Some benchmarks are measured with wrong approach

从 code/array/length-vs-size-vs-count.rb 开始，这是 issue 中明确链接的 benchmark，并检查其 Benchmark.ips 报告块。将当前基于块的方法与这里描述的 benchmark-ips 字符串报告形式进行比较。当受影响的小操作 benchmark 使用开销更低的方法并且其结果已更新时，即视为完成；issue 没有列出所有受影响的文件。

tl;dr When we measure small things we have to remove all overheads which have impact. Wrapping with a method and block call is quite big overhead for most of benchmarks in this repository.

Suppose you want compare performance of "2 + 3" vs "2 * 3" calls. If you do it with a approach used in this repository you will get this results:

require "benchmark/ips"

def slow
  2 * 2
end

def fast
  2 + 2
end

Benchmark.ips do |x|
  x.report("2 * 2") { slow }
  x.report("2 + 2") { fast }
  x.compare!
end

(simplified output)

Comparison:
               2 + 2:  8304760.0 i/s
               2 * 2:  7535516.6 i/s - 1.10x slower

But there is one problem. Calling a method + surrounding block ({ slow }) has bigger overhead than calling Fixnum#+ or Fixnum#* itself. Thea easiest way to observe it is to repeat benchmarked operations in one call. Like this:

require "benchmark/ips"

def slow
  2*2; 2*2; 2*2; 2*2; 2*2; 2*2; 2*2; 2*2; 2*2; 2*2;
end

def fast
  2+2; 2+2; 2+2; 2+2; 2+2; 2+2; 2+2; 2+2; 2+2; 2+2;
end

Benchmark.ips do |x|
  x.report("2 * 2") { slow }
  x.report("2 + 2") { fast }
  x.compare!
end

Comparison:
               2 + 2:  4680545.3 i/s
               2 * 2:  3468681.3 i/s - 1.35x slower

See how results changed? Writing our benchmarks in this way would be quite problematic. Fortunately benchmark-ips gem has answer for that. Benchmark::IPS::Job#report method allows to pass string which will be compiled before benchmark is run. Passing right string allows to measure it properly:

require "benchmark/ips"

Benchmark.ips do |x|
  x.report("2 * 2", "2 * 2;" * 1_000)
  x.report("2 + 2", "2 + 2;" * 1_000)
  x.compare!
end

Comparison:
               2 + 2:    91567.5 i/s
               2 * 2:    57994.4 i/s - 1.58x slower

This is greatly explained here: https://docs.omniref.com/ruby/2.2.1/symbols/Benchmark/bm#annotation=4095926&line=182

How does this affect fast-ruby benchmarks? All benchmarks that call small things are flawed. One of them is Array#length vs Array#size vs Array#count benchmark. Here is the original code and result obtained on my computer:

require 'benchmark/ips'

ARRAY = [*1..100]

Benchmark.ips do |x|
  x.report("Array#length") { ARRAY.length }
  x.report("Array#size") { ARRAY.size }
  x.report("Array#count") { ARRAY.count }
  x.compare!
end

Comparison:
          Array#size:  8679483.2 i/s
        Array#length:  8664450.7 i/s - 1.00x slower
         Array#count:  7237299.5 i/s - 1.20x slower

The same benchmark measure with described approach gives different numbers:

require 'benchmark/ips'

ARRAY = [*1..100]

Benchmark.ips do |x|
  x.report("Array#length", "ARRAY.length;" * 1_000)
  x.report("Array#size",   "ARRAY.size;"   * 1_000)
  x.report("Array#count",  "ARRAY.count;"  * 1_000)
  x.compare!
end

Comparison:
          Array#size:   113902.4 i/s
        Array#length:   113655.9 i/s - 1.00x slower
         Array#count:    28753.4 i/s - 3.96x slower

Difference: 1.20x slower vs 3.96x slower.

My guess is that it affects most of benchmarks.