xref: /llvm-project/llvm/lib/Target/AMDGPU/R600MachineScheduler.cpp (revision 44b30b453743e95d79ba69a7b9155e23ed4595e5) 
1 //===-- R600MachineScheduler.cpp - R600 Scheduler Interface -*- C++ -*-----===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// R600 Machine Scheduler interface
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "R600MachineScheduler.h"
16 #include "AMDGPUSubtarget.h"
17 #include "R600InstrInfo.h"
18 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/IR/LegacyPassManager.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Support/raw_ostream.h"
23 
24 using namespace llvm;
25 
26 #define DEBUG_TYPE "machine-scheduler"
27 
28 void R600SchedStrategy::initialize(ScheduleDAGMI *dag) {
29   assert(dag->hasVRegLiveness() && "R600SchedStrategy needs vreg liveness");
30   DAG = static_cast<ScheduleDAGMILive*>(dag);
31   const R600Subtarget &ST = DAG->MF.getSubtarget<R600Subtarget>();
32   TII = static_cast<const R600InstrInfo*>(DAG->TII);
33   TRI = static_cast<const R600RegisterInfo*>(DAG->TRI);
34   VLIW5 = !ST.hasCaymanISA();
35   MRI = &DAG->MRI;
36   CurInstKind = IDOther;
37   CurEmitted = 0;
38   OccupedSlotsMask = 31;
39   InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
40   InstKindLimit[IDOther] = 32;
41   InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
42   AluInstCount = 0;
43   FetchInstCount = 0;
44 }
45 
46 void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
47                                   std::vector<SUnit *> &QDst)
48 {
49   QDst.insert(QDst.end(), QSrc.begin(), QSrc.end());
50   QSrc.clear();
51 }
52 
53 static unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
54   assert (GPRCount && "GPRCount cannot be 0");
55   return 248 / GPRCount;
56 }
57 
58 SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
59   SUnit *SU = nullptr;
60   NextInstKind = IDOther;
61 
62   IsTopNode = false;
63 
64   // check if we might want to switch current clause type
65   bool AllowSwitchToAlu = (CurEmitted >= InstKindLimit[CurInstKind]) ||
66       (Available[CurInstKind].empty());
67   bool AllowSwitchFromAlu = (CurEmitted >= InstKindLimit[CurInstKind]) &&
68       (!Available[IDFetch].empty() || !Available[IDOther].empty());
69 
70   if (CurInstKind == IDAlu && !Available[IDFetch].empty()) {
71     // We use the heuristic provided by AMD Accelerated Parallel Processing
72     // OpenCL Programming Guide :
73     // The approx. number of WF that allows TEX inst to hide ALU inst is :
74     // 500 (cycles for TEX) / (AluFetchRatio * 8 (cycles for ALU))
75     float ALUFetchRationEstimate =
76         (AluInstCount + AvailablesAluCount() + Pending[IDAlu].size()) /
77         (FetchInstCount + Available[IDFetch].size());
78     if (ALUFetchRationEstimate == 0) {
79       AllowSwitchFromAlu = true;
80     } else {
81       unsigned NeededWF = 62.5f / ALUFetchRationEstimate;
82       LLVM_DEBUG(dbgs() << NeededWF << " approx. Wavefronts Required\n");
83       // We assume the local GPR requirements to be "dominated" by the requirement
84       // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
85       // after TEX are indeed likely to consume or generate values from/for the
86       // TEX clause.
87       // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
88       // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
89       // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
90       // (TODO : use RegisterPressure)
91       // If we are going too use too many GPR, we flush Fetch instruction to lower
92       // register pressure on 128 bits regs.
93       unsigned NearRegisterRequirement = 2 * Available[IDFetch].size();
94       if (NeededWF > getWFCountLimitedByGPR(NearRegisterRequirement))
95         AllowSwitchFromAlu = true;
96     }
97   }
98 
99   if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
100       (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
101     // try to pick ALU
102     SU = pickAlu();
103     if (!SU && !PhysicalRegCopy.empty()) {
104       SU = PhysicalRegCopy.front();
105       PhysicalRegCopy.erase(PhysicalRegCopy.begin());
106     }
107     if (SU) {
108       if (CurEmitted >= InstKindLimit[IDAlu])
109         CurEmitted = 0;
110       NextInstKind = IDAlu;
111     }
112   }
113 
114   if (!SU) {
115     // try to pick FETCH
116     SU = pickOther(IDFetch);
117     if (SU)
118       NextInstKind = IDFetch;
119   }
120 
121   // try to pick other
122   if (!SU) {
123     SU = pickOther(IDOther);
124     if (SU)
125       NextInstKind = IDOther;
126   }
127 
128   LLVM_DEBUG(if (SU) {
129     dbgs() << " ** Pick node **\n";
130     SU->dump(DAG);
131   } else {
132     dbgs() << "NO NODE \n";
133     for (unsigned i = 0; i < DAG->SUnits.size(); i++) {
134       const SUnit &S = DAG->SUnits[i];
135       if (!S.isScheduled)
136         S.dump(DAG);
137     }
138   });
139 
140   return SU;
141 }
142 
143 void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
144   if (NextInstKind != CurInstKind) {
145     LLVM_DEBUG(dbgs() << "Instruction Type Switch\n");
146     if (NextInstKind != IDAlu)
147       OccupedSlotsMask |= 31;
148     CurEmitted = 0;
149     CurInstKind = NextInstKind;
150   }
151 
152   if (CurInstKind == IDAlu) {
153     AluInstCount ++;
154     switch (getAluKind(SU)) {
155     case AluT_XYZW:
156       CurEmitted += 4;
157       break;
158     case AluDiscarded:
159       break;
160     default: {
161       ++CurEmitted;
162       for (MachineInstr::mop_iterator It = SU->getInstr()->operands_begin(),
163           E = SU->getInstr()->operands_end(); It != E; ++It) {
164         MachineOperand &MO = *It;
165         if (MO.isReg() && MO.getReg() == AMDGPU::ALU_LITERAL_X)
166           ++CurEmitted;
167       }
168     }
169     }
170   } else {
171     ++CurEmitted;
172   }
173 
174   LLVM_DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
175 
176   if (CurInstKind != IDFetch) {
177     MoveUnits(Pending[IDFetch], Available[IDFetch]);
178   } else
179     FetchInstCount++;
180 }
181 
182 static bool
183 isPhysicalRegCopy(MachineInstr *MI) {
184   if (MI->getOpcode() != AMDGPU::COPY)
185     return false;
186 
187   return !TargetRegisterInfo::isVirtualRegister(MI->getOperand(1).getReg());
188 }
189 
190 void R600SchedStrategy::releaseTopNode(SUnit *SU) {
191   LLVM_DEBUG(dbgs() << "Top Releasing "; SU->dump(DAG););
192 }
193 
194 void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
195   LLVM_DEBUG(dbgs() << "Bottom Releasing "; SU->dump(DAG););
196   if (isPhysicalRegCopy(SU->getInstr())) {
197     PhysicalRegCopy.push_back(SU);
198     return;
199   }
200 
201   int IK = getInstKind(SU);
202 
203   // There is no export clause, we can schedule one as soon as its ready
204   if (IK == IDOther)
205     Available[IDOther].push_back(SU);
206   else
207     Pending[IK].push_back(SU);
208 
209 }
210 
211 bool R600SchedStrategy::regBelongsToClass(unsigned Reg,
212                                           const TargetRegisterClass *RC) const {
213   if (!TargetRegisterInfo::isVirtualRegister(Reg)) {
214     return RC->contains(Reg);
215   } else {
216     return MRI->getRegClass(Reg) == RC;
217   }
218 }
219 
220 R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
221   MachineInstr *MI = SU->getInstr();
222 
223   if (TII->isTransOnly(*MI))
224     return AluTrans;
225 
226   switch (MI->getOpcode()) {
227   case AMDGPU::PRED_X:
228     return AluPredX;
229   case AMDGPU::INTERP_PAIR_XY:
230   case AMDGPU::INTERP_PAIR_ZW:
231   case AMDGPU::INTERP_VEC_LOAD:
232   case AMDGPU::DOT_4:
233     return AluT_XYZW;
234   case AMDGPU::COPY:
235     if (MI->getOperand(1).isUndef()) {
236       // MI will become a KILL, don't considers it in scheduling
237       return AluDiscarded;
238     }
239   default:
240     break;
241   }
242 
243   // Does the instruction take a whole IG ?
244   // XXX: Is it possible to add a helper function in R600InstrInfo that can
245   // be used here and in R600PacketizerList::isSoloInstruction() ?
246   if(TII->isVector(*MI) ||
247      TII->isCubeOp(MI->getOpcode()) ||
248      TII->isReductionOp(MI->getOpcode()) ||
249      MI->getOpcode() == AMDGPU::GROUP_BARRIER) {
250     return AluT_XYZW;
251   }
252 
253   if (TII->isLDSInstr(MI->getOpcode())) {
254     return AluT_X;
255   }
256 
257   // Is the result already assigned to a channel ?
258   unsigned DestSubReg = MI->getOperand(0).getSubReg();
259   switch (DestSubReg) {
260   case AMDGPU::sub0:
261     return AluT_X;
262   case AMDGPU::sub1:
263     return AluT_Y;
264   case AMDGPU::sub2:
265     return AluT_Z;
266   case AMDGPU::sub3:
267     return AluT_W;
268   default:
269     break;
270   }
271 
272   // Is the result already member of a X/Y/Z/W class ?
273   unsigned DestReg = MI->getOperand(0).getReg();
274   if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_XRegClass) ||
275       regBelongsToClass(DestReg, &AMDGPU::R600_AddrRegClass))
276     return AluT_X;
277   if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_YRegClass))
278     return AluT_Y;
279   if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_ZRegClass))
280     return AluT_Z;
281   if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_WRegClass))
282     return AluT_W;
283   if (regBelongsToClass(DestReg, &AMDGPU::R600_Reg128RegClass))
284     return AluT_XYZW;
285 
286   // LDS src registers cannot be used in the Trans slot.
287   if (TII->readsLDSSrcReg(*MI))
288     return AluT_XYZW;
289 
290   return AluAny;
291 }
292 
293 int R600SchedStrategy::getInstKind(SUnit* SU) {
294   int Opcode = SU->getInstr()->getOpcode();
295 
296   if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
297     return IDFetch;
298 
299   if (TII->isALUInstr(Opcode)) {
300     return IDAlu;
301   }
302 
303   switch (Opcode) {
304   case AMDGPU::PRED_X:
305   case AMDGPU::COPY:
306   case AMDGPU::CONST_COPY:
307   case AMDGPU::INTERP_PAIR_XY:
308   case AMDGPU::INTERP_PAIR_ZW:
309   case AMDGPU::INTERP_VEC_LOAD:
310   case AMDGPU::DOT_4:
311     return IDAlu;
312   default:
313     return IDOther;
314   }
315 }
316 
317 SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
318   if (Q.empty())
319     return nullptr;
320   for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
321       It != E; ++It) {
322     SUnit *SU = *It;
323     InstructionsGroupCandidate.push_back(SU->getInstr());
324     if (TII->fitsConstReadLimitations(InstructionsGroupCandidate) &&
325         (!AnyALU || !TII->isVectorOnly(*SU->getInstr()))) {
326       InstructionsGroupCandidate.pop_back();
327       Q.erase((It + 1).base());
328       return SU;
329     } else {
330       InstructionsGroupCandidate.pop_back();
331     }
332   }
333   return nullptr;
334 }
335 
336 void R600SchedStrategy::LoadAlu() {
337   std::vector<SUnit *> &QSrc = Pending[IDAlu];
338   for (unsigned i = 0, e = QSrc.size(); i < e; ++i) {
339     AluKind AK = getAluKind(QSrc[i]);
340     AvailableAlus[AK].push_back(QSrc[i]);
341   }
342   QSrc.clear();
343 }
344 
345 void R600SchedStrategy::PrepareNextSlot() {
346   LLVM_DEBUG(dbgs() << "New Slot\n");
347   assert (OccupedSlotsMask && "Slot wasn't filled");
348   OccupedSlotsMask = 0;
349 //  if (HwGen == R600Subtarget::NORTHERN_ISLANDS)
350 //    OccupedSlotsMask |= 16;
351   InstructionsGroupCandidate.clear();
352   LoadAlu();
353 }
354 
355 void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
356   int DstIndex = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
357   if (DstIndex == -1) {
358     return;
359   }
360   unsigned DestReg = MI->getOperand(DstIndex).getReg();
361   // PressureRegister crashes if an operand is def and used in the same inst
362   // and we try to constraint its regclass
363   for (MachineInstr::mop_iterator It = MI->operands_begin(),
364       E = MI->operands_end(); It != E; ++It) {
365     MachineOperand &MO = *It;
366     if (MO.isReg() && !MO.isDef() &&
367         MO.getReg() == DestReg)
368       return;
369   }
370   // Constrains the regclass of DestReg to assign it to Slot
371   switch (Slot) {
372   case 0:
373     MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_XRegClass);
374     break;
375   case 1:
376     MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_YRegClass);
377     break;
378   case 2:
379     MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_ZRegClass);
380     break;
381   case 3:
382     MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_WRegClass);
383     break;
384   }
385 }
386 
387 SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
388   static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
389   SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
390   if (SlotedSU)
391     return SlotedSU;
392   SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
393   if (UnslotedSU)
394     AssignSlot(UnslotedSU->getInstr(), Slot);
395   return UnslotedSU;
396 }
397 
398 unsigned R600SchedStrategy::AvailablesAluCount() const {
399   return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
400       AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
401       AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
402       AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
403       AvailableAlus[AluPredX].size();
404 }
405 
406 SUnit* R600SchedStrategy::pickAlu() {
407   while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
408     if (!OccupedSlotsMask) {
409       // Bottom up scheduling : predX must comes first
410       if (!AvailableAlus[AluPredX].empty()) {
411         OccupedSlotsMask |= 31;
412         return PopInst(AvailableAlus[AluPredX], false);
413       }
414       // Flush physical reg copies (RA will discard them)
415       if (!AvailableAlus[AluDiscarded].empty()) {
416         OccupedSlotsMask |= 31;
417         return PopInst(AvailableAlus[AluDiscarded], false);
418       }
419       // If there is a T_XYZW alu available, use it
420       if (!AvailableAlus[AluT_XYZW].empty()) {
421         OccupedSlotsMask |= 15;
422         return PopInst(AvailableAlus[AluT_XYZW], false);
423       }
424     }
425     bool TransSlotOccuped = OccupedSlotsMask & 16;
426     if (!TransSlotOccuped && VLIW5) {
427       if (!AvailableAlus[AluTrans].empty()) {
428         OccupedSlotsMask |= 16;
429         return PopInst(AvailableAlus[AluTrans], false);
430       }
431       SUnit *SU = AttemptFillSlot(3, true);
432       if (SU) {
433         OccupedSlotsMask |= 16;
434         return SU;
435       }
436     }
437     for (int Chan = 3; Chan > -1; --Chan) {
438       bool isOccupied = OccupedSlotsMask & (1 << Chan);
439       if (!isOccupied) {
440         SUnit *SU = AttemptFillSlot(Chan, false);
441         if (SU) {
442           OccupedSlotsMask |= (1 << Chan);
443           InstructionsGroupCandidate.push_back(SU->getInstr());
444           return SU;
445         }
446       }
447     }
448     PrepareNextSlot();
449   }
450   return nullptr;
451 }
452 
453 SUnit* R600SchedStrategy::pickOther(int QID) {
454   SUnit *SU = nullptr;
455   std::vector<SUnit *> &AQ = Available[QID];
456 
457   if (AQ.empty()) {
458     MoveUnits(Pending[QID], AQ);
459   }
460   if (!AQ.empty()) {
461     SU = AQ.back();
462     AQ.resize(AQ.size() - 1);
463   }
464   return SU;
465 }
466