walk: Add a cache for DirEntry metadata

[tavianator]

This is a reimplementation of #402, using once_cell::unsync::OnceCell instead of a hand-rolled Lazy type. It gives some performance benefits when multiple filters require metadata:

$ hyperfine ./fd-{before,after}" -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null"
Benchmark #1: ./fd-before -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null
  Time (mean ± σ):     449.3 ms ±   5.1 ms    [User: 246.9 ms, System: 354.8 ms]
  Range (min … max):   441.1 ms … 455.1 ms    10 runs
 
Benchmark #2: ./fd-after -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null
  Time (mean ± σ):     392.5 ms ±   3.8 ms    [User: 230.1 ms, System: 287.9 ms]
  Range (min … max):   388.0 ms … 399.3 ms    10 runs
 
Summary
  './fd-after -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null' ran
    1.14 ± 0.02 times faster than './fd-before -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null'

without much affect on perf in other cases:

$ hyperfine ./fd-{before,after}" -j1 --search-path=../../linux --size=+1k >/dev/null"
Benchmark #1: ./fd-before -j1 --search-path=../../linux --size=+1k >/dev/null
  Time (mean ± σ):     383.8 ms ±   5.6 ms    [User: 212.7 ms, System: 296.1 ms]
  Range (min … max):   377.5 ms … 393.3 ms    10 runs
 
Benchmark #2: ./fd-after -j1 --search-path=../../linux --size=+1k >/dev/null
  Time (mean ± σ):     385.2 ms ±   4.1 ms    [User: 228.4 ms, System: 280.5 ms]
  Range (min … max):   378.3 ms … 393.3 ms    10 runs
 
Summary
  './fd-before -j1 --search-path=../../linux --size=+1k >/dev/null' ran
    1.00 ± 0.02 times faster than './fd-after -j1 --search-path=../../linux --size=+1k >/dev/null'

strace confirms a reduction in statx() calls:

$ strace -fc ./fd-before -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ------------------
...
 18.32    0.927790           3    254795     24015 statx
...
$ strace -fc ./fd-after -j1 --search-path=../../linux --size=+1k --owner=tavianator >/dev/null
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ------------------
...
 18.92    0.954145           5    181209     24015 statx
...

Together with #853 it would be possible to share the Metadata with the colorization which would be an additional benefit, but it's not done in this PR.

[tavianator]

There is a minor behaviour change (which you can see from the changed tests): --changed-* now examines the symlink's modification time itself rather than always following links. I think the new behaviour is better and the old behaviour was arguably a bug.

[tavianator]

Perhaps I can make it build on Windows after 4 tries? :P

[sharkdp]

Cool - thank you very much!

[sharkdp]

I have two remarks:

• Can we please add an entry to CHANGELOG.md? For the caching itself and also for the bugfix concerning symlink modification times.
• It might be a good idea to wait with merging this until Some readability refactorings #833 has been merged. What do you guys think?

[tavianator]

Yeah I'm happy to wait for #833. I'll add the changelog.

[sharkdp]

Thank you. I merged #833.

[tavianator]

Rebased and added the CHANGELOG entry