aboutsummaryrefslogtreecommitdiffstats
path: root/lib/Target/R600/SILowerControlFlow.cpp
blob: 9702565c46257b7a2782803e6a2ded71e731d1e3 (plain)
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
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
//===-- SILowerControlFlow.cpp - Use predicates for control flow ----------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// \brief This pass lowers the pseudo control flow instructions to real
/// machine instructions.
///
/// All control flow is handled using predicated instructions and
/// a predicate stack.  Each Scalar ALU controls the operations of 64 Vector
/// ALUs.  The Scalar ALU can update the predicate for any of the Vector ALUs
/// by writting to the 64-bit EXEC register (each bit corresponds to a
/// single vector ALU).  Typically, for predicates, a vector ALU will write
/// to its bit of the VCC register (like EXEC VCC is 64-bits, one for each
/// Vector ALU) and then the ScalarALU will AND the VCC register with the
/// EXEC to update the predicates.
///
/// For example:
/// %VCC = V_CMP_GT_F32 %VGPR1, %VGPR2
/// %SGPR0 = SI_IF %VCC
///   %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
/// %SGPR0 = SI_ELSE %SGPR0
///   %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
/// SI_END_CF %SGPR0
///
/// becomes:
///
/// %SGPR0 = S_AND_SAVEEXEC_B64 %VCC  // Save and update the exec mask
/// %SGPR0 = S_XOR_B64 %SGPR0, %EXEC  // Clear live bits from saved exec mask
/// S_CBRANCH_EXECZ label0            // This instruction is an optional
///                                   // optimization which allows us to
///                                   // branch if all the bits of
///                                   // EXEC are zero.
/// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0 // Do the IF block of the branch
///
/// label0:
/// %SGPR0 = S_OR_SAVEEXEC_B64 %EXEC   // Restore the exec mask for the Then block
/// %EXEC = S_XOR_B64 %SGPR0, %EXEC    // Clear live bits from saved exec mask
/// S_BRANCH_EXECZ label1              // Use our branch optimization
///                                    // instruction again.
/// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR   // Do the THEN block
/// label1:
/// %EXEC = S_OR_B64 %EXEC, %SGPR0     // Re-enable saved exec mask bits
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Constants.h"

using namespace llvm;

namespace {

class SILowerControlFlowPass : public MachineFunctionPass {

private:
  static const unsigned SkipThreshold = 12;

  static char ID;
  const SIRegisterInfo *TRI;
  const SIInstrInfo *TII;

  bool shouldSkip(MachineBasicBlock *From, MachineBasicBlock *To);

  void Skip(MachineInstr &From, MachineOperand &To);
  void SkipIfDead(MachineInstr &MI);

  void If(MachineInstr &MI);
  void Else(MachineInstr &MI);
  void Break(MachineInstr &MI);
  void IfBreak(MachineInstr &MI);
  void ElseBreak(MachineInstr &MI);
  void Loop(MachineInstr &MI);
  void EndCf(MachineInstr &MI);

  void Kill(MachineInstr &MI);
  void Branch(MachineInstr &MI);

  void InitM0ForLDS(MachineBasicBlock::iterator MI);
  void LoadM0(MachineInstr &MI, MachineInstr *MovRel);
  void IndirectSrc(MachineInstr &MI);
  void IndirectDst(MachineInstr &MI);

public:
  SILowerControlFlowPass(TargetMachine &tm) :
    MachineFunctionPass(ID), TRI(nullptr), TII(nullptr) { }

  bool runOnMachineFunction(MachineFunction &MF) override;

  const char *getPassName() const override {
    return "SI Lower control flow instructions";
  }

};

} // End anonymous namespace

char SILowerControlFlowPass::ID = 0;

FunctionPass *llvm::createSILowerControlFlowPass(TargetMachine &tm) {
  return new SILowerControlFlowPass(tm);
}

bool SILowerControlFlowPass::shouldSkip(MachineBasicBlock *From,
                                        MachineBasicBlock *To) {

  unsigned NumInstr = 0;

  for (MachineBasicBlock *MBB = From; MBB != To && !MBB->succ_empty();
       MBB = *MBB->succ_begin()) {

    for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
         NumInstr < SkipThreshold && I != E; ++I) {

      if (I->isBundle() || !I->isBundled())
        if (++NumInstr >= SkipThreshold)
          return true;
    }
  }

  return false;
}

void SILowerControlFlowPass::Skip(MachineInstr &From, MachineOperand &To) {

  if (!shouldSkip(*From.getParent()->succ_begin(), To.getMBB()))
    return;

  DebugLoc DL = From.getDebugLoc();
  BuildMI(*From.getParent(), &From, DL, TII->get(AMDGPU::S_CBRANCH_EXECZ))
          .addOperand(To)
          .addReg(AMDGPU::EXEC);
}

void SILowerControlFlowPass::SkipIfDead(MachineInstr &MI) {

  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  if (MBB.getParent()->getInfo<SIMachineFunctionInfo>()->getShaderType() !=
      ShaderType::PIXEL ||
      !shouldSkip(&MBB, &MBB.getParent()->back()))
    return;

  MachineBasicBlock::iterator Insert = &MI;
  ++Insert;

  // If the exec mask is non-zero, skip the next two instructions
  BuildMI(MBB, Insert, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
          .addImm(3)
          .addReg(AMDGPU::EXEC);

  // Exec mask is zero: Export to NULL target...
  BuildMI(MBB, Insert, DL, TII->get(AMDGPU::EXP))
          .addImm(0)
          .addImm(0x09) // V_008DFC_SQ_EXP_NULL
          .addImm(0)
          .addImm(1)
          .addImm(1)
          .addReg(AMDGPU::VGPR0)
          .addReg(AMDGPU::VGPR0)
          .addReg(AMDGPU::VGPR0)
          .addReg(AMDGPU::VGPR0);

  // ... and terminate wavefront
  BuildMI(MBB, Insert, DL, TII->get(AMDGPU::S_ENDPGM));
}

void SILowerControlFlowPass::If(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  unsigned Reg = MI.getOperand(0).getReg();
  unsigned Vcc = MI.getOperand(1).getReg();

  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), Reg)
          .addReg(Vcc);

  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), Reg)
          .addReg(AMDGPU::EXEC)
          .addReg(Reg);

  Skip(MI, MI.getOperand(2));

  MI.eraseFromParent();
}

void SILowerControlFlowPass::Else(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Src = MI.getOperand(1).getReg();

  BuildMI(MBB, MBB.getFirstNonPHI(), DL,
          TII->get(AMDGPU::S_OR_SAVEEXEC_B64), Dst)
          .addReg(Src); // Saved EXEC

  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
          .addReg(AMDGPU::EXEC)
          .addReg(Dst);

  Skip(MI, MI.getOperand(2));

  MI.eraseFromParent();
}

void SILowerControlFlowPass::Break(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Src = MI.getOperand(1).getReg();
 
  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
          .addReg(AMDGPU::EXEC)
          .addReg(Src);

  MI.eraseFromParent();
}

void SILowerControlFlowPass::IfBreak(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Vcc = MI.getOperand(1).getReg();
  unsigned Src = MI.getOperand(2).getReg();
 
  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
          .addReg(Vcc)
          .addReg(Src);

  MI.eraseFromParent();
}

void SILowerControlFlowPass::ElseBreak(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Saved = MI.getOperand(1).getReg();
  unsigned Src = MI.getOperand(2).getReg();
 
  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
          .addReg(Saved)
          .addReg(Src);

  MI.eraseFromParent();
}

void SILowerControlFlowPass::Loop(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  unsigned Src = MI.getOperand(0).getReg();

  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ANDN2_B64), AMDGPU::EXEC)
          .addReg(AMDGPU::EXEC)
          .addReg(Src);

  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
          .addOperand(MI.getOperand(1))
          .addReg(AMDGPU::EXEC);

  MI.eraseFromParent();
}

void SILowerControlFlowPass::EndCf(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  unsigned Reg = MI.getOperand(0).getReg();

  BuildMI(MBB, MBB.getFirstNonPHI(), DL,
          TII->get(AMDGPU::S_OR_B64), AMDGPU::EXEC)
          .addReg(AMDGPU::EXEC)
          .addReg(Reg);

  MI.eraseFromParent();
}

void SILowerControlFlowPass::Branch(MachineInstr &MI) {
  if (MI.getOperand(0).getMBB() == MI.getParent()->getNextNode())
    MI.eraseFromParent();

  // If these aren't equal, this is probably an infinite loop.
}

void SILowerControlFlowPass::Kill(MachineInstr &MI) {
  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  const MachineOperand &Op = MI.getOperand(0);

#ifndef NDEBUG
  const SIMachineFunctionInfo *MFI
    = MBB.getParent()->getInfo<SIMachineFunctionInfo>();
  // Kill is only allowed in pixel / geometry shaders.
  assert(MFI->getShaderType() == ShaderType::PIXEL ||
         MFI->getShaderType() == ShaderType::GEOMETRY);
#endif

  // Clear this thread from the exec mask if the operand is negative
  if ((Op.isImm() || Op.isFPImm())) {
    // Constant operand: Set exec mask to 0 or do nothing
    if (Op.isImm() ? (Op.getImm() & 0x80000000) :
        Op.getFPImm()->isNegative()) {
      BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
              .addImm(0);
    }
  } else {
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
           .addImm(0)
           .addOperand(Op);
  }

  MI.eraseFromParent();
}

/// The m0 register stores the maximum allowable address for LDS reads and
/// writes.  Its value must be at least the size in bytes of LDS allocated by
/// the shader.  For simplicity, we set it to the maximum possible value.
void SILowerControlFlowPass::InitM0ForLDS(MachineBasicBlock::iterator MI) {
    BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),  TII->get(AMDGPU::S_MOV_B32),
            AMDGPU::M0).addImm(0xffffffff);
}

void SILowerControlFlowPass::LoadM0(MachineInstr &MI, MachineInstr *MovRel) {

  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();
  MachineBasicBlock::iterator I = MI;

  unsigned Save = MI.getOperand(1).getReg();
  unsigned Idx = MI.getOperand(3).getReg();

  if (AMDGPU::SReg_32RegClass.contains(Idx)) {
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
            .addReg(Idx);
    MBB.insert(I, MovRel);
  } else {

    assert(AMDGPU::SReg_64RegClass.contains(Save));
    assert(AMDGPU::VReg_32RegClass.contains(Idx));

    // Save the EXEC mask
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), Save)
            .addReg(AMDGPU::EXEC);

    // Read the next variant into VCC (lower 32 bits) <- also loop target
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
            AMDGPU::VCC_LO)
            .addReg(Idx);

    // Move index from VCC into M0
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
            .addReg(AMDGPU::VCC_LO);

    // Compare the just read M0 value to all possible Idx values
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e32), AMDGPU::VCC)
            .addReg(AMDGPU::M0)
            .addReg(Idx);

    // Update EXEC, save the original EXEC value to VCC
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), AMDGPU::VCC)
            .addReg(AMDGPU::VCC);

    // Do the actual move
    MBB.insert(I, MovRel);

    // Update EXEC, switch all done bits to 0 and all todo bits to 1
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
            .addReg(AMDGPU::EXEC)
            .addReg(AMDGPU::VCC);

    // Loop back to V_READFIRSTLANE_B32 if there are still variants to cover
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
            .addImm(-7)
            .addReg(AMDGPU::EXEC);

    // Restore EXEC
    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
            .addReg(Save);

  }
  // FIXME: Are there any values other than the LDS address clamp that need to
  // be stored in the m0 register and may be live for more than a few
  // instructions?  If so, we should save the m0 register at the beginning
  // of this function and restore it here.
  // FIXME: Add support for LDS direct loads.
  InitM0ForLDS(&MI);
  MI.eraseFromParent();
}

void SILowerControlFlowPass::IndirectSrc(MachineInstr &MI) {

  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Vec = MI.getOperand(2).getReg();
  unsigned Off = MI.getOperand(4).getImm();
  unsigned SubReg = TRI->getSubReg(Vec, AMDGPU::sub0);
  if (!SubReg)
    SubReg = Vec;

  MachineInstr *MovRel =
    BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
            .addReg(SubReg + Off)
            .addReg(AMDGPU::M0, RegState::Implicit)
            .addReg(Vec, RegState::Implicit);

  LoadM0(MI, MovRel);
}

void SILowerControlFlowPass::IndirectDst(MachineInstr &MI) {

  MachineBasicBlock &MBB = *MI.getParent();
  DebugLoc DL = MI.getDebugLoc();

  unsigned Dst = MI.getOperand(0).getReg();
  unsigned Off = MI.getOperand(4).getImm();
  unsigned Val = MI.getOperand(5).getReg();
  unsigned SubReg = TRI->getSubReg(Dst, AMDGPU::sub0);
  if (!SubReg)
    SubReg = Dst;

  MachineInstr *MovRel = 
    BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELD_B32_e32))
            .addReg(SubReg + Off, RegState::Define)
            .addReg(Val)
            .addReg(AMDGPU::M0, RegState::Implicit)
            .addReg(Dst, RegState::Implicit);

  LoadM0(MI, MovRel);
}

bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
  TII = static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
  TRI =
      static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
  SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

  bool HaveKill = false;
  bool NeedM0 = false;
  bool NeedWQM = false;
  bool NeedFlat = false;
  unsigned Depth = 0;

  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
       BI != BE; ++BI) {

    MachineBasicBlock &MBB = *BI;
    MachineBasicBlock::iterator I, Next;
    for (I = MBB.begin(); I != MBB.end(); I = Next) {
      Next = std::next(I);

      MachineInstr &MI = *I;
      if (TII->isDS(MI.getOpcode())) {
        NeedM0 = true;
        NeedWQM = true;
      }

      // Flat uses m0 in case it needs to access LDS.
      if (TII->isFLAT(MI.getOpcode())) {
        NeedM0 = true;
        NeedFlat = true;
      }

      switch (MI.getOpcode()) {
        default: break;
        case AMDGPU::SI_IF:
          ++Depth;
          If(MI);
          break;

        case AMDGPU::SI_ELSE:
          Else(MI);
          break;

        case AMDGPU::SI_BREAK:
          Break(MI);
          break;

        case AMDGPU::SI_IF_BREAK:
          IfBreak(MI);
          break;

        case AMDGPU::SI_ELSE_BREAK:
          ElseBreak(MI);
          break;

        case AMDGPU::SI_LOOP:
          ++Depth;
          Loop(MI);
          break;

        case AMDGPU::SI_END_CF:
          if (--Depth == 0 && HaveKill) {
            SkipIfDead(MI);
            HaveKill = false;
          }
          EndCf(MI);
          break;

        case AMDGPU::SI_KILL:
          if (Depth == 0)
            SkipIfDead(MI);
          else
            HaveKill = true;
          Kill(MI);
          break;

        case AMDGPU::S_BRANCH:
          Branch(MI);
          break;

        case AMDGPU::SI_INDIRECT_SRC:
          IndirectSrc(MI);
          break;

        case AMDGPU::SI_INDIRECT_DST_V1:
        case AMDGPU::SI_INDIRECT_DST_V2:
        case AMDGPU::SI_INDIRECT_DST_V4:
        case AMDGPU::SI_INDIRECT_DST_V8:
        case AMDGPU::SI_INDIRECT_DST_V16:
          IndirectDst(MI);
          break;

        case AMDGPU::V_INTERP_P1_F32:
        case AMDGPU::V_INTERP_P2_F32:
        case AMDGPU::V_INTERP_MOV_F32:
          NeedWQM = true;
          break;
      }
    }
  }

  if (NeedM0) {
    MachineBasicBlock &MBB = MF.front();
    // Initialize M0 to a value that won't cause LDS access to be discarded
    // due to offset clamping
    InitM0ForLDS(MBB.getFirstNonPHI());
  }

  if (NeedWQM && MFI->getShaderType() == ShaderType::PIXEL) {
    MachineBasicBlock &MBB = MF.front();
    BuildMI(MBB, MBB.getFirstNonPHI(), DebugLoc(), TII->get(AMDGPU::S_WQM_B64),
            AMDGPU::EXEC).addReg(AMDGPU::EXEC);
  }

  // FIXME: This seems inappropriate to do here.
  if (NeedFlat && MFI->IsKernel) {
    // Insert the prologue initializing the SGPRs pointing to the scratch space
    // for flat accesses.
    const MachineFrameInfo *FrameInfo = MF.getFrameInfo();

    // TODO: What to use with function calls?

    // FIXME: This is reporting stack size that is used in a scratch buffer
    // rather than registers as well.
    uint64_t StackSizeBytes = FrameInfo->getStackSize();

    int IndirectBegin
      = static_cast<const AMDGPUInstrInfo*>(TII)->getIndirectIndexBegin(MF);
    // Convert register index to 256-byte unit.
    uint64_t StackOffset = IndirectBegin < 0 ? 0 : (4 * IndirectBegin / 256);

    assert((StackSizeBytes < 0xffff) && StackOffset < 0xffff &&
           "Stack limits should be smaller than 16-bits");

    // Initialize the flat scratch register pair.
    // TODO: Can we use one s_mov_b64 here?

    // Offset is in units of 256-bytes.
    MachineBasicBlock &MBB = MF.front();
    DebugLoc NoDL;
    MachineBasicBlock::iterator Start = MBB.getFirstNonPHI();
    const MCInstrDesc &SMovK = TII->get(AMDGPU::S_MOVK_I32);

    assert(isInt<16>(StackOffset) && isInt<16>(StackSizeBytes));

    BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_LO)
      .addImm(StackOffset);

    // Documentation says size is "per-thread scratch size in bytes"
    BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_HI)
      .addImm(StackSizeBytes);
  }

  return true;
}