llvm.org GIT mirror llvm / 945c85d lib / Target / AMDGPU / SILoadStoreOptimizer.cpp
945c85d

Tree @945c85d (Download .tar.gz)

SILoadStoreOptimizer.cpp @945c85draw · 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
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
//===-- SILoadStoreOptimizer.cpp ------------------------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass tries to fuse DS instructions with close by immediate offsets.
// This will fuse operations such as
//  ds_read_b32 v0, v2 offset:16
//  ds_read_b32 v1, v2 offset:32
// ==>
//   ds_read2_b32 v[0:1], v2, offset0:4 offset1:8
//
//
// Future improvements:
//
// - This currently relies on the scheduler to place loads and stores next to
//   each other, and then only merges adjacent pairs of instructions. It would
//   be good to be more flexible with interleaved instructions, and possibly run
//   before scheduling. It currently missing stores of constants because loading
//   the constant into the data register is placed between the stores, although
//   this is arguably a scheduling problem.
//
// - Live interval recomputing seems inefficient. This currently only matches
//   one pair, and recomputes live intervals and moves on to the next pair. It
//   would be better to compute a list of all merges that need to occur.
//
// - With a list of instructions to process, we can also merge more. If a
//   cluster of loads have offsets that are too large to fit in the 8-bit
//   offsets, but are close enough to fit in the 8 bits, we can add to the base
//   pointer and use the new reduced offsets.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "SIRegisterInfo.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.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/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include <cassert>
#include <iterator>
#include <utility>

using namespace llvm;

#define DEBUG_TYPE "si-load-store-opt"

namespace {

class SILoadStoreOptimizer : public MachineFunctionPass {
private:
  const SIInstrInfo *TII = nullptr;
  const SIRegisterInfo *TRI = nullptr;
  MachineRegisterInfo *MRI = nullptr;
  AliasAnalysis *AA = nullptr;

  static bool offsetsCanBeCombined(unsigned Offset0,
                                   unsigned Offset1,
                                   unsigned EltSize);

  MachineBasicBlock::iterator findMatchingDSInst(
    MachineBasicBlock::iterator I,
    unsigned EltSize,
    SmallVectorImpl<MachineInstr*> &InstsToMove);

  MachineBasicBlock::iterator mergeRead2Pair(
    MachineBasicBlock::iterator I,
    MachineBasicBlock::iterator Paired,
    unsigned EltSize,
    ArrayRef<MachineInstr*> InstsToMove);

  MachineBasicBlock::iterator mergeWrite2Pair(
    MachineBasicBlock::iterator I,
    MachineBasicBlock::iterator Paired,
    unsigned EltSize,
    ArrayRef<MachineInstr*> InstsToMove);

public:
  static char ID;

  SILoadStoreOptimizer() : MachineFunctionPass(ID) {}

  SILoadStoreOptimizer(const TargetMachine &TM_) : MachineFunctionPass(ID) {
    initializeSILoadStoreOptimizerPass(*PassRegistry::getPassRegistry());
  }

  bool optimizeBlock(MachineBasicBlock &MBB);

  bool runOnMachineFunction(MachineFunction &MF) override;

  StringRef getPassName() const override { return "SI Load / Store Optimizer"; }

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.setPreservesCFG();
    AU.addRequired<AAResultsWrapperPass>();

    MachineFunctionPass::getAnalysisUsage(AU);
  }
};

} // end anonymous namespace.

INITIALIZE_PASS_BEGIN(SILoadStoreOptimizer, DEBUG_TYPE,
                      "SI Load / Store Optimizer", false, false)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(SILoadStoreOptimizer, DEBUG_TYPE,
                    "SI Load / Store Optimizer", false, false)

char SILoadStoreOptimizer::ID = 0;

char &llvm::SILoadStoreOptimizerID = SILoadStoreOptimizer::ID;

FunctionPass *llvm::createSILoadStoreOptimizerPass(TargetMachine &TM) {
  return new SILoadStoreOptimizer(TM);
}

static void moveInstsAfter(MachineBasicBlock::iterator I,
                           ArrayRef<MachineInstr*> InstsToMove) {
  MachineBasicBlock *MBB = I->getParent();
  ++I;
  for (MachineInstr *MI : InstsToMove) {
    MI->removeFromParent();
    MBB->insert(I, MI);
  }
}

static void addDefsToList(const MachineInstr &MI,
                          SmallVectorImpl<const MachineOperand *> &Defs) {
  for (const MachineOperand &Def : MI.defs()) {
    Defs.push_back(&Def);
  }
}

static bool memAccessesCanBeReordered(MachineBasicBlock::iterator A,
                                      MachineBasicBlock::iterator B,
                                      const SIInstrInfo *TII,
                                      AliasAnalysis * AA) {
  return (TII->areMemAccessesTriviallyDisjoint(*A, *B, AA) ||
    // RAW or WAR - cannot reorder
    // WAW - cannot reorder
    // RAR - safe to reorder
    !(A->mayStore() || B->mayStore()));
}

// Add MI and its defs to the lists if MI reads one of the defs that are
// already in the list. Returns true in that case.
static bool
addToListsIfDependent(MachineInstr &MI,
                      SmallVectorImpl<const MachineOperand *> &Defs,
                      SmallVectorImpl<MachineInstr*> &Insts) {
  for (const MachineOperand *Def : Defs) {
    bool ReadDef = MI.readsVirtualRegister(Def->getReg());
    // If ReadDef is true, then there is a use of Def between I
    // and the instruction that I will potentially be merged with. We
    // will need to move this instruction after the merged instructions.
    if (ReadDef) {
      Insts.push_back(&MI);
      addDefsToList(MI, Defs);
      return true;
    }
  }

  return false;
}

static bool
canMoveInstsAcrossMemOp(MachineInstr &MemOp,
                        ArrayRef<MachineInstr*> InstsToMove,
                        const SIInstrInfo *TII,
                        AliasAnalysis *AA) {
  assert(MemOp.mayLoadOrStore());

  for (MachineInstr *InstToMove : InstsToMove) {
    if (!InstToMove->mayLoadOrStore())
      continue;
    if (!memAccessesCanBeReordered(MemOp, *InstToMove, TII, AA))
        return false;
  }
  return true;
}

bool SILoadStoreOptimizer::offsetsCanBeCombined(unsigned Offset0,
                                                unsigned Offset1,
                                                unsigned Size) {
  // XXX - Would the same offset be OK? Is there any reason this would happen or
  // be useful?
  if (Offset0 == Offset1)
    return false;

  // This won't be valid if the offset isn't aligned.
  if ((Offset0 % Size != 0) || (Offset1 % Size != 0))
    return false;

  unsigned EltOffset0 = Offset0 / Size;
  unsigned EltOffset1 = Offset1 / Size;

  // Check if the new offsets fit in the reduced 8-bit range.
  if (isUInt<8>(EltOffset0) && isUInt<8>(EltOffset1))
    return true;

  // If the offset in elements doesn't fit in 8-bits, we might be able to use
  // the stride 64 versions.
  if ((EltOffset0 % 64 != 0) || (EltOffset1 % 64) != 0)
    return false;

  return isUInt<8>(EltOffset0 / 64) && isUInt<8>(EltOffset1 / 64);
}

MachineBasicBlock::iterator
SILoadStoreOptimizer::findMatchingDSInst(MachineBasicBlock::iterator I,
                                  unsigned EltSize,
                                  SmallVectorImpl<MachineInstr*> &InstsToMove) {
  MachineBasicBlock::iterator E = I->getParent()->end();
  MachineBasicBlock::iterator MBBI = I;
  ++MBBI;

  SmallVector<const MachineOperand *, 8> DefsToMove;
  addDefsToList(*I, DefsToMove);

  for ( ; MBBI != E; ++MBBI) {
    if (MBBI->getOpcode() != I->getOpcode()) {

      // This is not a matching DS instruction, but we can keep looking as
      // long as one of these conditions are met:
      // 1. It is safe to move I down past MBBI.
      // 2. It is safe to move MBBI down past the instruction that I will
      //    be merged into.

      if (MBBI->hasUnmodeledSideEffects())
        // We can't re-order this instruction with respect to other memory
        // opeations, so we fail both conditions mentioned above.
        return E;

      if (MBBI->mayLoadOrStore() &&
        !memAccessesCanBeReordered(*I, *MBBI, TII, AA)) {
        // We fail condition #1, but we may still be able to satisfy condition
        // #2.  Add this instruction to the move list and then we will check
        // if condition #2 holds once we have selected the matching instruction.
        InstsToMove.push_back(&*MBBI);
        addDefsToList(*MBBI, DefsToMove);
        continue;
      }

      // When we match I with another DS instruction we will be moving I down
      // to the location of the matched instruction any uses of I will need to
      // be moved down as well.
      addToListsIfDependent(*MBBI, DefsToMove, InstsToMove);
      continue;
    }

    // Don't merge volatiles.
    if (MBBI->hasOrderedMemoryRef())
      return E;

    // Handle a case like
    //   DS_WRITE_B32 addr, v, idx0
    //   w = DS_READ_B32 addr, idx0
    //   DS_WRITE_B32 addr, f(w), idx1
    // where the DS_READ_B32 ends up in InstsToMove and therefore prevents
    // merging of the two writes.
    if (addToListsIfDependent(*MBBI, DefsToMove, InstsToMove))
      continue;

    int AddrIdx = AMDGPU::getNamedOperandIdx(I->getOpcode(), AMDGPU::OpName::addr);
    const MachineOperand &AddrReg0 = I->getOperand(AddrIdx);
    const MachineOperand &AddrReg1 = MBBI->getOperand(AddrIdx);

    // Check same base pointer. Be careful of subregisters, which can occur with
    // vectors of pointers.
    if (AddrReg0.getReg() == AddrReg1.getReg() &&
        AddrReg0.getSubReg() == AddrReg1.getSubReg()) {
      int OffsetIdx = AMDGPU::getNamedOperandIdx(I->getOpcode(),
                                                 AMDGPU::OpName::offset);
      unsigned Offset0 = I->getOperand(OffsetIdx).getImm() & 0xffff;
      unsigned Offset1 = MBBI->getOperand(OffsetIdx).getImm() & 0xffff;

      // Check both offsets fit in the reduced range.
      // We also need to go through the list of instructions that we plan to
      // move and make sure they are all safe to move down past the merged
      // instruction.
      if (offsetsCanBeCombined(Offset0, Offset1, EltSize) &&
          canMoveInstsAcrossMemOp(*MBBI, InstsToMove, TII, AA))
        return MBBI;
    }

    // We've found a load/store that we couldn't merge for some reason.
    // We could potentially keep looking, but we'd need to make sure that
    // it was safe to move I and also all the instruction in InstsToMove
    // down past this instruction.
    if (!memAccessesCanBeReordered(*I, *MBBI, TII, AA) ||   // check if we can move I across MBBI
      !canMoveInstsAcrossMemOp(*MBBI, InstsToMove, TII, AA) // check if we can move all I's users
     )
      break;
  }
  return E;
}

MachineBasicBlock::iterator  SILoadStoreOptimizer::mergeRead2Pair(
  MachineBasicBlock::iterator I,
  MachineBasicBlock::iterator Paired,
  unsigned EltSize,
  ArrayRef<MachineInstr*> InstsToMove) {
  MachineBasicBlock *MBB = I->getParent();

  // Be careful, since the addresses could be subregisters themselves in weird
  // cases, like vectors of pointers.
  const MachineOperand *AddrReg = TII->getNamedOperand(*I, AMDGPU::OpName::addr);

  const MachineOperand *Dest0 = TII->getNamedOperand(*I, AMDGPU::OpName::vdst);
  const MachineOperand *Dest1 = TII->getNamedOperand(*Paired, AMDGPU::OpName::vdst);

  unsigned Offset0
    = TII->getNamedOperand(*I, AMDGPU::OpName::offset)->getImm() & 0xffff;
  unsigned Offset1
    = TII->getNamedOperand(*Paired, AMDGPU::OpName::offset)->getImm() & 0xffff;

  unsigned NewOffset0 = Offset0 / EltSize;
  unsigned NewOffset1 = Offset1 / EltSize;
  unsigned Opc = (EltSize == 4) ? AMDGPU::DS_READ2_B32 : AMDGPU::DS_READ2_B64;

  // Prefer the st64 form if we can use it, even if we can fit the offset in the
  // non st64 version. I'm not sure if there's any real reason to do this.
  bool UseST64 = (NewOffset0 % 64 == 0) && (NewOffset1 % 64 == 0);
  if (UseST64) {
    NewOffset0 /= 64;
    NewOffset1 /= 64;
    Opc = (EltSize == 4) ? AMDGPU::DS_READ2ST64_B32 : AMDGPU::DS_READ2ST64_B64;
  }

  unsigned SubRegIdx0 = (EltSize == 4) ? AMDGPU::sub0 : AMDGPU::sub0_sub1;
  unsigned SubRegIdx1 = (EltSize == 4) ? AMDGPU::sub1 : AMDGPU::sub2_sub3;

  if (NewOffset0 > NewOffset1) {
    // Canonicalize the merged instruction so the smaller offset comes first.
    std::swap(NewOffset0, NewOffset1);
    std::swap(SubRegIdx0, SubRegIdx1);
  }

  assert((isUInt<8>(NewOffset0) && isUInt<8>(NewOffset1)) &&
         (NewOffset0 != NewOffset1) &&
         "Computed offset doesn't fit");

  const MCInstrDesc &Read2Desc = TII->get(Opc);

  const TargetRegisterClass *SuperRC
    = (EltSize == 4) ? &AMDGPU::VReg_64RegClass : &AMDGPU::VReg_128RegClass;
  unsigned DestReg = MRI->createVirtualRegister(SuperRC);

  DebugLoc DL = I->getDebugLoc();
  MachineInstrBuilder Read2 = BuildMI(*MBB, Paired, DL, Read2Desc, DestReg)
                                  .add(*AddrReg)      // addr
                                  .addImm(NewOffset0) // offset0
                                  .addImm(NewOffset1) // offset1
                                  .addImm(0)          // gds
                                  .addMemOperand(*I->memoperands_begin())
                                  .addMemOperand(*Paired->memoperands_begin());
  (void)Read2;

  const MCInstrDesc &CopyDesc = TII->get(TargetOpcode::COPY);

  // Copy to the old destination registers.
  BuildMI(*MBB, Paired, DL, CopyDesc)
      .add(*Dest0) // Copy to same destination including flags and sub reg.
      .addReg(DestReg, 0, SubRegIdx0);
  MachineInstr *Copy1 = BuildMI(*MBB, Paired, DL, CopyDesc)
                            .add(*Dest1)
                            .addReg(DestReg, RegState::Kill, SubRegIdx1);

  moveInstsAfter(Copy1, InstsToMove);

  MachineBasicBlock::iterator Next = std::next(I);
  I->eraseFromParent();
  Paired->eraseFromParent();

  DEBUG(dbgs() << "Inserted read2: " << *Read2 << '\n');
  return Next;
}

MachineBasicBlock::iterator SILoadStoreOptimizer::mergeWrite2Pair(
  MachineBasicBlock::iterator I,
  MachineBasicBlock::iterator Paired,
  unsigned EltSize,
  ArrayRef<MachineInstr*> InstsToMove) {
  MachineBasicBlock *MBB = I->getParent();

  // Be sure to use .addOperand(), and not .addReg() with these. We want to be
  // sure we preserve the subregister index and any register flags set on them.
  const MachineOperand *Addr = TII->getNamedOperand(*I, AMDGPU::OpName::addr);
  const MachineOperand *Data0 = TII->getNamedOperand(*I, AMDGPU::OpName::data0);
  const MachineOperand *Data1
    = TII->getNamedOperand(*Paired, AMDGPU::OpName::data0);


  unsigned Offset0
    = TII->getNamedOperand(*I, AMDGPU::OpName::offset)->getImm() & 0xffff;
  unsigned Offset1
    = TII->getNamedOperand(*Paired, AMDGPU::OpName::offset)->getImm() & 0xffff;

  unsigned NewOffset0 = Offset0 / EltSize;
  unsigned NewOffset1 = Offset1 / EltSize;
  unsigned Opc = (EltSize == 4) ? AMDGPU::DS_WRITE2_B32 : AMDGPU::DS_WRITE2_B64;

  // Prefer the st64 form if we can use it, even if we can fit the offset in the
  // non st64 version. I'm not sure if there's any real reason to do this.
  bool UseST64 = (NewOffset0 % 64 == 0) && (NewOffset1 % 64 == 0);
  if (UseST64) {
    NewOffset0 /= 64;
    NewOffset1 /= 64;
    Opc = (EltSize == 4) ? AMDGPU::DS_WRITE2ST64_B32 : AMDGPU::DS_WRITE2ST64_B64;
  }

  if (NewOffset0 > NewOffset1) {
    // Canonicalize the merged instruction so the smaller offset comes first.
    std::swap(NewOffset0, NewOffset1);
    std::swap(Data0, Data1);
  }

  assert((isUInt<8>(NewOffset0) && isUInt<8>(NewOffset1)) &&
         (NewOffset0 != NewOffset1) &&
         "Computed offset doesn't fit");

  const MCInstrDesc &Write2Desc = TII->get(Opc);
  DebugLoc DL = I->getDebugLoc();

  MachineInstrBuilder Write2 = BuildMI(*MBB, Paired, DL, Write2Desc)
                                   .add(*Addr)         // addr
                                   .add(*Data0)        // data0
                                   .add(*Data1)        // data1
                                   .addImm(NewOffset0) // offset0
                                   .addImm(NewOffset1) // offset1
                                   .addImm(0)          // gds
                                   .addMemOperand(*I->memoperands_begin())
                                   .addMemOperand(*Paired->memoperands_begin());

  moveInstsAfter(Write2, InstsToMove);

  MachineBasicBlock::iterator Next = std::next(I);
  I->eraseFromParent();
  Paired->eraseFromParent();

  DEBUG(dbgs() << "Inserted write2 inst: " << *Write2 << '\n');
  return Next;
}

// Scan through looking for adjacent LDS operations with constant offsets from
// the same base register. We rely on the scheduler to do the hard work of
// clustering nearby loads, and assume these are all adjacent.
bool SILoadStoreOptimizer::optimizeBlock(MachineBasicBlock &MBB) {
  bool Modified = false;

  for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E;) {
    MachineInstr &MI = *I;

    // Don't combine if volatile.
    if (MI.hasOrderedMemoryRef()) {
      ++I;
      continue;
    }

    SmallVector<MachineInstr*, 8> InstsToMove;
    unsigned Opc = MI.getOpcode();
    if (Opc == AMDGPU::DS_READ_B32 || Opc == AMDGPU::DS_READ_B64) {
      unsigned Size = (Opc == AMDGPU::DS_READ_B64) ? 8 : 4;
      MachineBasicBlock::iterator Match = findMatchingDSInst(I, Size,
                                                             InstsToMove);
      if (Match != E) {
        Modified = true;
        I = mergeRead2Pair(I, Match, Size, InstsToMove);
      } else {
        ++I;
      }

      continue;
    } else if (Opc == AMDGPU::DS_WRITE_B32 || Opc == AMDGPU::DS_WRITE_B64) {
      unsigned Size = (Opc == AMDGPU::DS_WRITE_B64) ? 8 : 4;
      MachineBasicBlock::iterator Match = findMatchingDSInst(I, Size,
                                                             InstsToMove);
      if (Match != E) {
        Modified = true;
        I = mergeWrite2Pair(I, Match, Size, InstsToMove);
      } else {
        ++I;
      }

      continue;
    }

    ++I;
  }

  return Modified;
}

bool SILoadStoreOptimizer::runOnMachineFunction(MachineFunction &MF) {
  if (skipFunction(*MF.getFunction()))
    return false;

  const SISubtarget &STM = MF.getSubtarget<SISubtarget>();
  if (!STM.loadStoreOptEnabled())
    return false;

  TII = STM.getInstrInfo();
  TRI = &TII->getRegisterInfo();

  MRI = &MF.getRegInfo();
  AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();

  DEBUG(dbgs() << "Running SILoadStoreOptimizer\n");

  bool Modified = false;

  for (MachineBasicBlock &MBB : MF)
    Modified |= optimizeBlock(MBB);

  return Modified;
}