R600MachineScheduler.cpp revision 98ce62780ea7185ba710868bf83c8077e8d7f6d6 
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/// \brief R600 Machine Scheduler interface
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "misched"
16
17#include "R600MachineScheduler.h"
18#include "llvm/CodeGen/LiveIntervalAnalysis.h"
19#include "llvm/CodeGen/MachineRegisterInfo.h"
20#include "llvm/Pass.h"
21#include "llvm/PassManager.h"
22#include "llvm/Support/raw_ostream.h"
23
24using namespace llvm;
25
26void R600SchedStrategy::initialize(ScheduleDAGMI *dag) {
27
28  DAG = dag;
29  TII = static_cast<const R600InstrInfo*>(DAG->TII);
30  TRI = static_cast<const R600RegisterInfo*>(DAG->TRI);
31  VLIW5 = !DAG->MF.getTarget().getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
32  MRI = &DAG->MRI;
33  CurInstKind = IDOther;
34  CurEmitted = 0;
35  OccupedSlotsMask = 31;
36  InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
37  InstKindLimit[IDOther] = 32;
38
39  const AMDGPUSubtarget &ST = DAG->TM.getSubtarget<AMDGPUSubtarget>();
40  InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
41  AluInstCount = 0;
42  FetchInstCount = 0;
43}
44
45void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
46                                  std::vector<SUnit *> &QDst)
47{
48  QDst.insert(QDst.end(), QSrc.begin(), QSrc.end());
49  QSrc.clear();
50}
51
52static
53unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
54  assert (GPRCount && "GPRCount cannot be 0");
55  return 248 / GPRCount;
56}
57
58SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
59  SUnit *SU = 0;
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    unsigned NeededWF = 62.5f / ALUFetchRationEstimate;
79    DEBUG( dbgs() << NeededWF << " approx. Wavefronts Required\n" );
80    // We assume the local GPR requirements to be "dominated" by the requirement
81    // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
82    // after TEX are indeed likely to consume or generate values from/for the
83    // TEX clause.
84    // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
85    // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
86    // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
87    // (TODO : use RegisterPressure)
88    // If we are going too use too many GPR, we flush Fetch instruction to lower
89    // register pressure on 128 bits regs.
90    unsigned NearRegisterRequirement = 2 * Available[IDFetch].size();
91    if (NeededWF > getWFCountLimitedByGPR(NearRegisterRequirement))
92      AllowSwitchFromAlu = true;
93  }
94
95
96  // We want to scheduled AR defs as soon as possible to make sure they aren't
97  // put in a different ALU clause from their uses.
98  if (!SU && !UnscheduledARDefs.empty()) {
99      SU = UnscheduledARDefs[0];
100      UnscheduledARDefs.erase(UnscheduledARDefs.begin());
101      NextInstKind = IDAlu;
102  }
103
104  if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
105      (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
106    // try to pick ALU
107    SU = pickAlu();
108    if (!SU && !PhysicalRegCopy.empty()) {
109      SU = PhysicalRegCopy.front();
110      PhysicalRegCopy.erase(PhysicalRegCopy.begin());
111    }
112    if (SU) {
113      if (CurEmitted >= InstKindLimit[IDAlu])
114        CurEmitted = 0;
115      NextInstKind = IDAlu;
116    }
117  }
118
119  if (!SU) {
120    // try to pick FETCH
121    SU = pickOther(IDFetch);
122    if (SU)
123      NextInstKind = IDFetch;
124  }
125
126  // try to pick other
127  if (!SU) {
128    SU = pickOther(IDOther);
129    if (SU)
130      NextInstKind = IDOther;
131  }
132
133  // We want to schedule the AR uses as late as possible to make sure that
134  // the AR defs have been released.
135  if (!SU && !UnscheduledARUses.empty()) {
136      SU = UnscheduledARUses[0];
137      UnscheduledARUses.erase(UnscheduledARUses.begin());
138      NextInstKind = IDAlu;
139  }
140
141
142  DEBUG(
143      if (SU) {
144        dbgs() << " ** Pick node **\n";
145        SU->dump(DAG);
146      } else {
147        dbgs() << "NO NODE \n";
148        for (unsigned i = 0; i < DAG->SUnits.size(); i++) {
149          const SUnit &S = DAG->SUnits[i];
150          if (!S.isScheduled)
151            S.dump(DAG);
152        }
153      }
154  );
155
156  return SU;
157}
158
159void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
160  if (NextInstKind != CurInstKind) {
161    DEBUG(dbgs() << "Instruction Type Switch\n");
162    if (NextInstKind != IDAlu)
163      OccupedSlotsMask |= 31;
164    CurEmitted = 0;
165    CurInstKind = NextInstKind;
166  }
167
168  if (CurInstKind == IDAlu) {
169    AluInstCount ++;
170    switch (getAluKind(SU)) {
171    case AluT_XYZW:
172      CurEmitted += 4;
173      break;
174    case AluDiscarded:
175      break;
176    default: {
177      ++CurEmitted;
178      for (MachineInstr::mop_iterator It = SU->getInstr()->operands_begin(),
179          E = SU->getInstr()->operands_end(); It != E; ++It) {
180        MachineOperand &MO = *It;
181        if (MO.isReg() && MO.getReg() == AMDGPU::ALU_LITERAL_X)
182          ++CurEmitted;
183      }
184    }
185    }
186  } else {
187    ++CurEmitted;
188  }
189
190
191  DEBUG(dbgs() << CurEmitted << " Instructions Emitted in this clause\n");
192
193  if (CurInstKind != IDFetch) {
194    MoveUnits(Pending[IDFetch], Available[IDFetch]);
195  } else
196    FetchInstCount++;
197}
198
199static bool
200isPhysicalRegCopy(MachineInstr *MI) {
201  if (MI->getOpcode() != AMDGPU::COPY)
202    return false;
203
204  return !TargetRegisterInfo::isVirtualRegister(MI->getOperand(1).getReg());
205}
206
207void R600SchedStrategy::releaseTopNode(SUnit *SU) {
208  DEBUG(dbgs() << "Top Releasing ";SU->dump(DAG););
209}
210
211void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
212  DEBUG(dbgs() << "Bottom Releasing ";SU->dump(DAG););
213  if (isPhysicalRegCopy(SU->getInstr())) {
214    PhysicalRegCopy.push_back(SU);
215    return;
216  }
217
218  int IK = getInstKind(SU);
219
220  // Check for AR register defines
221  for (MachineInstr::const_mop_iterator I = SU->getInstr()->operands_begin(),
222                                        E = SU->getInstr()->operands_end();
223                                        I != E; ++I) {
224    if (I->isReg() && I->getReg() == AMDGPU::AR_X) {
225      if (I->isDef()) {
226        UnscheduledARDefs.push_back(SU);
227      } else {
228        UnscheduledARUses.push_back(SU);
229      }
230      return;
231    }
232  }
233
234  // There is no export clause, we can schedule one as soon as its ready
235  if (IK == IDOther)
236    Available[IDOther].push_back(SU);
237  else
238    Pending[IK].push_back(SU);
239
240}
241
242bool R600SchedStrategy::regBelongsToClass(unsigned Reg,
243                                          const TargetRegisterClass *RC) const {
244  if (!TargetRegisterInfo::isVirtualRegister(Reg)) {
245    return RC->contains(Reg);
246  } else {
247    return MRI->getRegClass(Reg) == RC;
248  }
249}
250
251R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
252  MachineInstr *MI = SU->getInstr();
253
254  if (TII->isTransOnly(MI))
255    return AluTrans;
256
257    switch (MI->getOpcode()) {
258    case AMDGPU::PRED_X:
259      return AluPredX;
260    case AMDGPU::INTERP_PAIR_XY:
261    case AMDGPU::INTERP_PAIR_ZW:
262    case AMDGPU::INTERP_VEC_LOAD:
263    case AMDGPU::DOT_4:
264      return AluT_XYZW;
265    case AMDGPU::COPY:
266      if (MI->getOperand(1).isUndef()) {
267        // MI will become a KILL, don't considers it in scheduling
268        return AluDiscarded;
269      }
270    default:
271      break;
272    }
273
274    // Does the instruction take a whole IG ?
275    // XXX: Is it possible to add a helper function in R600InstrInfo that can
276    // be used here and in R600PacketizerList::isSoloInstruction() ?
277    if(TII->isVector(*MI) ||
278        TII->isCubeOp(MI->getOpcode()) ||
279        TII->isReductionOp(MI->getOpcode()) ||
280        MI->getOpcode() == AMDGPU::GROUP_BARRIER) {
281      return AluT_XYZW;
282    }
283
284    if (TII->isLDSInstr(MI->getOpcode())) {
285      return AluT_X;
286    }
287
288    // Is the result already assigned to a channel ?
289    unsigned DestSubReg = MI->getOperand(0).getSubReg();
290    switch (DestSubReg) {
291    case AMDGPU::sub0:
292      return AluT_X;
293    case AMDGPU::sub1:
294      return AluT_Y;
295    case AMDGPU::sub2:
296      return AluT_Z;
297    case AMDGPU::sub3:
298      return AluT_W;
299    default:
300      break;
301    }
302
303    // Is the result already member of a X/Y/Z/W class ?
304    unsigned DestReg = MI->getOperand(0).getReg();
305    if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_XRegClass) ||
306        regBelongsToClass(DestReg, &AMDGPU::R600_AddrRegClass))
307      return AluT_X;
308    if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_YRegClass))
309      return AluT_Y;
310    if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_ZRegClass))
311      return AluT_Z;
312    if (regBelongsToClass(DestReg, &AMDGPU::R600_TReg32_WRegClass))
313      return AluT_W;
314    if (regBelongsToClass(DestReg, &AMDGPU::R600_Reg128RegClass))
315      return AluT_XYZW;
316
317    return AluAny;
318
319}
320
321int R600SchedStrategy::getInstKind(SUnit* SU) {
322  int Opcode = SU->getInstr()->getOpcode();
323
324  if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
325    return IDFetch;
326
327  if (TII->isALUInstr(Opcode)) {
328    return IDAlu;
329  }
330
331  switch (Opcode) {
332  case AMDGPU::PRED_X:
333  case AMDGPU::COPY:
334  case AMDGPU::CONST_COPY:
335  case AMDGPU::INTERP_PAIR_XY:
336  case AMDGPU::INTERP_PAIR_ZW:
337  case AMDGPU::INTERP_VEC_LOAD:
338  case AMDGPU::DOT_4:
339    return IDAlu;
340  default:
341    return IDOther;
342  }
343}
344
345SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
346  if (Q.empty())
347    return NULL;
348  for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
349      It != E; ++It) {
350    SUnit *SU = *It;
351    InstructionsGroupCandidate.push_back(SU->getInstr());
352    if (TII->fitsConstReadLimitations(InstructionsGroupCandidate)
353        && (!AnyALU || !TII->isVectorOnly(SU->getInstr()))
354    ) {
355      InstructionsGroupCandidate.pop_back();
356      Q.erase((It + 1).base());
357      return SU;
358    } else {
359      InstructionsGroupCandidate.pop_back();
360    }
361  }
362  return NULL;
363}
364
365void R600SchedStrategy::LoadAlu() {
366  std::vector<SUnit *> &QSrc = Pending[IDAlu];
367  for (unsigned i = 0, e = QSrc.size(); i < e; ++i) {
368    AluKind AK = getAluKind(QSrc[i]);
369    AvailableAlus[AK].push_back(QSrc[i]);
370  }
371  QSrc.clear();
372}
373
374void R600SchedStrategy::PrepareNextSlot() {
375  DEBUG(dbgs() << "New Slot\n");
376  assert (OccupedSlotsMask && "Slot wasn't filled");
377  OccupedSlotsMask = 0;
378//  if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
379//    OccupedSlotsMask |= 16;
380  InstructionsGroupCandidate.clear();
381  LoadAlu();
382}
383
384void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
385  int DstIndex = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
386  if (DstIndex == -1) {
387    return;
388  }
389  unsigned DestReg = MI->getOperand(DstIndex).getReg();
390  // PressureRegister crashes if an operand is def and used in the same inst
391  // and we try to constraint its regclass
392  for (MachineInstr::mop_iterator It = MI->operands_begin(),
393      E = MI->operands_end(); It != E; ++It) {
394    MachineOperand &MO = *It;
395    if (MO.isReg() && !MO.isDef() &&
396        MO.getReg() == DestReg)
397      return;
398  }
399  // Constrains the regclass of DestReg to assign it to Slot
400  switch (Slot) {
401  case 0:
402    MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_XRegClass);
403    break;
404  case 1:
405    MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_YRegClass);
406    break;
407  case 2:
408    MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_ZRegClass);
409    break;
410  case 3:
411    MRI->constrainRegClass(DestReg, &AMDGPU::R600_TReg32_WRegClass);
412    break;
413  }
414}
415
416SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
417  static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
418  SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
419  if (SlotedSU)
420    return SlotedSU;
421  SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
422  if (UnslotedSU)
423    AssignSlot(UnslotedSU->getInstr(), Slot);
424  return UnslotedSU;
425}
426
427unsigned R600SchedStrategy::AvailablesAluCount() const {
428  return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
429      AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
430      AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
431      AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
432      AvailableAlus[AluPredX].size();
433}
434
435SUnit* R600SchedStrategy::pickAlu() {
436  while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
437    if (!OccupedSlotsMask) {
438      // Bottom up scheduling : predX must comes first
439      if (!AvailableAlus[AluPredX].empty()) {
440        OccupedSlotsMask |= 31;
441        return PopInst(AvailableAlus[AluPredX], false);
442      }
443      // Flush physical reg copies (RA will discard them)
444      if (!AvailableAlus[AluDiscarded].empty()) {
445        OccupedSlotsMask |= 31;
446        return PopInst(AvailableAlus[AluDiscarded], false);
447      }
448      // If there is a T_XYZW alu available, use it
449      if (!AvailableAlus[AluT_XYZW].empty()) {
450        OccupedSlotsMask |= 15;
451        return PopInst(AvailableAlus[AluT_XYZW], false);
452      }
453    }
454    bool TransSlotOccuped = OccupedSlotsMask & 16;
455    if (!TransSlotOccuped && VLIW5) {
456      if (!AvailableAlus[AluTrans].empty()) {
457        OccupedSlotsMask |= 16;
458        return PopInst(AvailableAlus[AluTrans], false);
459      }
460      SUnit *SU = AttemptFillSlot(3, true);
461      if (SU) {
462        OccupedSlotsMask |= 16;
463        return SU;
464      }
465    }
466    for (int Chan = 3; Chan > -1; --Chan) {
467      bool isOccupied = OccupedSlotsMask & (1 << Chan);
468      if (!isOccupied) {
469        SUnit *SU = AttemptFillSlot(Chan, false);
470        if (SU) {
471          OccupedSlotsMask |= (1 << Chan);
472          InstructionsGroupCandidate.push_back(SU->getInstr());
473          return SU;
474        }
475      }
476    }
477    PrepareNextSlot();
478  }
479  return NULL;
480}
481
482SUnit* R600SchedStrategy::pickOther(int QID) {
483  SUnit *SU = 0;
484  std::vector<SUnit *> &AQ = Available[QID];
485
486  if (AQ.empty()) {
487    MoveUnits(Pending[QID], AQ);
488  }
489  if (!AQ.empty()) {
490    SU = AQ.back();
491    AQ.resize(AQ.size() - 1);
492  }
493  return SU;
494}
495
496