Feature #13750: Improve String#casecmp? and Symbol#casecmp? performance with ASCII string

• Status: Open
• Assignee: -
• Target version: -

[watson1978 (Shizuo Fujita)]

I think String#casecmp and String#casecmp? are similar methods. But they have different performance with ASCII strings.

It seems that String#casecmp handles ASCII string only, but it is faster than String#casecmp?.

This patch uses the code of String#casecmp on String#casecmp? for ASCII strings. However, it introduces a minor penalty for UTF8 strings due to detection of ASCII/UTF8 strings.

String#casecmp? ASCII -> 61.3 % up
String#casecmp? UTF8  ->  1.3 % down
Symbol#casecmp? ASCII -> 80.0 % up
Symbol#casecmp? UTF8  ->  4.0 % down

Before

Calculating -------------------------------------
      String#casecmp      5.961M (± 3.8%) i/s -     29.838M in   5.017907s
String#casecmp? ASCII
                          3.530M (± 8.6%) i/s -     17.554M in   5.034848s
String#casecmp? UTF8      1.252M (± 7.4%) i/s -      6.213M in   5.012168s
      Symbol#casecmp      8.555M (± 2.4%) i/s -     42.822M in   5.009280s
Symbol#casecmp? ASCII
                          4.235M (± 9.7%) i/s -     20.824M in   5.001368s
Symbol#casecmp? UTF8      1.329M (± 0.1%) i/s -      6.704M in   5.043725s

After

Calculating -------------------------------------
      String#casecmp      5.984M (± 6.4%) i/s -     29.829M in   5.020331s
String#casecmp? ASCII
                          5.658M (± 1.5%) i/s -     28.308M in   5.004547s
String#casecmp? UTF8      1.215M (± 4.3%) i/s -      6.132M in   5.060292s
      Symbol#casecmp      8.651M (± 0.9%) i/s -     43.313M in   5.007215s
Symbol#casecmp? ASCII
                          7.462M (± 0.5%) i/s -     37.489M in   5.023892s
Symbol#casecmp? UTF8      1.275M (± 0.2%) i/s -      6.444M in   5.052743s

Test code

require 'benchmark/ips'

Benchmark.ips do |x|
  x.report "String#casecmp" do |loop|
    loop.times { "aBcDeF".casecmp("abcdefg") }
  end
  x.report "String#casecmp? ASCII" do |loop|
    loop.times { "aBcDeF".casecmp?("abcdefg") }
  end
  x.report "String#casecmp? UTF8" do |loop|
    loop.times { "\u{e4 f6 fc}".casecmp?("\u{c4 d6 dc}") }
  end

  x.report "Symbol#casecmp" do |loop|
    loop.times { :aBcDeF.casecmp(:abcdefg) }
  end
  x.report "Symbol#casecmp? ASCII" do |loop|
    loop.times { :aBcDeF.casecmp?(:abcdefg) }
  end
  x.report "Symbol#casecmp? UTF8" do |loop|
    loop.times { :"\u{e4 f6 fc}".casecmp?(:"\u{c4 d6 dc}") }
  end
end

Patch

https://github.com/ruby/ruby/pull/1668

[watson1978 (Shizuo Fujita)]

Because String#casecmp? duplicates object at rb_str_downcase() every time, so String#casecmp? is slower than String#casecmp?

[jeremyevans0 (Jeremy Evans)]

• Tracker changed from Bug to Feature
• Backport deleted (2.2: UNKNOWN, 2.3: UNKNOWN, 2.4: UNKNOWN)

[the_spectator (Akshay Birajdar)]

@koic (Koichi ITO) Made a new attempt with patch https://github.com/ruby/ruby/pull/2941

[sawa (Tsuyoshi Sawada)]

• Description updated (diff)

[naruse (Yui NARUSE)]

When you avoid that case, you have a option around coderange: coderange is a cached information whether the string contains (1) only ASCII 7 bit characters (2) also has 8 bit characters (3) broken byte sequence (4) unknown. Some strings are already scanned its coderange and caches it in a string object, but others are not. Whether this casecmp? optimization uses the cache and not scan string if the cache doesn't exist, or scan if it doesn't have a cache. If you use the cache, I wonder whether strings in real applications have cache or not. If you scan, I wonder if it still gets faster.

Imagine casecmp? with following chatacters:

• "a" * 100000 + "A"
• "a" * 100000 + "a"
• "a" * 100000 + "À"
• "a" * 100000 + "à"
• "ab"
  Using rb_enc_str_asciionly_p enforces to scan whole strings if it doesn't have coderange cache, and it can be overhead.

To avoid such trade off, I think you need to implement　integrated casecmp with rb_str_casemap.