Small performance regression in ignore-0.4.x

[sharkdp]

What operating system are you using ripgrep on?

Arch Linux, x86_64

Describe your question, feature request, or bug.

I was running some benchmarks for fd and noticed that there was a small (~5%) performance regression when I updated from ignore-0.2.2 to ignore-0.4.x. I investigated some more and ran the same benchmark (fd -uu '[0-9]\.jpg$' in my home folder) multiple times while changing the version of the ignore crate:

There are two things to notice:

1. The timing for ignore-0.3.0 is not visisble because it's so fast (137 ms). This is due to a bug where not all results are shown (I didn't investigate further, but the git log seems to indicate that something was wrong with ignore-0.3.0).
2. There is a small (but statistically significant) performance regression that seems to have been introduced with commit 3cb4d13 (Custom ignore files). Interestingly, I'm running a search that doesn't even use ignore files (-uu is the same as in ripgrep).

What are the steps to reproduce the behavior?

I can reproduce this behavior with ripgrep itself:

▶ git checkout 51864c1; cargo install -f

...

▶ hyperfine --warmup 10 'rg -uu --files --iglob "*[0-9].jpg" ~'
Benchmark #1: rg -uu --files --iglob "*[0-9].jpg" ~

  Time (mean ± σ):     700.4 ms ±   5.3 ms    [User: 3.157 s, System: 2.310 s]
 
  Range (min … max):   691.4 ms … 709.0 ms

▶ git checkout 3cb4d13; cargo install -f

...

▶ hyperfine --warmup 10 'rg -uu --files --iglob "*[0-9].jpg" ~'
Benchmark #1: rg -uu --files --iglob "*[0-9].jpg" ~

  Time (mean ± σ):     725.4 ms ±   7.9 ms    [User: 3.318 s, System: 2.338 s]
 
  Range (min … max):   715.5 ms … 743.9 ms

(benchmarks performed with hyperfine on a warm disk cache)

Again, the difference is small (~4%) but significant.

[sharkdp]

This seems to be caused by the additional Gitignore instance custom_ignore_matcher that has to be constructed in Ignore::add_child_path.

(apparently, the gitignore matchers are fully constructed even though no ignore-matching is performed)

[BurntSushi]

@sharkdp What corpus were your tests above on? I finally had a chance to look at this, and I can't really reproduce your results.

I'm testing on the chromium repo:

$ git remote -v
origin  git:https://github.com/nwjs/chromium.src (fetch)
origin  git:https://github.com/nwjs/chromium.src (push)
$ git rev-parse --short=10 HEAD
3682723171

Testing ripgrep on current master:

$ rg-baseline --version
ripgrep 0.8.1 (rev 507801c1f2)
+SIMD +AVX
$ hyperfine -s basic --warmup 10 'rg-baseline -uu --files'
Benchmark #1: rg-baseline -uu --files
  Time (mean ± σ):     230.6 ms ±  10.2 ms    [User: 1.004 s, System: 0.349 s]
  Range (min … max):   211.7 ms … 243.1 ms

Testing ripgrep with this PR, rebased on current master:

$ rg --version
ripgrep 0.8.1 (rev 2c2162f7bc)
+SIMD +AVX
$ hyperfine -s basic --warmup 10 'rg -uu --files'
Benchmark #1: rg -uu --files
  Time (mean ± σ):     230.2 ms ±   9.6 ms    [User: 927.0 ms, System: 383.5 ms]
  Range (min … max):   210.4 ms … 246.0 ms

[sharkdp]

@sharkdp What corpus were your tests above on? I finally had a chance to look at this, and I can't really reproduce your results.

I was running this on my home folder - not very helpful, I'm sorry. Two things seem to be different in your case:

• I believe the associated PR will show a more significant speedup in cases where there is a deep directory tree with lots of .gitignore files.

• Could you please try to add the --iglob option (with some file pattern that occurs only rarely) to make sure that the execution time is not dominated by writing to standard out?

I know it's not a very scientific method, but could you try to run the following commands in your home folder (or some other large directory)? I was just re-running the benchmarks with the rebased branch (which is now included in #836), and I still see a 10% speedup:

> rg-baseline --version
ripgrep 0.8.1 (rev bf51058eb2)
-SIMD -AVX

> rg-pr --version                                               
ripgrep 0.8.1 (rev 92dcb44d10)
-SIMD -AVX

> hyperfine --warmup 10 'rg-baseline -uu --files --iglob "README*"' \
                        'rg-pr       -uu --files --iglob "README*"'
Benchmark #1: rg-baseline -uu --files --iglob "README*"

  Time (mean ± σ):     760.8 ms ±   3.5 ms    [User: 3.395 s, System: 2.511 s]
 
  Range (min … max):   756.1 ms … 766.5 ms
 
Benchmark #2: rg-pr -uu --files --iglob "README*"

  Time (mean ± σ):     692.0 ms ±  15.5 ms    [User: 2.844 s, System: 2.493 s]
 
  Range (min … max):   679.6 ms … 731.3 ms
 
Summary

'rg-pr -uu --files --iglob "README*"' ran
    1.10x faster than 'rg-baseline -uu --files --iglob "README*"'