xref: /llvm-project/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp (revision 64838ba3657614f80f8ec78b7967848ca0691411)
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/MachineFunction.h"
18 #include "llvm/CodeGen/ScheduleDAG.h"
19 #include "llvm/Support/TargetParser.h"
20 
21 using namespace llvm;
22 
23 namespace {
24 
25 struct MFMAPaddingRatioParser : public cl::parser<unsigned> {
26   MFMAPaddingRatioParser(cl::Option &O) : cl::parser<unsigned>(O) {}
27 
28   bool parse(cl::Option &O, StringRef ArgName, StringRef Arg, unsigned &Value) {
29     if (Arg.getAsInteger(0, Value))
30       return O.error("'" + Arg + "' value invalid for uint argument!");
31 
32     if (Value > 100)
33       return O.error("'" + Arg + "' value must be in the range [0, 100]!");
34 
35     return false;
36   }
37 };
38 
39 } // end anonymous namespace
40 
41 static cl::opt<unsigned, false, MFMAPaddingRatioParser>
42     MFMAPaddingRatio("amdgpu-mfma-padding-ratio", cl::init(0), cl::Hidden,
43                      cl::desc("Fill a percentage of the latency between "
44                               "neighboring MFMA with s_nops."));
45 
46 //===----------------------------------------------------------------------===//
47 // Hazard Recognizer Implementation
48 //===----------------------------------------------------------------------===//
49 
50 static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF,
51                                                  const GCNSubtarget &ST);
52 
53 GCNHazardRecognizer::GCNHazardRecognizer(const MachineFunction &MF) :
54   IsHazardRecognizerMode(false),
55   CurrCycleInstr(nullptr),
56   MF(MF),
57   ST(MF.getSubtarget<GCNSubtarget>()),
58   TII(*ST.getInstrInfo()),
59   TRI(TII.getRegisterInfo()),
60   ClauseUses(TRI.getNumRegUnits()),
61   ClauseDefs(TRI.getNumRegUnits()) {
62   MaxLookAhead = MF.getRegInfo().isPhysRegUsed(AMDGPU::AGPR0) ? 19 : 5;
63   TSchedModel.init(&ST);
64   RunLdsBranchVmemWARHazardFixup = shouldRunLdsBranchVmemWARHazardFixup(MF, ST);
65 }
66 
67 void GCNHazardRecognizer::Reset() {
68   EmittedInstrs.clear();
69 }
70 
71 void GCNHazardRecognizer::EmitInstruction(SUnit *SU) {
72   EmitInstruction(SU->getInstr());
73 }
74 
75 void GCNHazardRecognizer::EmitInstruction(MachineInstr *MI) {
76   CurrCycleInstr = MI;
77 }
78 
79 static bool isDivFMas(unsigned Opcode) {
80   return Opcode == AMDGPU::V_DIV_FMAS_F32_e64 || Opcode == AMDGPU::V_DIV_FMAS_F64_e64;
81 }
82 
83 static bool isSGetReg(unsigned Opcode) {
84   return Opcode == AMDGPU::S_GETREG_B32;
85 }
86 
87 static bool isSSetReg(unsigned Opcode) {
88   switch (Opcode) {
89   case AMDGPU::S_SETREG_B32:
90   case AMDGPU::S_SETREG_B32_mode:
91   case AMDGPU::S_SETREG_IMM32_B32:
92   case AMDGPU::S_SETREG_IMM32_B32_mode:
93     return true;
94   }
95   return false;
96 }
97 
98 static bool isRWLane(unsigned Opcode) {
99   return Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32;
100 }
101 
102 static bool isRFE(unsigned Opcode) {
103   return Opcode == AMDGPU::S_RFE_B64;
104 }
105 
106 static bool isSMovRel(unsigned Opcode) {
107   switch (Opcode) {
108   case AMDGPU::S_MOVRELS_B32:
109   case AMDGPU::S_MOVRELS_B64:
110   case AMDGPU::S_MOVRELD_B32:
111   case AMDGPU::S_MOVRELD_B64:
112     return true;
113   default:
114     return false;
115   }
116 }
117 
118 static bool isDGEMM(unsigned Opcode) {
119   return AMDGPU::getMAIIsDGEMM(Opcode);
120 }
121 
122 static bool isXDL(const GCNSubtarget &ST, const MachineInstr &MI) {
123   unsigned Opcode = MI.getOpcode();
124 
125   if (!SIInstrInfo::isMAI(MI) ||
126       isDGEMM(Opcode) ||
127       Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_e64 ||
128       Opcode == AMDGPU::V_ACCVGPR_READ_B32_e64)
129     return false;
130 
131   if (!ST.hasGFX940Insts())
132     return true;
133 
134   return AMDGPU::getMAIIsGFX940XDL(Opcode);
135 }
136 
137 static bool isSendMsgTraceDataOrGDS(const SIInstrInfo &TII,
138                                     const MachineInstr &MI) {
139   if (TII.isAlwaysGDS(MI.getOpcode()))
140     return true;
141 
142   switch (MI.getOpcode()) {
143   case AMDGPU::S_SENDMSG:
144   case AMDGPU::S_SENDMSGHALT:
145   case AMDGPU::S_TTRACEDATA:
146     return true;
147   // These DS opcodes don't support GDS.
148   case AMDGPU::DS_NOP:
149   case AMDGPU::DS_PERMUTE_B32:
150   case AMDGPU::DS_BPERMUTE_B32:
151     return false;
152   default:
153     if (TII.isDS(MI.getOpcode())) {
154       int GDS = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
155                                            AMDGPU::OpName::gds);
156       if (MI.getOperand(GDS).getImm())
157         return true;
158     }
159     return false;
160   }
161 }
162 
163 static bool isPermlane(const MachineInstr &MI) {
164   unsigned Opcode = MI.getOpcode();
165   return Opcode == AMDGPU::V_PERMLANE16_B32_e64 ||
166          Opcode == AMDGPU::V_PERMLANEX16_B32_e64;
167 }
168 
169 static unsigned getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) {
170   const MachineOperand *RegOp = TII->getNamedOperand(RegInstr,
171                                                      AMDGPU::OpName::simm16);
172   return RegOp->getImm() & AMDGPU::Hwreg::ID_MASK_;
173 }
174 
175 ScheduleHazardRecognizer::HazardType
176 GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
177   MachineInstr *MI = SU->getInstr();
178   // If we are not in "HazardRecognizerMode" and therefore not being run from
179   // the scheduler, track possible stalls from hazards but don't insert noops.
180   auto HazardType = IsHazardRecognizerMode ? NoopHazard : Hazard;
181 
182   if (MI->isBundle())
183    return NoHazard;
184 
185   if (SIInstrInfo::isSMRD(*MI) && checkSMRDHazards(MI) > 0)
186     return HazardType;
187 
188   if (ST.hasNSAtoVMEMBug() && checkNSAtoVMEMHazard(MI) > 0)
189     return HazardType;
190 
191   if (checkFPAtomicToDenormModeHazard(MI) > 0)
192     return HazardType;
193 
194   if (ST.hasNoDataDepHazard())
195     return NoHazard;
196 
197   // FIXME: Should flat be considered vmem?
198   if ((SIInstrInfo::isVMEM(*MI) ||
199        SIInstrInfo::isFLAT(*MI))
200       && checkVMEMHazards(MI) > 0)
201     return HazardType;
202 
203   if (SIInstrInfo::isVALU(*MI) && checkVALUHazards(MI) > 0)
204     return HazardType;
205 
206   if (SIInstrInfo::isDPP(*MI) && checkDPPHazards(MI) > 0)
207     return HazardType;
208 
209   if (isDivFMas(MI->getOpcode()) && checkDivFMasHazards(MI) > 0)
210     return HazardType;
211 
212   if (isRWLane(MI->getOpcode()) && checkRWLaneHazards(MI) > 0)
213     return HazardType;
214 
215   if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) ||
216        SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) ||
217        SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0)
218     return HazardType;
219 
220   if (isSGetReg(MI->getOpcode()) && checkGetRegHazards(MI) > 0)
221     return HazardType;
222 
223   if (isSSetReg(MI->getOpcode()) && checkSetRegHazards(MI) > 0)
224     return HazardType;
225 
226   if (isRFE(MI->getOpcode()) && checkRFEHazards(MI) > 0)
227     return HazardType;
228 
229   if (ST.hasReadM0MovRelInterpHazard() &&
230       (TII.isVINTRP(*MI) || isSMovRel(MI->getOpcode())) &&
231       checkReadM0Hazards(MI) > 0)
232     return HazardType;
233 
234   if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI) &&
235       checkReadM0Hazards(MI) > 0)
236     return HazardType;
237 
238   if (SIInstrInfo::isMAI(*MI) && checkMAIHazards(MI) > 0)
239     return HazardType;
240 
241   if ((SIInstrInfo::isVMEM(*MI) ||
242        SIInstrInfo::isFLAT(*MI) ||
243        SIInstrInfo::isDS(*MI)) && checkMAILdStHazards(MI) > 0)
244     return HazardType;
245 
246   if (MI->isInlineAsm() && checkInlineAsmHazards(MI) > 0)
247     return HazardType;
248 
249   return NoHazard;
250 }
251 
252 static void insertNoopsInBundle(MachineInstr *MI, const SIInstrInfo &TII,
253                                 unsigned Quantity) {
254   while (Quantity > 0) {
255     unsigned Arg = std::min(Quantity, 8u);
256     Quantity -= Arg;
257     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII.get(AMDGPU::S_NOP))
258         .addImm(Arg - 1);
259   }
260 }
261 
262 unsigned
263 GCNHazardRecognizer::getMFMAPipelineWaitStates(const MachineInstr &MI) const {
264   const MCSchedClassDesc *SC = TSchedModel.resolveSchedClass(&MI);
265   assert(TSchedModel.getWriteProcResBegin(SC) !=
266          TSchedModel.getWriteProcResEnd(SC));
267   return TSchedModel.getWriteProcResBegin(SC)->Cycles;
268 }
269 
270 void GCNHazardRecognizer::processBundle() {
271   MachineBasicBlock::instr_iterator MI = std::next(CurrCycleInstr->getIterator());
272   MachineBasicBlock::instr_iterator E = CurrCycleInstr->getParent()->instr_end();
273   // Check bundled MachineInstr's for hazards.
274   for (; MI != E && MI->isInsideBundle(); ++MI) {
275     CurrCycleInstr = &*MI;
276     unsigned WaitStates = PreEmitNoopsCommon(CurrCycleInstr);
277 
278     if (IsHazardRecognizerMode) {
279       fixHazards(CurrCycleInstr);
280 
281       insertNoopsInBundle(CurrCycleInstr, TII, WaitStates);
282     }
283 
284     // It’s unnecessary to track more than MaxLookAhead instructions. Since we
285     // include the bundled MI directly after, only add a maximum of
286     // (MaxLookAhead - 1) noops to EmittedInstrs.
287     for (unsigned i = 0, e = std::min(WaitStates, MaxLookAhead - 1); i < e; ++i)
288       EmittedInstrs.push_front(nullptr);
289 
290     EmittedInstrs.push_front(CurrCycleInstr);
291     EmittedInstrs.resize(MaxLookAhead);
292   }
293   CurrCycleInstr = nullptr;
294 }
295 
296 unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) {
297   IsHazardRecognizerMode = true;
298   CurrCycleInstr = MI;
299   unsigned W = PreEmitNoopsCommon(MI);
300   fixHazards(MI);
301   CurrCycleInstr = nullptr;
302   return W;
303 }
304 
305 unsigned GCNHazardRecognizer::PreEmitNoopsCommon(MachineInstr *MI) {
306   if (MI->isBundle())
307     return 0;
308 
309   int WaitStates = 0;
310 
311   if (SIInstrInfo::isSMRD(*MI))
312     return std::max(WaitStates, checkSMRDHazards(MI));
313 
314   if (ST.hasNSAtoVMEMBug())
315     WaitStates = std::max(WaitStates, checkNSAtoVMEMHazard(MI));
316 
317   WaitStates = std::max(WaitStates, checkFPAtomicToDenormModeHazard(MI));
318 
319   if (ST.hasNoDataDepHazard())
320     return WaitStates;
321 
322   if (SIInstrInfo::isVMEM(*MI) || SIInstrInfo::isFLAT(*MI))
323     WaitStates = std::max(WaitStates, checkVMEMHazards(MI));
324 
325   if (SIInstrInfo::isVALU(*MI))
326     WaitStates = std::max(WaitStates, checkVALUHazards(MI));
327 
328   if (SIInstrInfo::isDPP(*MI))
329     WaitStates = std::max(WaitStates, checkDPPHazards(MI));
330 
331   if (isDivFMas(MI->getOpcode()))
332     WaitStates = std::max(WaitStates, checkDivFMasHazards(MI));
333 
334   if (isRWLane(MI->getOpcode()))
335     WaitStates = std::max(WaitStates, checkRWLaneHazards(MI));
336 
337   if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) ||
338        SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) ||
339        SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0)
340     WaitStates = std::max(WaitStates, checkMAIVALUHazards(MI));
341 
342   if (MI->isInlineAsm())
343     return std::max(WaitStates, checkInlineAsmHazards(MI));
344 
345   if (isSGetReg(MI->getOpcode()))
346     return std::max(WaitStates, checkGetRegHazards(MI));
347 
348   if (isSSetReg(MI->getOpcode()))
349     return std::max(WaitStates, checkSetRegHazards(MI));
350 
351   if (isRFE(MI->getOpcode()))
352     return std::max(WaitStates, checkRFEHazards(MI));
353 
354   if (ST.hasReadM0MovRelInterpHazard() && (TII.isVINTRP(*MI) ||
355                                            isSMovRel(MI->getOpcode())))
356     return std::max(WaitStates, checkReadM0Hazards(MI));
357 
358   if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI))
359     return std::max(WaitStates, checkReadM0Hazards(MI));
360 
361   if (SIInstrInfo::isMAI(*MI))
362     return std::max(WaitStates, checkMAIHazards(MI));
363 
364   if (SIInstrInfo::isVMEM(*MI) ||
365       SIInstrInfo::isFLAT(*MI) ||
366       SIInstrInfo::isDS(*MI))
367     return std::max(WaitStates, checkMAILdStHazards(MI));
368 
369   return WaitStates;
370 }
371 
372 void GCNHazardRecognizer::EmitNoop() {
373   EmittedInstrs.push_front(nullptr);
374 }
375 
376 void GCNHazardRecognizer::AdvanceCycle() {
377   // When the scheduler detects a stall, it will call AdvanceCycle() without
378   // emitting any instructions.
379   if (!CurrCycleInstr) {
380     EmittedInstrs.push_front(nullptr);
381     return;
382   }
383 
384   if (CurrCycleInstr->isBundle()) {
385     processBundle();
386     return;
387   }
388 
389   unsigned NumWaitStates = TII.getNumWaitStates(*CurrCycleInstr);
390   if (!NumWaitStates) {
391     CurrCycleInstr = nullptr;
392     return;
393   }
394 
395   // Keep track of emitted instructions
396   EmittedInstrs.push_front(CurrCycleInstr);
397 
398   // Add a nullptr for each additional wait state after the first.  Make sure
399   // not to add more than getMaxLookAhead() items to the list, since we
400   // truncate the list to that size right after this loop.
401   for (unsigned i = 1, e = std::min(NumWaitStates, getMaxLookAhead());
402        i < e; ++i) {
403     EmittedInstrs.push_front(nullptr);
404   }
405 
406   // getMaxLookahead() is the largest number of wait states we will ever need
407   // to insert, so there is no point in keeping track of more than that many
408   // wait states.
409   EmittedInstrs.resize(getMaxLookAhead());
410 
411   CurrCycleInstr = nullptr;
412 }
413 
414 void GCNHazardRecognizer::RecedeCycle() {
415   llvm_unreachable("hazard recognizer does not support bottom-up scheduling.");
416 }
417 
418 //===----------------------------------------------------------------------===//
419 // Helper Functions
420 //===----------------------------------------------------------------------===//
421 
422 typedef function_ref<bool(const MachineInstr &, int WaitStates)> IsExpiredFn;
423 
424 // Returns a minimum wait states since \p I walking all predecessors.
425 // Only scans until \p IsExpired does not return true.
426 // Can only be run in a hazard recognizer mode.
427 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard,
428                               const MachineBasicBlock *MBB,
429                               MachineBasicBlock::const_reverse_instr_iterator I,
430                               int WaitStates, IsExpiredFn IsExpired,
431                               DenseSet<const MachineBasicBlock *> &Visited) {
432   for (auto E = MBB->instr_rend(); I != E; ++I) {
433     // Don't add WaitStates for parent BUNDLE instructions.
434     if (I->isBundle())
435       continue;
436 
437     if (IsHazard(*I))
438       return WaitStates;
439 
440     if (I->isInlineAsm())
441       continue;
442 
443     WaitStates += SIInstrInfo::getNumWaitStates(*I);
444 
445     if (IsExpired(*I, WaitStates))
446       return std::numeric_limits<int>::max();
447   }
448 
449   int MinWaitStates = std::numeric_limits<int>::max();
450   for (MachineBasicBlock *Pred : MBB->predecessors()) {
451     if (!Visited.insert(Pred).second)
452       continue;
453 
454     int W = getWaitStatesSince(IsHazard, Pred, Pred->instr_rbegin(),
455                                WaitStates, IsExpired, Visited);
456 
457     MinWaitStates = std::min(MinWaitStates, W);
458   }
459 
460   return MinWaitStates;
461 }
462 
463 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard,
464                               const MachineInstr *MI, IsExpiredFn IsExpired) {
465   DenseSet<const MachineBasicBlock *> Visited;
466   return getWaitStatesSince(IsHazard, MI->getParent(),
467                             std::next(MI->getReverseIterator()),
468                             0, IsExpired, Visited);
469 }
470 
471 int GCNHazardRecognizer::getWaitStatesSince(IsHazardFn IsHazard, int Limit) {
472   if (IsHazardRecognizerMode) {
473     auto IsExpiredFn = [Limit](const MachineInstr &, int WaitStates) {
474       return WaitStates >= Limit;
475     };
476     return ::getWaitStatesSince(IsHazard, CurrCycleInstr, IsExpiredFn);
477   }
478 
479   int WaitStates = 0;
480   for (MachineInstr *MI : EmittedInstrs) {
481     if (MI) {
482       if (IsHazard(*MI))
483         return WaitStates;
484 
485       if (MI->isInlineAsm())
486         continue;
487     }
488     ++WaitStates;
489 
490     if (WaitStates >= Limit)
491       break;
492   }
493   return std::numeric_limits<int>::max();
494 }
495 
496 int GCNHazardRecognizer::getWaitStatesSinceDef(unsigned Reg,
497                                                IsHazardFn IsHazardDef,
498                                                int Limit) {
499   const SIRegisterInfo *TRI = ST.getRegisterInfo();
500 
501   auto IsHazardFn = [IsHazardDef, TRI, Reg](const MachineInstr &MI) {
502     return IsHazardDef(MI) && MI.modifiesRegister(Reg, TRI);
503   };
504 
505   return getWaitStatesSince(IsHazardFn, Limit);
506 }
507 
508 int GCNHazardRecognizer::getWaitStatesSinceSetReg(IsHazardFn IsHazard,
509                                                   int Limit) {
510   auto IsHazardFn = [IsHazard](const MachineInstr &MI) {
511     return isSSetReg(MI.getOpcode()) && IsHazard(MI);
512   };
513 
514   return getWaitStatesSince(IsHazardFn, Limit);
515 }
516 
517 //===----------------------------------------------------------------------===//
518 // No-op Hazard Detection
519 //===----------------------------------------------------------------------===//
520 
521 static void addRegUnits(const SIRegisterInfo &TRI, BitVector &BV,
522                         MCRegister Reg) {
523   for (MCRegUnitIterator RUI(Reg, &TRI); RUI.isValid(); ++RUI)
524     BV.set(*RUI);
525 }
526 
527 static void addRegsToSet(const SIRegisterInfo &TRI,
528                          iterator_range<MachineInstr::const_mop_iterator> Ops,
529                          BitVector &Set) {
530   for (const MachineOperand &Op : Ops) {
531     if (Op.isReg())
532       addRegUnits(TRI, Set, Op.getReg().asMCReg());
533   }
534 }
535 
536 void GCNHazardRecognizer::addClauseInst(const MachineInstr &MI) {
537   // XXX: Do we need to worry about implicit operands
538   addRegsToSet(TRI, MI.defs(), ClauseDefs);
539   addRegsToSet(TRI, MI.uses(), ClauseUses);
540 }
541 
542 static bool breaksSMEMSoftClause(MachineInstr *MI) {
543   return !SIInstrInfo::isSMRD(*MI);
544 }
545 
546 static bool breaksVMEMSoftClause(MachineInstr *MI) {
547   return !SIInstrInfo::isVMEM(*MI) && !SIInstrInfo::isFLAT(*MI);
548 }
549 
550 int GCNHazardRecognizer::checkSoftClauseHazards(MachineInstr *MEM) {
551   // SMEM soft clause are only present on VI+, and only matter if xnack is
552   // enabled.
553   if (!ST.isXNACKEnabled())
554     return 0;
555 
556   bool IsSMRD = TII.isSMRD(*MEM);
557 
558   resetClause();
559 
560   // A soft-clause is any group of consecutive SMEM instructions.  The
561   // instructions in this group may return out of order and/or may be
562   // replayed (i.e. the same instruction issued more than once).
563   //
564   // In order to handle these situations correctly we need to make sure that
565   // when a clause has more than one instruction, no instruction in the clause
566   // writes to a register that is read by another instruction in the clause
567   // (including itself). If we encounter this situation, we need to break the
568   // clause by inserting a non SMEM instruction.
569 
570   for (MachineInstr *MI : EmittedInstrs) {
571     // When we hit a non-SMEM instruction then we have passed the start of the
572     // clause and we can stop.
573     if (!MI)
574       break;
575 
576     if (IsSMRD ? breaksSMEMSoftClause(MI) : breaksVMEMSoftClause(MI))
577       break;
578 
579     addClauseInst(*MI);
580   }
581 
582   if (ClauseDefs.none())
583     return 0;
584 
585   // We need to make sure not to put loads and stores in the same clause if they
586   // use the same address. For now, just start a new clause whenever we see a
587   // store.
588   if (MEM->mayStore())
589     return 1;
590 
591   addClauseInst(*MEM);
592 
593   // If the set of defs and uses intersect then we cannot add this instruction
594   // to the clause, so we have a hazard.
595   return ClauseDefs.anyCommon(ClauseUses) ? 1 : 0;
596 }
597 
598 int GCNHazardRecognizer::checkSMRDHazards(MachineInstr *SMRD) {
599   int WaitStatesNeeded = 0;
600 
601   WaitStatesNeeded = checkSoftClauseHazards(SMRD);
602 
603   // This SMRD hazard only affects SI.
604   if (!ST.hasSMRDReadVALUDefHazard())
605     return WaitStatesNeeded;
606 
607   // A read of an SGPR by SMRD instruction requires 4 wait states when the
608   // SGPR was written by a VALU instruction.
609   int SmrdSgprWaitStates = 4;
610   auto IsHazardDefFn = [this](const MachineInstr &MI) {
611     return TII.isVALU(MI);
612   };
613   auto IsBufferHazardDefFn = [this](const MachineInstr &MI) {
614     return TII.isSALU(MI);
615   };
616 
617   bool IsBufferSMRD = TII.isBufferSMRD(*SMRD);
618 
619   for (const MachineOperand &Use : SMRD->uses()) {
620     if (!Use.isReg())
621       continue;
622     int WaitStatesNeededForUse =
623         SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn,
624                                                    SmrdSgprWaitStates);
625     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
626 
627     // This fixes what appears to be undocumented hardware behavior in SI where
628     // s_mov writing a descriptor and s_buffer_load_dword reading the descriptor
629     // needs some number of nops in between. We don't know how many we need, but
630     // let's use 4. This wasn't discovered before probably because the only
631     // case when this happens is when we expand a 64-bit pointer into a full
632     // descriptor and use s_buffer_load_dword instead of s_load_dword, which was
633     // probably never encountered in the closed-source land.
634     if (IsBufferSMRD) {
635       int WaitStatesNeededForUse =
636         SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(),
637                                                    IsBufferHazardDefFn,
638                                                    SmrdSgprWaitStates);
639       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
640     }
641   }
642 
643   return WaitStatesNeeded;
644 }
645 
646 int GCNHazardRecognizer::checkVMEMHazards(MachineInstr* VMEM) {
647   if (!ST.hasVMEMReadSGPRVALUDefHazard())
648     return 0;
649 
650   int WaitStatesNeeded = checkSoftClauseHazards(VMEM);
651 
652   // A read of an SGPR by a VMEM instruction requires 5 wait states when the
653   // SGPR was written by a VALU Instruction.
654   const int VmemSgprWaitStates = 5;
655   auto IsHazardDefFn = [this](const MachineInstr &MI) {
656     return TII.isVALU(MI);
657   };
658   for (const MachineOperand &Use : VMEM->uses()) {
659     if (!Use.isReg() || TRI.isVectorRegister(MF.getRegInfo(), Use.getReg()))
660       continue;
661 
662     int WaitStatesNeededForUse =
663         VmemSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn,
664                                                    VmemSgprWaitStates);
665     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
666   }
667   return WaitStatesNeeded;
668 }
669 
670 int GCNHazardRecognizer::checkDPPHazards(MachineInstr *DPP) {
671   const SIRegisterInfo *TRI = ST.getRegisterInfo();
672   const SIInstrInfo *TII = ST.getInstrInfo();
673 
674   // Check for DPP VGPR read after VALU VGPR write and EXEC write.
675   int DppVgprWaitStates = 2;
676   int DppExecWaitStates = 5;
677   int WaitStatesNeeded = 0;
678   auto IsHazardDefFn = [TII](const MachineInstr &MI) {
679     return TII->isVALU(MI);
680   };
681 
682   for (const MachineOperand &Use : DPP->uses()) {
683     if (!Use.isReg() || !TRI->isVGPR(MF.getRegInfo(), Use.getReg()))
684       continue;
685     int WaitStatesNeededForUse =
686         DppVgprWaitStates - getWaitStatesSinceDef(
687                                 Use.getReg(),
688                                 [](const MachineInstr &) { return true; },
689                                 DppVgprWaitStates);
690     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
691   }
692 
693   WaitStatesNeeded = std::max(
694       WaitStatesNeeded,
695       DppExecWaitStates - getWaitStatesSinceDef(AMDGPU::EXEC, IsHazardDefFn,
696                                                 DppExecWaitStates));
697 
698   return WaitStatesNeeded;
699 }
700 
701 int GCNHazardRecognizer::checkDivFMasHazards(MachineInstr *DivFMas) {
702   const SIInstrInfo *TII = ST.getInstrInfo();
703 
704   // v_div_fmas requires 4 wait states after a write to vcc from a VALU
705   // instruction.
706   const int DivFMasWaitStates = 4;
707   auto IsHazardDefFn = [TII](const MachineInstr &MI) {
708     return TII->isVALU(MI);
709   };
710   int WaitStatesNeeded = getWaitStatesSinceDef(AMDGPU::VCC, IsHazardDefFn,
711                                                DivFMasWaitStates);
712 
713   return DivFMasWaitStates - WaitStatesNeeded;
714 }
715 
716 int GCNHazardRecognizer::checkGetRegHazards(MachineInstr *GetRegInstr) {
717   const SIInstrInfo *TII = ST.getInstrInfo();
718   unsigned GetRegHWReg = getHWReg(TII, *GetRegInstr);
719 
720   const int GetRegWaitStates = 2;
721   auto IsHazardFn = [TII, GetRegHWReg](const MachineInstr &MI) {
722     return GetRegHWReg == getHWReg(TII, MI);
723   };
724   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, GetRegWaitStates);
725 
726   return GetRegWaitStates - WaitStatesNeeded;
727 }
728 
729 int GCNHazardRecognizer::checkSetRegHazards(MachineInstr *SetRegInstr) {
730   const SIInstrInfo *TII = ST.getInstrInfo();
731   unsigned HWReg = getHWReg(TII, *SetRegInstr);
732 
733   const int SetRegWaitStates = ST.getSetRegWaitStates();
734   auto IsHazardFn = [TII, HWReg](const MachineInstr &MI) {
735     return HWReg == getHWReg(TII, MI);
736   };
737   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, SetRegWaitStates);
738   return SetRegWaitStates - WaitStatesNeeded;
739 }
740 
741 int GCNHazardRecognizer::createsVALUHazard(const MachineInstr &MI) {
742   if (!MI.mayStore())
743     return -1;
744 
745   const SIInstrInfo *TII = ST.getInstrInfo();
746   unsigned Opcode = MI.getOpcode();
747   const MCInstrDesc &Desc = MI.getDesc();
748 
749   int VDataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata);
750   int VDataRCID = -1;
751   if (VDataIdx != -1)
752     VDataRCID = Desc.OpInfo[VDataIdx].RegClass;
753 
754   if (TII->isMUBUF(MI) || TII->isMTBUF(MI)) {
755     // There is no hazard if the instruction does not use vector regs
756     // (like wbinvl1)
757     if (VDataIdx == -1)
758       return -1;
759     // For MUBUF/MTBUF instructions this hazard only exists if the
760     // instruction is not using a register in the soffset field.
761     const MachineOperand *SOffset =
762         TII->getNamedOperand(MI, AMDGPU::OpName::soffset);
763     // If we have no soffset operand, then assume this field has been
764     // hardcoded to zero.
765     if (AMDGPU::getRegBitWidth(VDataRCID) > 64 &&
766         (!SOffset || !SOffset->isReg()))
767       return VDataIdx;
768   }
769 
770   // MIMG instructions create a hazard if they don't use a 256-bit T# and
771   // the store size is greater than 8 bytes and they have more than two bits
772   // of their dmask set.
773   // All our MIMG definitions use a 256-bit T#, so we can skip checking for them.
774   if (TII->isMIMG(MI)) {
775     int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::srsrc);
776     assert(SRsrcIdx != -1 &&
777            AMDGPU::getRegBitWidth(Desc.OpInfo[SRsrcIdx].RegClass) == 256);
778     (void)SRsrcIdx;
779   }
780 
781   if (TII->isFLAT(MI)) {
782     int DataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata);
783     if (AMDGPU::getRegBitWidth(Desc.OpInfo[DataIdx].RegClass) > 64)
784       return DataIdx;
785   }
786 
787   return -1;
788 }
789 
790 int
791 GCNHazardRecognizer::checkVALUHazardsHelper(const MachineOperand &Def,
792                                             const MachineRegisterInfo &MRI) {
793   // Helper to check for the hazard where VMEM instructions that store more than
794   // 8 bytes can have there store data over written by the next instruction.
795   const SIRegisterInfo *TRI = ST.getRegisterInfo();
796 
797   const int VALUWaitStates = 1;
798   int WaitStatesNeeded = 0;
799 
800   if (!TRI->isVectorRegister(MRI, Def.getReg()))
801     return WaitStatesNeeded;
802   Register Reg = Def.getReg();
803   auto IsHazardFn = [this, Reg, TRI](const MachineInstr &MI) {
804     int DataIdx = createsVALUHazard(MI);
805     return DataIdx >= 0 &&
806            TRI->regsOverlap(MI.getOperand(DataIdx).getReg(), Reg);
807   };
808   int WaitStatesNeededForDef =
809     VALUWaitStates - getWaitStatesSince(IsHazardFn, VALUWaitStates);
810   WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForDef);
811 
812   return WaitStatesNeeded;
813 }
814 
815 int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) {
816   // This checks for the hazard where VMEM instructions that store more than
817   // 8 bytes can have there store data over written by the next instruction.
818   if (!ST.has12DWordStoreHazard())
819     return 0;
820 
821   const MachineRegisterInfo &MRI = MF.getRegInfo();
822   int WaitStatesNeeded = 0;
823 
824   for (const MachineOperand &Def : VALU->defs()) {
825     WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Def, MRI));
826   }
827 
828   return WaitStatesNeeded;
829 }
830 
831 int GCNHazardRecognizer::checkInlineAsmHazards(MachineInstr *IA) {
832   // This checks for hazards associated with inline asm statements.
833   // Since inline asms can contain just about anything, we use this
834   // to call/leverage other check*Hazard routines. Note that
835   // this function doesn't attempt to address all possible inline asm
836   // hazards (good luck), but is a collection of what has been
837   // problematic thus far.
838 
839   // see checkVALUHazards()
840   if (!ST.has12DWordStoreHazard())
841     return 0;
842 
843   const MachineRegisterInfo &MRI = MF.getRegInfo();
844   int WaitStatesNeeded = 0;
845 
846   for (unsigned I = InlineAsm::MIOp_FirstOperand, E = IA->getNumOperands();
847        I != E; ++I) {
848     const MachineOperand &Op = IA->getOperand(I);
849     if (Op.isReg() && Op.isDef()) {
850       WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Op, MRI));
851     }
852   }
853 
854   return WaitStatesNeeded;
855 }
856 
857 int GCNHazardRecognizer::checkRWLaneHazards(MachineInstr *RWLane) {
858   const SIInstrInfo *TII = ST.getInstrInfo();
859   const SIRegisterInfo *TRI = ST.getRegisterInfo();
860   const MachineRegisterInfo &MRI = MF.getRegInfo();
861 
862   const MachineOperand *LaneSelectOp =
863       TII->getNamedOperand(*RWLane, AMDGPU::OpName::src1);
864 
865   if (!LaneSelectOp->isReg() || !TRI->isSGPRReg(MRI, LaneSelectOp->getReg()))
866     return 0;
867 
868   Register LaneSelectReg = LaneSelectOp->getReg();
869   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVALU(MI); };
870 
871   const int RWLaneWaitStates = 4;
872   int WaitStatesSince = getWaitStatesSinceDef(LaneSelectReg, IsHazardFn,
873                                               RWLaneWaitStates);
874   return RWLaneWaitStates - WaitStatesSince;
875 }
876 
877 int GCNHazardRecognizer::checkRFEHazards(MachineInstr *RFE) {
878   if (!ST.hasRFEHazards())
879     return 0;
880 
881   const SIInstrInfo *TII = ST.getInstrInfo();
882 
883   const int RFEWaitStates = 1;
884 
885   auto IsHazardFn = [TII](const MachineInstr &MI) {
886     return getHWReg(TII, MI) == AMDGPU::Hwreg::ID_TRAPSTS;
887   };
888   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, RFEWaitStates);
889   return RFEWaitStates - WaitStatesNeeded;
890 }
891 
892 int GCNHazardRecognizer::checkReadM0Hazards(MachineInstr *MI) {
893   const SIInstrInfo *TII = ST.getInstrInfo();
894   const int SMovRelWaitStates = 1;
895   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isSALU(MI); };
896   return SMovRelWaitStates - getWaitStatesSinceDef(AMDGPU::M0, IsHazardFn,
897                                                    SMovRelWaitStates);
898 }
899 
900 void GCNHazardRecognizer::fixHazards(MachineInstr *MI) {
901   fixVMEMtoScalarWriteHazards(MI);
902   fixVcmpxPermlaneHazards(MI);
903   fixSMEMtoVectorWriteHazards(MI);
904   fixVcmpxExecWARHazard(MI);
905   fixLdsBranchVmemWARHazard(MI);
906 }
907 
908 bool GCNHazardRecognizer::fixVcmpxPermlaneHazards(MachineInstr *MI) {
909   if (!ST.hasVcmpxPermlaneHazard() || !isPermlane(*MI))
910     return false;
911 
912   const SIInstrInfo *TII = ST.getInstrInfo();
913   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVOPC(MI); };
914 
915   auto IsExpiredFn = [](const MachineInstr &MI, int) {
916     unsigned Opc = MI.getOpcode();
917     return SIInstrInfo::isVALU(MI) && Opc != AMDGPU::V_NOP_e32 &&
918            Opc != AMDGPU::V_NOP_e64 && Opc != AMDGPU::V_NOP_sdwa;
919   };
920 
921   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
922       std::numeric_limits<int>::max())
923     return false;
924 
925   // V_NOP will be discarded by SQ.
926   // Use V_MOV_B32 v?, v?. Register must be alive so use src0 of V_PERMLANE*
927   // which is always a VGPR and available.
928   auto *Src0 = TII->getNamedOperand(*MI, AMDGPU::OpName::src0);
929   Register Reg = Src0->getReg();
930   bool IsUndef = Src0->isUndef();
931   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
932           TII->get(AMDGPU::V_MOV_B32_e32))
933     .addReg(Reg, RegState::Define | (IsUndef ? RegState::Dead : 0))
934     .addReg(Reg, IsUndef ? RegState::Undef : RegState::Kill);
935 
936   return true;
937 }
938 
939 bool GCNHazardRecognizer::fixVMEMtoScalarWriteHazards(MachineInstr *MI) {
940   if (!ST.hasVMEMtoScalarWriteHazard())
941     return false;
942 
943   if (!SIInstrInfo::isSALU(*MI) && !SIInstrInfo::isSMRD(*MI))
944     return false;
945 
946   if (MI->getNumDefs() == 0)
947     return false;
948 
949   const SIRegisterInfo *TRI = ST.getRegisterInfo();
950 
951   auto IsHazardFn = [TRI, MI](const MachineInstr &I) {
952     if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isDS(I) &&
953         !SIInstrInfo::isFLAT(I))
954       return false;
955 
956     for (const MachineOperand &Def : MI->defs()) {
957       const MachineOperand *Op =
958           I.findRegisterUseOperand(Def.getReg(), false, TRI);
959       if (!Op)
960         continue;
961       return true;
962     }
963     return false;
964   };
965 
966   auto IsExpiredFn = [](const MachineInstr &MI, int) {
967     return SIInstrInfo::isVALU(MI) ||
968            (MI.getOpcode() == AMDGPU::S_WAITCNT &&
969             !MI.getOperand(0).getImm()) ||
970            (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
971             MI.getOperand(0).getImm() == 0xffe3);
972   };
973 
974   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
975       std::numeric_limits<int>::max())
976     return false;
977 
978   const SIInstrInfo *TII = ST.getInstrInfo();
979   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
980           TII->get(AMDGPU::S_WAITCNT_DEPCTR))
981       .addImm(0xffe3);
982   return true;
983 }
984 
985 bool GCNHazardRecognizer::fixSMEMtoVectorWriteHazards(MachineInstr *MI) {
986   if (!ST.hasSMEMtoVectorWriteHazard())
987     return false;
988 
989   if (!SIInstrInfo::isVALU(*MI))
990     return false;
991 
992   unsigned SDSTName;
993   switch (MI->getOpcode()) {
994   case AMDGPU::V_READLANE_B32:
995   case AMDGPU::V_READFIRSTLANE_B32:
996     SDSTName = AMDGPU::OpName::vdst;
997     break;
998   default:
999     SDSTName = AMDGPU::OpName::sdst;
1000     break;
1001   }
1002 
1003   const SIInstrInfo *TII = ST.getInstrInfo();
1004   const SIRegisterInfo *TRI = ST.getRegisterInfo();
1005   const AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST.getCPU());
1006   const MachineOperand *SDST = TII->getNamedOperand(*MI, SDSTName);
1007   if (!SDST) {
1008     for (const auto &MO : MI->implicit_operands()) {
1009       if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg()))) {
1010         SDST = &MO;
1011         break;
1012       }
1013     }
1014   }
1015 
1016   if (!SDST)
1017     return false;
1018 
1019   const Register SDSTReg = SDST->getReg();
1020   auto IsHazardFn = [SDSTReg, TRI](const MachineInstr &I) {
1021     return SIInstrInfo::isSMRD(I) && I.readsRegister(SDSTReg, TRI);
1022   };
1023 
1024   auto IsExpiredFn = [TII, IV](const MachineInstr &MI, int) {
1025     if (TII->isSALU(MI)) {
1026       switch (MI.getOpcode()) {
1027       case AMDGPU::S_SETVSKIP:
1028       case AMDGPU::S_VERSION:
1029       case AMDGPU::S_WAITCNT_VSCNT:
1030       case AMDGPU::S_WAITCNT_VMCNT:
1031       case AMDGPU::S_WAITCNT_EXPCNT:
1032         // These instructions cannot not mitigate the hazard.
1033         return false;
1034       case AMDGPU::S_WAITCNT_LGKMCNT:
1035         // Reducing lgkmcnt count to 0 always mitigates the hazard.
1036         return (MI.getOperand(1).getImm() == 0) &&
1037                (MI.getOperand(0).getReg() == AMDGPU::SGPR_NULL);
1038       case AMDGPU::S_WAITCNT: {
1039         const int64_t Imm = MI.getOperand(0).getImm();
1040         AMDGPU::Waitcnt Decoded = AMDGPU::decodeWaitcnt(IV, Imm);
1041         return (Decoded.LgkmCnt == 0);
1042       }
1043       default:
1044         // SOPP instructions cannot mitigate the hazard.
1045         if (TII->isSOPP(MI))
1046           return false;
1047         // At this point the SALU can be assumed to mitigate the hazard
1048         // because either:
1049         // (a) it is independent of the at risk SMEM (breaking chain),
1050         // or
1051         // (b) it is dependent on the SMEM, in which case an appropriate
1052         //     s_waitcnt lgkmcnt _must_ exist between it and the at risk
1053         //     SMEM instruction.
1054         return true;
1055       }
1056     }
1057     return false;
1058   };
1059 
1060   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1061       std::numeric_limits<int>::max())
1062     return false;
1063 
1064   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1065           TII->get(AMDGPU::S_MOV_B32), AMDGPU::SGPR_NULL)
1066       .addImm(0);
1067   return true;
1068 }
1069 
1070 bool GCNHazardRecognizer::fixVcmpxExecWARHazard(MachineInstr *MI) {
1071   if (!ST.hasVcmpxExecWARHazard() || !SIInstrInfo::isVALU(*MI))
1072     return false;
1073 
1074   const SIRegisterInfo *TRI = ST.getRegisterInfo();
1075   if (!MI->modifiesRegister(AMDGPU::EXEC, TRI))
1076     return false;
1077 
1078   auto IsHazardFn = [TRI](const MachineInstr &I) {
1079     if (SIInstrInfo::isVALU(I))
1080       return false;
1081     return I.readsRegister(AMDGPU::EXEC, TRI);
1082   };
1083 
1084   const SIInstrInfo *TII = ST.getInstrInfo();
1085   auto IsExpiredFn = [TII, TRI](const MachineInstr &MI, int) {
1086     if (SIInstrInfo::isVALU(MI)) {
1087       if (TII->getNamedOperand(MI, AMDGPU::OpName::sdst))
1088         return true;
1089       for (auto MO : MI.implicit_operands())
1090         if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg())))
1091           return true;
1092     }
1093     if (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
1094         (MI.getOperand(0).getImm() & 0xfffe) == 0xfffe)
1095       return true;
1096     return false;
1097   };
1098 
1099   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1100       std::numeric_limits<int>::max())
1101     return false;
1102 
1103   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1104           TII->get(AMDGPU::S_WAITCNT_DEPCTR))
1105     .addImm(0xfffe);
1106   return true;
1107 }
1108 
1109 static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF,
1110                                                  const GCNSubtarget &ST) {
1111   if (!ST.hasLdsBranchVmemWARHazard())
1112     return false;
1113 
1114   // Check if the necessary condition for the hazard is met: both LDS and VMEM
1115   // instructions need to appear in the same function.
1116   bool HasLds = false;
1117   bool HasVmem = false;
1118   for (auto &MBB : MF) {
1119     for (auto &MI : MBB) {
1120       HasLds |= SIInstrInfo::isDS(MI);
1121       HasVmem |=
1122           SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI);
1123       if (HasLds && HasVmem)
1124         return true;
1125     }
1126   }
1127   return false;
1128 }
1129 
1130 bool GCNHazardRecognizer::fixLdsBranchVmemWARHazard(MachineInstr *MI) {
1131   if (!RunLdsBranchVmemWARHazardFixup)
1132     return false;
1133 
1134   assert(ST.hasLdsBranchVmemWARHazard());
1135 
1136   auto IsHazardInst = [](const MachineInstr &MI) {
1137     if (SIInstrInfo::isDS(MI))
1138       return 1;
1139     if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI))
1140       return 2;
1141     return 0;
1142   };
1143 
1144   auto InstType = IsHazardInst(*MI);
1145   if (!InstType)
1146     return false;
1147 
1148   auto IsExpiredFn = [&IsHazardInst](const MachineInstr &I, int) {
1149     return IsHazardInst(I) || (I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT &&
1150                                I.getOperand(0).getReg() == AMDGPU::SGPR_NULL &&
1151                                !I.getOperand(1).getImm());
1152   };
1153 
1154   auto IsHazardFn = [InstType, &IsHazardInst](const MachineInstr &I) {
1155     if (!I.isBranch())
1156       return false;
1157 
1158     auto IsHazardFn = [InstType, IsHazardInst](const MachineInstr &I) {
1159       auto InstType2 = IsHazardInst(I);
1160       return InstType2 && InstType != InstType2;
1161     };
1162 
1163     auto IsExpiredFn = [InstType, &IsHazardInst](const MachineInstr &I, int) {
1164       auto InstType2 = IsHazardInst(I);
1165       if (InstType == InstType2)
1166         return true;
1167 
1168       return I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT &&
1169              I.getOperand(0).getReg() == AMDGPU::SGPR_NULL &&
1170              !I.getOperand(1).getImm();
1171     };
1172 
1173     return ::getWaitStatesSince(IsHazardFn, &I, IsExpiredFn) !=
1174            std::numeric_limits<int>::max();
1175   };
1176 
1177   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1178       std::numeric_limits<int>::max())
1179     return false;
1180 
1181   const SIInstrInfo *TII = ST.getInstrInfo();
1182   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1183           TII->get(AMDGPU::S_WAITCNT_VSCNT))
1184     .addReg(AMDGPU::SGPR_NULL, RegState::Undef)
1185     .addImm(0);
1186 
1187   return true;
1188 }
1189 
1190 int GCNHazardRecognizer::checkNSAtoVMEMHazard(MachineInstr *MI) {
1191   int NSAtoVMEMWaitStates = 1;
1192 
1193   if (!ST.hasNSAtoVMEMBug())
1194     return 0;
1195 
1196   if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isMTBUF(*MI))
1197     return 0;
1198 
1199   const SIInstrInfo *TII = ST.getInstrInfo();
1200   const auto *Offset = TII->getNamedOperand(*MI, AMDGPU::OpName::offset);
1201   if (!Offset || (Offset->getImm() & 6) == 0)
1202     return 0;
1203 
1204   auto IsHazardFn = [TII](const MachineInstr &I) {
1205     if (!SIInstrInfo::isMIMG(I))
1206       return false;
1207     const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(I.getOpcode());
1208     return Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA &&
1209            TII->getInstSizeInBytes(I) >= 16;
1210   };
1211 
1212   return NSAtoVMEMWaitStates - getWaitStatesSince(IsHazardFn, 1);
1213 }
1214 
1215 int GCNHazardRecognizer::checkFPAtomicToDenormModeHazard(MachineInstr *MI) {
1216   int FPAtomicToDenormModeWaitStates = 3;
1217 
1218   if (MI->getOpcode() != AMDGPU::S_DENORM_MODE)
1219     return 0;
1220 
1221   auto IsHazardFn = [](const MachineInstr &I) {
1222     if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isFLAT(I))
1223       return false;
1224     return SIInstrInfo::isFPAtomic(I);
1225   };
1226 
1227   auto IsExpiredFn = [](const MachineInstr &MI, int WaitStates) {
1228     if (WaitStates >= 3 || SIInstrInfo::isVALU(MI))
1229       return true;
1230 
1231     switch (MI.getOpcode()) {
1232     case AMDGPU::S_WAITCNT:
1233     case AMDGPU::S_WAITCNT_VSCNT:
1234     case AMDGPU::S_WAITCNT_VMCNT:
1235     case AMDGPU::S_WAITCNT_EXPCNT:
1236     case AMDGPU::S_WAITCNT_LGKMCNT:
1237     case AMDGPU::S_WAIT_IDLE:
1238       return true;
1239     default:
1240       break;
1241     }
1242 
1243     return false;
1244   };
1245 
1246   return FPAtomicToDenormModeWaitStates -
1247          ::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn);
1248 }
1249 
1250 int GCNHazardRecognizer::checkMAIHazards(MachineInstr *MI) {
1251   assert(SIInstrInfo::isMAI(*MI));
1252 
1253   return ST.hasGFX90AInsts() ? checkMAIHazards90A(MI) : checkMAIHazards908(MI);
1254 }
1255 
1256 int GCNHazardRecognizer::checkMFMAPadding(MachineInstr *MI) {
1257   // Early exit if no padding is requested.
1258   if (MFMAPaddingRatio == 0)
1259     return 0;
1260 
1261   auto IsMFMAFn = [](const MachineInstr &MI) {
1262     return SIInstrInfo::isMAI(MI) &&
1263            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1264            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1265   };
1266 
1267   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1268   if (!IsMFMAFn(*MI) || MFI->getOccupancy() < 2)
1269     return 0;
1270 
1271   int NeighborMFMALatency = 0;
1272   auto IsNeighboringMFMA = [&IsMFMAFn, &NeighborMFMALatency,
1273                             this](const MachineInstr &MI) {
1274     if (!IsMFMAFn(MI))
1275       return false;
1276 
1277     NeighborMFMALatency = this->getMFMAPipelineWaitStates(MI);
1278     return true;
1279   };
1280 
1281   const int MaxMFMAPipelineWaitStates = 16;
1282   int WaitStatesSinceNeighborMFMA =
1283       getWaitStatesSince(IsNeighboringMFMA, MaxMFMAPipelineWaitStates);
1284 
1285   int NeighborMFMAPaddingNeeded =
1286       (NeighborMFMALatency * MFMAPaddingRatio / 100) -
1287       WaitStatesSinceNeighborMFMA;
1288 
1289   return std::max(0, NeighborMFMAPaddingNeeded);
1290 }
1291 
1292 int GCNHazardRecognizer::checkMAIHazards908(MachineInstr *MI) {
1293   int WaitStatesNeeded = 0;
1294   unsigned Opc = MI->getOpcode();
1295 
1296   auto IsVALUFn = [](const MachineInstr &MI) {
1297     return SIInstrInfo::isVALU(MI);
1298   };
1299 
1300   if (Opc != AMDGPU::V_ACCVGPR_READ_B32_e64) { // MFMA or v_accvgpr_write
1301     const int LegacyVALUWritesVGPRWaitStates = 2;
1302     const int VALUWritesExecWaitStates = 4;
1303     const int MaxWaitStates = 4;
1304 
1305     int WaitStatesNeededForUse = VALUWritesExecWaitStates -
1306       getWaitStatesSinceDef(AMDGPU::EXEC, IsVALUFn, MaxWaitStates);
1307     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1308 
1309     if (WaitStatesNeeded < MaxWaitStates) {
1310       for (const MachineOperand &Use : MI->explicit_uses()) {
1311         const int MaxWaitStates = 2;
1312 
1313         if (!Use.isReg() || !TRI.isVGPR(MF.getRegInfo(), Use.getReg()))
1314           continue;
1315 
1316         int WaitStatesNeededForUse = LegacyVALUWritesVGPRWaitStates -
1317           getWaitStatesSinceDef(Use.getReg(), IsVALUFn, MaxWaitStates);
1318         WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1319 
1320         if (WaitStatesNeeded == MaxWaitStates)
1321           break;
1322       }
1323     }
1324   }
1325 
1326   auto IsMFMAFn = [](const MachineInstr &MI) {
1327     return SIInstrInfo::isMAI(MI) &&
1328            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1329            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1330   };
1331 
1332   for (const MachineOperand &Op : MI->explicit_operands()) {
1333     if (!Op.isReg() || !TRI.isAGPR(MF.getRegInfo(), Op.getReg()))
1334       continue;
1335 
1336     if (Op.isDef() && Opc != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1337       continue;
1338 
1339     const int MFMAWritesAGPROverlappedSrcABWaitStates = 4;
1340     const int MFMAWritesAGPROverlappedSrcCWaitStates = 2;
1341     const int MFMA4x4WritesAGPRAccVgprReadWaitStates = 4;
1342     const int MFMA16x16WritesAGPRAccVgprReadWaitStates = 10;
1343     const int MFMA32x32WritesAGPRAccVgprReadWaitStates = 18;
1344     const int MFMA4x4WritesAGPRAccVgprWriteWaitStates = 1;
1345     const int MFMA16x16WritesAGPRAccVgprWriteWaitStates = 7;
1346     const int MFMA32x32WritesAGPRAccVgprWriteWaitStates = 15;
1347     const int MaxWaitStates = 18;
1348     Register Reg = Op.getReg();
1349     unsigned HazardDefLatency = 0;
1350 
1351     auto IsOverlappedMFMAFn = [Reg, &IsMFMAFn, &HazardDefLatency,
1352                                this](const MachineInstr &MI) {
1353       if (!IsMFMAFn(MI))
1354         return false;
1355       Register DstReg = MI.getOperand(0).getReg();
1356       if (DstReg == Reg)
1357         return false;
1358       HazardDefLatency =
1359           std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI));
1360       return TRI.regsOverlap(DstReg, Reg);
1361     };
1362 
1363     int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsOverlappedMFMAFn,
1364                                                    MaxWaitStates);
1365     int NeedWaitStates = MFMAWritesAGPROverlappedSrcABWaitStates;
1366     int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
1367     int OpNo = MI->getOperandNo(&Op);
1368     if (OpNo == SrcCIdx) {
1369       NeedWaitStates = MFMAWritesAGPROverlappedSrcCWaitStates;
1370     } else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) {
1371       switch (HazardDefLatency) {
1372       case 2:  NeedWaitStates = MFMA4x4WritesAGPRAccVgprReadWaitStates;
1373                break;
1374       case 8:  NeedWaitStates = MFMA16x16WritesAGPRAccVgprReadWaitStates;
1375                break;
1376       case 16: LLVM_FALLTHROUGH;
1377       default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprReadWaitStates;
1378                break;
1379       }
1380     } else if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) {
1381       switch (HazardDefLatency) {
1382       case 2:  NeedWaitStates = MFMA4x4WritesAGPRAccVgprWriteWaitStates;
1383                break;
1384       case 8:  NeedWaitStates = MFMA16x16WritesAGPRAccVgprWriteWaitStates;
1385                break;
1386       case 16: LLVM_FALLTHROUGH;
1387       default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprWriteWaitStates;
1388                break;
1389       }
1390     }
1391 
1392     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1393     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1394 
1395     if (WaitStatesNeeded == MaxWaitStates)
1396       return WaitStatesNeeded; // Early exit.
1397 
1398     auto IsAccVgprWriteFn = [Reg, this](const MachineInstr &MI) {
1399       if (MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1400         return false;
1401       Register DstReg = MI.getOperand(0).getReg();
1402       return TRI.regsOverlap(Reg, DstReg);
1403     };
1404 
1405     const int AccVGPRWriteMFMAReadSrcCWaitStates = 1;
1406     const int AccVGPRWriteMFMAReadSrcABWaitStates = 3;
1407     const int AccVGPRWriteAccVgprReadWaitStates = 3;
1408     NeedWaitStates = AccVGPRWriteMFMAReadSrcABWaitStates;
1409     if (OpNo == SrcCIdx)
1410       NeedWaitStates = AccVGPRWriteMFMAReadSrcCWaitStates;
1411     else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64)
1412       NeedWaitStates = AccVGPRWriteAccVgprReadWaitStates;
1413 
1414     WaitStatesNeededForUse = NeedWaitStates -
1415       getWaitStatesSinceDef(Reg, IsAccVgprWriteFn, MaxWaitStates);
1416     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1417 
1418     if (WaitStatesNeeded == MaxWaitStates)
1419       return WaitStatesNeeded; // Early exit.
1420   }
1421 
1422   if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) {
1423     const int MFMA4x4ReadSrcCAccVgprWriteWaitStates = 0;
1424     const int MFMA16x16ReadSrcCAccVgprWriteWaitStates = 5;
1425     const int MFMA32x32ReadSrcCAccVgprWriteWaitStates = 13;
1426     const int MaxWaitStates = 13;
1427     Register DstReg = MI->getOperand(0).getReg();
1428     unsigned HazardDefLatency = 0;
1429 
1430     auto IsSrcCMFMAFn = [DstReg, &IsMFMAFn, &HazardDefLatency,
1431                          this](const MachineInstr &MI) {
1432       if (!IsMFMAFn(MI))
1433         return false;
1434       Register Reg = TII.getNamedOperand(MI, AMDGPU::OpName::src2)->getReg();
1435       HazardDefLatency =
1436           std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI));
1437       return TRI.regsOverlap(Reg, DstReg);
1438     };
1439 
1440     int WaitStatesSince = getWaitStatesSince(IsSrcCMFMAFn, MaxWaitStates);
1441     int NeedWaitStates;
1442     switch (HazardDefLatency) {
1443     case 2:  NeedWaitStates = MFMA4x4ReadSrcCAccVgprWriteWaitStates;
1444              break;
1445     case 8:  NeedWaitStates = MFMA16x16ReadSrcCAccVgprWriteWaitStates;
1446              break;
1447     case 16: LLVM_FALLTHROUGH;
1448     default: NeedWaitStates = MFMA32x32ReadSrcCAccVgprWriteWaitStates;
1449              break;
1450     }
1451 
1452     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSince;
1453     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1454   }
1455 
1456   // Pad neighboring MFMA with noops for better inter-wave performance.
1457   WaitStatesNeeded = std::max(WaitStatesNeeded, checkMFMAPadding(MI));
1458 
1459   return WaitStatesNeeded;
1460 }
1461 
1462 int GCNHazardRecognizer::checkMAIHazards90A(MachineInstr *MI) {
1463   int WaitStatesNeeded = 0;
1464   unsigned Opc = MI->getOpcode();
1465 
1466   auto IsMFMAFn = [](const MachineInstr &MI) {
1467     return SIInstrInfo::isMAI(MI) &&
1468            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1469            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1470   };
1471 
1472   auto IsLegacyVALUFn = [&IsMFMAFn](const MachineInstr &MI) {
1473     return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI);
1474   };
1475 
1476   auto IsLegacyVALUNotDotFn = [&IsMFMAFn](const MachineInstr &MI) {
1477     return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI) && !SIInstrInfo::isDOT(MI);
1478   };
1479 
1480   if (!IsMFMAFn(*MI))
1481     return WaitStatesNeeded;
1482 
1483   const int VALUWritesExecWaitStates = 4;
1484   int WaitStatesNeededForUse = VALUWritesExecWaitStates -
1485     getWaitStatesSinceDef(AMDGPU::EXEC, IsLegacyVALUFn,
1486                           VALUWritesExecWaitStates);
1487   WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1488 
1489   int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
1490 
1491   // Loop for both DGEMM and S/HGEMM 2nd instruction.
1492   for (const MachineOperand &Use : MI->explicit_uses()) {
1493     const int LegacyVALUNotDotWritesVGPRWaitStates = 2;
1494     const int SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates = 2;
1495     const int GFX940_XDL2PassWritesVGPROverlappedSMFMASrcCWaitStates = 3;
1496     const int GFX940_XDL4PassWritesVGPROverlappedSMFMASrcCWaitStates = 5;
1497     const int GFX940_SMFMA4PassWritesVGPROverlappedSMFMASrcCWaitStates = 4;
1498     const int GFX940_XDL8PassWritesVGPROverlappedSMFMASrcCWaitStates = 9;
1499     const int GFX940_SMFMA8PassWritesVGPROverlappedSMFMASrcCWaitStates = 8;
1500     const int GFX940_XDL16PassWritesVGPROverlappedSMFMASrcCWaitStates = 17;
1501     const int GFX940_SMFMA16PassWritesVGPROverlappedSMFMASrcCWaitStates = 16;
1502     const int SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates = 8;
1503     const int SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates = 16;
1504     const int SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates = 3;
1505     const int SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates = 9;
1506     const int SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates = 17;
1507     const int DMFMA16x16WritesVGPROverlappedSrcCWaitStates = 9;
1508     const int DMFMA4x4WritesVGPROverlappedSrcCWaitStates = 4;
1509     const int SMFMA4x4WritesVGPROverlappedSrcABWaitStates = 5;
1510     const int SMFMA16x16WritesVGPROverlappedSrcABWaitStates = 11;
1511     const int SMFMA32x32WritesVGPROverlappedSrcABWaitStates = 19;
1512     const int GFX940_SMFMA2PassWritesVGPROverlappedSrcABWaitStates = 4;
1513     const int GFX940_SMFMA4PassWritesVGPROverlappedSrcABWaitStates = 6;
1514     const int GFX940_SMFMA8PassWritesVGPROverlappedSrcABWaitStates = 10;
1515     const int GFX940_SMFMA16PassWritesVGPROverlappedSrcABWaitStates = 18;
1516     const int GFX940_XDL2PassWritesVGPROverlappedSrcABWaitStates = 5;
1517     const int GFX940_XDL4PassWritesVGPROverlappedSrcABWaitStates = 7;
1518     const int GFX940_XDL8PassWritesVGPROverlappedSrcABWaitStates = 11;
1519     const int GFX940_XDL16PassWritesVGPROverlappedSrcABWaitStates = 19;
1520     const int DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates = 6;
1521     const int DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates = 11;
1522     const int DMFMA4x4WritesVGPRFullSrcCWaitStates = 4;
1523     const int GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates = 2;
1524     const int MaxWaitStates = 19;
1525 
1526     if (!Use.isReg())
1527       continue;
1528     Register Reg = Use.getReg();
1529     bool FullReg;
1530     const MachineInstr *MI1;
1531 
1532     auto IsOverlappedMFMAFn = [Reg, &IsMFMAFn, &FullReg, &MI1,
1533                                this](const MachineInstr &MI) {
1534       if (!IsMFMAFn(MI))
1535         return false;
1536       Register DstReg = MI.getOperand(0).getReg();
1537       FullReg = (DstReg == Reg);
1538       MI1 = &MI;
1539       return TRI.regsOverlap(DstReg, Reg);
1540     };
1541 
1542     WaitStatesNeededForUse = LegacyVALUNotDotWritesVGPRWaitStates -
1543       getWaitStatesSinceDef(Reg, IsLegacyVALUNotDotFn, MaxWaitStates);
1544     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1545 
1546     int NumWaitStates =
1547         getWaitStatesSinceDef(Reg, IsOverlappedMFMAFn, MaxWaitStates);
1548     if (NumWaitStates == std::numeric_limits<int>::max())
1549       continue;
1550 
1551     int OpNo = MI->getOperandNo(&Use);
1552     unsigned Opc1 = MI1->getOpcode();
1553     int NeedWaitStates = 0;
1554     if (OpNo == SrcCIdx) {
1555       if (!isDGEMM(Opc) && (!ST.hasGFX940Insts() && isDGEMM(Opc1))) {
1556         NeedWaitStates = 0;
1557       } else if (FullReg) {
1558         if ((Opc == AMDGPU::V_MFMA_F64_4X4X4F64_e64 ||
1559              Opc == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64) &&
1560             (Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_e64 ||
1561              Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64))
1562           NeedWaitStates = DMFMA4x4WritesVGPRFullSrcCWaitStates;
1563         else if (ST.hasGFX940Insts() &&
1564                  TSchedModel.computeInstrLatency(MI1) == 2)
1565           NeedWaitStates = GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates;
1566       } else {
1567         switch (Opc1) {
1568         case AMDGPU::V_MFMA_F64_16X16X4F64_e64:
1569         case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64:
1570         case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64:
1571         case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64:
1572           if (!isXDL(ST, *MI))
1573             NeedWaitStates = DMFMA16x16WritesVGPROverlappedSrcCWaitStates;
1574           break;
1575         case AMDGPU::V_MFMA_F64_4X4X4F64_e64:
1576         case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64:
1577           if (!isXDL(ST, *MI))
1578             NeedWaitStates = DMFMA4x4WritesVGPROverlappedSrcCWaitStates;
1579           break;
1580         default:
1581           if (ST.hasGFX940Insts() && isXDL(ST, *MI) && !isXDL(ST, *MI1))
1582             break;
1583           switch (TSchedModel.computeInstrLatency(MI1)) {
1584           case 2:
1585             NeedWaitStates = ST.hasGFX940Insts()
1586               ? isXDL(ST, *MI1)
1587                 ? GFX940_XDL2PassWritesVGPROverlappedSMFMASrcCWaitStates
1588                 : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates
1589               : isDGEMM(Opc)
1590                 ? SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates
1591                 : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates;
1592             break;
1593           case 4:
1594             assert(ST.hasGFX940Insts());
1595             NeedWaitStates = isXDL(ST, *MI1)
1596               ? GFX940_XDL4PassWritesVGPROverlappedSMFMASrcCWaitStates
1597               : GFX940_SMFMA4PassWritesVGPROverlappedSMFMASrcCWaitStates;
1598             break;
1599           case 8:
1600             NeedWaitStates = ST.hasGFX940Insts()
1601               ? isXDL(ST, *MI1)
1602                 ? GFX940_XDL8PassWritesVGPROverlappedSMFMASrcCWaitStates
1603                 : GFX940_SMFMA8PassWritesVGPROverlappedSMFMASrcCWaitStates
1604               : isDGEMM(Opc)
1605                 ? SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates
1606                 : SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates;
1607             break;
1608           case 16: LLVM_FALLTHROUGH;
1609           default:
1610             NeedWaitStates = ST.hasGFX940Insts()
1611               ? isXDL(ST, *MI1)
1612                 ? GFX940_XDL16PassWritesVGPROverlappedSMFMASrcCWaitStates
1613                 : GFX940_SMFMA16PassWritesVGPROverlappedSMFMASrcCWaitStates
1614               : isDGEMM(Opc)
1615                 ? SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates
1616                 : SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates;
1617           }
1618         }
1619       }
1620     } else {
1621       switch (Opc1) {
1622       case AMDGPU::V_MFMA_F64_16X16X4F64_e64:
1623       case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64:
1624       case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64:
1625       case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64:
1626         NeedWaitStates = DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates;
1627         break;
1628       case AMDGPU::V_MFMA_F64_4X4X4F64_e64:
1629       case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64:
1630         NeedWaitStates = DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates;
1631         break;
1632       default:
1633         switch (TSchedModel.computeInstrLatency(MI1)) {
1634         case 2:
1635           NeedWaitStates = ST.hasGFX940Insts()
1636             ? isXDL(ST, *MI1)
1637               ? GFX940_XDL2PassWritesVGPROverlappedSrcABWaitStates
1638               : GFX940_SMFMA2PassWritesVGPROverlappedSrcABWaitStates
1639             : SMFMA4x4WritesVGPROverlappedSrcABWaitStates;
1640           break;
1641         case 4:
1642           assert(ST.hasGFX940Insts());
1643           NeedWaitStates = isXDL(ST, *MI1)
1644             ? GFX940_XDL4PassWritesVGPROverlappedSrcABWaitStates
1645             : GFX940_SMFMA4PassWritesVGPROverlappedSrcABWaitStates;
1646           break;
1647         case 8:
1648           NeedWaitStates = ST.hasGFX940Insts()
1649             ? isXDL(ST, *MI1)
1650               ? GFX940_XDL8PassWritesVGPROverlappedSrcABWaitStates
1651               : GFX940_SMFMA8PassWritesVGPROverlappedSrcABWaitStates
1652             : SMFMA16x16WritesVGPROverlappedSrcABWaitStates;
1653           break;
1654         case 16: LLVM_FALLTHROUGH;
1655         default:
1656           NeedWaitStates = ST.hasGFX940Insts()
1657             ? isXDL(ST, *MI1)
1658               ? GFX940_XDL16PassWritesVGPROverlappedSrcABWaitStates
1659               : GFX940_SMFMA16PassWritesVGPROverlappedSrcABWaitStates
1660             : SMFMA32x32WritesVGPROverlappedSrcABWaitStates;
1661         }
1662       }
1663     }
1664     if (WaitStatesNeeded >= NeedWaitStates)
1665       continue;
1666 
1667     WaitStatesNeededForUse = NeedWaitStates - NumWaitStates;
1668     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1669 
1670     if (WaitStatesNeeded == MaxWaitStates)
1671       break;
1672   }
1673 
1674   return WaitStatesNeeded;
1675 }
1676 
1677 int GCNHazardRecognizer::checkMAILdStHazards(MachineInstr *MI) {
1678   // On gfx90a+ relevant hazards are checked in checkMAIVALUHazards()
1679   if (!ST.hasMAIInsts() || ST.hasGFX90AInsts())
1680     return 0;
1681 
1682   int WaitStatesNeeded = 0;
1683 
1684   auto IsAccVgprReadFn = [](const MachineInstr &MI) {
1685     return MI.getOpcode() == AMDGPU::V_ACCVGPR_READ_B32_e64;
1686   };
1687 
1688   for (const MachineOperand &Op : MI->explicit_uses()) {
1689     if (!Op.isReg() || !TRI.isVGPR(MF.getRegInfo(), Op.getReg()))
1690       continue;
1691 
1692     Register Reg = Op.getReg();
1693 
1694     const int AccVgprReadLdStWaitStates = 2;
1695     const int VALUWriteAccVgprRdWrLdStDepVALUWaitStates = 1;
1696     const int MaxWaitStates = 2;
1697 
1698     int WaitStatesNeededForUse = AccVgprReadLdStWaitStates -
1699       getWaitStatesSinceDef(Reg, IsAccVgprReadFn, MaxWaitStates);
1700     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1701 
1702     if (WaitStatesNeeded == MaxWaitStates)
1703       return WaitStatesNeeded; // Early exit.
1704 
1705     auto IsVALUAccVgprRdWrCheckFn = [Reg, this](const MachineInstr &MI) {
1706       if (MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64 &&
1707           MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1708         return false;
1709       auto IsVALUFn = [](const MachineInstr &MI) {
1710         return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMAI(MI);
1711       };
1712       return getWaitStatesSinceDef(Reg, IsVALUFn, 2 /*MaxWaitStates*/) <
1713              std::numeric_limits<int>::max();
1714     };
1715 
1716     WaitStatesNeededForUse = VALUWriteAccVgprRdWrLdStDepVALUWaitStates -
1717       getWaitStatesSince(IsVALUAccVgprRdWrCheckFn, MaxWaitStates);
1718     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1719   }
1720 
1721   return WaitStatesNeeded;
1722 }
1723 
1724 int GCNHazardRecognizer::checkMAIVALUHazards(MachineInstr *MI) {
1725   if (!ST.hasGFX90AInsts())
1726     return 0;
1727 
1728   auto IsMFMAFn = [](const MachineInstr &MI) -> bool {
1729     return SIInstrInfo::isMAI(MI) &&
1730            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1731            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1732   };
1733 
1734   auto IsDGEMMFn = [](const MachineInstr &MI) -> bool {
1735     return isDGEMM(MI.getOpcode());
1736   };
1737 
1738   // This is checked in checkMAIHazards90A()
1739   if (IsMFMAFn(*MI))
1740     return 0;
1741 
1742   int WaitStatesNeeded = 0;
1743 
1744   bool IsMemOrExport = SIInstrInfo::isVMEM(*MI) ||
1745                        SIInstrInfo::isFLAT(*MI) ||
1746                        SIInstrInfo::isDS(*MI) ||
1747                        SIInstrInfo::isEXP(*MI);
1748   bool IsVALU = SIInstrInfo::isVALU(*MI);
1749 
1750   const MachineInstr *MFMA = nullptr;
1751   unsigned Reg;
1752   auto IsMFMAWriteFn = [&Reg, &IsMFMAFn, &MFMA, this](const MachineInstr &MI) {
1753     if (!IsMFMAFn(MI) || !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg))
1754       return false;
1755     MFMA = &MI;
1756     return true;
1757   };
1758 
1759   const MachineInstr *DOT = nullptr;
1760   auto IsDotWriteFn = [&Reg, &DOT, this](const MachineInstr &MI) {
1761     if (!SIInstrInfo::isDOT(MI) ||
1762         !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg))
1763       return false;
1764     DOT = &MI;
1765     return true;
1766   };
1767 
1768   int SrcCIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1769                                            AMDGPU::OpName::src2);
1770 
1771   if (IsMemOrExport || IsVALU) {
1772     const int SMFMA4x4WriteVgprVALUMemExpReadWaitStates = 5;
1773     const int SMFMA16x16WriteVgprVALUMemExpReadWaitStates = 11;
1774     const int SMFMA32x32WriteVgprVALUMemExpReadWaitStates = 19;
1775     const int GFX940_SMFMA2PassWriteVgprVALUMemExpReadWaitStates = 4;
1776     const int GFX940_SMFMA4PassWriteVgprVALUMemExpReadWaitStates = 6;
1777     const int GFX940_SMFMA8PassWriteVgprVALUMemExpReadWaitStates = 10;
1778     const int GFX940_SMFMA16PassWriteVgprVALUMemExpReadWaitStates = 18;
1779     const int GFX940_XDL2PassWriteVgprVALUMemExpReadWaitStates = 5;
1780     const int GFX940_XDL4PassWriteVgprVALUMemExpReadWaitStates = 7;
1781     const int GFX940_XDL8PassWriteVgprVALUMemExpReadWaitStates = 11;
1782     const int GFX940_XDL16PassWriteVgprVALUMemExpReadWaitStates = 19;
1783     const int DMFMA4x4WriteVgprMemExpReadWaitStates = 9;
1784     const int DMFMA16x16WriteVgprMemExpReadWaitStates = 18;
1785     const int DMFMA4x4WriteVgprVALUReadWaitStates = 6;
1786     const int DMFMA16x16WriteVgprVALUReadWaitStates = 11;
1787     const int DotWriteSameDotReadSrcAB = 3;
1788     const int DotWriteDifferentVALURead = 3;
1789     const int MaxWaitStates = 19;
1790 
1791     for (const MachineOperand &Use : MI->explicit_uses()) {
1792       if (!Use.isReg())
1793         continue;
1794       Reg = Use.getReg();
1795 
1796       DOT = nullptr;
1797       int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn,
1798                                                      MaxWaitStates);
1799       if (DOT) {
1800         int NeedWaitStates = 0;
1801         if (DOT->getOpcode() == MI->getOpcode()) {
1802           if (&Use - &MI->getOperand(0) != SrcCIdx)
1803             NeedWaitStates = DotWriteSameDotReadSrcAB;
1804         } else {
1805           NeedWaitStates = DotWriteDifferentVALURead;
1806         }
1807 
1808         int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1809         WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1810       }
1811 
1812       MFMA = nullptr;
1813       WaitStatesSinceDef =
1814           getWaitStatesSinceDef(Reg, IsMFMAWriteFn, MaxWaitStates);
1815       if (!MFMA)
1816         continue;
1817 
1818       unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA);
1819       int NeedWaitStates = MaxWaitStates;
1820       switch (HazardDefLatency) {
1821       case 2:
1822         NeedWaitStates =
1823           ST.hasGFX940Insts()
1824             ? isXDL(ST, *MFMA)
1825               ? GFX940_XDL2PassWriteVgprVALUMemExpReadWaitStates
1826               : GFX940_SMFMA2PassWriteVgprVALUMemExpReadWaitStates
1827             : SMFMA4x4WriteVgprVALUMemExpReadWaitStates;
1828         break;
1829       case 4:
1830         assert(isDGEMM(MFMA->getOpcode()) || ST.hasGFX940Insts());
1831         NeedWaitStates =
1832           isDGEMM(MFMA->getOpcode())
1833             ? IsMemOrExport ? DMFMA4x4WriteVgprMemExpReadWaitStates
1834                             : DMFMA4x4WriteVgprVALUReadWaitStates
1835             : isXDL(ST, *MFMA)
1836               ? GFX940_XDL4PassWriteVgprVALUMemExpReadWaitStates
1837               : GFX940_SMFMA4PassWriteVgprVALUMemExpReadWaitStates;
1838         break;
1839       case 8:
1840         NeedWaitStates =
1841           ST.hasGFX940Insts()
1842             ? isXDL(ST, *MFMA)
1843               ? GFX940_XDL8PassWriteVgprVALUMemExpReadWaitStates
1844               : GFX940_SMFMA8PassWriteVgprVALUMemExpReadWaitStates
1845             : SMFMA16x16WriteVgprVALUMemExpReadWaitStates;
1846         break;
1847       case 16: LLVM_FALLTHROUGH;
1848       default:
1849         NeedWaitStates =
1850           isDGEMM(MFMA->getOpcode())
1851             ? IsMemOrExport ? DMFMA16x16WriteVgprMemExpReadWaitStates
1852                             : DMFMA16x16WriteVgprVALUReadWaitStates
1853             : ST.hasGFX940Insts()
1854               ? isXDL(ST, *MFMA)
1855                 ? GFX940_XDL16PassWriteVgprVALUMemExpReadWaitStates
1856                 : GFX940_SMFMA16PassWriteVgprVALUMemExpReadWaitStates
1857               : SMFMA32x32WriteVgprVALUMemExpReadWaitStates;
1858         break;
1859       }
1860 
1861       int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1862       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1863 
1864       if (WaitStatesNeeded == MaxWaitStates)
1865         break;
1866     }
1867   }
1868 
1869   unsigned Opc = MI->getOpcode();
1870   const int DMFMAToFMA64WaitStates = 2;
1871   if ((Opc == AMDGPU::V_FMA_F64_e64 ||
1872        Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64 ||
1873        Opc == AMDGPU::V_FMAC_F64_dpp) &&
1874       WaitStatesNeeded < DMFMAToFMA64WaitStates) {
1875     int WaitStatesNeededForUse = DMFMAToFMA64WaitStates -
1876       getWaitStatesSince(IsDGEMMFn, DMFMAToFMA64WaitStates);
1877     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1878   }
1879 
1880   if (!IsVALU && !IsMemOrExport)
1881     return WaitStatesNeeded;
1882 
1883   for (const MachineOperand &Def : MI->defs()) {
1884     const int SMFMA4x4WriteVgprVALUWawWaitStates = 5;
1885     const int SMFMA16x16WriteVgprVALUWawWaitStates = 11;
1886     const int SMFMA32x32WriteVgprVALUWawWaitStates = 19;
1887     const int GFX940_SMFMA2PassWriteVgprVALUWawWaitStates = 4;
1888     const int GFX940_SMFMA4PassWriteVgprVALUWawWaitStates = 6;
1889     const int GFX940_SMFMA8PassWriteVgprVALUWawWaitStates = 10;
1890     const int GFX940_SMFMA16PassWriteVgprVALUWawWaitStates = 18;
1891     const int GFX940_XDL2PassWriteVgprVALUWawWaitStates = 5;
1892     const int GFX940_XDL4PassWriteVgprVALUWawWaitStates = 7;
1893     const int GFX940_XDL8PassWriteVgprVALUWawWaitStates = 11;
1894     const int GFX940_XDL16PassWriteVgprVALUWawWaitStates = 19;
1895     const int SMFMA4x4ReadVgprVALUWarWaitStates = 1;
1896     const int GFX940_XDL4PassReadVgprVALUWarWaitStates = 3;
1897     const int SMFMA16x16ReadVgprVALUWarWaitStates = 7;
1898     const int SMFMA32x32ReadVgprVALUWarWaitStates = 15;
1899     const int DMFMA4x4WriteVgprVALUWriteWaitStates = 6;
1900     const int DMFMA16x16WriteVgprVALUWriteWaitStates = 11;
1901     const int DotWriteDifferentVALUWrite = 3;
1902     const int MaxWaitStates = 19;
1903     const int MaxWarWaitStates = 15;
1904 
1905     Reg = Def.getReg();
1906 
1907     DOT = nullptr;
1908     int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn,
1909                                                    MaxWaitStates);
1910     if (DOT && DOT->getOpcode() != MI->getOpcode())
1911       WaitStatesNeeded = std::max(WaitStatesNeeded, DotWriteDifferentVALUWrite -
1912                                                     WaitStatesSinceDef);
1913 
1914     MFMA = nullptr;
1915     WaitStatesSinceDef =
1916         getWaitStatesSinceDef(Reg, IsMFMAWriteFn, MaxWaitStates);
1917     if (MFMA) {
1918       int NeedWaitStates = MaxWaitStates;
1919       switch (TSchedModel.computeInstrLatency(MFMA)) {
1920       case 2:
1921         NeedWaitStates = ST.hasGFX940Insts()
1922           ? isXDL(ST, *MFMA)
1923             ? GFX940_XDL2PassWriteVgprVALUWawWaitStates
1924             : GFX940_SMFMA2PassWriteVgprVALUWawWaitStates
1925           : SMFMA4x4WriteVgprVALUWawWaitStates;
1926         break;
1927       case 4:
1928         assert(isDGEMM(MFMA->getOpcode()) || ST.hasGFX940Insts());
1929         NeedWaitStates = isDGEMM(MFMA->getOpcode())
1930             ? DMFMA4x4WriteVgprVALUWriteWaitStates
1931             : isXDL(ST, *MFMA)
1932               ? GFX940_XDL4PassWriteVgprVALUWawWaitStates
1933               : GFX940_SMFMA4PassWriteVgprVALUWawWaitStates;
1934         break;
1935       case 8:
1936         NeedWaitStates = ST.hasGFX940Insts()
1937           ? isXDL(ST, *MFMA)
1938             ? GFX940_XDL8PassWriteVgprVALUWawWaitStates
1939             : GFX940_SMFMA8PassWriteVgprVALUWawWaitStates
1940           : SMFMA16x16WriteVgprVALUWawWaitStates;
1941         break;
1942       case 16: LLVM_FALLTHROUGH;
1943       default:
1944         NeedWaitStates = isDGEMM(MFMA->getOpcode())
1945                    ? DMFMA16x16WriteVgprVALUWriteWaitStates
1946                    : ST.hasGFX940Insts()
1947                      ? isXDL(ST, *MFMA)
1948                        ? GFX940_XDL16PassWriteVgprVALUWawWaitStates
1949                        : GFX940_SMFMA16PassWriteVgprVALUWawWaitStates
1950                    : SMFMA32x32WriteVgprVALUWawWaitStates;
1951         break;
1952       }
1953 
1954       int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1955       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1956 
1957       if (WaitStatesNeeded == MaxWaitStates)
1958         break;
1959     }
1960 
1961     auto IsSMFMAReadAsCFn = [&Reg, &IsMFMAFn, &MFMA,
1962                              this](const MachineInstr &MI) {
1963       if (!IsMFMAFn(MI) || isDGEMM(MI.getOpcode()) ||
1964           !MI.readsRegister(Reg, &TRI))
1965         return false;
1966 
1967       if (ST.hasGFX940Insts() && !isXDL(ST, MI))
1968         return false;
1969 
1970       const MachineOperand *SrcC =
1971           TII.getNamedOperand(MI, AMDGPU::OpName::src2);
1972       assert(SrcC);
1973       if (!SrcC->isReg() || !TRI.regsOverlap(SrcC->getReg(), Reg))
1974         return false;
1975 
1976       MFMA = &MI;
1977       return true;
1978     };
1979 
1980     MFMA = nullptr;
1981     int WaitStatesSinceUse = getWaitStatesSince(IsSMFMAReadAsCFn,
1982                                                 MaxWarWaitStates);
1983     if (!MFMA)
1984       continue;
1985 
1986     unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA);
1987     int NeedWaitStates = MaxWaitStates;
1988     switch (HazardDefLatency) {
1989     case 2:  NeedWaitStates = SMFMA4x4ReadVgprVALUWarWaitStates;
1990              break;
1991     case 4:  assert(ST.hasGFX940Insts());
1992              NeedWaitStates = GFX940_XDL4PassReadVgprVALUWarWaitStates;
1993              break;
1994     case 8:  NeedWaitStates = SMFMA16x16ReadVgprVALUWarWaitStates;
1995              break;
1996     case 16: LLVM_FALLTHROUGH;
1997     default: NeedWaitStates = SMFMA32x32ReadVgprVALUWarWaitStates;
1998              break;
1999     }
2000 
2001     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceUse;
2002     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
2003   }
2004 
2005   return WaitStatesNeeded;
2006 }
2007 
2008 bool GCNHazardRecognizer::ShouldPreferAnother(SUnit *SU) {
2009   if (!SU->isInstr())
2010     return false;
2011 
2012   const MachineInstr *MAI = nullptr;
2013   auto IsMFMAFn = [&MAI](const MachineInstr &MI) {
2014     MAI = nullptr;
2015     if (SIInstrInfo::isMAI(MI) &&
2016         MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
2017         MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64)
2018       MAI = &MI;
2019     return MAI != nullptr;
2020   };
2021 
2022   MachineInstr *MI = SU->getInstr();
2023   if (IsMFMAFn(*MI)) {
2024     int W = getWaitStatesSince(IsMFMAFn, 16);
2025     if (MAI)
2026       return W < (int)TSchedModel.computeInstrLatency(MAI);
2027   }
2028 
2029   return false;
2030 }
2031