Implement HTML.escape(string) in terms of HTML.escape(string, io) #4674

chastell · 2017-07-04T21:45:27Z

While working on my PolyConf slides 😉 I discovered that HTML.escape(string) is implemeted in terms of String#gsub, rather than in terms of HTML.escape(string, io). The below benchmark (with string examples straight from this wonderful source) suggests implementing HTML.escape(string) in terms of HTML.escape(string, io) is significantly faster while not sacrificing much readability.

require "benchmark"
require "html"

strings = {
  no_html_short: "A paragraph without HTML characters that need to be escaped.",
  no_html_long: "Another paragraph without characters that need to be escaped. This paragraph is a bit longer, as sometimes there can be large paragraphs that don't any special characters, e.g., in novels or whatever.",
  html_short: "Here->An <<example>> of rust codefn foo(u: &u32) -> &u32 {u}",
  html_long: "A somewhat longer paragraph containing a character that needs to be escaped, because e.g. the author mentions the movie Fast&Furious in it. This paragraph is also quite long to match the non-html one.",
}

strings.each do |name, string|
  Benchmark.ips do |bench|
    bench.report("HTML.escape_io #{name}") do
      io = IO::Memory.new
      HTML.escape(string, io)
      io.to_s
    end
    bench.report("HTML.escape #{name}") do
      HTML.escape(string)
    end
  end
end

On current master (d24f79c), realigned for readability:

HTML.escape_io no_html_short   1.34M (743.83ns) (± 2.78%)       fastest
   HTML.escape no_html_short   1.02M (976.64ns) (± 3.55%)  1.31× slower
HTML.escape_io no_html_long  373.33k (  2.68µs) (± 1.70%)       fastest
   HTML.escape no_html_long  312.29k (   3.2µs) (± 1.73%)  1.20× slower
HTML.escape_io html_short      1.26M (793.38ns) (± 3.16%)       fastest
   HTML.escape html_short      1.03M (970.21ns) (± 2.64%)  1.22× slower
HTML.escape_io html_long     369.11k (  2.71µs) (± 2.09%)       fastest
   HTML.escape html_long     322.86k (   3.1µs) (± 1.50%)  1.14× slower

I’m not 100% sure what are the ramifications of choosing IO::Memory#to_s over String#gsub’s solution based on String.build(bytesize) (I’d guess potentially using up to twice the memory?), though, so do feel free to correct me on the usefulness of this, of course. ❤️

straight-shoota · 2017-07-04T22:11:48Z

The overhead seems to come from using String::Builder instead of IO::Memory which seems odd since it's purpose is to build strings...

straight-shoota · 2017-07-04T22:24:16Z

I've taken a look at the link where the samples come from... are these benchmarks seriously showing that in Rust a regex is faster than a plain loop with simple replacements??

akzhan · 2017-07-05T05:28:54Z

@straight-shoota big improvement by byte vs char replacement. good catch.

chastell · 2017-07-05T10:30:59Z

@straight-shoota

The overhead seems to come from using String::Builder instead of IO::Memory

It might also come from the longer callstack of HTML.escape(string) → String#gsub(hash) → String#gsub(&block) → String.build(capacity) → String::Builder.build(capacity).

It might be also that in the particular case of replacing HTML (single-byte) special characters with (multibyte) entities the String.build(bytesize) almost always allocates too few bytes upfront (but then IO::Memory doesn’t preallocate anything…).

are these benchmarks seriously showing that in Rust a regex is faster than a plain loop with simple replacements??

Yup. 🤷

sdogruyol · 2017-07-05T10:52:49Z

I'm really curious to learn how come Rust is faster in Regex @chastell ?

RX14 · 2017-07-05T12:07:27Z

@sdogruyol See the linked reddit thread here.

sdogruyol · 2017-07-05T12:08:17Z

thank you @RX14 👍

chastell · 2017-07-06T09:12:33Z

Sooo… anything else I should do with this PR? 😉

(I’ll happily look into String::Builder performance separately, this is super interesting!)

asterite · 2017-08-18T02:13:53Z

src/html.cr

@@ -30,7 +30,9 @@ module HTML
  # HTML.escape("Crystal & You") # => "Crystal &amp; You"
  # ```
  def self.escape(string : String) : String
-    string.gsub(SUBSTITUTIONS)
+    io = IO::Memory.new


It's better (more idiomatic) to use String.build { |io| escape(string, io) } here.

IO::Memory is somehow faster as the OP shows...

String.build is the idiomatic way to build strings. If alternatives are faster, then String.build should be improved (global overall benefit), instead of using alternatives in stdlib (duplications).

asterite · 2017-08-18T10:25:29Z

I'm getting different results:

HTML.escape_io no_html_short 780.35k (  1.28µs) (± 8.47%)  1.39× slower
   HTML.escape no_html_short   1.09M (920.18ns) (± 5.61%)       fastest
HTML.escape_io no_html_long 252.86k (  3.95µs) (± 9.09%)  1.16× slower
   HTML.escape no_html_long 293.18k (  3.41µs) (± 3.37%)       fastest
HTML.escape_io html_short 711.76k (   1.4µs) (± 8.08%)  1.19× slower
   HTML.escape html_short 846.82k (  1.18µs) (± 5.36%)       fastest
HTML.escape_io html_long 271.04k (  3.69µs) (± 1.76%)  1.03× slower
   HTML.escape html_long 279.72k (  3.58µs) (± 8.31%)       fastest

String::Builder can be a bit fast (it's missing overriding write_byte for performance), but I still see it faster than IO::Memory...

This is all on a MacBook Pro.

RX14 · 2017-08-18T10:38:51Z

Same here (linux, i5-3570k):

HTML.escape_io no_html_short   1.45M (691.44ns) (± 2.51%)  1.19× slower
   HTML.escape no_html_short   1.72M (580.67ns) (± 2.84%)       fastest
 HTML.escape_io no_html_long 409.16k (  2.44µs) (± 2.04%)  1.13× slower
    HTML.escape no_html_long 461.91k (  2.16µs) (± 4.16%)       fastest
   HTML.escape_io html_short    1.3M (770.96ns) (± 2.82%)  1.18× slower
      HTML.escape html_short   1.53M (652.86ns) (± 2.79%)       fastest
    HTML.escape_io html_long 419.95k (  2.38µs) (± 1.94%)  1.08× slower
       HTML.escape html_long  452.9k (  2.21µs) (± 3.69%)       fastest

asterite · 2017-09-28T20:22:11Z

Closing this because the PR doesn't seem to optimize the code (at least for me and @RX14 it results in slower code). It's also longer code.

Implement HTML.escape(string) in terms of HTML.escape(string, io)

a152a42

asterite requested changes Aug 18, 2017

View reviewed changes

asterite closed this Sep 28, 2017

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement HTML.escape(string) in terms of HTML.escape(string, io) #4674

Implement HTML.escape(string) in terms of HTML.escape(string, io) #4674

chastell commented Jul 4, 2017

straight-shoota commented Jul 4, 2017

straight-shoota commented Jul 4, 2017

akzhan commented Jul 5, 2017

chastell commented Jul 5, 2017

sdogruyol commented Jul 5, 2017

RX14 commented Jul 5, 2017

sdogruyol commented Jul 5, 2017

chastell commented Jul 6, 2017

asterite Aug 18, 2017

straight-shoota Aug 18, 2017

ysbaddaden Aug 18, 2017

asterite commented Aug 18, 2017

RX14 commented Aug 18, 2017 •

edited

Loading

asterite commented Sep 28, 2017

Implement HTML.escape(string) in terms of HTML.escape(string, io) #4674

Implement HTML.escape(string) in terms of HTML.escape(string, io) #4674

Conversation

chastell commented Jul 4, 2017

straight-shoota commented Jul 4, 2017

straight-shoota commented Jul 4, 2017

akzhan commented Jul 5, 2017

chastell commented Jul 5, 2017

sdogruyol commented Jul 5, 2017

RX14 commented Jul 5, 2017

sdogruyol commented Jul 5, 2017

chastell commented Jul 6, 2017

asterite Aug 18, 2017

Choose a reason for hiding this comment

straight-shoota Aug 18, 2017

Choose a reason for hiding this comment

ysbaddaden Aug 18, 2017

Choose a reason for hiding this comment

asterite commented Aug 18, 2017

RX14 commented Aug 18, 2017 • edited Loading

asterite commented Sep 28, 2017

RX14 commented Aug 18, 2017 •

edited

Loading