Fixes #2832

This reverts commit 0df27c98.

Reverted due to failing test, see #2845.

Replace split on comma with a regexp that will reverse all ASC/DESC specifically

The join_nodes must be passed to the JoinDependency initializer and therefore counted by the alias tracker. This is because the association_joins are aliased on initialization and then the tables are cached, so it is no use to alias the join_nodes later. Fixes #2556.

This reverts commit 9d396ee8, reversing
changes made to fa2bfd83.

Reason: the change broke the build.

This reverts commit 04cc446d.

I reverted it because apparently we want to use: select([:a, :b])
instead of select(:a, :b), but there was no tests for that form.

Make it the responsibility of the connection to hold onto an ARel visitor for generating SQL. This improves the code architecture generally, and solves some problems with marshalling. Adapter authors please take note: you now need to define an Adapter.visitor_for method, but it degrades gracefully with a deprecation warning for now.

This reverts commit 9d396ee8, reversing
changes made to fa2bfd83.

Reason: the change broke the build.

This reverts commit 04cc446d.

I reverted it because apparently we want to use: select([:a, :b])
instead of select(:a, :b), but there was no tests for that form.

having raises NoMethodError: undefined method `empty?' when a Fixnum or Date/Time were passed via varargs

This caused that `find_each` was producing extra db call taking all the records from db, and was less efficient than `ActiveRecord::Base#all`.

In order make Relation behavior closer to Array
Made Relation#sum to accept block and delegate it to Array#sum

default create_with_value to a hash so we can eliminate conditionals, add test surrounding create_with(nil) behavior

Currently if the code that calls .find_in_batches modifies the yielded array in place then .find_in_batches can enter an infinite loop searching with ruby object ids in the database instead of the primary key of records in the database. This happens because it naively assumes the yielded array hasn't been modified before calling #id on the last object in the array. And ruby (1.8 at least) alias' #id to #object_id so an integer is still returned no matter what the last object is.

By moving finding the #id of the last object before yielding the array it means the calling code can do whatever it wants to the array in terms of modifying it in place, and .find_in_batches doesn't care.

Set reverse_order_value when asked to reverse_order().
Do the actual reversal in build_arel.

Ensure that calculations properly override the select value even if it's set in the default scope. Fixes #1395.

Addresses an inconsistency in the ActiveRecord::Base.method_missing handling of dynamic finder methods and the passing of the &block parameter for :find_by_attributes.

- If doing a count on a relation that has an :include and a :join, it does a distinct even though it should not.
- When doing a count on a relation that has an :include, it always falls back to a old style left join when performing the count.  Looks like it was broken here:
https://github.com/rails/rails/commit/b9599502c9e738a5a1513e75d08f8d40ed408265

if a query contains a limit or an offset, ActiveRecord::FinderMethods#find_last had inconsistent behavior.

If the records were loaded, it returned the last record in the cached list.
If they were not, it reversed the order of the query and changed the limit to one.
If the earlier limit was less than the total matching the query in the db,
it would return a different record than if the records had been cached.

This commit changes find_last so that it loads the records when getting the
last record on a query containing a limit or an offset, which makes the behavior consistent.