xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/R600MachineScheduler.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===-- R600MachineScheduler.cpp - R600 Scheduler Interface -*- C++ -*-----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// R600 Machine Scheduler interface
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "R600MachineScheduler.h"
15 #include "MCTargetDesc/R600MCTargetDesc.h"
16 #include "R600Subtarget.h"
17 
18 using namespace llvm;
19 
20 #define DEBUG_TYPE "machine-scheduler"
21 
initialize(ScheduleDAGMI * dag)22 void R600SchedStrategy::initialize(ScheduleDAGMI *dag) {
23   assert(dag->hasVRegLiveness() && "R600SchedStrategy needs vreg liveness");
24   DAG = static_cast<ScheduleDAGMILive*>(dag);
25   const R600Subtarget &ST = DAG->MF.getSubtarget<R600Subtarget>();
26   TII = static_cast<const R600InstrInfo*>(DAG->TII);
27   TRI = static_cast<const R600RegisterInfo*>(DAG->TRI);
28   VLIW5 = !ST.hasCaymanISA();
29   MRI = &DAG->MRI;
30   CurInstKind = IDOther;
31   CurEmitted = 0;
32   OccupiedSlotsMask = 31;
33   InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
34   InstKindLimit[IDOther] = 32;
35   InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
36   AluInstCount = 0;
37   FetchInstCount = 0;
38 }
39 
MoveUnits(std::vector<SUnit * > & QSrc,std::vector<SUnit * > & QDst)40 void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
41                                   std::vector<SUnit *> &QDst)
42 {
43   llvm::append_range(QDst, QSrc);
44   QSrc.clear();
45 }
46 
getWFCountLimitedByGPR(unsigned GPRCount)47 static unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
48   assert (GPRCount && "GPRCount cannot be 0");
49   return 248 / GPRCount;
50 }
51 
pickNode(bool & IsTopNode)52 SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
53   SUnit *SU = nullptr;
54   NextInstKind = IDOther;
55 
56   IsTopNode = false;
57 
58   // check if we might want to switch current clause type
59   bool AllowSwitchToAlu = (CurEmitted >= InstKindLimit[CurInstKind]) ||
60       (Available[CurInstKind].empty());
61   bool AllowSwitchFromAlu = (CurEmitted >= InstKindLimit[CurInstKind]) &&
62       (!Available[IDFetch].empty() || !Available[IDOther].empty());
63 
64   if (CurInstKind == IDAlu && !Available[IDFetch].empty()) {
65     // We use the heuristic provided by AMD Accelerated Parallel Processing
66     // OpenCL Programming Guide :
67     // The approx. number of WF that allows TEX inst to hide ALU inst is :
68     // 500 (cycles for TEX) / (AluFetchRatio * 8 (cycles for ALU))
69     float ALUFetchRationEstimate =
70         (AluInstCount + AvailablesAluCount() + Pending[IDAlu].size()) /
71         (FetchInstCount + Available[IDFetch].size());
72     if (ALUFetchRationEstimate == 0) {
73       AllowSwitchFromAlu = true;
74     } else {
75       unsigned NeededWF = 62.5f / ALUFetchRationEstimate;
76       LLVM_DEBUG(dbgs() << NeededWF << " approx. Wavefronts Required\n");
77       // We assume the local GPR requirements to be "dominated" by the requirement
78       // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
79       // after TEX are indeed likely to consume or generate values from/for the
80       // TEX clause.
81       // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
82       // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
83       // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
84       // (TODO : use RegisterPressure)
85       // If we are going too use too many GPR, we flush Fetch instruction to lower
86       // register pressure on 128 bits regs.
87       unsigned NearRegisterRequirement = 2 * Available[IDFetch].size();
88       if (NeededWF > getWFCountLimitedByGPR(NearRegisterRequirement))
89         AllowSwitchFromAlu = true;
90     }
91   }
92 
93   if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
94       (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
95     // try to pick ALU
96     SU = pickAlu();
97     if (!SU && !PhysicalRegCopy.empty()) {
98       SU = PhysicalRegCopy.front();
99       PhysicalRegCopy.erase(PhysicalRegCopy.begin());
100     }
101     if (SU) {
102       if (CurEmitted >= InstKindLimit[IDAlu])
103         CurEmitted = 0;
104       NextInstKind = IDAlu;
105     }
106   }
107 
108   if (!SU) {
109     // try to pick FETCH
110     SU = pickOther(IDFetch);
111     if (SU)
112       NextInstKind = IDFetch;
113   }
114 
115   // try to pick other
116   if (!SU) {
117     SU = pickOther(IDOther);
118     if (SU)
119       NextInstKind = IDOther;
120   }
121 
122   LLVM_DEBUG(if (SU) {
123     dbgs() << " ** Pick node **\n";
124     DAG->dumpNode(*SU);
125   } else {
126     dbgs() << "NO NODE \n";
127     for (const SUnit &S : DAG->SUnits)
128       if (!S.isScheduled)
129         DAG->dumpNode(S);
130   });
131 
132   return SU;
133 }
134 
schedNode(SUnit * SU,bool IsTopNode)135 void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
136   if (NextInstKind != CurInstKind) {
137     LLVM_DEBUG(dbgs() << "Instruction Type Switch\n");
138     if (NextInstKind != IDAlu)
139       OccupiedSlotsMask |= 31;
140     CurEmitted = 0;
141     CurInstKind = NextInstKind;
142   }
143 
144   if (CurInstKind == IDAlu) {
145     AluInstCount ++;
146     switch (getAluKind(SU)) {
147     case AluT_XYZW:
148       CurEmitted += 4;
149       break;
150     case AluDiscarded:
151       break;
152     default: {
153       ++CurEmitted;
154       for (MachineInstr::mop_iterator It = SU->getInstr()->operands_begin(),
155           E = SU->getInstr()->operands_end(); It != E; ++It) {
156         MachineOperand &MO = *It;
157         if (MO.isReg() && MO.getReg() == R600::ALU_LITERAL_X)
158           ++CurEmitted;
159       }
160     }
161     }
162   } else {
163     ++CurEmitted;
164   }
165 
166   LLVM_DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
167 
168   if (CurInstKind != IDFetch) {
169     MoveUnits(Pending[IDFetch], Available[IDFetch]);
170   } else
171     FetchInstCount++;
172 }
173 
174 static bool
isPhysicalRegCopy(MachineInstr * MI)175 isPhysicalRegCopy(MachineInstr *MI) {
176   if (MI->getOpcode() != R600::COPY)
177     return false;
178 
179   return !MI->getOperand(1).getReg().isVirtual();
180 }
181 
releaseTopNode(SUnit * SU)182 void R600SchedStrategy::releaseTopNode(SUnit *SU) {
183   LLVM_DEBUG(dbgs() << "Top Releasing "; DAG->dumpNode(*SU));
184 }
185 
releaseBottomNode(SUnit * SU)186 void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
187   LLVM_DEBUG(dbgs() << "Bottom Releasing "; DAG->dumpNode(*SU));
188   if (isPhysicalRegCopy(SU->getInstr())) {
189     PhysicalRegCopy.push_back(SU);
190     return;
191   }
192 
193   int IK = getInstKind(SU);
194 
195   // There is no export clause, we can schedule one as soon as its ready
196   if (IK == IDOther)
197     Available[IDOther].push_back(SU);
198   else
199     Pending[IK].push_back(SU);
200 
201 }
202 
regBelongsToClass(Register Reg,const TargetRegisterClass * RC) const203 bool R600SchedStrategy::regBelongsToClass(Register Reg,
204                                           const TargetRegisterClass *RC) const {
205   if (!Reg.isVirtual())
206     return RC->contains(Reg);
207   return MRI->getRegClass(Reg) == RC;
208 }
209 
getAluKind(SUnit * SU) const210 R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
211   MachineInstr *MI = SU->getInstr();
212 
213   if (TII->isTransOnly(*MI))
214     return AluTrans;
215 
216   switch (MI->getOpcode()) {
217   case R600::PRED_X:
218     return AluPredX;
219   case R600::INTERP_PAIR_XY:
220   case R600::INTERP_PAIR_ZW:
221   case R600::INTERP_VEC_LOAD:
222   case R600::DOT_4:
223     return AluT_XYZW;
224   case R600::COPY:
225     if (MI->getOperand(1).isUndef()) {
226       // MI will become a KILL, don't considers it in scheduling
227       return AluDiscarded;
228     }
229     break;
230   default:
231     break;
232   }
233 
234   // Does the instruction take a whole IG ?
235   // XXX: Is it possible to add a helper function in R600InstrInfo that can
236   // be used here and in R600PacketizerList::isSoloInstruction() ?
237   if(TII->isVector(*MI) ||
238      TII->isCubeOp(MI->getOpcode()) ||
239      TII->isReductionOp(MI->getOpcode()) ||
240      MI->getOpcode() == R600::GROUP_BARRIER) {
241     return AluT_XYZW;
242   }
243 
244   if (TII->isLDSInstr(MI->getOpcode())) {
245     return AluT_X;
246   }
247 
248   // Is the result already assigned to a channel ?
249   unsigned DestSubReg = MI->getOperand(0).getSubReg();
250   switch (DestSubReg) {
251   case R600::sub0:
252     return AluT_X;
253   case R600::sub1:
254     return AluT_Y;
255   case R600::sub2:
256     return AluT_Z;
257   case R600::sub3:
258     return AluT_W;
259   default:
260     break;
261   }
262 
263   // Is the result already member of a X/Y/Z/W class ?
264   Register DestReg = MI->getOperand(0).getReg();
265   if (regBelongsToClass(DestReg, &R600::R600_TReg32_XRegClass) ||
266       regBelongsToClass(DestReg, &R600::R600_AddrRegClass))
267     return AluT_X;
268   if (regBelongsToClass(DestReg, &R600::R600_TReg32_YRegClass))
269     return AluT_Y;
270   if (regBelongsToClass(DestReg, &R600::R600_TReg32_ZRegClass))
271     return AluT_Z;
272   if (regBelongsToClass(DestReg, &R600::R600_TReg32_WRegClass))
273     return AluT_W;
274   if (regBelongsToClass(DestReg, &R600::R600_Reg128RegClass))
275     return AluT_XYZW;
276 
277   // LDS src registers cannot be used in the Trans slot.
278   if (TII->readsLDSSrcReg(*MI))
279     return AluT_XYZW;
280 
281   return AluAny;
282 }
283 
getInstKind(SUnit * SU)284 int R600SchedStrategy::getInstKind(SUnit* SU) {
285   int Opcode = SU->getInstr()->getOpcode();
286 
287   if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
288     return IDFetch;
289 
290   if (TII->isALUInstr(Opcode)) {
291     return IDAlu;
292   }
293 
294   switch (Opcode) {
295   case R600::PRED_X:
296   case R600::COPY:
297   case R600::CONST_COPY:
298   case R600::INTERP_PAIR_XY:
299   case R600::INTERP_PAIR_ZW:
300   case R600::INTERP_VEC_LOAD:
301   case R600::DOT_4:
302     return IDAlu;
303   default:
304     return IDOther;
305   }
306 }
307 
PopInst(std::vector<SUnit * > & Q,bool AnyALU)308 SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
309   if (Q.empty())
310     return nullptr;
311   for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
312       It != E; ++It) {
313     SUnit *SU = *It;
314     InstructionsGroupCandidate.push_back(SU->getInstr());
315     if (TII->fitsConstReadLimitations(InstructionsGroupCandidate) &&
316         (!AnyALU || !TII->isVectorOnly(*SU->getInstr()))) {
317       InstructionsGroupCandidate.pop_back();
318       Q.erase((It + 1).base());
319       return SU;
320     }
321     InstructionsGroupCandidate.pop_back();
322   }
323   return nullptr;
324 }
325 
LoadAlu()326 void R600SchedStrategy::LoadAlu() {
327   std::vector<SUnit *> &QSrc = Pending[IDAlu];
328   for (SUnit *SU : QSrc) {
329     AluKind AK = getAluKind(SU);
330     AvailableAlus[AK].push_back(SU);
331   }
332   QSrc.clear();
333 }
334 
PrepareNextSlot()335 void R600SchedStrategy::PrepareNextSlot() {
336   LLVM_DEBUG(dbgs() << "New Slot\n");
337   assert(OccupiedSlotsMask && "Slot wasn't filled");
338   OccupiedSlotsMask = 0;
339   //  if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
340   //    OccupiedSlotsMask |= 16;
341   InstructionsGroupCandidate.clear();
342   LoadAlu();
343 }
344 
AssignSlot(MachineInstr * MI,unsigned Slot)345 void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
346   int DstIndex = TII->getOperandIdx(MI->getOpcode(), R600::OpName::dst);
347   if (DstIndex == -1) {
348     return;
349   }
350   Register DestReg = MI->getOperand(DstIndex).getReg();
351   // PressureRegister crashes if an operand is def and used in the same inst
352   // and we try to constraint its regclass
353   for (MachineInstr::mop_iterator It = MI->operands_begin(),
354       E = MI->operands_end(); It != E; ++It) {
355     MachineOperand &MO = *It;
356     if (MO.isReg() && !MO.isDef() &&
357         MO.getReg() == DestReg)
358       return;
359   }
360   // Constrains the regclass of DestReg to assign it to Slot
361   switch (Slot) {
362   case 0:
363     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_XRegClass);
364     break;
365   case 1:
366     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_YRegClass);
367     break;
368   case 2:
369     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_ZRegClass);
370     break;
371   case 3:
372     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_WRegClass);
373     break;
374   }
375 }
376 
AttemptFillSlot(unsigned Slot,bool AnyAlu)377 SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
378   static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
379   SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
380   if (SlotedSU)
381     return SlotedSU;
382   SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
383   if (UnslotedSU)
384     AssignSlot(UnslotedSU->getInstr(), Slot);
385   return UnslotedSU;
386 }
387 
AvailablesAluCount() const388 unsigned R600SchedStrategy::AvailablesAluCount() const {
389   return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
390       AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
391       AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
392       AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
393       AvailableAlus[AluPredX].size();
394 }
395 
pickAlu()396 SUnit* R600SchedStrategy::pickAlu() {
397   while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
398     if (!OccupiedSlotsMask) {
399       // Bottom up scheduling : predX must comes first
400       if (!AvailableAlus[AluPredX].empty()) {
401         OccupiedSlotsMask |= 31;
402         return PopInst(AvailableAlus[AluPredX], false);
403       }
404       // Flush physical reg copies (RA will discard them)
405       if (!AvailableAlus[AluDiscarded].empty()) {
406         OccupiedSlotsMask |= 31;
407         return PopInst(AvailableAlus[AluDiscarded], false);
408       }
409       // If there is a T_XYZW alu available, use it
410       if (!AvailableAlus[AluT_XYZW].empty()) {
411         OccupiedSlotsMask |= 15;
412         return PopInst(AvailableAlus[AluT_XYZW], false);
413       }
414     }
415     bool TransSlotOccupied = OccupiedSlotsMask & 16;
416     if (!TransSlotOccupied && VLIW5) {
417       if (!AvailableAlus[AluTrans].empty()) {
418         OccupiedSlotsMask |= 16;
419         return PopInst(AvailableAlus[AluTrans], false);
420       }
421       SUnit *SU = AttemptFillSlot(3, true);
422       if (SU) {
423         OccupiedSlotsMask |= 16;
424         return SU;
425       }
426     }
427     for (int Chan = 3; Chan > -1; --Chan) {
428       bool isOccupied = OccupiedSlotsMask & (1 << Chan);
429       if (!isOccupied) {
430         SUnit *SU = AttemptFillSlot(Chan, false);
431         if (SU) {
432           OccupiedSlotsMask |= (1 << Chan);
433           InstructionsGroupCandidate.push_back(SU->getInstr());
434           return SU;
435         }
436       }
437     }
438     PrepareNextSlot();
439   }
440   return nullptr;
441 }
442 
pickOther(int QID)443 SUnit* R600SchedStrategy::pickOther(int QID) {
444   SUnit *SU = nullptr;
445   std::vector<SUnit *> &AQ = Available[QID];
446 
447   if (AQ.empty()) {
448     MoveUnits(Pending[QID], AQ);
449   }
450   if (!AQ.empty()) {
451     SU = AQ.back();
452     AQ.pop_back();
453   }
454   return SU;
455 }
456