Cache filename for sorting in fill_todo

[Kobzol]

Alternative to #144.

This reduces the duration to glob **/*.rs in a rustc checkout on my PC from ~1000ms to ~820ms:

Benchmark 1: ./local
  Time (mean ± σ):     822.0 ms ±  11.2 ms    [User: 272.5 ms, System: 549.2 ms]
  Range (min … max):   811.8 ms … 849.6 ms    10 runs
 
Benchmark 2: ./upstream
  Time (mean ± σ):      1.019 s ±  0.006 s    [User: 0.468 s, System: 0.551 s]
  Range (min … max):    1.005 s …  1.026 s    10 runs
 
Summary
  ./local ran
    1.24 ± 0.02 times faster than ./upstream

[the8472]

"still"? afaict this adds an allocation. So it's surprising that allocating an extra OsString is still cheaper than getting the filename from the path later.

If you keep the whole DirEntry you could call file_name_ref on cfg(unix) to avoid that allocation too.

[Kobzol]

It is a bit surprising, but the benchmark is quite clear on the results. It's one allocation vs potentially tens/hundreds/thousands of iterations of components, which seemingly isn't cheap, perf.-wise.

file_name_ref is unstable, as far as I can see?

[the8472]

file_name_ref is unstable, as far as I can see?

Right, too bad.

Regarding the overhead, maybe wait until my rust PR lands and then benchmark again? The file_name optimization looks like it just might give us that 20% speedup too.

[Kobzol]

Fair enough, once it gets into nightly, I'll benchmark it against a previous nightly to see if there's a big difference.

[Kobzol]

With rust-lang/rust#148084 landed:

• Upstream (0.3.3)

/pr/pe/ru/gl/globtest [cache-filename]$ hyperfine ./globtest-2025-11-09 ./globtest-2025-11-10
Benchmark 1: ./globtest-2025-11-09
  Time (mean ± σ):     761.6 ms ±   3.6 ms    [User: 359.8 ms, System: 401.6 ms]
  Range (min … max):   756.5 ms … 768.4 ms    10 runs
 
Benchmark 2: ./globtest-2025-11-10
  Time (mean ± σ):     738.1 ms ±   4.2 ms    [User: 342.1 ms, System: 395.8 ms]
  Range (min … max):   734.4 ms … 748.6 ms    10 runs
 
Summary
  ./globtest-2025-11-10 ran
    1.03 ± 0.01 times faster than ./globtest-2025-11-09

• This PR

/pr/pe/ru/gl/globtest [cache-filename]$ hyperfine ./globtest-2025-11-09 ./globtest-2025-11-10
Benchmark 1: ./globtest-2025-11-09
  Time (mean ± σ):     614.1 ms ±   4.4 ms    [User: 212.6 ms, System: 401.2 ms]
  Range (min … max):   608.1 ms … 621.7 ms    10 runs
 
Benchmark 2: ./globtest-2025-11-10
  Time (mean ± σ):     606.5 ms ±   1.6 ms    [User: 203.3 ms, System: 403.0 ms]
  Range (min … max):   603.8 ms … 609.8 ms    10 runs
 
Summary
  ./globtest-2025-11-10 ran
    1.01 ± 0.01 times faster than ./globtest-2025-11-09

Still seems worth landing (~600ms vs ~730ms).

[the8472]

Are those results comparable to the previous ones, i.e. did we get a speedup from the path changes and this PR will provide additional ones on top?

[Kobzol]

This should be the difference with/without your PR with 0.3.3 of glob:

Benchmark 1: ./globtest-2025-11-09
  Time (mean ± σ):     761.6 ms ±   3.6 ms    [User: 359.8 ms, System: 401.6 ms]
  Range (min … max):   756.5 ms … 768.4 ms    10 runs
 
Benchmark 2: ./globtest-2025-11-10
  Time (mean ± σ):     738.1 ms ±   4.2 ms    [User: 342.1 ms, System: 395.8 ms]
  Range (min … max):   734.4 ms … 748.6 ms    10 runs

It looks like your changes helped, but this PR provides an additional improvement of around 10-15%.

[the8472]

Thanks, it helps arguing the compile time impact. Also not as big as it looked from micro-benchmarks, but so it goes.