xref: /llvm-project/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp (revision cad9de71d7119e2b8ece14c8243a77bb0ef59a16)
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 Opcode == AMDGPU::V_MFMA_F64_4X4X4F64_e64 ||
120          Opcode == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64 ||
121          Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_e64 ||
122          Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64 ||
123          Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64 ||
124          Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64;
125 }
126 
127 static bool isXDL(const GCNSubtarget &ST, const MachineInstr &MI) {
128   unsigned Opcode = MI.getOpcode();
129 
130   if (!SIInstrInfo::isMAI(MI) ||
131       isDGEMM(Opcode) ||
132       Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_e64 ||
133       Opcode == AMDGPU::V_ACCVGPR_READ_B32_e64)
134     return false;
135 
136   if (!ST.hasGFX940Insts())
137     return true;
138 
139   return AMDGPU::getMAIIsGFX940XDL(Opcode);
140 }
141 
142 static bool isSendMsgTraceDataOrGDS(const SIInstrInfo &TII,
143                                     const MachineInstr &MI) {
144   if (TII.isAlwaysGDS(MI.getOpcode()))
145     return true;
146 
147   switch (MI.getOpcode()) {
148   case AMDGPU::S_SENDMSG:
149   case AMDGPU::S_SENDMSGHALT:
150   case AMDGPU::S_TTRACEDATA:
151     return true;
152   // These DS opcodes don't support GDS.
153   case AMDGPU::DS_NOP:
154   case AMDGPU::DS_PERMUTE_B32:
155   case AMDGPU::DS_BPERMUTE_B32:
156     return false;
157   default:
158     if (TII.isDS(MI.getOpcode())) {
159       int GDS = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
160                                            AMDGPU::OpName::gds);
161       if (MI.getOperand(GDS).getImm())
162         return true;
163     }
164     return false;
165   }
166 }
167 
168 static bool isPermlane(const MachineInstr &MI) {
169   unsigned Opcode = MI.getOpcode();
170   return Opcode == AMDGPU::V_PERMLANE16_B32_e64 ||
171          Opcode == AMDGPU::V_PERMLANEX16_B32_e64;
172 }
173 
174 static unsigned getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) {
175   const MachineOperand *RegOp = TII->getNamedOperand(RegInstr,
176                                                      AMDGPU::OpName::simm16);
177   return RegOp->getImm() & AMDGPU::Hwreg::ID_MASK_;
178 }
179 
180 ScheduleHazardRecognizer::HazardType
181 GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
182   MachineInstr *MI = SU->getInstr();
183   // If we are not in "HazardRecognizerMode" and therefore not being run from
184   // the scheduler, track possible stalls from hazards but don't insert noops.
185   auto HazardType = IsHazardRecognizerMode ? NoopHazard : Hazard;
186 
187   if (MI->isBundle())
188    return NoHazard;
189 
190   if (SIInstrInfo::isSMRD(*MI) && checkSMRDHazards(MI) > 0)
191     return HazardType;
192 
193   if (ST.hasNSAtoVMEMBug() && checkNSAtoVMEMHazard(MI) > 0)
194     return HazardType;
195 
196   if (checkFPAtomicToDenormModeHazard(MI) > 0)
197     return HazardType;
198 
199   if (ST.hasNoDataDepHazard())
200     return NoHazard;
201 
202   // FIXME: Should flat be considered vmem?
203   if ((SIInstrInfo::isVMEM(*MI) ||
204        SIInstrInfo::isFLAT(*MI))
205       && checkVMEMHazards(MI) > 0)
206     return HazardType;
207 
208   if (SIInstrInfo::isVALU(*MI) && checkVALUHazards(MI) > 0)
209     return HazardType;
210 
211   if (SIInstrInfo::isDPP(*MI) && checkDPPHazards(MI) > 0)
212     return HazardType;
213 
214   if (isDivFMas(MI->getOpcode()) && checkDivFMasHazards(MI) > 0)
215     return HazardType;
216 
217   if (isRWLane(MI->getOpcode()) && checkRWLaneHazards(MI) > 0)
218     return HazardType;
219 
220   if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) ||
221        SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) ||
222        SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0)
223     return HazardType;
224 
225   if (isSGetReg(MI->getOpcode()) && checkGetRegHazards(MI) > 0)
226     return HazardType;
227 
228   if (isSSetReg(MI->getOpcode()) && checkSetRegHazards(MI) > 0)
229     return HazardType;
230 
231   if (isRFE(MI->getOpcode()) && checkRFEHazards(MI) > 0)
232     return HazardType;
233 
234   if (ST.hasReadM0MovRelInterpHazard() &&
235       (TII.isVINTRP(*MI) || isSMovRel(MI->getOpcode())) &&
236       checkReadM0Hazards(MI) > 0)
237     return HazardType;
238 
239   if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI) &&
240       checkReadM0Hazards(MI) > 0)
241     return HazardType;
242 
243   if (SIInstrInfo::isMAI(*MI) && checkMAIHazards(MI) > 0)
244     return HazardType;
245 
246   if ((SIInstrInfo::isVMEM(*MI) ||
247        SIInstrInfo::isFLAT(*MI) ||
248        SIInstrInfo::isDS(*MI)) && checkMAILdStHazards(MI) > 0)
249     return HazardType;
250 
251   if (MI->isInlineAsm() && checkInlineAsmHazards(MI) > 0)
252     return HazardType;
253 
254   return NoHazard;
255 }
256 
257 static void insertNoopsInBundle(MachineInstr *MI, const SIInstrInfo &TII,
258                                 unsigned Quantity) {
259   while (Quantity > 0) {
260     unsigned Arg = std::min(Quantity, 8u);
261     Quantity -= Arg;
262     BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII.get(AMDGPU::S_NOP))
263         .addImm(Arg - 1);
264   }
265 }
266 
267 unsigned
268 GCNHazardRecognizer::getMFMAPipelineWaitStates(const MachineInstr &MI) const {
269   const MCSchedClassDesc *SC = TSchedModel.resolveSchedClass(&MI);
270   assert(TSchedModel.getWriteProcResBegin(SC) !=
271          TSchedModel.getWriteProcResEnd(SC));
272   return TSchedModel.getWriteProcResBegin(SC)->Cycles;
273 }
274 
275 void GCNHazardRecognizer::processBundle() {
276   MachineBasicBlock::instr_iterator MI = std::next(CurrCycleInstr->getIterator());
277   MachineBasicBlock::instr_iterator E = CurrCycleInstr->getParent()->instr_end();
278   // Check bundled MachineInstr's for hazards.
279   for (; MI != E && MI->isInsideBundle(); ++MI) {
280     CurrCycleInstr = &*MI;
281     unsigned WaitStates = PreEmitNoopsCommon(CurrCycleInstr);
282 
283     if (IsHazardRecognizerMode) {
284       fixHazards(CurrCycleInstr);
285 
286       insertNoopsInBundle(CurrCycleInstr, TII, WaitStates);
287     }
288 
289     // It’s unnecessary to track more than MaxLookAhead instructions. Since we
290     // include the bundled MI directly after, only add a maximum of
291     // (MaxLookAhead - 1) noops to EmittedInstrs.
292     for (unsigned i = 0, e = std::min(WaitStates, MaxLookAhead - 1); i < e; ++i)
293       EmittedInstrs.push_front(nullptr);
294 
295     EmittedInstrs.push_front(CurrCycleInstr);
296     EmittedInstrs.resize(MaxLookAhead);
297   }
298   CurrCycleInstr = nullptr;
299 }
300 
301 unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) {
302   IsHazardRecognizerMode = true;
303   CurrCycleInstr = MI;
304   unsigned W = PreEmitNoopsCommon(MI);
305   fixHazards(MI);
306   CurrCycleInstr = nullptr;
307   return W;
308 }
309 
310 unsigned GCNHazardRecognizer::PreEmitNoopsCommon(MachineInstr *MI) {
311   if (MI->isBundle())
312     return 0;
313 
314   int WaitStates = 0;
315 
316   if (SIInstrInfo::isSMRD(*MI))
317     return std::max(WaitStates, checkSMRDHazards(MI));
318 
319   if (ST.hasNSAtoVMEMBug())
320     WaitStates = std::max(WaitStates, checkNSAtoVMEMHazard(MI));
321 
322   WaitStates = std::max(WaitStates, checkFPAtomicToDenormModeHazard(MI));
323 
324   if (ST.hasNoDataDepHazard())
325     return WaitStates;
326 
327   if (SIInstrInfo::isVMEM(*MI) || SIInstrInfo::isFLAT(*MI))
328     WaitStates = std::max(WaitStates, checkVMEMHazards(MI));
329 
330   if (SIInstrInfo::isVALU(*MI))
331     WaitStates = std::max(WaitStates, checkVALUHazards(MI));
332 
333   if (SIInstrInfo::isDPP(*MI))
334     WaitStates = std::max(WaitStates, checkDPPHazards(MI));
335 
336   if (isDivFMas(MI->getOpcode()))
337     WaitStates = std::max(WaitStates, checkDivFMasHazards(MI));
338 
339   if (isRWLane(MI->getOpcode()))
340     WaitStates = std::max(WaitStates, checkRWLaneHazards(MI));
341 
342   if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) ||
343        SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) ||
344        SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0)
345     WaitStates = std::max(WaitStates, checkMAIVALUHazards(MI));
346 
347   if (MI->isInlineAsm())
348     return std::max(WaitStates, checkInlineAsmHazards(MI));
349 
350   if (isSGetReg(MI->getOpcode()))
351     return std::max(WaitStates, checkGetRegHazards(MI));
352 
353   if (isSSetReg(MI->getOpcode()))
354     return std::max(WaitStates, checkSetRegHazards(MI));
355 
356   if (isRFE(MI->getOpcode()))
357     return std::max(WaitStates, checkRFEHazards(MI));
358 
359   if (ST.hasReadM0MovRelInterpHazard() && (TII.isVINTRP(*MI) ||
360                                            isSMovRel(MI->getOpcode())))
361     return std::max(WaitStates, checkReadM0Hazards(MI));
362 
363   if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI))
364     return std::max(WaitStates, checkReadM0Hazards(MI));
365 
366   if (SIInstrInfo::isMAI(*MI))
367     return std::max(WaitStates, checkMAIHazards(MI));
368 
369   if (SIInstrInfo::isVMEM(*MI) ||
370       SIInstrInfo::isFLAT(*MI) ||
371       SIInstrInfo::isDS(*MI))
372     return std::max(WaitStates, checkMAILdStHazards(MI));
373 
374   return WaitStates;
375 }
376 
377 void GCNHazardRecognizer::EmitNoop() {
378   EmittedInstrs.push_front(nullptr);
379 }
380 
381 void GCNHazardRecognizer::AdvanceCycle() {
382   // When the scheduler detects a stall, it will call AdvanceCycle() without
383   // emitting any instructions.
384   if (!CurrCycleInstr) {
385     EmittedInstrs.push_front(nullptr);
386     return;
387   }
388 
389   if (CurrCycleInstr->isBundle()) {
390     processBundle();
391     return;
392   }
393 
394   unsigned NumWaitStates = TII.getNumWaitStates(*CurrCycleInstr);
395   if (!NumWaitStates) {
396     CurrCycleInstr = nullptr;
397     return;
398   }
399 
400   // Keep track of emitted instructions
401   EmittedInstrs.push_front(CurrCycleInstr);
402 
403   // Add a nullptr for each additional wait state after the first.  Make sure
404   // not to add more than getMaxLookAhead() items to the list, since we
405   // truncate the list to that size right after this loop.
406   for (unsigned i = 1, e = std::min(NumWaitStates, getMaxLookAhead());
407        i < e; ++i) {
408     EmittedInstrs.push_front(nullptr);
409   }
410 
411   // getMaxLookahead() is the largest number of wait states we will ever need
412   // to insert, so there is no point in keeping track of more than that many
413   // wait states.
414   EmittedInstrs.resize(getMaxLookAhead());
415 
416   CurrCycleInstr = nullptr;
417 }
418 
419 void GCNHazardRecognizer::RecedeCycle() {
420   llvm_unreachable("hazard recognizer does not support bottom-up scheduling.");
421 }
422 
423 //===----------------------------------------------------------------------===//
424 // Helper Functions
425 //===----------------------------------------------------------------------===//
426 
427 typedef function_ref<bool(const MachineInstr &, int WaitStates)> IsExpiredFn;
428 
429 // Returns a minimum wait states since \p I walking all predecessors.
430 // Only scans until \p IsExpired does not return true.
431 // Can only be run in a hazard recognizer mode.
432 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard,
433                               const MachineBasicBlock *MBB,
434                               MachineBasicBlock::const_reverse_instr_iterator I,
435                               int WaitStates, IsExpiredFn IsExpired,
436                               DenseSet<const MachineBasicBlock *> &Visited) {
437   for (auto E = MBB->instr_rend(); I != E; ++I) {
438     // Don't add WaitStates for parent BUNDLE instructions.
439     if (I->isBundle())
440       continue;
441 
442     if (IsHazard(*I))
443       return WaitStates;
444 
445     if (I->isInlineAsm())
446       continue;
447 
448     WaitStates += SIInstrInfo::getNumWaitStates(*I);
449 
450     if (IsExpired(*I, WaitStates))
451       return std::numeric_limits<int>::max();
452   }
453 
454   int MinWaitStates = std::numeric_limits<int>::max();
455   for (MachineBasicBlock *Pred : MBB->predecessors()) {
456     if (!Visited.insert(Pred).second)
457       continue;
458 
459     int W = getWaitStatesSince(IsHazard, Pred, Pred->instr_rbegin(),
460                                WaitStates, IsExpired, Visited);
461 
462     MinWaitStates = std::min(MinWaitStates, W);
463   }
464 
465   return MinWaitStates;
466 }
467 
468 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard,
469                               const MachineInstr *MI, IsExpiredFn IsExpired) {
470   DenseSet<const MachineBasicBlock *> Visited;
471   return getWaitStatesSince(IsHazard, MI->getParent(),
472                             std::next(MI->getReverseIterator()),
473                             0, IsExpired, Visited);
474 }
475 
476 int GCNHazardRecognizer::getWaitStatesSince(IsHazardFn IsHazard, int Limit) {
477   if (IsHazardRecognizerMode) {
478     auto IsExpiredFn = [Limit](const MachineInstr &, int WaitStates) {
479       return WaitStates >= Limit;
480     };
481     return ::getWaitStatesSince(IsHazard, CurrCycleInstr, IsExpiredFn);
482   }
483 
484   int WaitStates = 0;
485   for (MachineInstr *MI : EmittedInstrs) {
486     if (MI) {
487       if (IsHazard(*MI))
488         return WaitStates;
489 
490       if (MI->isInlineAsm())
491         continue;
492     }
493     ++WaitStates;
494 
495     if (WaitStates >= Limit)
496       break;
497   }
498   return std::numeric_limits<int>::max();
499 }
500 
501 int GCNHazardRecognizer::getWaitStatesSinceDef(unsigned Reg,
502                                                IsHazardFn IsHazardDef,
503                                                int Limit) {
504   const SIRegisterInfo *TRI = ST.getRegisterInfo();
505 
506   auto IsHazardFn = [IsHazardDef, TRI, Reg](const MachineInstr &MI) {
507     return IsHazardDef(MI) && MI.modifiesRegister(Reg, TRI);
508   };
509 
510   return getWaitStatesSince(IsHazardFn, Limit);
511 }
512 
513 int GCNHazardRecognizer::getWaitStatesSinceSetReg(IsHazardFn IsHazard,
514                                                   int Limit) {
515   auto IsHazardFn = [IsHazard](const MachineInstr &MI) {
516     return isSSetReg(MI.getOpcode()) && IsHazard(MI);
517   };
518 
519   return getWaitStatesSince(IsHazardFn, Limit);
520 }
521 
522 //===----------------------------------------------------------------------===//
523 // No-op Hazard Detection
524 //===----------------------------------------------------------------------===//
525 
526 static void addRegUnits(const SIRegisterInfo &TRI, BitVector &BV,
527                         MCRegister Reg) {
528   for (MCRegUnitIterator RUI(Reg, &TRI); RUI.isValid(); ++RUI)
529     BV.set(*RUI);
530 }
531 
532 static void addRegsToSet(const SIRegisterInfo &TRI,
533                          iterator_range<MachineInstr::const_mop_iterator> Ops,
534                          BitVector &Set) {
535   for (const MachineOperand &Op : Ops) {
536     if (Op.isReg())
537       addRegUnits(TRI, Set, Op.getReg().asMCReg());
538   }
539 }
540 
541 void GCNHazardRecognizer::addClauseInst(const MachineInstr &MI) {
542   // XXX: Do we need to worry about implicit operands
543   addRegsToSet(TRI, MI.defs(), ClauseDefs);
544   addRegsToSet(TRI, MI.uses(), ClauseUses);
545 }
546 
547 static bool breaksSMEMSoftClause(MachineInstr *MI) {
548   return !SIInstrInfo::isSMRD(*MI);
549 }
550 
551 static bool breaksVMEMSoftClause(MachineInstr *MI) {
552   return !SIInstrInfo::isVMEM(*MI) && !SIInstrInfo::isFLAT(*MI);
553 }
554 
555 int GCNHazardRecognizer::checkSoftClauseHazards(MachineInstr *MEM) {
556   // SMEM soft clause are only present on VI+, and only matter if xnack is
557   // enabled.
558   if (!ST.isXNACKEnabled())
559     return 0;
560 
561   bool IsSMRD = TII.isSMRD(*MEM);
562 
563   resetClause();
564 
565   // A soft-clause is any group of consecutive SMEM instructions.  The
566   // instructions in this group may return out of order and/or may be
567   // replayed (i.e. the same instruction issued more than once).
568   //
569   // In order to handle these situations correctly we need to make sure that
570   // when a clause has more than one instruction, no instruction in the clause
571   // writes to a register that is read by another instruction in the clause
572   // (including itself). If we encounter this situation, we need to break the
573   // clause by inserting a non SMEM instruction.
574 
575   for (MachineInstr *MI : EmittedInstrs) {
576     // When we hit a non-SMEM instruction then we have passed the start of the
577     // clause and we can stop.
578     if (!MI)
579       break;
580 
581     if (IsSMRD ? breaksSMEMSoftClause(MI) : breaksVMEMSoftClause(MI))
582       break;
583 
584     addClauseInst(*MI);
585   }
586 
587   if (ClauseDefs.none())
588     return 0;
589 
590   // We need to make sure not to put loads and stores in the same clause if they
591   // use the same address. For now, just start a new clause whenever we see a
592   // store.
593   if (MEM->mayStore())
594     return 1;
595 
596   addClauseInst(*MEM);
597 
598   // If the set of defs and uses intersect then we cannot add this instruction
599   // to the clause, so we have a hazard.
600   return ClauseDefs.anyCommon(ClauseUses) ? 1 : 0;
601 }
602 
603 int GCNHazardRecognizer::checkSMRDHazards(MachineInstr *SMRD) {
604   int WaitStatesNeeded = 0;
605 
606   WaitStatesNeeded = checkSoftClauseHazards(SMRD);
607 
608   // This SMRD hazard only affects SI.
609   if (!ST.hasSMRDReadVALUDefHazard())
610     return WaitStatesNeeded;
611 
612   // A read of an SGPR by SMRD instruction requires 4 wait states when the
613   // SGPR was written by a VALU instruction.
614   int SmrdSgprWaitStates = 4;
615   auto IsHazardDefFn = [this](const MachineInstr &MI) {
616     return TII.isVALU(MI);
617   };
618   auto IsBufferHazardDefFn = [this](const MachineInstr &MI) {
619     return TII.isSALU(MI);
620   };
621 
622   bool IsBufferSMRD = TII.isBufferSMRD(*SMRD);
623 
624   for (const MachineOperand &Use : SMRD->uses()) {
625     if (!Use.isReg())
626       continue;
627     int WaitStatesNeededForUse =
628         SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn,
629                                                    SmrdSgprWaitStates);
630     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
631 
632     // This fixes what appears to be undocumented hardware behavior in SI where
633     // s_mov writing a descriptor and s_buffer_load_dword reading the descriptor
634     // needs some number of nops in between. We don't know how many we need, but
635     // let's use 4. This wasn't discovered before probably because the only
636     // case when this happens is when we expand a 64-bit pointer into a full
637     // descriptor and use s_buffer_load_dword instead of s_load_dword, which was
638     // probably never encountered in the closed-source land.
639     if (IsBufferSMRD) {
640       int WaitStatesNeededForUse =
641         SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(),
642                                                    IsBufferHazardDefFn,
643                                                    SmrdSgprWaitStates);
644       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
645     }
646   }
647 
648   return WaitStatesNeeded;
649 }
650 
651 int GCNHazardRecognizer::checkVMEMHazards(MachineInstr* VMEM) {
652   if (!ST.hasVMEMReadSGPRVALUDefHazard())
653     return 0;
654 
655   int WaitStatesNeeded = checkSoftClauseHazards(VMEM);
656 
657   // A read of an SGPR by a VMEM instruction requires 5 wait states when the
658   // SGPR was written by a VALU Instruction.
659   const int VmemSgprWaitStates = 5;
660   auto IsHazardDefFn = [this](const MachineInstr &MI) {
661     return TII.isVALU(MI);
662   };
663   for (const MachineOperand &Use : VMEM->uses()) {
664     if (!Use.isReg() || TRI.isVectorRegister(MF.getRegInfo(), Use.getReg()))
665       continue;
666 
667     int WaitStatesNeededForUse =
668         VmemSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn,
669                                                    VmemSgprWaitStates);
670     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
671   }
672   return WaitStatesNeeded;
673 }
674 
675 int GCNHazardRecognizer::checkDPPHazards(MachineInstr *DPP) {
676   const SIRegisterInfo *TRI = ST.getRegisterInfo();
677   const SIInstrInfo *TII = ST.getInstrInfo();
678 
679   // Check for DPP VGPR read after VALU VGPR write and EXEC write.
680   int DppVgprWaitStates = 2;
681   int DppExecWaitStates = 5;
682   int WaitStatesNeeded = 0;
683   auto IsHazardDefFn = [TII](const MachineInstr &MI) {
684     return TII->isVALU(MI);
685   };
686 
687   for (const MachineOperand &Use : DPP->uses()) {
688     if (!Use.isReg() || !TRI->isVGPR(MF.getRegInfo(), Use.getReg()))
689       continue;
690     int WaitStatesNeededForUse =
691         DppVgprWaitStates - getWaitStatesSinceDef(
692                                 Use.getReg(),
693                                 [](const MachineInstr &) { return true; },
694                                 DppVgprWaitStates);
695     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
696   }
697 
698   WaitStatesNeeded = std::max(
699       WaitStatesNeeded,
700       DppExecWaitStates - getWaitStatesSinceDef(AMDGPU::EXEC, IsHazardDefFn,
701                                                 DppExecWaitStates));
702 
703   return WaitStatesNeeded;
704 }
705 
706 int GCNHazardRecognizer::checkDivFMasHazards(MachineInstr *DivFMas) {
707   const SIInstrInfo *TII = ST.getInstrInfo();
708 
709   // v_div_fmas requires 4 wait states after a write to vcc from a VALU
710   // instruction.
711   const int DivFMasWaitStates = 4;
712   auto IsHazardDefFn = [TII](const MachineInstr &MI) {
713     return TII->isVALU(MI);
714   };
715   int WaitStatesNeeded = getWaitStatesSinceDef(AMDGPU::VCC, IsHazardDefFn,
716                                                DivFMasWaitStates);
717 
718   return DivFMasWaitStates - WaitStatesNeeded;
719 }
720 
721 int GCNHazardRecognizer::checkGetRegHazards(MachineInstr *GetRegInstr) {
722   const SIInstrInfo *TII = ST.getInstrInfo();
723   unsigned GetRegHWReg = getHWReg(TII, *GetRegInstr);
724 
725   const int GetRegWaitStates = 2;
726   auto IsHazardFn = [TII, GetRegHWReg](const MachineInstr &MI) {
727     return GetRegHWReg == getHWReg(TII, MI);
728   };
729   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, GetRegWaitStates);
730 
731   return GetRegWaitStates - WaitStatesNeeded;
732 }
733 
734 int GCNHazardRecognizer::checkSetRegHazards(MachineInstr *SetRegInstr) {
735   const SIInstrInfo *TII = ST.getInstrInfo();
736   unsigned HWReg = getHWReg(TII, *SetRegInstr);
737 
738   const int SetRegWaitStates = ST.getSetRegWaitStates();
739   auto IsHazardFn = [TII, HWReg](const MachineInstr &MI) {
740     return HWReg == getHWReg(TII, MI);
741   };
742   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, SetRegWaitStates);
743   return SetRegWaitStates - WaitStatesNeeded;
744 }
745 
746 int GCNHazardRecognizer::createsVALUHazard(const MachineInstr &MI) {
747   if (!MI.mayStore())
748     return -1;
749 
750   const SIInstrInfo *TII = ST.getInstrInfo();
751   unsigned Opcode = MI.getOpcode();
752   const MCInstrDesc &Desc = MI.getDesc();
753 
754   int VDataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata);
755   int VDataRCID = -1;
756   if (VDataIdx != -1)
757     VDataRCID = Desc.OpInfo[VDataIdx].RegClass;
758 
759   if (TII->isMUBUF(MI) || TII->isMTBUF(MI)) {
760     // There is no hazard if the instruction does not use vector regs
761     // (like wbinvl1)
762     if (VDataIdx == -1)
763       return -1;
764     // For MUBUF/MTBUF instructions this hazard only exists if the
765     // instruction is not using a register in the soffset field.
766     const MachineOperand *SOffset =
767         TII->getNamedOperand(MI, AMDGPU::OpName::soffset);
768     // If we have no soffset operand, then assume this field has been
769     // hardcoded to zero.
770     if (AMDGPU::getRegBitWidth(VDataRCID) > 64 &&
771         (!SOffset || !SOffset->isReg()))
772       return VDataIdx;
773   }
774 
775   // MIMG instructions create a hazard if they don't use a 256-bit T# and
776   // the store size is greater than 8 bytes and they have more than two bits
777   // of their dmask set.
778   // All our MIMG definitions use a 256-bit T#, so we can skip checking for them.
779   if (TII->isMIMG(MI)) {
780     int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::srsrc);
781     assert(SRsrcIdx != -1 &&
782            AMDGPU::getRegBitWidth(Desc.OpInfo[SRsrcIdx].RegClass) == 256);
783     (void)SRsrcIdx;
784   }
785 
786   if (TII->isFLAT(MI)) {
787     int DataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata);
788     if (AMDGPU::getRegBitWidth(Desc.OpInfo[DataIdx].RegClass) > 64)
789       return DataIdx;
790   }
791 
792   return -1;
793 }
794 
795 int
796 GCNHazardRecognizer::checkVALUHazardsHelper(const MachineOperand &Def,
797                                             const MachineRegisterInfo &MRI) {
798   // Helper to check for the hazard where VMEM instructions that store more than
799   // 8 bytes can have there store data over written by the next instruction.
800   const SIRegisterInfo *TRI = ST.getRegisterInfo();
801 
802   const int VALUWaitStates = 1;
803   int WaitStatesNeeded = 0;
804 
805   if (!TRI->isVectorRegister(MRI, Def.getReg()))
806     return WaitStatesNeeded;
807   Register Reg = Def.getReg();
808   auto IsHazardFn = [this, Reg, TRI](const MachineInstr &MI) {
809     int DataIdx = createsVALUHazard(MI);
810     return DataIdx >= 0 &&
811            TRI->regsOverlap(MI.getOperand(DataIdx).getReg(), Reg);
812   };
813   int WaitStatesNeededForDef =
814     VALUWaitStates - getWaitStatesSince(IsHazardFn, VALUWaitStates);
815   WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForDef);
816 
817   return WaitStatesNeeded;
818 }
819 
820 int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) {
821   // This checks for the hazard where VMEM instructions that store more than
822   // 8 bytes can have there store data over written by the next instruction.
823   if (!ST.has12DWordStoreHazard())
824     return 0;
825 
826   const MachineRegisterInfo &MRI = MF.getRegInfo();
827   int WaitStatesNeeded = 0;
828 
829   for (const MachineOperand &Def : VALU->defs()) {
830     WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Def, MRI));
831   }
832 
833   return WaitStatesNeeded;
834 }
835 
836 int GCNHazardRecognizer::checkInlineAsmHazards(MachineInstr *IA) {
837   // This checks for hazards associated with inline asm statements.
838   // Since inline asms can contain just about anything, we use this
839   // to call/leverage other check*Hazard routines. Note that
840   // this function doesn't attempt to address all possible inline asm
841   // hazards (good luck), but is a collection of what has been
842   // problematic thus far.
843 
844   // see checkVALUHazards()
845   if (!ST.has12DWordStoreHazard())
846     return 0;
847 
848   const MachineRegisterInfo &MRI = MF.getRegInfo();
849   int WaitStatesNeeded = 0;
850 
851   for (unsigned I = InlineAsm::MIOp_FirstOperand, E = IA->getNumOperands();
852        I != E; ++I) {
853     const MachineOperand &Op = IA->getOperand(I);
854     if (Op.isReg() && Op.isDef()) {
855       WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Op, MRI));
856     }
857   }
858 
859   return WaitStatesNeeded;
860 }
861 
862 int GCNHazardRecognizer::checkRWLaneHazards(MachineInstr *RWLane) {
863   const SIInstrInfo *TII = ST.getInstrInfo();
864   const SIRegisterInfo *TRI = ST.getRegisterInfo();
865   const MachineRegisterInfo &MRI = MF.getRegInfo();
866 
867   const MachineOperand *LaneSelectOp =
868       TII->getNamedOperand(*RWLane, AMDGPU::OpName::src1);
869 
870   if (!LaneSelectOp->isReg() || !TRI->isSGPRReg(MRI, LaneSelectOp->getReg()))
871     return 0;
872 
873   Register LaneSelectReg = LaneSelectOp->getReg();
874   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVALU(MI); };
875 
876   const int RWLaneWaitStates = 4;
877   int WaitStatesSince = getWaitStatesSinceDef(LaneSelectReg, IsHazardFn,
878                                               RWLaneWaitStates);
879   return RWLaneWaitStates - WaitStatesSince;
880 }
881 
882 int GCNHazardRecognizer::checkRFEHazards(MachineInstr *RFE) {
883   if (!ST.hasRFEHazards())
884     return 0;
885 
886   const SIInstrInfo *TII = ST.getInstrInfo();
887 
888   const int RFEWaitStates = 1;
889 
890   auto IsHazardFn = [TII](const MachineInstr &MI) {
891     return getHWReg(TII, MI) == AMDGPU::Hwreg::ID_TRAPSTS;
892   };
893   int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, RFEWaitStates);
894   return RFEWaitStates - WaitStatesNeeded;
895 }
896 
897 int GCNHazardRecognizer::checkReadM0Hazards(MachineInstr *MI) {
898   const SIInstrInfo *TII = ST.getInstrInfo();
899   const int SMovRelWaitStates = 1;
900   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isSALU(MI); };
901   return SMovRelWaitStates - getWaitStatesSinceDef(AMDGPU::M0, IsHazardFn,
902                                                    SMovRelWaitStates);
903 }
904 
905 void GCNHazardRecognizer::fixHazards(MachineInstr *MI) {
906   fixVMEMtoScalarWriteHazards(MI);
907   fixVcmpxPermlaneHazards(MI);
908   fixSMEMtoVectorWriteHazards(MI);
909   fixVcmpxExecWARHazard(MI);
910   fixLdsBranchVmemWARHazard(MI);
911 }
912 
913 bool GCNHazardRecognizer::fixVcmpxPermlaneHazards(MachineInstr *MI) {
914   if (!ST.hasVcmpxPermlaneHazard() || !isPermlane(*MI))
915     return false;
916 
917   const SIInstrInfo *TII = ST.getInstrInfo();
918   auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVOPC(MI); };
919 
920   auto IsExpiredFn = [](const MachineInstr &MI, int) {
921     unsigned Opc = MI.getOpcode();
922     return SIInstrInfo::isVALU(MI) && Opc != AMDGPU::V_NOP_e32 &&
923            Opc != AMDGPU::V_NOP_e64 && Opc != AMDGPU::V_NOP_sdwa;
924   };
925 
926   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
927       std::numeric_limits<int>::max())
928     return false;
929 
930   // V_NOP will be discarded by SQ.
931   // Use V_MOV_B32 v?, v?. Register must be alive so use src0 of V_PERMLANE*
932   // which is always a VGPR and available.
933   auto *Src0 = TII->getNamedOperand(*MI, AMDGPU::OpName::src0);
934   Register Reg = Src0->getReg();
935   bool IsUndef = Src0->isUndef();
936   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
937           TII->get(AMDGPU::V_MOV_B32_e32))
938     .addReg(Reg, RegState::Define | (IsUndef ? RegState::Dead : 0))
939     .addReg(Reg, IsUndef ? RegState::Undef : RegState::Kill);
940 
941   return true;
942 }
943 
944 bool GCNHazardRecognizer::fixVMEMtoScalarWriteHazards(MachineInstr *MI) {
945   if (!ST.hasVMEMtoScalarWriteHazard())
946     return false;
947 
948   if (!SIInstrInfo::isSALU(*MI) && !SIInstrInfo::isSMRD(*MI))
949     return false;
950 
951   if (MI->getNumDefs() == 0)
952     return false;
953 
954   const SIRegisterInfo *TRI = ST.getRegisterInfo();
955 
956   auto IsHazardFn = [TRI, MI](const MachineInstr &I) {
957     if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isDS(I) &&
958         !SIInstrInfo::isFLAT(I))
959       return false;
960 
961     for (const MachineOperand &Def : MI->defs()) {
962       const MachineOperand *Op =
963           I.findRegisterUseOperand(Def.getReg(), false, TRI);
964       if (!Op)
965         continue;
966       return true;
967     }
968     return false;
969   };
970 
971   auto IsExpiredFn = [](const MachineInstr &MI, int) {
972     return SIInstrInfo::isVALU(MI) ||
973            (MI.getOpcode() == AMDGPU::S_WAITCNT &&
974             !MI.getOperand(0).getImm()) ||
975            (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
976             MI.getOperand(0).getImm() == 0xffe3);
977   };
978 
979   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
980       std::numeric_limits<int>::max())
981     return false;
982 
983   const SIInstrInfo *TII = ST.getInstrInfo();
984   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
985           TII->get(AMDGPU::S_WAITCNT_DEPCTR))
986       .addImm(0xffe3);
987   return true;
988 }
989 
990 bool GCNHazardRecognizer::fixSMEMtoVectorWriteHazards(MachineInstr *MI) {
991   if (!ST.hasSMEMtoVectorWriteHazard())
992     return false;
993 
994   if (!SIInstrInfo::isVALU(*MI))
995     return false;
996 
997   unsigned SDSTName;
998   switch (MI->getOpcode()) {
999   case AMDGPU::V_READLANE_B32:
1000   case AMDGPU::V_READFIRSTLANE_B32:
1001     SDSTName = AMDGPU::OpName::vdst;
1002     break;
1003   default:
1004     SDSTName = AMDGPU::OpName::sdst;
1005     break;
1006   }
1007 
1008   const SIInstrInfo *TII = ST.getInstrInfo();
1009   const SIRegisterInfo *TRI = ST.getRegisterInfo();
1010   const AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST.getCPU());
1011   const MachineOperand *SDST = TII->getNamedOperand(*MI, SDSTName);
1012   if (!SDST) {
1013     for (const auto &MO : MI->implicit_operands()) {
1014       if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg()))) {
1015         SDST = &MO;
1016         break;
1017       }
1018     }
1019   }
1020 
1021   if (!SDST)
1022     return false;
1023 
1024   const Register SDSTReg = SDST->getReg();
1025   auto IsHazardFn = [SDSTReg, TRI](const MachineInstr &I) {
1026     return SIInstrInfo::isSMRD(I) && I.readsRegister(SDSTReg, TRI);
1027   };
1028 
1029   auto IsExpiredFn = [TII, IV](const MachineInstr &MI, int) {
1030     if (TII->isSALU(MI)) {
1031       switch (MI.getOpcode()) {
1032       case AMDGPU::S_SETVSKIP:
1033       case AMDGPU::S_VERSION:
1034       case AMDGPU::S_WAITCNT_VSCNT:
1035       case AMDGPU::S_WAITCNT_VMCNT:
1036       case AMDGPU::S_WAITCNT_EXPCNT:
1037         // These instructions cannot not mitigate the hazard.
1038         return false;
1039       case AMDGPU::S_WAITCNT_LGKMCNT:
1040         // Reducing lgkmcnt count to 0 always mitigates the hazard.
1041         return (MI.getOperand(1).getImm() == 0) &&
1042                (MI.getOperand(0).getReg() == AMDGPU::SGPR_NULL);
1043       case AMDGPU::S_WAITCNT: {
1044         const int64_t Imm = MI.getOperand(0).getImm();
1045         AMDGPU::Waitcnt Decoded = AMDGPU::decodeWaitcnt(IV, Imm);
1046         return (Decoded.LgkmCnt == 0);
1047       }
1048       default:
1049         // SOPP instructions cannot mitigate the hazard.
1050         if (TII->isSOPP(MI))
1051           return false;
1052         // At this point the SALU can be assumed to mitigate the hazard
1053         // because either:
1054         // (a) it is independent of the at risk SMEM (breaking chain),
1055         // or
1056         // (b) it is dependent on the SMEM, in which case an appropriate
1057         //     s_waitcnt lgkmcnt _must_ exist between it and the at risk
1058         //     SMEM instruction.
1059         return true;
1060       }
1061     }
1062     return false;
1063   };
1064 
1065   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1066       std::numeric_limits<int>::max())
1067     return false;
1068 
1069   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1070           TII->get(AMDGPU::S_MOV_B32), AMDGPU::SGPR_NULL)
1071       .addImm(0);
1072   return true;
1073 }
1074 
1075 bool GCNHazardRecognizer::fixVcmpxExecWARHazard(MachineInstr *MI) {
1076   if (!ST.hasVcmpxExecWARHazard() || !SIInstrInfo::isVALU(*MI))
1077     return false;
1078 
1079   const SIRegisterInfo *TRI = ST.getRegisterInfo();
1080   if (!MI->modifiesRegister(AMDGPU::EXEC, TRI))
1081     return false;
1082 
1083   auto IsHazardFn = [TRI](const MachineInstr &I) {
1084     if (SIInstrInfo::isVALU(I))
1085       return false;
1086     return I.readsRegister(AMDGPU::EXEC, TRI);
1087   };
1088 
1089   const SIInstrInfo *TII = ST.getInstrInfo();
1090   auto IsExpiredFn = [TII, TRI](const MachineInstr &MI, int) {
1091     if (SIInstrInfo::isVALU(MI)) {
1092       if (TII->getNamedOperand(MI, AMDGPU::OpName::sdst))
1093         return true;
1094       for (auto MO : MI.implicit_operands())
1095         if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg())))
1096           return true;
1097     }
1098     if (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR &&
1099         (MI.getOperand(0).getImm() & 0xfffe) == 0xfffe)
1100       return true;
1101     return false;
1102   };
1103 
1104   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1105       std::numeric_limits<int>::max())
1106     return false;
1107 
1108   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1109           TII->get(AMDGPU::S_WAITCNT_DEPCTR))
1110     .addImm(0xfffe);
1111   return true;
1112 }
1113 
1114 static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF,
1115                                                  const GCNSubtarget &ST) {
1116   if (!ST.hasLdsBranchVmemWARHazard())
1117     return false;
1118 
1119   // Check if the necessary condition for the hazard is met: both LDS and VMEM
1120   // instructions need to appear in the same function.
1121   bool HasLds = false;
1122   bool HasVmem = false;
1123   for (auto &MBB : MF) {
1124     for (auto &MI : MBB) {
1125       HasLds |= SIInstrInfo::isDS(MI);
1126       HasVmem |=
1127           SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI);
1128       if (HasLds && HasVmem)
1129         return true;
1130     }
1131   }
1132   return false;
1133 }
1134 
1135 bool GCNHazardRecognizer::fixLdsBranchVmemWARHazard(MachineInstr *MI) {
1136   if (!RunLdsBranchVmemWARHazardFixup)
1137     return false;
1138 
1139   assert(ST.hasLdsBranchVmemWARHazard());
1140 
1141   auto IsHazardInst = [](const MachineInstr &MI) {
1142     if (SIInstrInfo::isDS(MI))
1143       return 1;
1144     if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI))
1145       return 2;
1146     return 0;
1147   };
1148 
1149   auto InstType = IsHazardInst(*MI);
1150   if (!InstType)
1151     return false;
1152 
1153   auto IsExpiredFn = [&IsHazardInst](const MachineInstr &I, int) {
1154     return IsHazardInst(I) || (I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT &&
1155                                I.getOperand(0).getReg() == AMDGPU::SGPR_NULL &&
1156                                !I.getOperand(1).getImm());
1157   };
1158 
1159   auto IsHazardFn = [InstType, &IsHazardInst](const MachineInstr &I) {
1160     if (!I.isBranch())
1161       return false;
1162 
1163     auto IsHazardFn = [InstType, IsHazardInst](const MachineInstr &I) {
1164       auto InstType2 = IsHazardInst(I);
1165       return InstType2 && InstType != InstType2;
1166     };
1167 
1168     auto IsExpiredFn = [InstType, &IsHazardInst](const MachineInstr &I, int) {
1169       auto InstType2 = IsHazardInst(I);
1170       if (InstType == InstType2)
1171         return true;
1172 
1173       return I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT &&
1174              I.getOperand(0).getReg() == AMDGPU::SGPR_NULL &&
1175              !I.getOperand(1).getImm();
1176     };
1177 
1178     return ::getWaitStatesSince(IsHazardFn, &I, IsExpiredFn) !=
1179            std::numeric_limits<int>::max();
1180   };
1181 
1182   if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) ==
1183       std::numeric_limits<int>::max())
1184     return false;
1185 
1186   const SIInstrInfo *TII = ST.getInstrInfo();
1187   BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1188           TII->get(AMDGPU::S_WAITCNT_VSCNT))
1189     .addReg(AMDGPU::SGPR_NULL, RegState::Undef)
1190     .addImm(0);
1191 
1192   return true;
1193 }
1194 
1195 int GCNHazardRecognizer::checkNSAtoVMEMHazard(MachineInstr *MI) {
1196   int NSAtoVMEMWaitStates = 1;
1197 
1198   if (!ST.hasNSAtoVMEMBug())
1199     return 0;
1200 
1201   if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isMTBUF(*MI))
1202     return 0;
1203 
1204   const SIInstrInfo *TII = ST.getInstrInfo();
1205   const auto *Offset = TII->getNamedOperand(*MI, AMDGPU::OpName::offset);
1206   if (!Offset || (Offset->getImm() & 6) == 0)
1207     return 0;
1208 
1209   auto IsHazardFn = [TII](const MachineInstr &I) {
1210     if (!SIInstrInfo::isMIMG(I))
1211       return false;
1212     const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(I.getOpcode());
1213     return Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA &&
1214            TII->getInstSizeInBytes(I) >= 16;
1215   };
1216 
1217   return NSAtoVMEMWaitStates - getWaitStatesSince(IsHazardFn, 1);
1218 }
1219 
1220 int GCNHazardRecognizer::checkFPAtomicToDenormModeHazard(MachineInstr *MI) {
1221   int FPAtomicToDenormModeWaitStates = 3;
1222 
1223   if (MI->getOpcode() != AMDGPU::S_DENORM_MODE)
1224     return 0;
1225 
1226   auto IsHazardFn = [](const MachineInstr &I) {
1227     if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isFLAT(I))
1228       return false;
1229     return SIInstrInfo::isFPAtomic(I);
1230   };
1231 
1232   auto IsExpiredFn = [](const MachineInstr &MI, int WaitStates) {
1233     if (WaitStates >= 3 || SIInstrInfo::isVALU(MI))
1234       return true;
1235 
1236     switch (MI.getOpcode()) {
1237     case AMDGPU::S_WAITCNT:
1238     case AMDGPU::S_WAITCNT_VSCNT:
1239     case AMDGPU::S_WAITCNT_VMCNT:
1240     case AMDGPU::S_WAITCNT_EXPCNT:
1241     case AMDGPU::S_WAITCNT_LGKMCNT:
1242     case AMDGPU::S_WAIT_IDLE:
1243       return true;
1244     default:
1245       break;
1246     }
1247 
1248     return false;
1249   };
1250 
1251   return FPAtomicToDenormModeWaitStates -
1252          ::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn);
1253 }
1254 
1255 int GCNHazardRecognizer::checkMAIHazards(MachineInstr *MI) {
1256   assert(SIInstrInfo::isMAI(*MI));
1257 
1258   return ST.hasGFX90AInsts() ? checkMAIHazards90A(MI) : checkMAIHazards908(MI);
1259 }
1260 
1261 int GCNHazardRecognizer::checkMFMAPadding(MachineInstr *MI) {
1262   // Early exit if no padding is requested.
1263   if (MFMAPaddingRatio == 0)
1264     return 0;
1265 
1266   auto IsMFMAFn = [](const MachineInstr &MI) {
1267     return SIInstrInfo::isMAI(MI) &&
1268            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1269            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1270   };
1271 
1272   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1273   if (!IsMFMAFn(*MI) || MFI->getOccupancy() < 2)
1274     return 0;
1275 
1276   int NeighborMFMALatency = 0;
1277   auto IsNeighboringMFMA = [&IsMFMAFn, &NeighborMFMALatency,
1278                             this](const MachineInstr &MI) {
1279     if (!IsMFMAFn(MI))
1280       return false;
1281 
1282     NeighborMFMALatency = this->getMFMAPipelineWaitStates(MI);
1283     return true;
1284   };
1285 
1286   const int MaxMFMAPipelineWaitStates = 16;
1287   int WaitStatesSinceNeighborMFMA =
1288       getWaitStatesSince(IsNeighboringMFMA, MaxMFMAPipelineWaitStates);
1289 
1290   int NeighborMFMAPaddingNeeded =
1291       (NeighborMFMALatency * MFMAPaddingRatio / 100) -
1292       WaitStatesSinceNeighborMFMA;
1293 
1294   return std::max(0, NeighborMFMAPaddingNeeded);
1295 }
1296 
1297 int GCNHazardRecognizer::checkMAIHazards908(MachineInstr *MI) {
1298   int WaitStatesNeeded = 0;
1299   unsigned Opc = MI->getOpcode();
1300 
1301   auto IsVALUFn = [](const MachineInstr &MI) {
1302     return SIInstrInfo::isVALU(MI);
1303   };
1304 
1305   if (Opc != AMDGPU::V_ACCVGPR_READ_B32_e64) { // MFMA or v_accvgpr_write
1306     const int LegacyVALUWritesVGPRWaitStates = 2;
1307     const int VALUWritesExecWaitStates = 4;
1308     const int MaxWaitStates = 4;
1309 
1310     int WaitStatesNeededForUse = VALUWritesExecWaitStates -
1311       getWaitStatesSinceDef(AMDGPU::EXEC, IsVALUFn, MaxWaitStates);
1312     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1313 
1314     if (WaitStatesNeeded < MaxWaitStates) {
1315       for (const MachineOperand &Use : MI->explicit_uses()) {
1316         const int MaxWaitStates = 2;
1317 
1318         if (!Use.isReg() || !TRI.isVGPR(MF.getRegInfo(), Use.getReg()))
1319           continue;
1320 
1321         int WaitStatesNeededForUse = LegacyVALUWritesVGPRWaitStates -
1322           getWaitStatesSinceDef(Use.getReg(), IsVALUFn, MaxWaitStates);
1323         WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1324 
1325         if (WaitStatesNeeded == MaxWaitStates)
1326           break;
1327       }
1328     }
1329   }
1330 
1331   auto IsMFMAFn = [](const MachineInstr &MI) {
1332     return SIInstrInfo::isMAI(MI) &&
1333            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1334            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1335   };
1336 
1337   for (const MachineOperand &Op : MI->explicit_operands()) {
1338     if (!Op.isReg() || !TRI.isAGPR(MF.getRegInfo(), Op.getReg()))
1339       continue;
1340 
1341     if (Op.isDef() && Opc != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1342       continue;
1343 
1344     const int MFMAWritesAGPROverlappedSrcABWaitStates = 4;
1345     const int MFMAWritesAGPROverlappedSrcCWaitStates = 2;
1346     const int MFMA4x4WritesAGPRAccVgprReadWaitStates = 4;
1347     const int MFMA16x16WritesAGPRAccVgprReadWaitStates = 10;
1348     const int MFMA32x32WritesAGPRAccVgprReadWaitStates = 18;
1349     const int MFMA4x4WritesAGPRAccVgprWriteWaitStates = 1;
1350     const int MFMA16x16WritesAGPRAccVgprWriteWaitStates = 7;
1351     const int MFMA32x32WritesAGPRAccVgprWriteWaitStates = 15;
1352     const int MaxWaitStates = 18;
1353     Register Reg = Op.getReg();
1354     unsigned HazardDefLatency = 0;
1355 
1356     auto IsOverlappedMFMAFn = [Reg, &IsMFMAFn, &HazardDefLatency,
1357                                this](const MachineInstr &MI) {
1358       if (!IsMFMAFn(MI))
1359         return false;
1360       Register DstReg = MI.getOperand(0).getReg();
1361       if (DstReg == Reg)
1362         return false;
1363       HazardDefLatency =
1364           std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI));
1365       return TRI.regsOverlap(DstReg, Reg);
1366     };
1367 
1368     int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsOverlappedMFMAFn,
1369                                                    MaxWaitStates);
1370     int NeedWaitStates = MFMAWritesAGPROverlappedSrcABWaitStates;
1371     int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
1372     int OpNo = MI->getOperandNo(&Op);
1373     if (OpNo == SrcCIdx) {
1374       NeedWaitStates = MFMAWritesAGPROverlappedSrcCWaitStates;
1375     } else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) {
1376       switch (HazardDefLatency) {
1377       case 2:  NeedWaitStates = MFMA4x4WritesAGPRAccVgprReadWaitStates;
1378                break;
1379       case 8:  NeedWaitStates = MFMA16x16WritesAGPRAccVgprReadWaitStates;
1380                break;
1381       case 16: LLVM_FALLTHROUGH;
1382       default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprReadWaitStates;
1383                break;
1384       }
1385     } else if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) {
1386       switch (HazardDefLatency) {
1387       case 2:  NeedWaitStates = MFMA4x4WritesAGPRAccVgprWriteWaitStates;
1388                break;
1389       case 8:  NeedWaitStates = MFMA16x16WritesAGPRAccVgprWriteWaitStates;
1390                break;
1391       case 16: LLVM_FALLTHROUGH;
1392       default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprWriteWaitStates;
1393                break;
1394       }
1395     }
1396 
1397     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1398     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1399 
1400     if (WaitStatesNeeded == MaxWaitStates)
1401       return WaitStatesNeeded; // Early exit.
1402 
1403     auto IsAccVgprWriteFn = [Reg, this](const MachineInstr &MI) {
1404       if (MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1405         return false;
1406       Register DstReg = MI.getOperand(0).getReg();
1407       return TRI.regsOverlap(Reg, DstReg);
1408     };
1409 
1410     const int AccVGPRWriteMFMAReadSrcCWaitStates = 1;
1411     const int AccVGPRWriteMFMAReadSrcABWaitStates = 3;
1412     const int AccVGPRWriteAccVgprReadWaitStates = 3;
1413     NeedWaitStates = AccVGPRWriteMFMAReadSrcABWaitStates;
1414     if (OpNo == SrcCIdx)
1415       NeedWaitStates = AccVGPRWriteMFMAReadSrcCWaitStates;
1416     else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64)
1417       NeedWaitStates = AccVGPRWriteAccVgprReadWaitStates;
1418 
1419     WaitStatesNeededForUse = NeedWaitStates -
1420       getWaitStatesSinceDef(Reg, IsAccVgprWriteFn, MaxWaitStates);
1421     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1422 
1423     if (WaitStatesNeeded == MaxWaitStates)
1424       return WaitStatesNeeded; // Early exit.
1425   }
1426 
1427   if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) {
1428     const int MFMA4x4ReadSrcCAccVgprWriteWaitStates = 0;
1429     const int MFMA16x16ReadSrcCAccVgprWriteWaitStates = 5;
1430     const int MFMA32x32ReadSrcCAccVgprWriteWaitStates = 13;
1431     const int MaxWaitStates = 13;
1432     Register DstReg = MI->getOperand(0).getReg();
1433     unsigned HazardDefLatency = 0;
1434 
1435     auto IsSrcCMFMAFn = [DstReg, &IsMFMAFn, &HazardDefLatency,
1436                          this](const MachineInstr &MI) {
1437       if (!IsMFMAFn(MI))
1438         return false;
1439       Register Reg = TII.getNamedOperand(MI, AMDGPU::OpName::src2)->getReg();
1440       HazardDefLatency =
1441           std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI));
1442       return TRI.regsOverlap(Reg, DstReg);
1443     };
1444 
1445     int WaitStatesSince = getWaitStatesSince(IsSrcCMFMAFn, MaxWaitStates);
1446     int NeedWaitStates;
1447     switch (HazardDefLatency) {
1448     case 2:  NeedWaitStates = MFMA4x4ReadSrcCAccVgprWriteWaitStates;
1449              break;
1450     case 8:  NeedWaitStates = MFMA16x16ReadSrcCAccVgprWriteWaitStates;
1451              break;
1452     case 16: LLVM_FALLTHROUGH;
1453     default: NeedWaitStates = MFMA32x32ReadSrcCAccVgprWriteWaitStates;
1454              break;
1455     }
1456 
1457     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSince;
1458     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1459   }
1460 
1461   // Pad neighboring MFMA with noops for better inter-wave performance.
1462   WaitStatesNeeded = std::max(WaitStatesNeeded, checkMFMAPadding(MI));
1463 
1464   return WaitStatesNeeded;
1465 }
1466 
1467 int GCNHazardRecognizer::checkMAIHazards90A(MachineInstr *MI) {
1468   int WaitStatesNeeded = 0;
1469   unsigned Opc = MI->getOpcode();
1470 
1471   auto IsMFMAFn = [](const MachineInstr &MI) {
1472     return SIInstrInfo::isMAI(MI) &&
1473            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1474            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1475   };
1476 
1477   auto IsLegacyVALUFn = [&IsMFMAFn](const MachineInstr &MI) {
1478     return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI);
1479   };
1480 
1481   auto IsLegacyVALUNotDotFn = [&IsMFMAFn](const MachineInstr &MI) {
1482     return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI) && !SIInstrInfo::isDOT(MI);
1483   };
1484 
1485   if (!IsMFMAFn(*MI))
1486     return WaitStatesNeeded;
1487 
1488   const int VALUWritesExecWaitStates = 4;
1489   int WaitStatesNeededForUse = VALUWritesExecWaitStates -
1490     getWaitStatesSinceDef(AMDGPU::EXEC, IsLegacyVALUFn,
1491                           VALUWritesExecWaitStates);
1492   WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1493 
1494   int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
1495 
1496   // Loop for both DGEMM and S/HGEMM 2nd instruction.
1497   for (const MachineOperand &Use : MI->explicit_uses()) {
1498     const int LegacyVALUNotDotWritesVGPRWaitStates = 2;
1499     const int SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates = 2;
1500     const int GFX940_XDL2PassWritesVGPROverlappedSMFMASrcCWaitStates = 3;
1501     const int GFX940_XDL4PassWritesVGPROverlappedSMFMASrcCWaitStates = 5;
1502     const int GFX940_SMFMA4PassWritesVGPROverlappedSMFMASrcCWaitStates = 4;
1503     const int GFX940_XDL8PassWritesVGPROverlappedSMFMASrcCWaitStates = 9;
1504     const int GFX940_SMFMA8PassWritesVGPROverlappedSMFMASrcCWaitStates = 8;
1505     const int GFX940_XDL16PassWritesVGPROverlappedSMFMASrcCWaitStates = 17;
1506     const int GFX940_SMFMA16PassWritesVGPROverlappedSMFMASrcCWaitStates = 16;
1507     const int SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates = 8;
1508     const int SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates = 16;
1509     const int SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates = 3;
1510     const int SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates = 9;
1511     const int SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates = 17;
1512     const int DMFMA16x16WritesVGPROverlappedSrcCWaitStates = 9;
1513     const int DMFMA4x4WritesVGPROverlappedSrcCWaitStates = 4;
1514     const int SMFMA4x4WritesVGPROverlappedSrcABWaitStates = 5;
1515     const int SMFMA16x16WritesVGPROverlappedSrcABWaitStates = 11;
1516     const int SMFMA32x32WritesVGPROverlappedSrcABWaitStates = 19;
1517     const int GFX940_SMFMA2PassWritesVGPROverlappedSrcABWaitStates = 4;
1518     const int GFX940_SMFMA4PassWritesVGPROverlappedSrcABWaitStates = 6;
1519     const int GFX940_SMFMA8PassWritesVGPROverlappedSrcABWaitStates = 10;
1520     const int GFX940_SMFMA16PassWritesVGPROverlappedSrcABWaitStates = 18;
1521     const int GFX940_XDL2PassWritesVGPROverlappedSrcABWaitStates = 5;
1522     const int GFX940_XDL4PassWritesVGPROverlappedSrcABWaitStates = 7;
1523     const int GFX940_XDL8PassWritesVGPROverlappedSrcABWaitStates = 11;
1524     const int GFX940_XDL16PassWritesVGPROverlappedSrcABWaitStates = 19;
1525     const int DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates = 6;
1526     const int DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates = 11;
1527     const int DMFMA4x4WritesVGPRFullSrcCWaitStates = 4;
1528     const int GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates = 2;
1529     const int MaxWaitStates = 19;
1530 
1531     if (!Use.isReg())
1532       continue;
1533     Register Reg = Use.getReg();
1534     bool FullReg;
1535     const MachineInstr *MI1;
1536 
1537     auto IsOverlappedMFMAFn = [Reg, &IsMFMAFn, &FullReg, &MI1,
1538                                this](const MachineInstr &MI) {
1539       if (!IsMFMAFn(MI))
1540         return false;
1541       Register DstReg = MI.getOperand(0).getReg();
1542       FullReg = (DstReg == Reg);
1543       MI1 = &MI;
1544       return TRI.regsOverlap(DstReg, Reg);
1545     };
1546 
1547     WaitStatesNeededForUse = LegacyVALUNotDotWritesVGPRWaitStates -
1548       getWaitStatesSinceDef(Reg, IsLegacyVALUNotDotFn, MaxWaitStates);
1549     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1550 
1551     int NumWaitStates =
1552         getWaitStatesSinceDef(Reg, IsOverlappedMFMAFn, MaxWaitStates);
1553     if (NumWaitStates == std::numeric_limits<int>::max())
1554       continue;
1555 
1556     int OpNo = MI->getOperandNo(&Use);
1557     unsigned Opc1 = MI1->getOpcode();
1558     int NeedWaitStates = 0;
1559     if (OpNo == SrcCIdx) {
1560       if (!isDGEMM(Opc) && (!ST.hasGFX940Insts() && isDGEMM(Opc1))) {
1561         NeedWaitStates = 0;
1562       } else if (FullReg) {
1563         if ((Opc == AMDGPU::V_MFMA_F64_4X4X4F64_e64 ||
1564              Opc == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64) &&
1565             (Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_e64 ||
1566              Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64))
1567           NeedWaitStates = DMFMA4x4WritesVGPRFullSrcCWaitStates;
1568         else if (ST.hasGFX940Insts() &&
1569                  TSchedModel.computeInstrLatency(MI1) == 2)
1570           NeedWaitStates = GFX940_SMFMA4x4WritesVGPRFullSrcCWaitStates;
1571       } else {
1572         switch (Opc1) {
1573         case AMDGPU::V_MFMA_F64_16X16X4F64_e64:
1574         case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64:
1575         case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64:
1576         case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64:
1577           if (!isXDL(ST, *MI))
1578             NeedWaitStates = DMFMA16x16WritesVGPROverlappedSrcCWaitStates;
1579           break;
1580         case AMDGPU::V_MFMA_F64_4X4X4F64_e64:
1581         case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64:
1582           if (!isXDL(ST, *MI))
1583             NeedWaitStates = DMFMA4x4WritesVGPROverlappedSrcCWaitStates;
1584           break;
1585         default:
1586           if (ST.hasGFX940Insts() && isXDL(ST, *MI) && !isXDL(ST, *MI1))
1587             break;
1588           switch (TSchedModel.computeInstrLatency(MI1)) {
1589           case 2:
1590             NeedWaitStates = ST.hasGFX940Insts()
1591               ? isXDL(ST, *MI1)
1592                 ? GFX940_XDL2PassWritesVGPROverlappedSMFMASrcCWaitStates
1593                 : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates
1594               : isDGEMM(Opc)
1595                 ? SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates
1596                 : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates;
1597             break;
1598           case 4:
1599             assert(ST.hasGFX940Insts());
1600             NeedWaitStates = isXDL(ST, *MI1)
1601               ? GFX940_XDL4PassWritesVGPROverlappedSMFMASrcCWaitStates
1602               : GFX940_SMFMA4PassWritesVGPROverlappedSMFMASrcCWaitStates;
1603             break;
1604           case 8:
1605             NeedWaitStates = ST.hasGFX940Insts()
1606               ? isXDL(ST, *MI1)
1607                 ? GFX940_XDL8PassWritesVGPROverlappedSMFMASrcCWaitStates
1608                 : GFX940_SMFMA8PassWritesVGPROverlappedSMFMASrcCWaitStates
1609               : isDGEMM(Opc)
1610                 ? SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates
1611                 : SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates;
1612             break;
1613           case 16: LLVM_FALLTHROUGH;
1614           default:
1615             NeedWaitStates = ST.hasGFX940Insts()
1616               ? isXDL(ST, *MI1)
1617                 ? GFX940_XDL16PassWritesVGPROverlappedSMFMASrcCWaitStates
1618                 : GFX940_SMFMA16PassWritesVGPROverlappedSMFMASrcCWaitStates
1619               : isDGEMM(Opc)
1620                 ? SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates
1621                 : SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates;
1622           }
1623         }
1624       }
1625     } else {
1626       switch (Opc1) {
1627       case AMDGPU::V_MFMA_F64_16X16X4F64_e64:
1628       case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64:
1629       case AMDGPU::V_MFMA_F64_16X16X4F64_mac_e64:
1630       case AMDGPU::V_MFMA_F64_16X16X4F64_mac_vgprcd_e64:
1631         NeedWaitStates = DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates;
1632         break;
1633       case AMDGPU::V_MFMA_F64_4X4X4F64_e64:
1634       case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64:
1635         NeedWaitStates = DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates;
1636         break;
1637       default:
1638         switch (TSchedModel.computeInstrLatency(MI1)) {
1639         case 2:
1640           NeedWaitStates = ST.hasGFX940Insts()
1641             ? isXDL(ST, *MI1)
1642               ? GFX940_XDL2PassWritesVGPROverlappedSrcABWaitStates
1643               : GFX940_SMFMA2PassWritesVGPROverlappedSrcABWaitStates
1644             : SMFMA4x4WritesVGPROverlappedSrcABWaitStates;
1645           break;
1646         case 4:
1647           assert(ST.hasGFX940Insts());
1648           NeedWaitStates = isXDL(ST, *MI1)
1649             ? GFX940_XDL4PassWritesVGPROverlappedSrcABWaitStates
1650             : GFX940_SMFMA4PassWritesVGPROverlappedSrcABWaitStates;
1651           break;
1652         case 8:
1653           NeedWaitStates = ST.hasGFX940Insts()
1654             ? isXDL(ST, *MI1)
1655               ? GFX940_XDL8PassWritesVGPROverlappedSrcABWaitStates
1656               : GFX940_SMFMA8PassWritesVGPROverlappedSrcABWaitStates
1657             : SMFMA16x16WritesVGPROverlappedSrcABWaitStates;
1658           break;
1659         case 16: LLVM_FALLTHROUGH;
1660         default:
1661           NeedWaitStates = ST.hasGFX940Insts()
1662             ? isXDL(ST, *MI1)
1663               ? GFX940_XDL16PassWritesVGPROverlappedSrcABWaitStates
1664               : GFX940_SMFMA16PassWritesVGPROverlappedSrcABWaitStates
1665             : SMFMA32x32WritesVGPROverlappedSrcABWaitStates;
1666         }
1667       }
1668     }
1669     if (WaitStatesNeeded >= NeedWaitStates)
1670       continue;
1671 
1672     WaitStatesNeededForUse = NeedWaitStates - NumWaitStates;
1673     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1674 
1675     if (WaitStatesNeeded == MaxWaitStates)
1676       break;
1677   }
1678 
1679   return WaitStatesNeeded;
1680 }
1681 
1682 int GCNHazardRecognizer::checkMAILdStHazards(MachineInstr *MI) {
1683   // On gfx90a+ relevant hazards are checked in checkMAIVALUHazards()
1684   if (!ST.hasMAIInsts() || ST.hasGFX90AInsts())
1685     return 0;
1686 
1687   int WaitStatesNeeded = 0;
1688 
1689   auto IsAccVgprReadFn = [](const MachineInstr &MI) {
1690     return MI.getOpcode() == AMDGPU::V_ACCVGPR_READ_B32_e64;
1691   };
1692 
1693   for (const MachineOperand &Op : MI->explicit_uses()) {
1694     if (!Op.isReg() || !TRI.isVGPR(MF.getRegInfo(), Op.getReg()))
1695       continue;
1696 
1697     Register Reg = Op.getReg();
1698 
1699     const int AccVgprReadLdStWaitStates = 2;
1700     const int VALUWriteAccVgprRdWrLdStDepVALUWaitStates = 1;
1701     const int MaxWaitStates = 2;
1702 
1703     int WaitStatesNeededForUse = AccVgprReadLdStWaitStates -
1704       getWaitStatesSinceDef(Reg, IsAccVgprReadFn, MaxWaitStates);
1705     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1706 
1707     if (WaitStatesNeeded == MaxWaitStates)
1708       return WaitStatesNeeded; // Early exit.
1709 
1710     auto IsVALUAccVgprRdWrCheckFn = [Reg, this](const MachineInstr &MI) {
1711       if (MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64 &&
1712           MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
1713         return false;
1714       auto IsVALUFn = [](const MachineInstr &MI) {
1715         return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMAI(MI);
1716       };
1717       return getWaitStatesSinceDef(Reg, IsVALUFn, 2 /*MaxWaitStates*/) <
1718              std::numeric_limits<int>::max();
1719     };
1720 
1721     WaitStatesNeededForUse = VALUWriteAccVgprRdWrLdStDepVALUWaitStates -
1722       getWaitStatesSince(IsVALUAccVgprRdWrCheckFn, MaxWaitStates);
1723     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1724   }
1725 
1726   return WaitStatesNeeded;
1727 }
1728 
1729 int GCNHazardRecognizer::checkMAIVALUHazards(MachineInstr *MI) {
1730   if (!ST.hasGFX90AInsts())
1731     return 0;
1732 
1733   auto IsMFMAFn = [](const MachineInstr &MI) -> bool {
1734     return SIInstrInfo::isMAI(MI) &&
1735            MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
1736            MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64;
1737   };
1738 
1739   auto IsDGEMMFn = [](const MachineInstr &MI) -> bool {
1740     return isDGEMM(MI.getOpcode());
1741   };
1742 
1743   // This is checked in checkMAIHazards90A()
1744   if (IsMFMAFn(*MI))
1745     return 0;
1746 
1747   int WaitStatesNeeded = 0;
1748 
1749   bool IsMemOrExport = SIInstrInfo::isVMEM(*MI) ||
1750                        SIInstrInfo::isFLAT(*MI) ||
1751                        SIInstrInfo::isDS(*MI) ||
1752                        SIInstrInfo::isEXP(*MI);
1753   bool IsVALU = SIInstrInfo::isVALU(*MI);
1754 
1755   const MachineInstr *MFMA = nullptr;
1756   unsigned Reg;
1757   auto IsMFMAWriteFn = [&Reg, &IsMFMAFn, &MFMA, this](const MachineInstr &MI) {
1758     if (!IsMFMAFn(MI) || !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg))
1759       return false;
1760     MFMA = &MI;
1761     return true;
1762   };
1763 
1764   const MachineInstr *DOT = nullptr;
1765   auto IsDotWriteFn = [&Reg, &DOT, this](const MachineInstr &MI) {
1766     if (!SIInstrInfo::isDOT(MI) ||
1767         !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg))
1768       return false;
1769     DOT = &MI;
1770     return true;
1771   };
1772 
1773   int SrcCIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(),
1774                                            AMDGPU::OpName::src2);
1775 
1776   if (IsMemOrExport || IsVALU) {
1777     const int SMFMA4x4WriteVgprVALUMemExpReadWaitStates = 5;
1778     const int SMFMA16x16WriteVgprVALUMemExpReadWaitStates = 11;
1779     const int SMFMA32x32WriteVgprVALUMemExpReadWaitStates = 19;
1780     const int GFX940_SMFMA2PassWriteVgprVALUMemExpReadWaitStates = 4;
1781     const int GFX940_SMFMA4PassWriteVgprVALUMemExpReadWaitStates = 6;
1782     const int GFX940_SMFMA8PassWriteVgprVALUMemExpReadWaitStates = 10;
1783     const int GFX940_SMFMA16PassWriteVgprVALUMemExpReadWaitStates = 18;
1784     const int GFX940_XDL2PassWriteVgprVALUMemExpReadWaitStates = 5;
1785     const int GFX940_XDL4PassWriteVgprVALUMemExpReadWaitStates = 7;
1786     const int GFX940_XDL8PassWriteVgprVALUMemExpReadWaitStates = 11;
1787     const int GFX940_XDL16PassWriteVgprVALUMemExpReadWaitStates = 19;
1788     const int DMFMA4x4WriteVgprMemExpReadWaitStates = 9;
1789     const int DMFMA16x16WriteVgprMemExpReadWaitStates = 18;
1790     const int DMFMA4x4WriteVgprVALUReadWaitStates = 6;
1791     const int DMFMA16x16WriteVgprVALUReadWaitStates = 11;
1792     const int DotWriteSameDotReadSrcAB = 3;
1793     const int DotWriteDifferentVALURead = 3;
1794     const int MaxWaitStates = 19;
1795 
1796     for (const MachineOperand &Use : MI->explicit_uses()) {
1797       if (!Use.isReg())
1798         continue;
1799       Reg = Use.getReg();
1800 
1801       DOT = nullptr;
1802       int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn,
1803                                                      MaxWaitStates);
1804       if (DOT) {
1805         int NeedWaitStates = 0;
1806         if (DOT->getOpcode() == MI->getOpcode()) {
1807           if (&Use - &MI->getOperand(0) != SrcCIdx)
1808             NeedWaitStates = DotWriteSameDotReadSrcAB;
1809         } else {
1810           NeedWaitStates = DotWriteDifferentVALURead;
1811         }
1812 
1813         int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1814         WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1815       }
1816 
1817       MFMA = nullptr;
1818       WaitStatesSinceDef =
1819           getWaitStatesSinceDef(Reg, IsMFMAWriteFn, MaxWaitStates);
1820       if (!MFMA)
1821         continue;
1822 
1823       unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA);
1824       int NeedWaitStates = MaxWaitStates;
1825       switch (HazardDefLatency) {
1826       case 2:
1827         NeedWaitStates =
1828           ST.hasGFX940Insts()
1829             ? isXDL(ST, *MFMA)
1830               ? GFX940_XDL2PassWriteVgprVALUMemExpReadWaitStates
1831               : GFX940_SMFMA2PassWriteVgprVALUMemExpReadWaitStates
1832             : SMFMA4x4WriteVgprVALUMemExpReadWaitStates;
1833         break;
1834       case 4:
1835         assert(isDGEMM(MFMA->getOpcode()) || ST.hasGFX940Insts());
1836         NeedWaitStates =
1837           isDGEMM(MFMA->getOpcode())
1838             ? IsMemOrExport ? DMFMA4x4WriteVgprMemExpReadWaitStates
1839                             : DMFMA4x4WriteVgprVALUReadWaitStates
1840             : isXDL(ST, *MFMA)
1841               ? GFX940_XDL4PassWriteVgprVALUMemExpReadWaitStates
1842               : GFX940_SMFMA4PassWriteVgprVALUMemExpReadWaitStates;
1843         break;
1844       case 8:
1845         NeedWaitStates =
1846           ST.hasGFX940Insts()
1847             ? isXDL(ST, *MFMA)
1848               ? GFX940_XDL8PassWriteVgprVALUMemExpReadWaitStates
1849               : GFX940_SMFMA8PassWriteVgprVALUMemExpReadWaitStates
1850             : SMFMA16x16WriteVgprVALUMemExpReadWaitStates;
1851         break;
1852       case 16: LLVM_FALLTHROUGH;
1853       default:
1854         NeedWaitStates =
1855           isDGEMM(MFMA->getOpcode())
1856             ? IsMemOrExport ? DMFMA16x16WriteVgprMemExpReadWaitStates
1857                             : DMFMA16x16WriteVgprVALUReadWaitStates
1858             : ST.hasGFX940Insts()
1859               ? isXDL(ST, *MFMA)
1860                 ? GFX940_XDL16PassWriteVgprVALUMemExpReadWaitStates
1861                 : GFX940_SMFMA16PassWriteVgprVALUMemExpReadWaitStates
1862               : SMFMA32x32WriteVgprVALUMemExpReadWaitStates;
1863         break;
1864       }
1865 
1866       int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1867       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1868 
1869       if (WaitStatesNeeded == MaxWaitStates)
1870         break;
1871     }
1872   }
1873 
1874   unsigned Opc = MI->getOpcode();
1875   const int DMFMAToFMA64WaitStates = 2;
1876   if ((Opc == AMDGPU::V_FMA_F64_e64 ||
1877        Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64 ||
1878        Opc == AMDGPU::V_FMAC_F64_dpp) &&
1879       WaitStatesNeeded < DMFMAToFMA64WaitStates) {
1880     int WaitStatesNeededForUse = DMFMAToFMA64WaitStates -
1881       getWaitStatesSince(IsDGEMMFn, DMFMAToFMA64WaitStates);
1882     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1883   }
1884 
1885   if (!IsVALU && !IsMemOrExport)
1886     return WaitStatesNeeded;
1887 
1888   for (const MachineOperand &Def : MI->defs()) {
1889     const int SMFMA4x4WriteVgprVALUWawWaitStates = 5;
1890     const int SMFMA16x16WriteVgprVALUWawWaitStates = 11;
1891     const int SMFMA32x32WriteVgprVALUWawWaitStates = 19;
1892     const int GFX940_SMFMA2PassWriteVgprVALUWawWaitStates = 4;
1893     const int GFX940_SMFMA4PassWriteVgprVALUWawWaitStates = 6;
1894     const int GFX940_SMFMA8PassWriteVgprVALUWawWaitStates = 10;
1895     const int GFX940_SMFMA16PassWriteVgprVALUWawWaitStates = 18;
1896     const int GFX940_XDL2PassWriteVgprVALUWawWaitStates = 5;
1897     const int GFX940_XDL4PassWriteVgprVALUWawWaitStates = 7;
1898     const int GFX940_XDL8PassWriteVgprVALUWawWaitStates = 11;
1899     const int GFX940_XDL16PassWriteVgprVALUWawWaitStates = 19;
1900     const int SMFMA4x4ReadVgprVALUWarWaitStates = 1;
1901     const int GFX940_XDL4PassReadVgprVALUWarWaitStates = 3;
1902     const int SMFMA16x16ReadVgprVALUWarWaitStates = 7;
1903     const int SMFMA32x32ReadVgprVALUWarWaitStates = 15;
1904     const int DMFMA4x4WriteVgprVALUWriteWaitStates = 6;
1905     const int DMFMA16x16WriteVgprVALUWriteWaitStates = 11;
1906     const int DotWriteDifferentVALUWrite = 3;
1907     const int MaxWaitStates = 19;
1908     const int MaxWarWaitStates = 15;
1909 
1910     Reg = Def.getReg();
1911 
1912     DOT = nullptr;
1913     int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn,
1914                                                    MaxWaitStates);
1915     if (DOT && DOT->getOpcode() != MI->getOpcode())
1916       WaitStatesNeeded = std::max(WaitStatesNeeded, DotWriteDifferentVALUWrite -
1917                                                     WaitStatesSinceDef);
1918 
1919     MFMA = nullptr;
1920     WaitStatesSinceDef =
1921         getWaitStatesSinceDef(Reg, IsMFMAWriteFn, MaxWaitStates);
1922     if (MFMA) {
1923       int NeedWaitStates = MaxWaitStates;
1924       switch (TSchedModel.computeInstrLatency(MFMA)) {
1925       case 2:
1926         NeedWaitStates = ST.hasGFX940Insts()
1927           ? isXDL(ST, *MFMA)
1928             ? GFX940_XDL2PassWriteVgprVALUWawWaitStates
1929             : GFX940_SMFMA2PassWriteVgprVALUWawWaitStates
1930           : SMFMA4x4WriteVgprVALUWawWaitStates;
1931         break;
1932       case 4:
1933         assert(isDGEMM(MFMA->getOpcode()) || ST.hasGFX940Insts());
1934         NeedWaitStates = isDGEMM(MFMA->getOpcode())
1935             ? DMFMA4x4WriteVgprVALUWriteWaitStates
1936             : isXDL(ST, *MFMA)
1937               ? GFX940_XDL4PassWriteVgprVALUWawWaitStates
1938               : GFX940_SMFMA4PassWriteVgprVALUWawWaitStates;
1939         break;
1940       case 8:
1941         NeedWaitStates = ST.hasGFX940Insts()
1942           ? isXDL(ST, *MFMA)
1943             ? GFX940_XDL8PassWriteVgprVALUWawWaitStates
1944             : GFX940_SMFMA8PassWriteVgprVALUWawWaitStates
1945           : SMFMA16x16WriteVgprVALUWawWaitStates;
1946         break;
1947       case 16: LLVM_FALLTHROUGH;
1948       default:
1949         NeedWaitStates = isDGEMM(MFMA->getOpcode())
1950                    ? DMFMA16x16WriteVgprVALUWriteWaitStates
1951                    : ST.hasGFX940Insts()
1952                      ? isXDL(ST, *MFMA)
1953                        ? GFX940_XDL16PassWriteVgprVALUWawWaitStates
1954                        : GFX940_SMFMA16PassWriteVgprVALUWawWaitStates
1955                    : SMFMA32x32WriteVgprVALUWawWaitStates;
1956         break;
1957       }
1958 
1959       int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef;
1960       WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
1961 
1962       if (WaitStatesNeeded == MaxWaitStates)
1963         break;
1964     }
1965 
1966     auto IsSMFMAReadAsCFn = [&Reg, &IsMFMAFn, &MFMA,
1967                              this](const MachineInstr &MI) {
1968       if (!IsMFMAFn(MI) || isDGEMM(MI.getOpcode()) ||
1969           !MI.readsRegister(Reg, &TRI))
1970         return false;
1971 
1972       if (ST.hasGFX940Insts() && !isXDL(ST, MI))
1973         return false;
1974 
1975       const MachineOperand *SrcC =
1976           TII.getNamedOperand(MI, AMDGPU::OpName::src2);
1977       assert(SrcC);
1978       if (!SrcC->isReg() || !TRI.regsOverlap(SrcC->getReg(), Reg))
1979         return false;
1980 
1981       MFMA = &MI;
1982       return true;
1983     };
1984 
1985     MFMA = nullptr;
1986     int WaitStatesSinceUse = getWaitStatesSince(IsSMFMAReadAsCFn,
1987                                                 MaxWarWaitStates);
1988     if (!MFMA)
1989       continue;
1990 
1991     unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA);
1992     int NeedWaitStates = MaxWaitStates;
1993     switch (HazardDefLatency) {
1994     case 2:  NeedWaitStates = SMFMA4x4ReadVgprVALUWarWaitStates;
1995              break;
1996     case 4:  assert(ST.hasGFX940Insts());
1997              NeedWaitStates = GFX940_XDL4PassReadVgprVALUWarWaitStates;
1998              break;
1999     case 8:  NeedWaitStates = SMFMA16x16ReadVgprVALUWarWaitStates;
2000              break;
2001     case 16: LLVM_FALLTHROUGH;
2002     default: NeedWaitStates = SMFMA32x32ReadVgprVALUWarWaitStates;
2003              break;
2004     }
2005 
2006     int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceUse;
2007     WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse);
2008   }
2009 
2010   return WaitStatesNeeded;
2011 }
2012 
2013 bool GCNHazardRecognizer::ShouldPreferAnother(SUnit *SU) {
2014   if (!SU->isInstr())
2015     return false;
2016 
2017   const MachineInstr *MAI = nullptr;
2018   auto IsMFMAFn = [&MAI](const MachineInstr &MI) {
2019     MAI = nullptr;
2020     if (SIInstrInfo::isMAI(MI) &&
2021         MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
2022         MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64)
2023       MAI = &MI;
2024     return MAI != nullptr;
2025   };
2026 
2027   MachineInstr *MI = SU->getInstr();
2028   if (IsMFMAFn(*MI)) {
2029     int W = getWaitStatesSince(IsMFMAFn, 16);
2030     if (MAI)
2031       return W < (int)TSchedModel.computeInstrLatency(MAI);
2032   }
2033 
2034   return false;
2035 }
2036