Speed up homogeneous AR lists / reduce allocations

[tenderlove]

This commit speeds up allocating homogeneous lists of AR objects. We
can know if the result set contains an STI column before initializing
every AR object, so this change pulls the "does this result set contain
an STI column?" test up, then uses a specialized instantiation function.
This way we only have to check for an STI column once rather than N
times.

This change also introduces a new initialization function that is meant
for use when allocating AR objects that come from the database. Doing
this allows us to eliminate one hash allocation per AR instance.

Here is a benchmark:

require 'active_record'
require 'benchmark/ips'

ActiveRecord::Base.establish_connection adapter: "sqlite3", database: ":memory:"

ActiveRecord::Migration.verbose = false

ActiveRecord::Schema.define do
  create_table :users, force: true do |t|
    t.string :name
    t.timestamps null: false
  end
end

class User < ActiveRecord::Base; end

2000.times do
  User.create!(name: "Gorby")
end

Benchmark.ips do |x|
  x.report("find") do
    User.limit(2000).to_a
  end
end

Results:

Before:

[aaron@TC activerecord (master)]$ be ruby -I lib:~/git/allocation_tracer/lib speed.rb
Warming up --------------------------------------
                find     5.000  i/100ms
Calculating -------------------------------------
                find     56.192  (± 3.6%) i/s -    285.000  in   5.080940s

After:

[aaron@TC activerecord (homogeneous-allocation)]$ be ruby -I lib:~/git/allocation_tracer/lib speed.rb
Warming up --------------------------------------
                find     7.000  i/100ms
Calculating -------------------------------------
                find     72.204  (± 2.8%) i/s -    364.000  in   5.044592s

[utilum]

double space after # database. redundant?

[tenderlove]

I'm not sure. We learned double space after punctuation in school because everything was fixed width font at that time. These days we don't need it because most stuff (except code) isn't fixed width. Since code is fixed width, I go with double space but I'm not sure what our style guide is in this case. @fxn?

[rafaelfranca]

Some questions about the documentation, but the implementation looks good.

[rafaelfranca]

I think it is +instantiate_instance_of+ instead of instantiate here. But I guess we can remove that part of the documentation if we rephrase it.

[rafaelfranca]

To follow the same convention on this same class we should do:

# Returns true if this result set include the column named +name+

[tenderlove]

@rafaelfranca how about Returns truthy if this result set include the column named +name+. I don't want to guarantee the return value is true or false if we decide to change implementation (I remember there was some controversy because we changed a method to return a truthy value). @fxn thoughts?

[rafaelfranca]

true in our guideline means truthy. +true+ means the instance of TrueClass.

http://guides.rubyonrails.org/api_documentation_guidelines.html#booleans

[fxn]

I am not really happy with our convention (that I defined myself many years ago), which is to use fixed-width font for the singletons, and regular font to mean "true" in the Ruby sense of the word (i.e., "truthy"). My thoughts by then were that Ruby defines what is true and false, not what is truthy and falsey, so I stuck with that jargon for respect.

In retrospect, that choice was too subtle and time has shown it doesn't communicate well in practice. It has also the problem that in sample code like

user.blank? # => true

you cannot distinguish if that is regular or fixed-width, therefore people assume the singleton true sometimes and that has made us change some return contracts, to be stricter.

I suggested to change this convention in Basecamp a while back, but it was agreed to leave it like it is.

[rafaelfranca]

Do we want/expect people to call this method on their applications? If not this should be # :nodoc:

[tenderlove]

I don't think we want people to call this. I'll :nodoc: it.

[rafaelfranca]

👍