llvm.org GIT mirror llvm / release_80 lib / Target / AMDGPU / R600OptimizeVectorRegisters.cpp
release_80

Tree @release_80 (Download .tar.gz)

R600OptimizeVectorRegisters.cpp @release_80raw · 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
407
408
//===- R600MergeVectorRegisters.cpp ---------------------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This pass merges inputs of swizzeable instructions into vector sharing
/// common data and/or have enough undef subreg using swizzle abilities.
///
/// For instance let's consider the following pseudo code :
/// %5 = REG_SEQ %1, sub0, %2, sub1, %3, sub2, undef, sub3
/// ...
/// %7 = REG_SEQ %1, sub0, %3, sub1, undef, sub2, %4, sub3
/// (swizzable Inst) %7, SwizzleMask : sub0, sub1, sub2, sub3
///
/// is turned into :
/// %5 = REG_SEQ %1, sub0, %2, sub1, %3, sub2, undef, sub3
/// ...
/// %7 = INSERT_SUBREG %4, sub3
/// (swizzable Inst) %7, SwizzleMask : sub0, sub2, sub1, sub3
///
/// This allow regalloc to reduce register pressure for vector registers and
/// to reduce MOV count.
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "R600Defines.h"
#include "R600InstrInfo.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
#include <utility>
#include <vector>

using namespace llvm;

#define DEBUG_TYPE "vec-merger"

static bool
isImplicitlyDef(MachineRegisterInfo &MRI, unsigned Reg) {
  for (MachineRegisterInfo::def_instr_iterator It = MRI.def_instr_begin(Reg),
      E = MRI.def_instr_end(); It != E; ++It) {
    return (*It).isImplicitDef();
  }
  if (MRI.isReserved(Reg)) {
    return false;
  }
  llvm_unreachable("Reg without a def");
  return false;
}

namespace {

class RegSeqInfo {
public:
  MachineInstr *Instr;
  DenseMap<unsigned, unsigned> RegToChan;
  std::vector<unsigned> UndefReg;

  RegSeqInfo(MachineRegisterInfo &MRI, MachineInstr *MI) : Instr(MI) {
    assert(MI->getOpcode() == R600::REG_SEQUENCE);
    for (unsigned i = 1, e = Instr->getNumOperands(); i < e; i+=2) {
      MachineOperand &MO = Instr->getOperand(i);
      unsigned Chan = Instr->getOperand(i + 1).getImm();
      if (isImplicitlyDef(MRI, MO.getReg()))
        UndefReg.push_back(Chan);
      else
        RegToChan[MO.getReg()] = Chan;
    }
  }

  RegSeqInfo() = default;

  bool operator==(const RegSeqInfo &RSI) const {
    return RSI.Instr == Instr;
  }
};

class R600VectorRegMerger : public MachineFunctionPass {
private:
  using InstructionSetMap = DenseMap<unsigned, std::vector<MachineInstr *>>;

  MachineRegisterInfo *MRI;
  const R600InstrInfo *TII = nullptr;
  DenseMap<MachineInstr *, RegSeqInfo> PreviousRegSeq;
  InstructionSetMap PreviousRegSeqByReg;
  InstructionSetMap PreviousRegSeqByUndefCount;

  bool canSwizzle(const MachineInstr &MI) const;
  bool areAllUsesSwizzeable(unsigned Reg) const;
  void SwizzleInput(MachineInstr &,
      const std::vector<std::pair<unsigned, unsigned>> &RemapChan) const;
  bool tryMergeVector(const RegSeqInfo *Untouched, RegSeqInfo *ToMerge,
      std::vector<std::pair<unsigned, unsigned>> &Remap) const;
  bool tryMergeUsingCommonSlot(RegSeqInfo &RSI, RegSeqInfo &CompatibleRSI,
      std::vector<std::pair<unsigned, unsigned>> &RemapChan);
  bool tryMergeUsingFreeSlot(RegSeqInfo &RSI, RegSeqInfo &CompatibleRSI,
      std::vector<std::pair<unsigned, unsigned>> &RemapChan);
  MachineInstr *RebuildVector(RegSeqInfo *MI, const RegSeqInfo *BaseVec,
      const std::vector<std::pair<unsigned, unsigned>> &RemapChan) const;
  void RemoveMI(MachineInstr *);
  void trackRSI(const RegSeqInfo &RSI);

public:
  static char ID;

  R600VectorRegMerger() : MachineFunctionPass(ID) {}

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<MachineDominatorTree>();
    AU.addPreserved<MachineDominatorTree>();
    AU.addRequired<MachineLoopInfo>();
    AU.addPreserved<MachineLoopInfo>();
    MachineFunctionPass::getAnalysisUsage(AU);
  }

  StringRef getPassName() const override {
    return "R600 Vector Registers Merge Pass";
  }

  bool runOnMachineFunction(MachineFunction &Fn) override;
};

} // end anonymous namespace

INITIALIZE_PASS_BEGIN(R600VectorRegMerger, DEBUG_TYPE,
                     "R600 Vector Reg Merger", false, false)
INITIALIZE_PASS_END(R600VectorRegMerger, DEBUG_TYPE,
                    "R600 Vector Reg Merger", false, false)

char R600VectorRegMerger::ID = 0;

char &llvm::R600VectorRegMergerID = R600VectorRegMerger::ID;

bool R600VectorRegMerger::canSwizzle(const MachineInstr &MI)
    const {
  if (TII->get(MI.getOpcode()).TSFlags & R600_InstFlag::TEX_INST)
    return true;
  switch (MI.getOpcode()) {
  case R600::R600_ExportSwz:
  case R600::EG_ExportSwz:
    return true;
  default:
    return false;
  }
}

bool R600VectorRegMerger::tryMergeVector(const RegSeqInfo *Untouched,
    RegSeqInfo *ToMerge, std::vector< std::pair<unsigned, unsigned>> &Remap)
    const {
  unsigned CurrentUndexIdx = 0;
  for (DenseMap<unsigned, unsigned>::iterator It = ToMerge->RegToChan.begin(),
      E = ToMerge->RegToChan.end(); It != E; ++It) {
    DenseMap<unsigned, unsigned>::const_iterator PosInUntouched =
        Untouched->RegToChan.find((*It).first);
    if (PosInUntouched != Untouched->RegToChan.end()) {
      Remap.push_back(std::pair<unsigned, unsigned>
          ((*It).second, (*PosInUntouched).second));
      continue;
    }
    if (CurrentUndexIdx >= Untouched->UndefReg.size())
      return false;
    Remap.push_back(std::pair<unsigned, unsigned>
        ((*It).second, Untouched->UndefReg[CurrentUndexIdx++]));
  }

  return true;
}

static
unsigned getReassignedChan(
    const std::vector<std::pair<unsigned, unsigned>> &RemapChan,
    unsigned Chan) {
  for (unsigned j = 0, je = RemapChan.size(); j < je; j++) {
    if (RemapChan[j].first == Chan)
      return RemapChan[j].second;
  }
  llvm_unreachable("Chan wasn't reassigned");
}

MachineInstr *R600VectorRegMerger::RebuildVector(
    RegSeqInfo *RSI, const RegSeqInfo *BaseRSI,
    const std::vector<std::pair<unsigned, unsigned>> &RemapChan) const {
  unsigned Reg = RSI->Instr->getOperand(0).getReg();
  MachineBasicBlock::iterator Pos = RSI->Instr;
  MachineBasicBlock &MBB = *Pos->getParent();
  DebugLoc DL = Pos->getDebugLoc();

  unsigned SrcVec = BaseRSI->Instr->getOperand(0).getReg();
  DenseMap<unsigned, unsigned> UpdatedRegToChan = BaseRSI->RegToChan;
  std::vector<unsigned> UpdatedUndef = BaseRSI->UndefReg;
  for (DenseMap<unsigned, unsigned>::iterator It = RSI->RegToChan.begin(),
      E = RSI->RegToChan.end(); It != E; ++It) {
    unsigned DstReg = MRI->createVirtualRegister(&R600::R600_Reg128RegClass);
    unsigned SubReg = (*It).first;
    unsigned Swizzle = (*It).second;
    unsigned Chan = getReassignedChan(RemapChan, Swizzle);

    MachineInstr *Tmp = BuildMI(MBB, Pos, DL, TII->get(R600::INSERT_SUBREG),
        DstReg)
        .addReg(SrcVec)
        .addReg(SubReg)
        .addImm(Chan);
    UpdatedRegToChan[SubReg] = Chan;
    std::vector<unsigned>::iterator ChanPos = llvm::find(UpdatedUndef, Chan);
    if (ChanPos != UpdatedUndef.end())
      UpdatedUndef.erase(ChanPos);
    assert(!is_contained(UpdatedUndef, Chan) &&
           "UpdatedUndef shouldn't contain Chan more than once!");
    LLVM_DEBUG(dbgs() << "    ->"; Tmp->dump(););
    (void)Tmp;
    SrcVec = DstReg;
  }
  MachineInstr *NewMI =
      BuildMI(MBB, Pos, DL, TII->get(R600::COPY), Reg).addReg(SrcVec);
  LLVM_DEBUG(dbgs() << "    ->"; NewMI->dump(););

  LLVM_DEBUG(dbgs() << "  Updating Swizzle:\n");
  for (MachineRegisterInfo::use_instr_iterator It = MRI->use_instr_begin(Reg),
      E = MRI->use_instr_end(); It != E; ++It) {
    LLVM_DEBUG(dbgs() << "    "; (*It).dump(); dbgs() << "    ->");
    SwizzleInput(*It, RemapChan);
    LLVM_DEBUG((*It).dump());
  }
  RSI->Instr->eraseFromParent();

  // Update RSI
  RSI->Instr = NewMI;
  RSI->RegToChan = UpdatedRegToChan;
  RSI->UndefReg = UpdatedUndef;

  return NewMI;
}

void R600VectorRegMerger::RemoveMI(MachineInstr *MI) {
  for (InstructionSetMap::iterator It = PreviousRegSeqByReg.begin(),
      E = PreviousRegSeqByReg.end(); It != E; ++It) {
    std::vector<MachineInstr *> &MIs = (*It).second;
    MIs.erase(llvm::find(MIs, MI), MIs.end());
  }
  for (InstructionSetMap::iterator It = PreviousRegSeqByUndefCount.begin(),
      E = PreviousRegSeqByUndefCount.end(); It != E; ++It) {
    std::vector<MachineInstr *> &MIs = (*It).second;
    MIs.erase(llvm::find(MIs, MI), MIs.end());
  }
}

void R600VectorRegMerger::SwizzleInput(MachineInstr &MI,
    const std::vector<std::pair<unsigned, unsigned>> &RemapChan) const {
  unsigned Offset;
  if (TII->get(MI.getOpcode()).TSFlags & R600_InstFlag::TEX_INST)
    Offset = 2;
  else
    Offset = 3;
  for (unsigned i = 0; i < 4; i++) {
    unsigned Swizzle = MI.getOperand(i + Offset).getImm() + 1;
    for (unsigned j = 0, e = RemapChan.size(); j < e; j++) {
      if (RemapChan[j].first == Swizzle) {
        MI.getOperand(i + Offset).setImm(RemapChan[j].second - 1);
        break;
      }
    }
  }
}

bool R600VectorRegMerger::areAllUsesSwizzeable(unsigned Reg) const {
  for (MachineRegisterInfo::use_instr_iterator It = MRI->use_instr_begin(Reg),
      E = MRI->use_instr_end(); It != E; ++It) {
    if (!canSwizzle(*It))
      return false;
  }
  return true;
}

bool R600VectorRegMerger::tryMergeUsingCommonSlot(RegSeqInfo &RSI,
    RegSeqInfo &CompatibleRSI,
    std::vector<std::pair<unsigned, unsigned>> &RemapChan) {
  for (MachineInstr::mop_iterator MOp = RSI.Instr->operands_begin(),
      MOE = RSI.Instr->operands_end(); MOp != MOE; ++MOp) {
    if (!MOp->isReg())
      continue;
    if (PreviousRegSeqByReg[MOp->getReg()].empty())
      continue;
    for (MachineInstr *MI : PreviousRegSeqByReg[MOp->getReg()]) {
      CompatibleRSI = PreviousRegSeq[MI];
      if (RSI == CompatibleRSI)
        continue;
      if (tryMergeVector(&CompatibleRSI, &RSI, RemapChan))
        return true;
    }
  }
  return false;
}

bool R600VectorRegMerger::tryMergeUsingFreeSlot(RegSeqInfo &RSI,
    RegSeqInfo &CompatibleRSI,
    std::vector<std::pair<unsigned, unsigned>> &RemapChan) {
  unsigned NeededUndefs = 4 - RSI.UndefReg.size();
  if (PreviousRegSeqByUndefCount[NeededUndefs].empty())
    return false;
  std::vector<MachineInstr *> &MIs =
      PreviousRegSeqByUndefCount[NeededUndefs];
  CompatibleRSI = PreviousRegSeq[MIs.back()];
  tryMergeVector(&CompatibleRSI, &RSI, RemapChan);
  return true;
}

void R600VectorRegMerger::trackRSI(const RegSeqInfo &RSI) {
  for (DenseMap<unsigned, unsigned>::const_iterator
  It = RSI.RegToChan.begin(), E = RSI.RegToChan.end(); It != E; ++It) {
    PreviousRegSeqByReg[(*It).first].push_back(RSI.Instr);
  }
  PreviousRegSeqByUndefCount[RSI.UndefReg.size()].push_back(RSI.Instr);
  PreviousRegSeq[RSI.Instr] = RSI;
}

bool R600VectorRegMerger::runOnMachineFunction(MachineFunction &Fn) {
  if (skipFunction(Fn.getFunction()))
    return false;

  const R600Subtarget &ST = Fn.getSubtarget<R600Subtarget>();
  TII = ST.getInstrInfo();
  MRI = &Fn.getRegInfo();

  for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
       MBB != MBBe; ++MBB) {
    MachineBasicBlock *MB = &*MBB;
    PreviousRegSeq.clear();
    PreviousRegSeqByReg.clear();
    PreviousRegSeqByUndefCount.clear();

    for (MachineBasicBlock::iterator MII = MB->begin(), MIIE = MB->end();
         MII != MIIE; ++MII) {
      MachineInstr &MI = *MII;
      if (MI.getOpcode() != R600::REG_SEQUENCE) {
        if (TII->get(MI.getOpcode()).TSFlags & R600_InstFlag::TEX_INST) {
          unsigned Reg = MI.getOperand(1).getReg();
          for (MachineRegisterInfo::def_instr_iterator
               It = MRI->def_instr_begin(Reg), E = MRI->def_instr_end();
               It != E; ++It) {
            RemoveMI(&(*It));
          }
        }
        continue;
      }

      RegSeqInfo RSI(*MRI, &MI);

      // All uses of MI are swizzeable ?
      unsigned Reg = MI.getOperand(0).getReg();
      if (!areAllUsesSwizzeable(Reg))
        continue;

      LLVM_DEBUG({
        dbgs() << "Trying to optimize ";
        MI.dump();
      });

      RegSeqInfo CandidateRSI;
      std::vector<std::pair<unsigned, unsigned>> RemapChan;
      LLVM_DEBUG(dbgs() << "Using common slots...\n";);
      if (tryMergeUsingCommonSlot(RSI, CandidateRSI, RemapChan)) {
        // Remove CandidateRSI mapping
        RemoveMI(CandidateRSI.Instr);
        MII = RebuildVector(&RSI, &CandidateRSI, RemapChan);
        trackRSI(RSI);
        continue;
      }
      LLVM_DEBUG(dbgs() << "Using free slots...\n";);
      RemapChan.clear();
      if (tryMergeUsingFreeSlot(RSI, CandidateRSI, RemapChan)) {
        RemoveMI(CandidateRSI.Instr);
        MII = RebuildVector(&RSI, &CandidateRSI, RemapChan);
        trackRSI(RSI);
        continue;
      }
      //Failed to merge
      trackRSI(RSI);
    }
  }
  return false;
}

llvm::FunctionPass *llvm::createR600VectorRegMerger() {
  return new R600VectorRegMerger();
}