Feature #17115: Optimize String#casecmp? for ASCII strings

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

[byroot (Jean Boussier)]

Patch: https://github.com/ruby/ruby/pull/3369

casecmp? is a kind of performance trap as it's much slower than using a case insensitive regexp or just casecmp == 0.

str = "Connection"
cmp = "connection"
Benchmark.ips do |x|
  x.report('/\A\z/i.match?') { /\Afoo\Z/i.match?(str) }
  x.report('casecmp?') { cmp.casecmp?(str) }
  x.report('casecmp') { cmp.casecmp(str) == 0 }
  x.compare!
end
Calculating -------------------------------------
      /\A\z/i.match?     11.447M (± 1.3%) i/s -     57.814M in   5.051489s
            casecmp?      6.197M (± 0.9%) i/s -     31.138M in   5.025252s
             casecmp     12.753M (± 1.2%) i/s -     64.636M in   5.069195s

Comparison:
             casecmp: 12752791.6 i/s
      /\A\z/i.match?: 11446996.1 i/s - 1.11x  (± 0.00) slower
            casecmp?:  6196886.0 i/s - 2.06x  (± 0.00) slower

This is because, unlike the others, it is sensitive to unicode case folding.

However, there are cases where fast case insensitive equality check of known ASCII strings is useful. For instance, matching HTTP headers.

This patch checks if both strings use a single byte encoding, and if so then does a simple iterative comparison with TOLOWER(). This makes casecmp? slightly faster than casecmp == 0 when both strings are ASCII.

|                        |compare-ruby|built-ruby|
|:-----------------------|-----------:|---------:|
|casecmp-1               |     11.618M|   10.757M|
|                        |       1.08x|         -|
|casecmp-10              |      1.849M|    1.723M|
|                        |       1.07x|         -|
|casecmp-100             |    204.490k|  186.798k|
|                        |       1.09x|         -|
|casecmp-1000            |     20.413k|   20.184k|
|                        |       1.01x|         -|
|casecmp-nonascii1       |     19.541M|   20.100M|
|                        |           -|     1.03x|
|casecmp-nonascii10      |     19.489M|   19.914M|
|                        |           -|     1.02x|
|casecmp-nonascii100     |     19.479M|   20.155M|
|                        |           -|     1.03x|
|casecmp-nonascii1000    |     19.462M|   20.064M|
|                        |           -|     1.03x|
|casecmp_p-1             |      2.214M|   12.030M|
|                        |           -|     5.43x|
|casecmp_p-10            |      1.373M|    2.150M|
|                        |           -|     1.57x|
|casecmp_p-100           |    249.292k|  231.041k|
|                        |       1.08x|         -|
|casecmp_p-1000          |     16.173k|   23.592k|
|                        |           -|     1.46x|
|casecmp_p-nonascii1     |    651.921k|  650.572k|
|                        |       1.00x|         -|
|casecmp_p-nonascii10    |    108.253k|  109.006k|
|                        |           -|     1.01x|
|casecmp_p-nonascii100   |     11.749k|   11.889k|
|                        |           -|     1.01x|
|casecmp_p-nonascii1000  |      1.140k|    1.138k|
|  

[byroot (Jean Boussier)]

• Description updated (diff)

[Dan0042 (Daniel DeLorme)]

In the benchmark you'd need to change the regexp from /\Afoo\Z/i to /\Aconnection\z/i; if you do so you'll find the regexp performance is similar to casecmp?

+1 for special-casing ASCII strings though.

Related: #13750, #14055

[sawa (Tsuyoshi Sawada)]

• Description updated (diff)