llvm.org GIT mirror llvm / release_60 lib / CodeGen / TargetSchedule.cpp
release_60

Tree @release_60 (Download .tar.gz)

TargetSchedule.cpp @release_60raw · history · blame

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
//===- llvm/Target/TargetSchedule.cpp - Sched Machine Model ---------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements a wrapper around MCSchedModel that allows the interface
// to benefit from information currently only available in TargetInstrInfo.
//
//===----------------------------------------------------------------------===//

#include "llvm/CodeGen/TargetSchedule.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/MC/MCSchedule.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
#include <cassert>
#include <cstdint>

using namespace llvm;

static cl::opt<bool> EnableSchedModel("schedmodel", cl::Hidden, cl::init(true),
  cl::desc("Use TargetSchedModel for latency lookup"));

static cl::opt<bool> EnableSchedItins("scheditins", cl::Hidden, cl::init(true),
  cl::desc("Use InstrItineraryData for latency lookup"));

bool TargetSchedModel::hasInstrSchedModel() const {
  return EnableSchedModel && SchedModel.hasInstrSchedModel();
}

bool TargetSchedModel::hasInstrItineraries() const {
  return EnableSchedItins && !InstrItins.isEmpty();
}

static unsigned gcd(unsigned Dividend, unsigned Divisor) {
  // Dividend and Divisor will be naturally swapped as needed.
  while (Divisor) {
    unsigned Rem = Dividend % Divisor;
    Dividend = Divisor;
    Divisor = Rem;
  };
  return Dividend;
}

static unsigned lcm(unsigned A, unsigned B) {
  unsigned LCM = (uint64_t(A) * B) / gcd(A, B);
  assert((LCM >= A && LCM >= B) && "LCM overflow");
  return LCM;
}

void TargetSchedModel::init(const MCSchedModel &sm,
                            const TargetSubtargetInfo *sti,
                            const TargetInstrInfo *tii) {
  SchedModel = sm;
  STI = sti;
  TII = tii;
  STI->initInstrItins(InstrItins);

  unsigned NumRes = SchedModel.getNumProcResourceKinds();
  ResourceFactors.resize(NumRes);
  ResourceLCM = SchedModel.IssueWidth;
  for (unsigned Idx = 0; Idx < NumRes; ++Idx) {
    unsigned NumUnits = SchedModel.getProcResource(Idx)->NumUnits;
    if (NumUnits > 0)
      ResourceLCM = lcm(ResourceLCM, NumUnits);
  }
  MicroOpFactor = ResourceLCM / SchedModel.IssueWidth;
  for (unsigned Idx = 0; Idx < NumRes; ++Idx) {
    unsigned NumUnits = SchedModel.getProcResource(Idx)->NumUnits;
    ResourceFactors[Idx] = NumUnits ? (ResourceLCM / NumUnits) : 0;
  }
}

/// Returns true only if instruction is specified as single issue.
bool TargetSchedModel::mustBeginGroup(const MachineInstr *MI,
                                     const MCSchedClassDesc *SC) const {
  if (hasInstrSchedModel()) {
    if (!SC)
      SC = resolveSchedClass(MI);
    if (SC->isValid())
      return SC->BeginGroup;
  }
  return false;
}

bool TargetSchedModel::mustEndGroup(const MachineInstr *MI,
                                     const MCSchedClassDesc *SC) const {
  if (hasInstrSchedModel()) {
    if (!SC)
      SC = resolveSchedClass(MI);
    if (SC->isValid())
      return SC->EndGroup;
  }
  return false;
}

unsigned TargetSchedModel::getNumMicroOps(const MachineInstr *MI,
                                          const MCSchedClassDesc *SC) const {
  if (hasInstrItineraries()) {
    int UOps = InstrItins.getNumMicroOps(MI->getDesc().getSchedClass());
    return (UOps >= 0) ? UOps : TII->getNumMicroOps(&InstrItins, *MI);
  }
  if (hasInstrSchedModel()) {
    if (!SC)
      SC = resolveSchedClass(MI);
    if (SC->isValid())
      return SC->NumMicroOps;
  }
  return MI->isTransient() ? 0 : 1;
}

// The machine model may explicitly specify an invalid latency, which
// effectively means infinite latency. Since users of the TargetSchedule API
// don't know how to handle this, we convert it to a very large latency that is
// easy to distinguish when debugging the DAG but won't induce overflow.
static unsigned capLatency(int Cycles) {
  return Cycles >= 0 ? Cycles : 1000;
}

/// Return the MCSchedClassDesc for this instruction. Some SchedClasses require
/// evaluation of predicates that depend on instruction operands or flags.
const MCSchedClassDesc *TargetSchedModel::
resolveSchedClass(const MachineInstr *MI) const {
  // Get the definition's scheduling class descriptor from this machine model.
  unsigned SchedClass = MI->getDesc().getSchedClass();
  const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClass);
  if (!SCDesc->isValid())
    return SCDesc;

#ifndef NDEBUG
  unsigned NIter = 0;
#endif
  while (SCDesc->isVariant()) {
    assert(++NIter < 6 && "Variants are nested deeper than the magic number");

    SchedClass = STI->resolveSchedClass(SchedClass, MI, this);
    SCDesc = SchedModel.getSchedClassDesc(SchedClass);
  }
  return SCDesc;
}

/// Find the def index of this operand. This index maps to the machine model and
/// is independent of use operands. Def operands may be reordered with uses or
/// merged with uses without affecting the def index (e.g. before/after
/// regalloc). However, an instruction's def operands must never be reordered
/// with respect to each other.
static unsigned findDefIdx(const MachineInstr *MI, unsigned DefOperIdx) {
  unsigned DefIdx = 0;
  for (unsigned i = 0; i != DefOperIdx; ++i) {
    const MachineOperand &MO = MI->getOperand(i);
    if (MO.isReg() && MO.isDef())
      ++DefIdx;
  }
  return DefIdx;
}

/// Find the use index of this operand. This is independent of the instruction's
/// def operands.
///
/// Note that uses are not determined by the operand's isUse property, which
/// is simply the inverse of isDef. Here we consider any readsReg operand to be
/// a "use". The machine model allows an operand to be both a Def and Use.
static unsigned findUseIdx(const MachineInstr *MI, unsigned UseOperIdx) {
  unsigned UseIdx = 0;
  for (unsigned i = 0; i != UseOperIdx; ++i) {
    const MachineOperand &MO = MI->getOperand(i);
    if (MO.isReg() && MO.readsReg() && !MO.isDef())
      ++UseIdx;
  }
  return UseIdx;
}

// Top-level API for clients that know the operand indices.
unsigned TargetSchedModel::computeOperandLatency(
  const MachineInstr *DefMI, unsigned DefOperIdx,
  const MachineInstr *UseMI, unsigned UseOperIdx) const {

  if (!hasInstrSchedModel() && !hasInstrItineraries())
    return TII->defaultDefLatency(SchedModel, *DefMI);

  if (hasInstrItineraries()) {
    int OperLatency = 0;
    if (UseMI) {
      OperLatency = TII->getOperandLatency(&InstrItins, *DefMI, DefOperIdx,
                                           *UseMI, UseOperIdx);
    }
    else {
      unsigned DefClass = DefMI->getDesc().getSchedClass();
      OperLatency = InstrItins.getOperandCycle(DefClass, DefOperIdx);
    }
    if (OperLatency >= 0)
      return OperLatency;

    // No operand latency was found.
    unsigned InstrLatency = TII->getInstrLatency(&InstrItins, *DefMI);

    // Expected latency is the max of the stage latency and itinerary props.
    // Rather than directly querying InstrItins stage latency, we call a TII
    // hook to allow subtargets to specialize latency. This hook is only
    // applicable to the InstrItins model. InstrSchedModel should model all
    // special cases without TII hooks.
    InstrLatency =
        std::max(InstrLatency, TII->defaultDefLatency(SchedModel, *DefMI));
    return InstrLatency;
  }
  // hasInstrSchedModel()
  const MCSchedClassDesc *SCDesc = resolveSchedClass(DefMI);
  unsigned DefIdx = findDefIdx(DefMI, DefOperIdx);
  if (DefIdx < SCDesc->NumWriteLatencyEntries) {
    // Lookup the definition's write latency in SubtargetInfo.
    const MCWriteLatencyEntry *WLEntry =
      STI->getWriteLatencyEntry(SCDesc, DefIdx);
    unsigned WriteID = WLEntry->WriteResourceID;
    unsigned Latency = capLatency(WLEntry->Cycles);
    if (!UseMI)
      return Latency;

    // Lookup the use's latency adjustment in SubtargetInfo.
    const MCSchedClassDesc *UseDesc = resolveSchedClass(UseMI);
    if (UseDesc->NumReadAdvanceEntries == 0)
      return Latency;
    unsigned UseIdx = findUseIdx(UseMI, UseOperIdx);
    int Advance = STI->getReadAdvanceCycles(UseDesc, UseIdx, WriteID);
    if (Advance > 0 && (unsigned)Advance > Latency) // unsigned wrap
      return 0;
    return Latency - Advance;
  }
  // If DefIdx does not exist in the model (e.g. implicit defs), then return
  // unit latency (defaultDefLatency may be too conservative).
#ifndef NDEBUG
  if (SCDesc->isValid() && !DefMI->getOperand(DefOperIdx).isImplicit()
      && !DefMI->getDesc().OpInfo[DefOperIdx].isOptionalDef()
      && SchedModel.isComplete()) {
    errs() << "DefIdx " << DefIdx << " exceeds machine model writes for "
           << *DefMI << " (Try with MCSchedModel.CompleteModel set to false)";
    llvm_unreachable("incomplete machine model");
  }
#endif
  // FIXME: Automatically giving all implicit defs defaultDefLatency is
  // undesirable. We should only do it for defs that are known to the MC
  // desc like flags. Truly implicit defs should get 1 cycle latency.
  return DefMI->isTransient() ? 0 : TII->defaultDefLatency(SchedModel, *DefMI);
}

unsigned
TargetSchedModel::computeInstrLatency(const MCSchedClassDesc &SCDesc) const {
  unsigned Latency = 0;
  for (unsigned DefIdx = 0, DefEnd = SCDesc.NumWriteLatencyEntries;
       DefIdx != DefEnd; ++DefIdx) {
    // Lookup the definition's write latency in SubtargetInfo.
    const MCWriteLatencyEntry *WLEntry =
      STI->getWriteLatencyEntry(&SCDesc, DefIdx);
    Latency = std::max(Latency, capLatency(WLEntry->Cycles));
  }
  return Latency;
}

unsigned TargetSchedModel::computeInstrLatency(unsigned Opcode) const {
  assert(hasInstrSchedModel() && "Only call this function with a SchedModel");

  unsigned SCIdx = TII->get(Opcode).getSchedClass();
  const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SCIdx);

  if (SCDesc->isValid() && !SCDesc->isVariant())
    return computeInstrLatency(*SCDesc);

  if (SCDesc->isValid()) {
    assert (!SCDesc->isVariant() && "No MI sched latency: SCDesc->isVariant()");
    return computeInstrLatency(*SCDesc);
  }
  return 0;
}

unsigned
TargetSchedModel::computeInstrLatency(const MachineInstr *MI,
                                      bool UseDefaultDefLatency) const {
  // For the itinerary model, fall back to the old subtarget hook.
  // Allow subtargets to compute Bundle latencies outside the machine model.
  if (hasInstrItineraries() || MI->isBundle() ||
      (!hasInstrSchedModel() && !UseDefaultDefLatency))
    return TII->getInstrLatency(&InstrItins, *MI);

  if (hasInstrSchedModel()) {
    const MCSchedClassDesc *SCDesc = resolveSchedClass(MI);
    if (SCDesc->isValid())
      return computeInstrLatency(*SCDesc);
  }
  return TII->defaultDefLatency(SchedModel, *MI);
}

unsigned TargetSchedModel::
computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx,
                     const MachineInstr *DepMI) const {
  if (!SchedModel.isOutOfOrder())
    return 1;

  // Out-of-order processor can dispatch WAW dependencies in the same cycle.

  // Treat predication as a data dependency for out-of-order cpus. In-order
  // cpus do not need to treat predicated writes specially.
  //
  // TODO: The following hack exists because predication passes do not
  // correctly append imp-use operands, and readsReg() strangely returns false
  // for predicated defs.
  unsigned Reg = DefMI->getOperand(DefOperIdx).getReg();
  const MachineFunction &MF = *DefMI->getMF();
  const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
  if (!DepMI->readsRegister(Reg, TRI) && TII->isPredicated(*DepMI))
    return computeInstrLatency(DefMI);

  // If we have a per operand scheduling model, check if this def is writing
  // an unbuffered resource. If so, it treated like an in-order cpu.
  if (hasInstrSchedModel()) {
    const MCSchedClassDesc *SCDesc = resolveSchedClass(DefMI);
    if (SCDesc->isValid()) {
      for (const MCWriteProcResEntry *PRI = STI->getWriteProcResBegin(SCDesc),
             *PRE = STI->getWriteProcResEnd(SCDesc); PRI != PRE; ++PRI) {
        if (!SchedModel.getProcResource(PRI->ProcResourceIdx)->BufferSize)
          return 1;
      }
    }
  }
  return 0;
}

static Optional<double>
getRThroughputFromItineraries(unsigned schedClass,
                              const InstrItineraryData *IID){
  Optional<double> Throughput;

  for (const InstrStage *IS = IID->beginStage(schedClass),
                        *E = IID->endStage(schedClass);
       IS != E; ++IS) {
    if (IS->getCycles()) {
      double Temp = countPopulation(IS->getUnits()) * 1.0 / IS->getCycles();
      Throughput = Throughput.hasValue()
                        ? std::min(Throughput.getValue(), Temp)
                        : Temp;
    }
  }
  if (Throughput.hasValue())
    // We need reciprocal throughput that's why we return such value.
    return 1 / Throughput.getValue();
  return Throughput;
}

static Optional<double>
getRThroughputFromInstrSchedModel(const MCSchedClassDesc *SCDesc,
                                  const TargetSubtargetInfo *STI,
                                  const MCSchedModel &SchedModel) {
  Optional<double> Throughput;

  for (const MCWriteProcResEntry *WPR = STI->getWriteProcResBegin(SCDesc),
                                 *WEnd = STI->getWriteProcResEnd(SCDesc);
       WPR != WEnd; ++WPR) {
    if (WPR->Cycles) {
      unsigned NumUnits =
          SchedModel.getProcResource(WPR->ProcResourceIdx)->NumUnits;
      double Temp = NumUnits * 1.0 / WPR->Cycles;
      Throughput = Throughput.hasValue()
                       ? std::min(Throughput.getValue(), Temp)
                       : Temp;
    }
  }
  if (Throughput.hasValue())
    // We need reciprocal throughput that's why we return such value.
    return 1 / Throughput.getValue();
  return Throughput;
}

Optional<double>
TargetSchedModel::computeInstrRThroughput(const MachineInstr *MI) const {
  if (hasInstrItineraries())
    return getRThroughputFromItineraries(MI->getDesc().getSchedClass(),
                                         getInstrItineraries());
  if (hasInstrSchedModel())
    return getRThroughputFromInstrSchedModel(resolveSchedClass(MI), STI,
                                             SchedModel);
  return Optional<double>();
}

Optional<double>
TargetSchedModel::computeInstrRThroughput(unsigned Opcode) const {
  unsigned SchedClass = TII->get(Opcode).getSchedClass();
  if (hasInstrItineraries())
    return getRThroughputFromItineraries(SchedClass, getInstrItineraries());
  if (hasInstrSchedModel()) {
    const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClass);
    if (SCDesc->isValid() && !SCDesc->isVariant())
      return getRThroughputFromInstrSchedModel(SCDesc, STI, SchedModel);
  }
  return Optional<double>();
}