Separate IOSource#ensure_buffer from IOSource#match.

[naitoh]

Why?

It would affect performance to do a read check in IOSource#match every time, Separate read processing from IOSource#ensure_buffer.

Use IOSource#ensure_buffer in the following cases where @source.buffer is empty.

1. at the start of pull_event
2. If a trailing '>' pattern matches, as in @source.match(/\s*>/um).

Benchmark

RUBYLIB= BUNDLER_ORIG_RUBYLIB= /Users/naitoh/.rbenv/versions/3.3.0/bin/ruby -v -S benchmark-driver /Users/naitoh/ghq/github.com/naitoh/rexml/benchmark/parse.yaml ruby 3.3.0 (2023-12-25 revision 5124f9ac75) [arm64-darwin22] Calculating ------------------------------------- before after before(YJIT) after(YJIT) dom 10.278 10.986 16.430 16.941 i/s - 100.000 times in 9.729858s 9.102574s 6.086579s 5.902885s sax 30.166 30.496 49.851 51.596 i/s - 100.000 times in 3.315008s 3.279069s 2.005961s 1.938123s pull 35.459 36.380 60.266 63.134 i/s - 100.000 times in 2.820181s 2.748745s 1.659301s 1.583928s stream 33.762 34.636 55.173 55.859 i/s - 100.000 times in 2.961948s 2.887131s 1.812485s 1.790218s Comparison: dom after(YJIT): 16.9 i/s before(YJIT): 16.4 i/s - 1.03x slower after: 11.0 i/s - 1.54x slower before: 10.3 i/s - 1.65x slower sax after(YJIT): 51.6 i/s before(YJIT): 49.9 i/s - 1.04x slower after: 30.5 i/s - 1.69x slower before: 30.2 i/s - 1.71x slower pull after(YJIT): 63.1 i/s before(YJIT): 60.3 i/s - 1.05x slower after: 36.4 i/s - 1.74x slower before: 35.5 i/s - 1.78x slower stream after(YJIT): 55.9 i/s before(YJIT): 55.2 i/s - 1.01x slower after: 34.6 i/s - 1.61x slower before: 33.8 i/s - 1.65x slower 

• YJIT=ON : 1.01x - 1.05x faster
• YJIT=OFF : 1.01x - 1.06x faster

[kou]

Can we give it better name?
I feel that next_read doesn't imply "read if buffer is empty".

ensure_buffered?

[naitoh]

@kou

Can we give it better name? I feel that next_read doesn't imply "read if buffer is empty".

ensure_buffered?

I changed it to ensure_buffer.
What do you think?

[kou]

OK. Let's use it.