[async] Use llvm::SmallVector/llvm:SmallSet for latest state readers

[k-ye]

This is to address #1936 (review). But TBH, I saw no performance difference... Still good to keep the data structures consistent, though.

I feel like it's fine to keep lastest_state_owners_ as is because it's not shown as a bottleneck, and it's not multi-leveled.

Related issue = #1936 (review)

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1951 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1951   +/-   ##
=======================================
  Coverage   43.65%   43.65%           
=======================================
  Files          45       45           
  Lines        6226     6226           
  Branches     1107     1107           
=======================================
  Hits         2718     2718           
  Misses       3334     3334           
  Partials      174      174           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b675bee...9349748. Read the comment docs.

[xumingkuan]

async_mgpcg.py on kun:
Before:

473.875 ms 63.79%  rebuild_graph         [44 x  10.770 ms]
    403.036 ms 85.05%  insert_task           [208655 x   1.932 us]
         15.607 ms  3.87%  get_task_meta         [208655 x  74.796 ns]
        117.734 ms 29.21%  insert_task meta->input_states [583601 x 201.738 ns]
             50.695 ms 43.06%  insert_edge           [583601 x  86.866 ns]
             67.039 ms 56.94%  [unaccounted]
         80.928 ms 20.08%  insert_task meta->output_states [217798 x 371.572 ns]
             36.551 ms 45.17%  insert_edge           [522732 x  69.923 ns]
             44.377 ms 54.83%  [unaccounted]
         71.740 ms 17.80%  insert_task latest_state_readers_ [583601 x 122.927 ns]
        117.027 ms 29.04%  [unaccounted]
      1.597 ms  0.34%  reid_nodes            [44 x  36.305 us]
      1.465 ms  0.31%  reid_pending_nodes    [44 x  33.303 us]
     67.777 ms 14.30%  [unaccounted]

After:

446.719 ms 64.48%  rebuild_graph         [44 x  10.153 ms]
    379.415 ms 84.93%  insert_task           [208655 x   1.818 us]
         15.268 ms  4.02%  get_task_meta         [208655 x  73.172 ns]
        113.590 ms 29.94%  insert_task meta->input_states [583601 x 194.636 ns]
             48.635 ms 42.82%  insert_edge           [583601 x  83.336 ns]
             64.955 ms 57.18%  [unaccounted]
         82.662 ms 21.79%  insert_task meta->output_states [217798 x 379.535 ns]
             34.935 ms 42.26%  insert_edge           [522732 x  66.832 ns]
             47.727 ms 57.74%  [unaccounted]
         57.767 ms 15.23%  insert_task latest_state_readers_ [583601 x  98.984 ns] (time reduced!)
        110.128 ms 29.03%  [unaccounted]
      1.542 ms  0.35%  reid_nodes            [44 x  35.042 us]
      1.466 ms  0.33%  reid_pending_nodes    [44 x  33.324 us]
     64.296 ms 14.39%  [unaccounted]

If we change StateToNodesMap from llvm::SmallVector<std::pair<AsyncState, llvm::SmallSet<Node *, 8>>, 4> to llvm::SmallVector<std::pair<AsyncState, llvm::SmallSet<Node *, 8>>, 100>:

0.771  s 69.32%  rebuild_graph         [44 x  17.521 ms]
  636.297 ms 82.54%  insert_task           [208655 x   3.050 us]
       17.767 ms  2.79%  get_task_meta         [208655 x  85.151 ns]
      127.272 ms 20.00%  insert_task meta->input_states [583601 x 218.081 ns]
           50.515 ms 39.69%  insert_edge           [583601 x  86.558 ns]
           76.757 ms 60.31%  [unaccounted]
       86.888 ms 13.66%  insert_task meta->output_states [217798 x 398.940 ns]
           37.968 ms 43.70%  insert_edge           [522732 x  72.635 ns]
           48.920 ms 56.30%  [unaccounted]
       59.681 ms  9.38%  insert_task latest_state_readers_ [583601 x 102.263 ns]
      344.689 ms 54.17%  [unaccounted]
    8.575 ms  1.11%  reid_nodes            [44 x 194.880 us]
    2.866 ms  0.37%  reid_pending_nodes    [44 x  65.137 us]
  123.183 ms 15.98%  [unaccounted]

to llvm::SmallVector<std::pair<AsyncState, llvm::SmallSet<Node *, 16>>, 4>:

487.942 ms 65.44%  rebuild_graph         [44 x  11.090 ms]
    412.135 ms 84.46%  insert_task           [208655 x   1.975 us]
         15.255 ms  3.70%  get_task_meta         [208655 x  73.113 ns]
        122.649 ms 29.76%  insert_task meta->input_states [583601 x 210.160 ns]
             51.231 ms 41.77%  insert_edge           [583601 x  87.784 ns]
             71.419 ms 58.23%  [unaccounted]
         84.822 ms 20.58%  insert_task meta->output_states [217798 x 389.453 ns]
             36.829 ms 43.42%  insert_edge           [522732 x  70.454 ns]
             47.994 ms 56.58%  [unaccounted]
         61.341 ms 14.88%  insert_task latest_state_readers_ [583601 x 105.108 ns]
        128.067 ms 31.07%  [unaccounted]
      1.702 ms  0.35%  reid_nodes            [44 x  38.678 us]
      1.576 ms  0.32%  reid_pending_nodes    [44 x  35.817 us]
     72.529 ms 14.86%  [unaccounted]

[xumingkuan]

LGTM, thanks!

[k-ye]

Thanks for doing the benchmark! So it actually did accelerate things a bit...