && !TII.isXDL(MI)) |
2868 | return false; |
2869 | |
2870 | const MachineOperand *SrcC = |
2871 | TII.getNamedOperand(MI, OperandName: AMDGPU::OpName::src2); |
2872 | assert(SrcC); |
2873 | if (!SrcC->isReg() || !TRI.regsOverlap(RegA: SrcC->getReg(), RegB: Reg)) |
2874 | return false; |
2875 | |
2876 | MFMA = &MI; |
2877 | return true; |
2878 | }; |
2879 | |
2880 | MFMA = nullptr; |
2881 | int WaitStatesSinceUse = getWaitStatesSince(IsHazard: IsSMFMAReadAsCFn, |
2882 | Limit: MaxWarWaitStates); |
2883 | if (!MFMA) |
2884 | continue; |
2885 | |
2886 | unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MI: MFMA); |
2887 | int NeedWaitStates = MaxWaitStates; |
2888 | switch (HazardDefLatency) { |
2889 | case 2: NeedWaitStates = SMFMA4x4ReadVgprVALUWarWaitStates; |
2890 | break; |
2891 | case 4: assert(ST.hasGFX940Insts()); |
2892 | NeedWaitStates = GFX940_XDL4PassReadVgprVALUWarWaitStates; |
2893 | break; |
2894 | case 8: NeedWaitStates = SMFMA16x16ReadVgprVALUWarWaitStates; |
2895 | break; |
2896 | case 16: [[fallthrough]]; |
2897 | default: NeedWaitStates = SMFMA32x32ReadVgprVALUWarWaitStates; |
2898 | break; |
2899 | } |
2900 | |
2901 | int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceUse; |
2902 | WaitStatesNeeded = std::max(a: WaitStatesNeeded, b: WaitStatesNeededForUse); |
2903 | } |
2904 | |
2905 | return WaitStatesNeeded; |
2906 | } |
2907 | |
2908 | bool GCNHazardRecognizer::ShouldPreferAnother(SUnit *SU) { |
2909 | if (!SU->isInstr()) |
2910 | return false; |
2911 | |
2912 | const MachineInstr *MAI = nullptr; |
2913 | |
2914 | auto IsMFMAFn = [&MAI](const MachineInstr &MI) { |
2915 | MAI = nullptr; |
2916 | if (SIInstrInfo::isMFMA(MI)) |
2917 | MAI = &MI; |
2918 | return MAI != nullptr; |
2919 | }; |
2920 | |
2921 | MachineInstr *MI = SU->getInstr(); |
2922 | if (IsMFMAFn(*MI)) { |
2923 | int W = getWaitStatesSince(IsHazard: IsMFMAFn, Limit: 16); |
2924 | if (MAI) |
2925 | return W < (int)TSchedModel.computeInstrLatency(MI: MAI); |
2926 | } |
2927 | |
2928 | return false; |
2929 | } |
2930 | |
2931 | // Adjust global offsets for instructions bundled with S_GETPC_B64 after |
2932 | // insertion of a new instruction. |
2933 | static void updateGetPCBundle(MachineInstr *NewMI) { |
2934 | if (!NewMI->isBundled()) |
2935 | return; |
2936 | |
2937 | // Find start of bundle. |
2938 | auto I = NewMI->getIterator(); |
2939 | while (I->isBundledWithPred()) |
2940 | I--; |
2941 | if (I->isBundle()) |
2942 | I++; |
2943 | |
2944 | // Bail if this is not an S_GETPC bundle. |
2945 | if (I->getOpcode() != AMDGPU::S_GETPC_B64) |
2946 | return; |
2947 | |
2948 | // Update offsets of any references in the bundle. |
2949 | const unsigned NewBytes = 4; |
2950 | assert(NewMI->getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
2951 | "Unexpected instruction insertion in bundle" ); |
2952 | auto NextMI = std::next(x: NewMI->getIterator()); |
2953 | auto End = NewMI->getParent()->end(); |
2954 | while (NextMI != End && NextMI->isBundledWithPred()) { |
2955 | for (auto &Operand : NextMI->operands()) { |
2956 | if (Operand.isGlobal()) |
2957 | Operand.setOffset(Operand.getOffset() + NewBytes); |
2958 | } |
2959 | NextMI++; |
2960 | } |
2961 | } |
2962 | |
2963 | bool GCNHazardRecognizer::fixVALUMaskWriteHazard(MachineInstr *MI) { |
2964 | if (!ST.hasVALUMaskWriteHazard()) |
2965 | return false; |
2966 | assert(!ST.hasExtendedWaitCounts()); |
2967 | |
2968 | if (!ST.isWave64() || !SIInstrInfo::isSALU(MI: *MI)) |
2969 | return false; |
2970 | |
2971 | // The hazard sequence is three instructions: |
2972 | // 1. VALU reads SGPR as mask |
2973 | // 2. SALU writes SGPR |
2974 | // 3. SALU reads SGPR |
2975 | // The hazard can expire if the distance between 2 and 3 is sufficient. |
2976 | // In practice this happens <10% of the time, hence this always assumes |
2977 | // the hazard exists if 1 and 2 are present to avoid searching. |
2978 | |
2979 | const MachineOperand *SDSTOp = TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::sdst); |
2980 | if (!SDSTOp || !SDSTOp->isReg()) |
2981 | return false; |
2982 | |
2983 | const Register HazardReg = SDSTOp->getReg(); |
2984 | if (HazardReg == AMDGPU::EXEC || |
2985 | HazardReg == AMDGPU::EXEC_LO || |
2986 | HazardReg == AMDGPU::EXEC_HI || |
2987 | HazardReg == AMDGPU::M0) |
2988 | return false; |
2989 | |
2990 | auto IsHazardFn = [HazardReg, this](const MachineInstr &I) { |
2991 | switch (I.getOpcode()) { |
2992 | case AMDGPU::V_ADDC_U32_e32: |
2993 | case AMDGPU::V_ADDC_U32_dpp: |
2994 | case AMDGPU::V_CNDMASK_B16_t16_e32: |
2995 | case AMDGPU::V_CNDMASK_B16_fake16_e32: |
2996 | case AMDGPU::V_CNDMASK_B16_t16_dpp: |
2997 | case AMDGPU::V_CNDMASK_B16_fake16_dpp: |
2998 | case AMDGPU::V_CNDMASK_B32_e32: |
2999 | case AMDGPU::V_CNDMASK_B32_dpp: |
3000 | case AMDGPU::V_DIV_FMAS_F32_e64: |
3001 | case AMDGPU::V_DIV_FMAS_F64_e64: |
3002 | case AMDGPU::V_SUBB_U32_e32: |
3003 | case AMDGPU::V_SUBB_U32_dpp: |
3004 | case AMDGPU::V_SUBBREV_U32_e32: |
3005 | case AMDGPU::V_SUBBREV_U32_dpp: |
3006 | // These implicitly read VCC as mask source. |
3007 | return HazardReg == AMDGPU::VCC || |
3008 | HazardReg == AMDGPU::VCC_LO || |
3009 | HazardReg == AMDGPU::VCC_HI; |
3010 | case AMDGPU::V_ADDC_U32_e64: |
3011 | case AMDGPU::V_ADDC_U32_e64_dpp: |
3012 | case AMDGPU::V_CNDMASK_B16_t16_e64: |
3013 | case AMDGPU::V_CNDMASK_B16_fake16_e64: |
3014 | case AMDGPU::V_CNDMASK_B16_t16_e64_dpp: |
3015 | case AMDGPU::V_CNDMASK_B16_fake16_e64_dpp: |
3016 | case AMDGPU::V_CNDMASK_B32_e64: |
3017 | case AMDGPU::V_CNDMASK_B32_e64_dpp: |
3018 | case AMDGPU::V_SUBB_U32_e64: |
3019 | case AMDGPU::V_SUBB_U32_e64_dpp: |
3020 | case AMDGPU::V_SUBBREV_U32_e64: |
3021 | case AMDGPU::V_SUBBREV_U32_e64_dpp: { |
3022 | // Only check mask register overlaps. |
3023 | const MachineOperand *SSRCOp = TII.getNamedOperand(MI: I, OperandName: AMDGPU::OpName::src2); |
3024 | assert(SSRCOp); |
3025 | return TRI.regsOverlap(RegA: SSRCOp->getReg(), RegB: HazardReg); |
3026 | } |
3027 | default: |
3028 | return false; |
3029 | } |
3030 | }; |
3031 | |
3032 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
3033 | auto IsExpiredFn = [&MRI, this](const MachineInstr &I, int) { |
3034 | // s_waitcnt_depctr sa_sdst(0) mitigates hazard. |
3035 | if (I.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && |
3036 | AMDGPU::DepCtr::decodeFieldSaSdst(Encoded: I.getOperand(i: 0).getImm()) == 0) |
3037 | return true; |
3038 | |
3039 | // VALU access to any SGPR or literal constant other than HazardReg |
3040 | // mitigates hazard. No need to check HazardReg here as this will |
3041 | // only be called when !IsHazardFn. |
3042 | if (!SIInstrInfo::isVALU(MI: I)) |
3043 | return false; |
3044 | for (int OpNo = 0, End = I.getNumOperands(); OpNo < End; ++OpNo) { |
3045 | const MachineOperand &Op = I.getOperand(i: OpNo); |
3046 | if (Op.isReg()) { |
3047 | Register OpReg = Op.getReg(); |
3048 | // Only consider uses |
3049 | if (!Op.isUse()) |
3050 | continue; |
3051 | // Ignore EXEC |
3052 | if (OpReg == AMDGPU::EXEC || |
3053 | OpReg == AMDGPU::EXEC_LO || |
3054 | OpReg == AMDGPU::EXEC_HI) |
3055 | continue; |
3056 | // Ignore all implicit uses except VCC |
3057 | if (Op.isImplicit()) { |
3058 | if (OpReg == AMDGPU::VCC || |
3059 | OpReg == AMDGPU::VCC_LO || |
3060 | OpReg == AMDGPU::VCC_HI) |
3061 | return true; |
3062 | continue; |
3063 | } |
3064 | if (TRI.isSGPRReg(MRI, Reg: OpReg)) |
3065 | return true; |
3066 | } else { |
3067 | const MCInstrDesc &InstDesc = I.getDesc(); |
3068 | const MCOperandInfo &OpInfo = InstDesc.operands()[OpNo]; |
3069 | if (!TII.isInlineConstant(MO: Op, OpInfo)) |
3070 | return true; |
3071 | } |
3072 | } |
3073 | return false; |
3074 | }; |
3075 | |
3076 | // Check for hazard |
3077 | if (::getWaitStatesSince(IsHazard: IsHazardFn, MI, IsExpired: IsExpiredFn) == |
3078 | std::numeric_limits<int>::max()) |
3079 | return false; |
3080 | |
3081 | auto NextMI = std::next(x: MI->getIterator()); |
3082 | |
3083 | // Add s_waitcnt_depctr sa_sdst(0) after SALU write. |
3084 | auto NewMI = BuildMI(BB&: *MI->getParent(), I: NextMI, MIMD: MI->getDebugLoc(), |
3085 | MCID: TII.get(Opcode: AMDGPU::S_WAITCNT_DEPCTR)) |
3086 | .addImm(Val: AMDGPU::DepCtr::encodeFieldSaSdst(SaSdst: 0)); |
3087 | |
3088 | // SALU write may be s_getpc in a bundle. |
3089 | updateGetPCBundle(NewMI); |
3090 | |
3091 | return true; |
3092 | } |
3093 | |
3094 | static bool ensureEntrySetPrio(MachineFunction *MF, int Priority, |
3095 | const SIInstrInfo &TII) { |
3096 | MachineBasicBlock &EntryMBB = MF->front(); |
3097 | if (EntryMBB.begin() != EntryMBB.end()) { |
3098 | auto &EntryMI = *EntryMBB.begin(); |
3099 | if (EntryMI.getOpcode() == AMDGPU::S_SETPRIO && |
3100 | EntryMI.getOperand(i: 0).getImm() >= Priority) |
3101 | return false; |
3102 | } |
3103 | |
3104 | BuildMI(BB&: EntryMBB, I: EntryMBB.begin(), MIMD: DebugLoc(), MCID: TII.get(Opcode: AMDGPU::S_SETPRIO)) |
3105 | .addImm(Val: Priority); |
3106 | return true; |
3107 | } |
3108 | |
3109 | bool GCNHazardRecognizer::fixRequiredExportPriority(MachineInstr *MI) { |
3110 | if (!ST.hasRequiredExportPriority()) |
3111 | return false; |
3112 | |
3113 | // Assume the following shader types will never have exports, |
3114 | // and avoid adding or adjusting S_SETPRIO. |
3115 | MachineBasicBlock *MBB = MI->getParent(); |
3116 | MachineFunction *MF = MBB->getParent(); |
3117 | auto CC = MF->getFunction().getCallingConv(); |
3118 | switch (CC) { |
3119 | case CallingConv::AMDGPU_CS: |
3120 | case CallingConv::AMDGPU_CS_Chain: |
3121 | case CallingConv::AMDGPU_CS_ChainPreserve: |
3122 | case CallingConv::AMDGPU_KERNEL: |
3123 | return false; |
3124 | default: |
3125 | break; |
3126 | } |
3127 | |
3128 | const int MaxPriority = 3; |
3129 | const int NormalPriority = 2; |
3130 | const int PostExportPriority = 0; |
3131 | |
3132 | auto It = MI->getIterator(); |
3133 | switch (MI->getOpcode()) { |
3134 | case AMDGPU::S_ENDPGM: |
3135 | case AMDGPU::S_ENDPGM_SAVED: |
3136 | case AMDGPU::S_ENDPGM_ORDERED_PS_DONE: |
3137 | case AMDGPU::SI_RETURN_TO_EPILOG: |
3138 | // Ensure shader with calls raises priority at entry. |
3139 | // This ensures correct priority if exports exist in callee. |
3140 | if (MF->getFrameInfo().hasCalls()) |
3141 | return ensureEntrySetPrio(MF, Priority: NormalPriority, TII); |
3142 | return false; |
3143 | case AMDGPU::S_SETPRIO: { |
3144 | // Raise minimum priority unless in workaround. |
3145 | auto &PrioOp = MI->getOperand(i: 0); |
3146 | int Prio = PrioOp.getImm(); |
3147 | bool InWA = (Prio == PostExportPriority) && |
3148 | (It != MBB->begin() && TII.isEXP(MI: *std::prev(x: It))); |
3149 | if (InWA || Prio >= NormalPriority) |
3150 | return false; |
3151 | PrioOp.setImm(std::min(a: Prio + NormalPriority, b: MaxPriority)); |
3152 | return true; |
3153 | } |
3154 | default: |
3155 | if (!TII.isEXP(MI: *MI)) |
3156 | return false; |
3157 | break; |
3158 | } |
3159 | |
3160 | // Check entry priority at each export (as there will only be a few). |
3161 | // Note: amdgpu_gfx can only be a callee, so defer to caller setprio. |
3162 | bool Changed = false; |
3163 | if (CC != CallingConv::AMDGPU_Gfx) |
3164 | Changed = ensureEntrySetPrio(MF, Priority: NormalPriority, TII); |
3165 | |
3166 | auto NextMI = std::next(x: It); |
3167 | bool EndOfShader = false; |
3168 | if (NextMI != MBB->end()) { |
3169 | // Only need WA at end of sequence of exports. |
3170 | if (TII.isEXP(MI: *NextMI)) |
3171 | return Changed; |
3172 | // Assume appropriate S_SETPRIO after export means WA already applied. |
3173 | if (NextMI->getOpcode() == AMDGPU::S_SETPRIO && |
3174 | NextMI->getOperand(i: 0).getImm() == PostExportPriority) |
3175 | return Changed; |
3176 | EndOfShader = NextMI->getOpcode() == AMDGPU::S_ENDPGM; |
3177 | } |
3178 | |
3179 | const DebugLoc &DL = MI->getDebugLoc(); |
3180 | |
3181 | // Lower priority. |
3182 | BuildMI(BB&: *MBB, I: NextMI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_SETPRIO)) |
3183 | .addImm(Val: PostExportPriority); |
3184 | |
3185 | if (!EndOfShader) { |
3186 | // Wait for exports to complete. |
3187 | BuildMI(BB&: *MBB, I: NextMI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_WAITCNT_EXPCNT)) |
3188 | .addReg(RegNo: AMDGPU::SGPR_NULL) |
3189 | .addImm(Val: 0); |
3190 | } |
3191 | |
3192 | BuildMI(BB&: *MBB, I: NextMI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_NOP)).addImm(Val: 0); |
3193 | BuildMI(BB&: *MBB, I: NextMI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_NOP)).addImm(Val: 0); |
3194 | |
3195 | if (!EndOfShader) { |
3196 | // Return to normal (higher) priority. |
3197 | BuildMI(BB&: *MBB, I: NextMI, MIMD: DL, MCID: TII.get(Opcode: AMDGPU::S_SETPRIO)) |
3198 | .addImm(Val: NormalPriority); |
3199 | } |
3200 | |
3201 | return true; |
3202 | } |
3203 | |