Optimize performance of Buffer#line_begins

[bbatsov]

Replaced push/reverse with unshift. This is generally about 10 times
faster according to my simple benchmarks.

[coveralls]

Coverage decreased (-0%) when pulling 6b184e7 on bbatsov:optimize-line-begins into beb08d0 on whitequark:master.

[whitequark]

This is really strange; unshift must relocate the complete contents of the array. I'd like to see the benchmarks you have used.

[bbatsov]

The benchmark I used was simply pushing 10k strings into an array and reversing; pushing + destructive reverse; unshift;

require 'benchmark'

a = []

time = Benchmark.realtime do
  10000.times do |i|
    a << "top #{i}"
    a = a.reverse
  end
end

puts time

This took 1.1 seconds. With reverse! it was 0.9. With unshift - 0.12. Benchmark is fairly primitive, but the results seem sensible to me. It simply seems that reverse is pretty expensive.

[whitequark]

See also https://gist.github.com/whitequark/a70f0d6fcaaf3a1cda0a and http://rxr.whitequark.org/mri/source/array.c#1031. This result is somewhat strange to me.

@bbatsov, does rubycop spend significant time in this function?

[bbatsov]

Haven't checked. I was just reading Parser's source and noticed this could be optimized.