xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/R600MachineScheduler.cpp (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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/AMDGPUMCTargetDesc.h"
16 #include "R600Subtarget.h"
17 
18 using namespace llvm;
19 
20 #define DEBUG_TYPE "machine-scheduler"
21 
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   OccupedSlotsMask = 31;
33   InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
34   InstKindLimit[IDOther] = 32;
35   InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
36   AluInstCount = 0;
37   FetchInstCount = 0;
38 }
39 
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 
47 static unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
48   assert (GPRCount && "GPRCount cannot be 0");
49   return 248 / GPRCount;
50 }
51 
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 (unsigned i = 0; i < DAG->SUnits.size(); i++) {
128       const SUnit &S = DAG->SUnits[i];
129       if (!S.isScheduled)
130         DAG->dumpNode(S);
131     }
132   });
133 
134   return SU;
135 }
136 
137 void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
138   if (NextInstKind != CurInstKind) {
139     LLVM_DEBUG(dbgs() << "Instruction Type Switch\n");
140     if (NextInstKind != IDAlu)
141       OccupedSlotsMask |= 31;
142     CurEmitted = 0;
143     CurInstKind = NextInstKind;
144   }
145 
146   if (CurInstKind == IDAlu) {
147     AluInstCount ++;
148     switch (getAluKind(SU)) {
149     case AluT_XYZW:
150       CurEmitted += 4;
151       break;
152     case AluDiscarded:
153       break;
154     default: {
155       ++CurEmitted;
156       for (MachineInstr::mop_iterator It = SU->getInstr()->operands_begin(),
157           E = SU->getInstr()->operands_end(); It != E; ++It) {
158         MachineOperand &MO = *It;
159         if (MO.isReg() && MO.getReg() == R600::ALU_LITERAL_X)
160           ++CurEmitted;
161       }
162     }
163     }
164   } else {
165     ++CurEmitted;
166   }
167 
168   LLVM_DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
169 
170   if (CurInstKind != IDFetch) {
171     MoveUnits(Pending[IDFetch], Available[IDFetch]);
172   } else
173     FetchInstCount++;
174 }
175 
176 static bool
177 isPhysicalRegCopy(MachineInstr *MI) {
178   if (MI->getOpcode() != R600::COPY)
179     return false;
180 
181   return !MI->getOperand(1).getReg().isVirtual();
182 }
183 
184 void R600SchedStrategy::releaseTopNode(SUnit *SU) {
185   LLVM_DEBUG(dbgs() << "Top Releasing "; DAG->dumpNode(*SU));
186 }
187 
188 void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
189   LLVM_DEBUG(dbgs() << "Bottom Releasing "; DAG->dumpNode(*SU));
190   if (isPhysicalRegCopy(SU->getInstr())) {
191     PhysicalRegCopy.push_back(SU);
192     return;
193   }
194 
195   int IK = getInstKind(SU);
196 
197   // There is no export clause, we can schedule one as soon as its ready
198   if (IK == IDOther)
199     Available[IDOther].push_back(SU);
200   else
201     Pending[IK].push_back(SU);
202 
203 }
204 
205 bool R600SchedStrategy::regBelongsToClass(Register Reg,
206                                           const TargetRegisterClass *RC) const {
207   if (!Reg.isVirtual()) {
208     return RC->contains(Reg);
209   } else {
210     return MRI->getRegClass(Reg) == RC;
211   }
212 }
213 
214 R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
215   MachineInstr *MI = SU->getInstr();
216 
217   if (TII->isTransOnly(*MI))
218     return AluTrans;
219 
220   switch (MI->getOpcode()) {
221   case R600::PRED_X:
222     return AluPredX;
223   case R600::INTERP_PAIR_XY:
224   case R600::INTERP_PAIR_ZW:
225   case R600::INTERP_VEC_LOAD:
226   case R600::DOT_4:
227     return AluT_XYZW;
228   case R600::COPY:
229     if (MI->getOperand(1).isUndef()) {
230       // MI will become a KILL, don't considers it in scheduling
231       return AluDiscarded;
232     }
233     break;
234   default:
235     break;
236   }
237 
238   // Does the instruction take a whole IG ?
239   // XXX: Is it possible to add a helper function in R600InstrInfo that can
240   // be used here and in R600PacketizerList::isSoloInstruction() ?
241   if(TII->isVector(*MI) ||
242      TII->isCubeOp(MI->getOpcode()) ||
243      TII->isReductionOp(MI->getOpcode()) ||
244      MI->getOpcode() == R600::GROUP_BARRIER) {
245     return AluT_XYZW;
246   }
247 
248   if (TII->isLDSInstr(MI->getOpcode())) {
249     return AluT_X;
250   }
251 
252   // Is the result already assigned to a channel ?
253   unsigned DestSubReg = MI->getOperand(0).getSubReg();
254   switch (DestSubReg) {
255   case R600::sub0:
256     return AluT_X;
257   case R600::sub1:
258     return AluT_Y;
259   case R600::sub2:
260     return AluT_Z;
261   case R600::sub3:
262     return AluT_W;
263   default:
264     break;
265   }
266 
267   // Is the result already member of a X/Y/Z/W class ?
268   Register DestReg = MI->getOperand(0).getReg();
269   if (regBelongsToClass(DestReg, &R600::R600_TReg32_XRegClass) ||
270       regBelongsToClass(DestReg, &R600::R600_AddrRegClass))
271     return AluT_X;
272   if (regBelongsToClass(DestReg, &R600::R600_TReg32_YRegClass))
273     return AluT_Y;
274   if (regBelongsToClass(DestReg, &R600::R600_TReg32_ZRegClass))
275     return AluT_Z;
276   if (regBelongsToClass(DestReg, &R600::R600_TReg32_WRegClass))
277     return AluT_W;
278   if (regBelongsToClass(DestReg, &R600::R600_Reg128RegClass))
279     return AluT_XYZW;
280 
281   // LDS src registers cannot be used in the Trans slot.
282   if (TII->readsLDSSrcReg(*MI))
283     return AluT_XYZW;
284 
285   return AluAny;
286 }
287 
288 int R600SchedStrategy::getInstKind(SUnit* SU) {
289   int Opcode = SU->getInstr()->getOpcode();
290 
291   if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
292     return IDFetch;
293 
294   if (TII->isALUInstr(Opcode)) {
295     return IDAlu;
296   }
297 
298   switch (Opcode) {
299   case R600::PRED_X:
300   case R600::COPY:
301   case R600::CONST_COPY:
302   case R600::INTERP_PAIR_XY:
303   case R600::INTERP_PAIR_ZW:
304   case R600::INTERP_VEC_LOAD:
305   case R600::DOT_4:
306     return IDAlu;
307   default:
308     return IDOther;
309   }
310 }
311 
312 SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
313   if (Q.empty())
314     return nullptr;
315   for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
316       It != E; ++It) {
317     SUnit *SU = *It;
318     InstructionsGroupCandidate.push_back(SU->getInstr());
319     if (TII->fitsConstReadLimitations(InstructionsGroupCandidate) &&
320         (!AnyALU || !TII->isVectorOnly(*SU->getInstr()))) {
321       InstructionsGroupCandidate.pop_back();
322       Q.erase((It + 1).base());
323       return SU;
324     } else {
325       InstructionsGroupCandidate.pop_back();
326     }
327   }
328   return nullptr;
329 }
330 
331 void R600SchedStrategy::LoadAlu() {
332   std::vector<SUnit *> &QSrc = Pending[IDAlu];
333   for (unsigned i = 0, e = QSrc.size(); i < e; ++i) {
334     AluKind AK = getAluKind(QSrc[i]);
335     AvailableAlus[AK].push_back(QSrc[i]);
336   }
337   QSrc.clear();
338 }
339 
340 void R600SchedStrategy::PrepareNextSlot() {
341   LLVM_DEBUG(dbgs() << "New Slot\n");
342   assert (OccupedSlotsMask && "Slot wasn't filled");
343   OccupedSlotsMask = 0;
344 //  if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
345 //    OccupedSlotsMask |= 16;
346   InstructionsGroupCandidate.clear();
347   LoadAlu();
348 }
349 
350 void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
351   int DstIndex = TII->getOperandIdx(MI->getOpcode(), R600::OpName::dst);
352   if (DstIndex == -1) {
353     return;
354   }
355   Register DestReg = MI->getOperand(DstIndex).getReg();
356   // PressureRegister crashes if an operand is def and used in the same inst
357   // and we try to constraint its regclass
358   for (MachineInstr::mop_iterator It = MI->operands_begin(),
359       E = MI->operands_end(); It != E; ++It) {
360     MachineOperand &MO = *It;
361     if (MO.isReg() && !MO.isDef() &&
362         MO.getReg() == DestReg)
363       return;
364   }
365   // Constrains the regclass of DestReg to assign it to Slot
366   switch (Slot) {
367   case 0:
368     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_XRegClass);
369     break;
370   case 1:
371     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_YRegClass);
372     break;
373   case 2:
374     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_ZRegClass);
375     break;
376   case 3:
377     MRI->constrainRegClass(DestReg, &R600::R600_TReg32_WRegClass);
378     break;
379   }
380 }
381 
382 SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
383   static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
384   SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
385   if (SlotedSU)
386     return SlotedSU;
387   SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
388   if (UnslotedSU)
389     AssignSlot(UnslotedSU->getInstr(), Slot);
390   return UnslotedSU;
391 }
392 
393 unsigned R600SchedStrategy::AvailablesAluCount() const {
394   return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
395       AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
396       AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
397       AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
398       AvailableAlus[AluPredX].size();
399 }
400 
401 SUnit* R600SchedStrategy::pickAlu() {
402   while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
403     if (!OccupedSlotsMask) {
404       // Bottom up scheduling : predX must comes first
405       if (!AvailableAlus[AluPredX].empty()) {
406         OccupedSlotsMask |= 31;
407         return PopInst(AvailableAlus[AluPredX], false);
408       }
409       // Flush physical reg copies (RA will discard them)
410       if (!AvailableAlus[AluDiscarded].empty()) {
411         OccupedSlotsMask |= 31;
412         return PopInst(AvailableAlus[AluDiscarded], false);
413       }
414       // If there is a T_XYZW alu available, use it
415       if (!AvailableAlus[AluT_XYZW].empty()) {
416         OccupedSlotsMask |= 15;
417         return PopInst(AvailableAlus[AluT_XYZW], false);
418       }
419     }
420     bool TransSlotOccuped = OccupedSlotsMask & 16;
421     if (!TransSlotOccuped && VLIW5) {
422       if (!AvailableAlus[AluTrans].empty()) {
423         OccupedSlotsMask |= 16;
424         return PopInst(AvailableAlus[AluTrans], false);
425       }
426       SUnit *SU = AttemptFillSlot(3, true);
427       if (SU) {
428         OccupedSlotsMask |= 16;
429         return SU;
430       }
431     }
432     for (int Chan = 3; Chan > -1; --Chan) {
433       bool isOccupied = OccupedSlotsMask & (1 << Chan);
434       if (!isOccupied) {
435         SUnit *SU = AttemptFillSlot(Chan, false);
436         if (SU) {
437           OccupedSlotsMask |= (1 << Chan);
438           InstructionsGroupCandidate.push_back(SU->getInstr());
439           return SU;
440         }
441       }
442     }
443     PrepareNextSlot();
444   }
445   return nullptr;
446 }
447 
448 SUnit* R600SchedStrategy::pickOther(int QID) {
449   SUnit *SU = nullptr;
450   std::vector<SUnit *> &AQ = Available[QID];
451 
452   if (AQ.empty()) {
453     MoveUnits(Pending[QID], AQ);
454   }
455   if (!AQ.empty()) {
456     SU = AQ.back();
457     AQ.pop_back();
458   }
459   return SU;
460 }
461