Do not pre-compute MountInfo to reduce readlink calls

[tertsdiepraam]

More speedups! In my previous PR I mentioned being baffled by the readlink and getcwd and I figured out where they came from: File::mount_point_info, which calls std::fs::canonicalize. The easy fix was to not call that when it's not necessary, which turns out to be most of the time!

Here's a benchmark:

❯ hyperfine -L eza eza-old,eza-new "./{eza} -R"
Benchmark 1: ./eza-old -R
  Time (mean ± σ):     744.0 ms ±   9.3 ms    [User: 275.2 ms, System: 465.3 ms]
  Range (min … max):   734.2 ms … 766.0 ms    10 runs
 
Benchmark 2: ./eza-new -R
  Time (mean ± σ):     313.3 ms ±   3.0 ms    [User: 142.2 ms, System: 169.4 ms]
  Range (min … max):   309.7 ms … 318.8 ms    10 runs
 
Summary
  ./eza-new -R ran
    2.37 ± 0.04 times faster than ./eza-old -R

[daviessm]

Thanks for doing this! I obviously didn't check the performance when I added this functionality. It seemed logical at the time.

[tertsdiepraam]

Yeah it does seem logical. It feels strange that canonicalize would be that expensive (even though it has to work like it does). I find that in Rust it's sometimes hard to tell which functions do syscalls. I only found this by looking at flame graphs.

[PThorpe92]

#558 this always made me curious as to what exactly was the cause of this (obviously not to say it's one thing in particular, in fact this may not have even been merged at the time of the issue). But I agree it's definitely not always obvious.

Big thanks again. Looks like we are heading for a significant perf improvement :D

[cafkafk]

This is so nice, thanks a lot :3

re: it being hard to guess what slows down eza, I've though for a while that we should include some performance testing step in reviews, but I haven't really found the time to figure out how we should do it.

[tertsdiepraam]

To be clear, I wasn't blaming eza for that. It's just hard in general. At uutils we've been thinking about tracking the number of syscalls, maybe that could work for you? We haven't set that up yet, because it doesn't feel super important in the end.

[tertsdiepraam]

To be clear, I wasn't blaming eza for that. It's just hard in general. At uutils we've been thinking about tracking the number of syscalls, maybe that could work for you? We haven't set that up yet, because it doesn't feel super important in the end.

[cafkafk]

At uutils we've been thinking about tracking the number of syscalls, maybe that could work for you?

That sounds like an okay solution actually, we should do this.

Also NW didn't think you were blaming eza!