File: AMDGPUMarkLastScratchLoad.cpp

package info (click to toggle)
llvm-toolchain-19 1%3A19.1.7-3
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid, trixie
  • size: 1,998,520 kB
  • sloc: cpp: 6,951,680; ansic: 1,486,157; asm: 913,598; python: 232,024; f90: 80,126; objc: 75,281; lisp: 37,276; pascal: 16,990; sh: 10,009; ml: 5,058; perl: 4,724; awk: 3,523; makefile: 3,167; javascript: 2,504; xml: 892; fortran: 664; cs: 573
file content (142 lines) | stat: -rw-r--r-- 4,625 bytes parent folder | download | duplicates (3)
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
//===-- AMDGPUMarkLastScratchLoad.cpp -------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// Mark scratch load/spill instructions which are guaranteed to be the last time
// this scratch slot is used so it can be evicted from caches.
//
// TODO: Handle general stack accesses not just spilling.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LiveStacks.h"
#include "llvm/CodeGen/MachineOperand.h"

using namespace llvm;

#define DEBUG_TYPE "amdgpu-mark-last-scratch-load"

namespace {

class AMDGPUMarkLastScratchLoad : public MachineFunctionPass {
private:
  LiveStacks *LS = nullptr;
  LiveIntervals *LIS = nullptr;
  SlotIndexes *SI = nullptr;
  const SIInstrInfo *SII = nullptr;

public:
  static char ID;

  AMDGPUMarkLastScratchLoad() : MachineFunctionPass(ID) {
    initializeAMDGPUMarkLastScratchLoadPass(*PassRegistry::getPassRegistry());
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<SlotIndexesWrapperPass>();
    AU.addRequired<LiveIntervalsWrapperPass>();
    AU.addRequired<LiveStacks>();
    AU.setPreservesAll();
    MachineFunctionPass::getAnalysisUsage(AU);
  }

  StringRef getPassName() const override {
    return "AMDGPU Mark Last Scratch Load";
  }
};

} // end anonymous namespace

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

  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  if (ST.getGeneration() < AMDGPUSubtarget::GFX12)
    return false;

  LS = &getAnalysis<LiveStacks>();
  LIS = &getAnalysis<LiveIntervalsWrapperPass>().getLIS();
  SI = &getAnalysis<SlotIndexesWrapperPass>().getSI();
  SII = ST.getInstrInfo();
  SlotIndexes &Slots = *LIS->getSlotIndexes();

  const unsigned NumSlots = LS->getNumIntervals();
  if (NumSlots == 0) {
    LLVM_DEBUG(dbgs() << "No live slots, skipping\n");
    return false;
  }

  LLVM_DEBUG(dbgs() << LS->getNumIntervals() << " intervals\n");

  bool Changed = false;

  for (auto &[SS, LI] : *LS) {
    for (const LiveRange::Segment &Segment : LI.segments) {

      // Ignore segments that run to the end of basic block because in this case
      // slot is still live at the end of it.
      if (Segment.end.isBlock())
        continue;

      const int FrameIndex = Register::stackSlot2Index(LI.reg());
      MachineInstr *LastLoad = nullptr;

      MachineInstr *MISegmentEnd = SI->getInstructionFromIndex(Segment.end);

      // If there is no instruction at this slot because it was deleted take the
      // instruction from the next slot.
      if (!MISegmentEnd) {
        SlotIndex NextSlot = Slots.getNextNonNullIndex(Segment.end);
        MISegmentEnd = SI->getInstructionFromIndex(NextSlot);
      }

      MachineInstr *MISegmentStart = SI->getInstructionFromIndex(Segment.start);
      MachineBasicBlock *BB = MISegmentEnd->getParent();

      // Start iteration backwards from segment end until the start of basic
      // block or start of segment if it is in the same basic block.
      auto End = BB->rend();
      if (MISegmentStart && MISegmentStart->getParent() == BB)
        End = MISegmentStart->getReverseIterator();

      for (auto MI = MISegmentEnd->getReverseIterator(); MI != End; ++MI) {
        int LoadFI = 0;

        if (SII->isLoadFromStackSlot(*MI, LoadFI) && LoadFI == FrameIndex) {
          LastLoad = &*MI;
          break;
        }
      }

      if (LastLoad && !LastLoad->memoperands_empty()) {
        MachineMemOperand *MMO = *LastLoad->memoperands_begin();
        MMO->setFlags(MOLastUse);
        Changed = true;
        LLVM_DEBUG(dbgs() << "  Found last load: " << *LastLoad);
      }
    }
  }

  return Changed;
}

char AMDGPUMarkLastScratchLoad::ID = 0;

char &llvm::AMDGPUMarkLastScratchLoadID = AMDGPUMarkLastScratchLoad::ID;

INITIALIZE_PASS_BEGIN(AMDGPUMarkLastScratchLoad, DEBUG_TYPE,
                      "AMDGPU Mark last scratch load", false, false)
INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass)
INITIALIZE_PASS_DEPENDENCY(LiveStacks)
INITIALIZE_PASS_END(AMDGPUMarkLastScratchLoad, DEBUG_TYPE,
                    "AMDGPU Mark last scratch load", false, false)