The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.
package DBIx::Class::ResultClass::HashRefInflator;

use strict;
use warnings;

=head1 NAME

DBIx::Class::ResultClass::HashRefInflator

=head1 SYNOPSIS

 use DBIx::Class::ResultClass::HashRefInflator;

 my $rs = $schema->resultset('CD');
 $rs->result_class('DBIx::Class::ResultClass::HashRefInflator');
 while (my $hashref = $rs->next) {
    ...
 }

=head1 DESCRIPTION

DBIx::Class is faster than older ORMs like Class::DBI but it still isn't 
designed primarily for speed. Sometimes you need to quickly retrieve the data
from a massive resultset, while skipping the creation of fancy row objects.
Specifying this class as a C<result_class> for a resultset will change C<< $rs->next >>
to return a plain data hash-ref (or a list of such hash-refs if C<< $rs->all >> is used).

There are two ways of applying this class to a resultset:

=over

=item *

Specify C<< $rs->result_class >> on a specific resultset to affect only that
resultset (and any chained off of it); or

=item *

Specify C<< __PACKAGE__->result_class >> on your source object to force all
uses of that result source to be inflated to hash-refs - this approach is not
recommended.

=back

=cut

##############
# NOTE
#
# Generally people use this to gain as much speed as possible. If a new &mk_hash is
# implemented, it should be benchmarked using the maint/benchmark_hashrefinflator.pl
# script (in addition to passing all tests of course :). Additional instructions are
# provided in the script itself.
#

# This coderef is a simple recursive function
# Arguments: ($me, $prefetch) from inflate_result() below
my $mk_hash;
$mk_hash = sub {
    if (ref $_[0] eq 'ARRAY') {     # multi relationship
        return [ map { $mk_hash->(@$_) || () } (@_) ];
    }
    else {
        my $hash = {
            # the main hash could be an undef if we are processing a skipped-over join
            $_[0] ? %{$_[0]} : (),

            # the second arg is a hash of arrays for each prefetched relation
            map
                { $_ => $mk_hash->( @{$_[1]->{$_}} ) }
                ( $_[1] ? (keys %{$_[1]}) : () )
        };

        # if there is at least one defined column consider the resultset real
        # (and not an emtpy has_many rel containing one empty hashref)
        for (values %$hash) {
            return $hash if defined $_;
        }

        return undef;
    }
};

=head1 METHODS

=head2 inflate_result

Inflates the result and prefetched data into a hash-ref (invoked by L<DBIx::Class::ResultSet>)

=cut

##################################################################################
# inflate_result is invoked as:
# HRI->inflate_result ($resultsource_instance, $main_data_hashref, $prefetch_data_hashref)
sub inflate_result {
    return $mk_hash->($_[2], $_[3]);
}


=head1 CAVEATS

=over

=item *

This will not work for relationships that have been prefetched. Consider the
following:

 my $artist = $artitsts_rs->search({}, {prefetch => 'cds' })->first;

 my $cds = $artist->cds;
 $cds->result_class('DBIx::Class::ResultClass::HashRefInflator');
 my $first = $cds->first; 

C<$first> will B<not> be a hashref, it will be a normal CD row since 
HashRefInflator only affects resultsets at inflation time, and prefetch causes
relations to be inflated when the master C<$artist> row is inflated.

=back

=cut

1;