X-Git-Url: http://git.shadowcat.co.uk/gitweb/gitweb.cgi?a=blobdiff_plain;f=lib%2FDBIx%2FClass%2FStorage%2FDBIHacks.pm;h=14410b700ddb47db9f593c7e811bb60773a12fef;hb=d52c4a75538bd15d1fa531a63c6ec45b8943dfd0;hp=d4f4058b347f78f0abd39c27d84ab02a075f1e60;hpb=975b573aff1f568f24951ca574fc636de396e506;p=dbsrgits%2FDBIx-Class.git diff --git a/lib/DBIx/Class/Storage/DBIHacks.pm b/lib/DBIx/Class/Storage/DBIHacks.pm index d4f4058..14410b7 100644 --- a/lib/DBIx/Class/Storage/DBIHacks.pm +++ b/lib/DBIx/Class/Storage/DBIHacks.pm @@ -2,9 +2,24 @@ package #hide from PAUSE DBIx::Class::Storage::DBIHacks; # -# This module contains code that should never have seen the light of day, -# does not belong in the Storage, or is otherwise unfit for public -# display. The arrival of SQLA2 should immediately obsolete 90% of this +# This module contains code supporting a battery of special cases and tests for +# many corner cases pushing the envelope of what DBIC can do. When work on +# these utilities began in mid 2009 (51a296b402c) it wasn't immediately obvious +# that these pieces, despite their misleading on-first-sighe-flakiness, will +# become part of the generic query rewriting machinery of DBIC, allowing it to +# both generate and process queries representing incredibly complex sets with +# reasonable efficiency. +# +# Now (end of 2015), more than 6 years later the routines in this class have +# stabilized enough, and are meticulously covered with tests, to a point where +# an effort to formalize them into user-facing APIs might be worthwhile. +# +# An implementor working on publicizing and/or replacing the routines with a +# more modern SQL generation framework should keep in mind that pretty much all +# existing tests are constructed on the basis of real-world code used in +# production somewhere. +# +# Please hack on this responsibly ;) # use strict; @@ -15,7 +30,9 @@ use mro 'c3'; use List::Util 'first'; use Scalar::Util 'blessed'; -use Sub::Name 'subname'; +use DBIx::Class::_Util qw(UNRESOLVABLE_CONDITION serialize); +use SQL::Abstract qw(is_plain_value is_literal_value); +use DBIx::Class::Carp; use namespace::clean; # @@ -23,30 +40,47 @@ use namespace::clean; # {from} specs, aiding the RDBMS query optimizer # sub _prune_unused_joins { - my $self = shift; - my ($from, $select, $where, $attrs) = @_; + my ($self, $attrs) = @_; - return $from unless $self->_use_join_optimizer; + # only standard {from} specs are supported, and we could be disabled in general + return ($attrs->{from}, {}) unless ( + ref $attrs->{from} eq 'ARRAY' + and + @{$attrs->{from}} > 1 + and + ref $attrs->{from}[0] eq 'HASH' + and + ref $attrs->{from}[1] eq 'ARRAY' + and + $self->_use_join_optimizer + ); - if (ref $from ne 'ARRAY' || ref $from->[0] ne 'HASH' || ref $from->[1] ne 'ARRAY') { - return $from; # only standard {from} specs are supported - } + my $orig_aliastypes = + $attrs->{_precalculated_aliastypes} + || + $self->_resolve_aliastypes_from_select_args($attrs) + ; - my $aliastypes = $self->_resolve_aliastypes_from_select_args(@_); + my $new_aliastypes = { %$orig_aliastypes }; - my $orig_joins = delete $aliastypes->{joining}; - my $orig_multiplying = $aliastypes->{multiplying}; + # we will be recreating this entirely + my @reclassify = 'joining'; # a grouped set will not be affected by amount of rows. Thus any - # {multiplying} joins can go - delete $aliastypes->{multiplying} + # purely multiplicator classifications can go + # (will be reintroduced below if needed by something else) + push @reclassify, qw(multiplying premultiplied) if $attrs->{_force_prune_multiplying_joins} or $attrs->{group_by}; - my @newfrom = $from->[0]; # FROM head is always present + # nuke what will be recalculated + delete @{$new_aliastypes}{@reclassify}; - my %need_joins; + my @newfrom = $attrs->{from}[0]; # FROM head is always present - for (values %$aliastypes) { + # recalculate what we need once the multipliers are potentially gone + # ignore premultiplies, since they do not add any value to anything + my %need_joins; + for ( @{$new_aliastypes}{grep { $_ ne 'premultiplied' } keys %$new_aliastypes }) { # add all requested aliases $need_joins{$_} = 1 for keys %$_; @@ -54,7 +88,7 @@ sub _prune_unused_joins { $need_joins{$_} = 1 for map { values %$_ } map { @{$_->{-parents}} } values %$_; } - for my $j (@{$from}[1..$#$from]) { + for my $j (@{$attrs->{from}}[1..$#{$attrs->{from}}]) { push @newfrom, $j if ( (! defined $j->[0]{-alias}) # legacy crap || @@ -62,11 +96,16 @@ sub _prune_unused_joins { ); } - return ( \@newfrom, { - multiplying => { map { $need_joins{$_} ? ($_ => $orig_multiplying->{$_}) : () } keys %$orig_multiplying }, - %$aliastypes, - joining => { map { $_ => $orig_joins->{$_} } keys %need_joins }, - } ); + # we have a new set of joiners - for everything we nuked pull the classification + # off the original stack + for my $ctype (@reclassify) { + $new_aliastypes->{$ctype} = { map + { $need_joins{$_} ? ( $_ => $orig_aliastypes->{$ctype}{$_} ) : () } + keys %{$orig_aliastypes->{$ctype}} + } + } + + return ( \@newfrom, $new_aliastypes ); } # @@ -74,19 +113,26 @@ sub _prune_unused_joins { # SELECT me.*, other.* FROM ( SELECT me.* FROM ... ) JOIN other ON ... # sub _adjust_select_args_for_complex_prefetch { - my ($self, $from, $select, $where, $attrs) = @_; + my ($self, $attrs) = @_; - $self->throw_exception ('Complex prefetches are not supported on resultsets with a custom from attribute') - if (ref $from ne 'ARRAY' || ref $from->[0] ne 'HASH' || ref $from->[1] ne 'ARRAY'); + $self->throw_exception ('Complex prefetches are not supported on resultsets with a custom from attribute') unless ( + ref $attrs->{from} eq 'ARRAY' + and + @{$attrs->{from}} > 1 + and + ref $attrs->{from}[0] eq 'HASH' + and + ref $attrs->{from}[1] eq 'ARRAY' + ); my $root_alias = $attrs->{alias}; # generate inner/outer attribute lists, remove stuff that doesn't apply my $outer_attrs = { %$attrs }; - delete @{$outer_attrs}{qw(where bind rows offset group_by _grouped_by_distinct having)}; + delete @{$outer_attrs}{qw(from bind rows offset group_by _grouped_by_distinct having)}; - my $inner_attrs = { %$attrs }; - delete @{$inner_attrs}{qw(from for collapse select as _related_results_construction)}; + my $inner_attrs = { %$attrs, _simple_passthrough_construction => 1 }; + delete @{$inner_attrs}{qw(for collapse select as)}; # there is no point of ordering the insides if there is no limit delete $inner_attrs->{order_by} if ( @@ -98,13 +144,12 @@ sub _adjust_select_args_for_complex_prefetch { # generate the inner/outer select lists # for inside we consider only stuff *not* brought in by the prefetch # on the outside we substitute any function for its alias - my $outer_select = [ @$select ]; - my $inner_select; + $outer_attrs->{select} = [ @{$attrs->{select}} ]; my ($root_node, $root_node_offset); - for my $i (0 .. $#$from) { - my $node = $from->[$i]; + for my $i (0 .. $#{$inner_attrs->{from}}) { + my $node = $inner_attrs->{from}[$i]; my $h = (ref $node eq 'HASH') ? $node : (ref $node eq 'ARRAY' and ref $node->[0] eq 'HASH') ? $node->[0] : next @@ -121,11 +166,11 @@ sub _adjust_select_args_for_complex_prefetch { unless $root_node; # use the heavy duty resolver to take care of aliased/nonaliased naming - my $colinfo = $self->_resolve_column_info($from); + my $colinfo = $self->_resolve_column_info($inner_attrs->{from}); my $selected_root_columns; - for my $i (0 .. $#$outer_select) { - my $sel = $outer_select->[$i]; + for my $i (0 .. $#{$outer_attrs->{select}}) { + my $sel = $outer_attrs->{select}->[$i]; next if ( $colinfo->{$sel} and $colinfo->{$sel}{-source_alias} ne $root_alias @@ -133,30 +178,38 @@ sub _adjust_select_args_for_complex_prefetch { if (ref $sel eq 'HASH' ) { $sel->{-as} ||= $attrs->{as}[$i]; - $outer_select->[$i] = join ('.', $root_alias, ($sel->{-as} || "inner_column_$i") ); + $outer_attrs->{select}->[$i] = join ('.', $root_alias, ($sel->{-as} || "inner_column_$i") ); } elsif (! ref $sel and my $ci = $colinfo->{$sel}) { $selected_root_columns->{$ci->{-colname}} = 1; } - push @$inner_select, $sel; + push @{$inner_attrs->{select}}, $sel; push @{$inner_attrs->{as}}, $attrs->{as}[$i]; } - # We will need to fetch all native columns in the inner subquery, which may + my $inner_aliastypes = $self->_resolve_aliastypes_from_select_args($inner_attrs); + + # In the inner subq we will need to fetch *only* native columns which may # be a part of an *outer* join condition, or an order_by (which needs to be - # preserved outside) + # preserved outside), or wheres. In other words everything but the inner + # selector # We can not just fetch everything because a potential has_many restricting # join collapse *will not work* on heavy data types. - my $connecting_aliastypes = $self->_resolve_aliastypes_from_select_args( - $from, - [], - $where, - $inner_attrs - ); - for (sort map { keys %{$_->{-seen_columns}||{}} } map { values %$_ } values %$connecting_aliastypes) { + # essentially a map of all non-selecting seen columns + # the sort is there for a nicer select list + for ( + sort + map + { keys %{$_->{-seen_columns}||{}} } + map + { values %{$inner_aliastypes->{$_}} } + grep + { $_ ne 'selecting' } + keys %$inner_aliastypes + ) { my $ci = $colinfo->{$_} or next; if ( $ci->{-source_alias} eq $root_alias @@ -164,150 +217,67 @@ sub _adjust_select_args_for_complex_prefetch { ! $selected_root_columns->{$ci->{-colname}}++ ) { # adding it to both to keep limits not supporting dark selectors happy - push @$inner_select, $ci->{-fq_colname}; + push @{$inner_attrs->{select}}, $ci->{-fq_colname}; push @{$inner_attrs->{as}}, $ci->{-fq_colname}; } } - # construct the inner $from and lock it in a subquery + # construct the inner {from} and lock it in a subquery # we need to prune first, because this will determine if we need a group_by below # throw away all non-selecting, non-restricting multijoins - # (since we def. do not care about multiplication those inside the subquery) + # (since we def. do not care about multiplication of the contents of the subquery) my $inner_subq = do { - # must use it here regardless of user requests + # must use it here regardless of user requests (vastly gentler on optimizer) local $self->{_use_join_optimizer} = 1; # throw away multijoins since we def. do not care about those inside the subquery - my ($inner_from, $inner_aliastypes) = $self->_prune_unused_joins ($from, $inner_select, $where, { - %$inner_attrs, _force_prune_multiplying_joins => 1 + # $inner_aliastypes *will* be redefined at this point + ($inner_attrs->{from}, $inner_aliastypes ) = $self->_prune_unused_joins ({ + %$inner_attrs, + _force_prune_multiplying_joins => 1, + _precalculated_aliastypes => $inner_aliastypes, }); - # uh-oh a multiplier (which is not us) left in, this is a problem + # uh-oh a multiplier (which is not us) left in, this is a problem for limits + # we will need to add a group_by to collapse the resultset for proper counts if ( - $inner_aliastypes->{multiplying} + grep { $_ ne $root_alias } keys %{ $inner_aliastypes->{multiplying} || {} } and # if there are user-supplied groups - assume user knows wtf they are up to ( ! $inner_aliastypes->{grouping} or $inner_attrs->{_grouped_by_distinct} ) - and - my @multipliers = grep { $_ ne $root_alias } keys %{$inner_aliastypes->{multiplying}} ) { - # if none of the multipliers came from an order_by (guaranteed to have been combined - # with a limit) - easy - just slap a group_by to simulate a collape and be on our way - if ( - ! $inner_aliastypes->{ordering} - or - ! first { $inner_aliastypes->{ordering}{$_} } @multipliers - ) { - - my $unprocessed_order_chunks; - ($inner_attrs->{group_by}, $unprocessed_order_chunks) = $self->_group_over_selection ({ - %$inner_attrs, - from => $inner_from, - select => $inner_select, - }); - - $self->throw_exception ( - 'A required group_by clause could not be constructed automatically due to a complex ' - . 'order_by criteria. Either order_by columns only (no functions) or construct a suitable ' - . 'group_by by hand' - ) if $unprocessed_order_chunks; + my $cur_sel = { map { $_ => 1 } @{$inner_attrs->{select}} }; + + # *possibly* supplement the main selection with pks if not already + # there, as they will have to be a part of the group_by to collapse + # things properly + my $inner_select_with_extras; + my @pks = map { "$root_alias.$_" } $root_node->{-rsrc}->primary_columns + or $self->throw_exception( sprintf + 'Unable to perform complex limited prefetch off %s without declared primary key', + $root_node->{-rsrc}->source_name, + ); + for my $col (@pks) { + push @{ $inner_select_with_extras ||= [ @{$inner_attrs->{select}} ] }, $col + unless $cur_sel->{$col}++; } - else { - # We need to order by external columns and group at the same time - # so we can calculate the proper limit - # This doesn't really make sense in SQL, however from DBICs point - # of view is rather valid (order the leftmost objects by whatever - # criteria and get the offset/rows many). There is a way around - # this however in SQL - we simply tae the direction of each piece - # of the foreign order and convert them to MIN(X) for ASC or MAX(X) - # for DESC, and group_by the root columns. The end result should be - # exactly what we expect - - # supplement the main selection with pks if not already there, - # as they will have to be a part of the group_by to colapse - # things properly - my $cur_sel = { map { $_ => 1 } @$inner_select }; - - my @pks = map { "$root_alias.$_" } $root_node->{-rsrc}->primary_columns - or $self->throw_exception( sprintf - 'Unable to perform complex limited prefetch off %s without declared primary key', - $root_node->{-rsrc}->source_name, - ); - for my $col (@pks) { - push @$inner_select, $col - unless $cur_sel->{$col}++; - } - # wrap any part of the order_by that "responds" to an ordering alias - # into a MIN/MAX - # FIXME - this code is a joke, will need to be completely rewritten in - # the DQ branch. But I need to push a POC here, otherwise the - # pesky tests won't pass - my $sql_maker = $self->sql_maker; - my ($lquote, $rquote, $sep) = map { quotemeta $_ } ($sql_maker->_quote_chars, $sql_maker->name_sep); - my $own_re = qr/ $lquote \Q$root_alias\E $rquote $sep | \b \Q$root_alias\E $sep /x; - my @order_chunks = map { ref $_ eq 'ARRAY' ? $_ : [ $_ ] } $sql_maker->_order_by_chunks($attrs->{order_by}); - my @new_order = map { \$_ } @order_chunks; - my $inner_columns_info = $self->_resolve_column_info($inner_from); - - # loop through and replace stuff that is not "ours" with a min/max func - # everything is a literal at this point, since we are likely properly - # quoted and stuff - for my $i (0 .. $#new_order) { - my $chunk = $order_chunks[$i][0]; - - # skip ourselves - next if $chunk =~ $own_re; - - ($chunk, my $is_desc) = $sql_maker->_split_order_chunk($chunk); - - # maybe our own unqualified column - my $ord_bit = ( - $lquote and $sep and $chunk =~ /^ $lquote ([^$sep]+) $rquote $/x - ) ? $1 : $chunk; - - next if ( - $ord_bit - and - $inner_columns_info->{$ord_bit} - and - $inner_columns_info->{$ord_bit}{-source_alias} eq $root_alias - ); - - $new_order[$i] = \[ - sprintf( - '%s(%s)%s', - ($is_desc ? 'MAX' : 'MIN'), - $chunk, - ($is_desc ? ' DESC' : ''), - ), - @ {$order_chunks[$i]} [ 1 .. $#{$order_chunks[$i]} ] - ]; - } - - $inner_attrs->{order_by} = \@new_order; - - # do not care about leftovers here - it will be all the functions - # we just created - ($inner_attrs->{group_by}) = $self->_group_over_selection ({ - %$inner_attrs, - from => $inner_from, - select => $inner_select, - }); - } + ($inner_attrs->{group_by}, $inner_attrs->{order_by}) = $self->_group_over_selection({ + %$inner_attrs, + $inner_select_with_extras ? ( select => $inner_select_with_extras ) : (), + _aliastypes => $inner_aliastypes, + }); } - # we already optimized $inner_from above + # we already optimized $inner_attrs->{from} above # and already local()ized $self->{_use_join_optimizer} = 0; # generate the subquery $self->_select_args_to_query ( - $inner_from, - $inner_select, - $where, + @{$inner_attrs}{qw(from select where)}, $inner_attrs, ); }; @@ -323,22 +293,25 @@ sub _adjust_select_args_for_complex_prefetch { # result by tackling yet another group_by to the outside of the query # work on a shallow copy - $from = [ @$from ]; + my @orig_from = @{$attrs->{from}}; - my @outer_from; + + $outer_attrs->{from} = \ my @outer_from; # we may not be the head if ($root_node_offset) { - # first generate the outer_from, up and including the substitution point - @outer_from = splice @$from, 0, $root_node_offset; + # first generate the outer_from, up to the substitution point + @outer_from = splice @orig_from, 0, $root_node_offset; + # substitute the subq at the right spot push @outer_from, [ { -alias => $root_alias, -rsrc => $root_node->{-rsrc}, $root_alias => $inner_subq, }, - @{$from->[0]}[1 .. $#{$from->[0]}], + # preserve attrs from what is now the head of the from after the splice + @{$orig_from[0]}[1 .. $#{$orig_from[0]}], ]; } else { @@ -349,12 +322,12 @@ sub _adjust_select_args_for_complex_prefetch { }; } - shift @$from; # what we just replaced above + shift @orig_from; # what we just replaced above # scan the *remaining* from spec against different attributes, and see which joins are needed # in what role my $outer_aliastypes = $outer_attrs->{_aliastypes} = - $self->_resolve_aliastypes_from_select_args( $from, $outer_select, $where, $outer_attrs ); + $self->_resolve_aliastypes_from_select_args({ %$outer_attrs, from => \@orig_from }); # unroll parents my ($outer_select_chain, @outer_nonselecting_chains) = map { +{ @@ -362,10 +335,8 @@ sub _adjust_select_args_for_complex_prefetch { } } qw/selecting restricting grouping ordering/; # see what's left - throw away if not selecting/restricting - # also throw in a group_by if a non-selecting multiplier, - # to guard against cross-join explosions - my $need_outer_group_by; - while (my $j = shift @$from) { + my $may_need_outer_group_by; + while (my $j = shift @orig_from) { my $alias = $j->[0]{-alias}; if ( @@ -375,82 +346,108 @@ sub _adjust_select_args_for_complex_prefetch { } elsif (first { $_->{$alias} } @outer_nonselecting_chains ) { push @outer_from, $j; - $need_outer_group_by ||= $outer_aliastypes->{multiplying}{$alias} ? 1 : 0; + $may_need_outer_group_by ||= $outer_aliastypes->{multiplying}{$alias} ? 1 : 0; } } - if ( $need_outer_group_by and $attrs->{_grouped_by_distinct} ) { - my $unprocessed_order_chunks; - ($outer_attrs->{group_by}, $unprocessed_order_chunks) = $self->_group_over_selection ({ + # also throw in a synthetic group_by if a non-selecting multiplier, + # to guard against cross-join explosions + # the logic is somewhat fragile, but relies on the idea that if a user supplied + # a group by on their own - they know what they were doing + if ( $may_need_outer_group_by and $attrs->{_grouped_by_distinct} ) { + ($outer_attrs->{group_by}, $outer_attrs->{order_by}) = $self->_group_over_selection ({ %$outer_attrs, from => \@outer_from, - select => $outer_select, }); - - $self->throw_exception ( - 'A required group_by clause could not be constructed automatically due to a complex ' - . 'order_by criteria. Either order_by columns only (no functions) or construct a suitable ' - . 'group_by by hand' - ) if $unprocessed_order_chunks; - } - # This is totally horrific - the $where ends up in both the inner and outer query - # Unfortunately not much can be done until SQLA2 introspection arrives, and even - # then if where conditions apply to the *right* side of the prefetch, you may have - # to both filter the inner select (e.g. to apply a limit) and then have to re-filter - # the outer select to exclude joins you didin't want in the first place + # FIXME: The {where} ends up in both the inner and outer query, i.e. *twice* + # + # This is rather horrific, and while we currently *do* have enough + # introspection tooling available to attempt a stab at properly deciding + # whether or not to include the where condition on the outside, the + # machinery is still too slow to apply it here. + # Thus for the time being we do not attempt any sanitation of the where + # clause and just pass it through on both sides of the subquery. This *will* + # be addressed at a later stage, most likely after folding the SQL generator + # into SQLMaker proper # # OTOH it can be seen as a plus: (notes that this query would make a DBA cry ;) - return (\@outer_from, $outer_select, $where, $outer_attrs); + # + return $outer_attrs; } +# This is probably the ickiest, yet most relied upon part of the codebase: +# this is the place where we take arbitrary SQL input and break it into its +# constituent parts, making sure we know which *sources* are used in what +# *capacity* ( selecting / restricting / grouping / ordering / joining, etc ) +# Although the method is pretty horrific, the worst thing that can happen is +# for a classification failure, which in turn will result in a vocal exception, +# and will lead to a relatively prompt fix. +# The code has been slowly improving and is covered with a formiddable battery +# of tests, so can be considered "reliably stable" at this point (Oct 2015). +# +# A note to implementors attempting to "replace" this - keep in mind that while +# there are multiple optimization avenues, the actual "scan literal elements" +# part *MAY NEVER BE REMOVED*, even if it is limited only ot the (future) AST +# nodes that are deemed opaque (i.e. contain literal expressions). The use of +# blackbox literals is at this point firmly a user-facing API, and is one of +# *the* reasons DBIC remains as flexible as it is. In other words, when working +# on this keep in mind that the following is widespread and *encouraged* way +# of using DBIC in the wild when push comes to shove: +# +# $rs->search( {}, { +# select => \[ $random, @stuff], +# from => \[ $random, @stuff ], +# where => \[ $random, @stuff ], +# group_by => \[ $random, @stuff ], +# order_by => \[ $random, @stuff ], +# } ) # -# I KNOW THIS SUCKS! GET SQLA2 OUT THE DOOR SO THIS CAN DIE! +# Various incarnations of the above are reflected in many of the tests. If one +# gets to fail, you get to fix it. A "this is crazy, nobody does that" is not +# acceptable going forward. # -# Due to a lack of SQLA2 we fall back to crude scans of all the -# select/where/order/group attributes, in order to determine what -# aliases are neded to fulfill the query. This information is used -# throughout the code to prune unnecessary JOINs from the queries -# in an attempt to reduce the execution time. -# Although the method is pretty horrific, the worst thing that can -# happen is for it to fail due to some scalar SQL, which in turn will -# result in a vocal exception. sub _resolve_aliastypes_from_select_args { - my ( $self, $from, $select, $where, $attrs ) = @_; + my ( $self, $attrs ) = @_; $self->throw_exception ('Unable to analyze custom {from}') - if ref $from ne 'ARRAY'; + if ref $attrs->{from} ne 'ARRAY'; # what we will return my $aliases_by_type; # see what aliases are there to work with + # and record who is a multiplier and who is premultiplied my $alias_list; - for (@$from) { - my $j = $_; + for my $node (@{$attrs->{from}}) { + + my $j = $node; $j = $j->[0] if ref $j eq 'ARRAY'; my $al = $j->{-alias} or next; $alias_list->{$al} = $j; - $aliases_by_type->{multiplying}{$al} ||= { -parents => $j->{-join_path}||[] } if ( + + $aliases_by_type->{multiplying}{$al} ||= { -parents => $j->{-join_path}||[] } # not array == {from} head == can't be multiplying - ( ref($_) eq 'ARRAY' and ! $j->{-is_single} ) - or - # a parent of ours is already a multiplier - ( grep { $aliases_by_type->{multiplying}{$_} } @{ $j->{-join_path}||[] } ) - ); + if ref($node) eq 'ARRAY' and ! $j->{-is_single}; + + $aliases_by_type->{premultiplied}{$al} ||= { -parents => $j->{-join_path}||[] } + # parts of the path that are not us but are multiplying + if grep { $aliases_by_type->{multiplying}{$_} } + grep { $_ ne $al } + map { values %$_ } + @{ $j->{-join_path}||[] } } # get a column to source/alias map (including unambiguous unqualified ones) - my $colinfo = $self->_resolve_column_info ($from); + my $colinfo = $self->_resolve_column_info ($attrs->{from}); # set up a botched SQLA my $sql_maker = $self->sql_maker; # these are throw away results, do not pollute the bind stack - local $sql_maker->{select_bind}; local $sql_maker->{where_bind}; local $sql_maker->{group_bind}; local $sql_maker->{having_bind}; @@ -477,7 +474,7 @@ sub _resolve_aliastypes_from_select_args { # generate sql chunks my $to_scan = { restricting => [ - $sql_maker->_recurse_where ($where), + ($sql_maker->_recurse_where ($attrs->{where}))[0], $sql_maker->_parse_rs_attrs ({ having => $attrs->{having} }), ], grouping => [ @@ -485,88 +482,152 @@ sub _resolve_aliastypes_from_select_args { ], joining => [ $sql_maker->_recurse_from ( - ref $from->[0] eq 'ARRAY' ? $from->[0][0] : $from->[0], - @{$from}[1 .. $#$from], + ref $attrs->{from}[0] eq 'ARRAY' ? $attrs->{from}[0][0] : $attrs->{from}[0], + @{$attrs->{from}}[1 .. $#{$attrs->{from}}], ), ], selecting => [ - $sql_maker->_recurse_fields ($select), + # kill all selectors which look like a proper subquery + # this is a sucky heuristic *BUT* - if we get it wrong the query will simply + # fail to run, so we are relatively safe + grep + { $_ !~ / \A \s* \( \s* SELECT \s+ .+? \s+ FROM \s+ .+? \) \s* \z /xsi } + map + { ($sql_maker->_recurse_fields($_))[0] } + @{$attrs->{select}} ], - ordering => [ - map { $_->[0] } $self->_extract_order_criteria ($attrs->{order_by}, $sql_maker), + ordering => [ map + { + ( my $sql = (ref $_ ? $_->[0] : $_) ) =~ s/ \s+ (?: ASC | DESC ) \s* \z //xi; + $sql; + } + $sql_maker->_order_by_chunks( $attrs->{order_by} ), ], }; - # throw away empty chunks - $_ = [ map { $_ || () } @$_ ] for values %$to_scan; + # we will be bulk-scanning anyway - pieces will not matter in that case, + # thus join everything up + # throw away empty-string chunks, and make sure no binds snuck in + # note that we operate over @{$to_scan->{$type}}, hence the + # semi-mindbending ... map ... for values ... + ( $_ = join ' ', map { + + ( ! defined $_ ) ? () + : ( length ref $_ ) ? (require Data::Dumper::Concise && $self->throw_exception( + "Unexpected ref in scan-plan: " . Data::Dumper::Concise::Dumper($_) + )) + : ( $_ =~ /^\s*$/ ) ? () + : $_ + + } @$_ ) for values %$to_scan; + + # throw away empty to-scan's + ( + length $to_scan->{$_} + or + delete $to_scan->{$_} + ) for keys %$to_scan; + - # first see if we have any exact matches (qualified or unqualified) - for my $type (keys %$to_scan) { - for my $piece (@{$to_scan->{$type}}) { - if ($colinfo->{$piece} and my $alias = $colinfo->{$piece}{-source_alias}) { - $aliases_by_type->{$type}{$alias} ||= { -parents => $alias_list->{$alias}{-join_path}||[] }; - $aliases_by_type->{$type}{$alias}{-seen_columns}{$colinfo->{$piece}{-fq_colname}} = $piece; - } - } - } - # now loop through all fully qualified columns and get the corresponding - # alias (should work even if they are in scalarrefs) - for my $alias (keys %$alias_list) { - my $al_re = qr/ - $lquote $alias $rquote $sep (?: $lquote ([^$rquote]+) $rquote )? + # these will be used for matching in the loop below + my $all_aliases = join ' | ', map { quotemeta $_ } keys %$alias_list; + my $fq_col_re = qr/ + $lquote ( $all_aliases ) $rquote $sep (?: $lquote ([^$rquote]+) $rquote )? + | + \b ( $all_aliases ) \. ( [^\s\)\($rquote]+ )? + /x; + + + my $all_unq_columns = join ' | ', + map + { quotemeta $_ } + grep + # using a regex here shows up on profiles, boggle + { index( $_, '.') < 0 } + keys %$colinfo + ; + my $unq_col_re = $all_unq_columns + ? qr/ + $lquote ( $all_unq_columns ) $rquote | - \b $alias \. ([^\s\)\($rquote]+)? - /x; - - for my $type (keys %$to_scan) { - for my $piece (@{$to_scan->{$type}}) { - if (my @matches = $piece =~ /$al_re/g) { - $aliases_by_type->{$type}{$alias} ||= { -parents => $alias_list->{$alias}{-join_path}||[] }; - $aliases_by_type->{$type}{$alias}{-seen_columns}{"$alias.$_"} = "$alias.$_" - for grep { defined $_ } @matches; - } + (?: \A | \s ) ( $all_unq_columns ) (?: \s | \z ) + /x + : undef + ; + + + # the actual scan, per type + for my $type (keys %$to_scan) { + + + # now loop through all fully qualified columns and get the corresponding + # alias (should work even if they are in scalarrefs) + # + # The regex captures in multiples of 4, with one of the two pairs being + # undef. There may be a *lot* of matches, hence the convoluted loop + my @matches = $to_scan->{$type} =~ /$fq_col_re/g; + my $i = 0; + while( $i < $#matches ) { + + if ( + defined $matches[$i] + ) { + $aliases_by_type->{$type}{$matches[$i]} ||= { -parents => $alias_list->{$matches[$i]}{-join_path}||[] }; + + $aliases_by_type->{$type}{$matches[$i]}{-seen_columns}{"$matches[$i].$matches[$i+1]"} = "$matches[$i].$matches[$i+1]" + if defined $matches[$i+1]; + + $i += 2; } + + $i += 2; } - } - # now loop through unqualified column names, and try to locate them within - # the chunks - for my $col (keys %$colinfo) { - next if $col =~ / \. /x; # if column is qualified it was caught by the above - my $col_re = qr/ $lquote ($col) $rquote /x; + # now loop through unqualified column names, and try to locate them within + # the chunks, if there are any unqualified columns in the 1st place + next unless $unq_col_re; - for my $type (keys %$to_scan) { - for my $piece (@{$to_scan->{$type}}) { - if ( my @matches = $piece =~ /$col_re/g) { - my $alias = $colinfo->{$col}{-source_alias}; - $aliases_by_type->{$type}{$alias} ||= { -parents => $alias_list->{$alias}{-join_path}||[] }; - $aliases_by_type->{$type}{$alias}{-seen_columns}{"$alias.$_"} = $_ - for grep { defined $_ } @matches; - } - } + # The regex captures in multiples of 2, one of the two being undef + for ( $to_scan->{$type} =~ /$unq_col_re/g ) { + defined $_ or next; + my $alias = $colinfo->{$_}{-source_alias} or next; + $aliases_by_type->{$type}{$alias} ||= { -parents => $alias_list->{$alias}{-join_path}||[] }; + $aliases_by_type->{$type}{$alias}{-seen_columns}{"$alias.$_"} = $_ } } + # Add any non-left joins to the restriction list (such joins are indeed restrictions) - for my $j (values %$alias_list) { - my $alias = $j->{-alias} or next; - $aliases_by_type->{restricting}{$alias} ||= { -parents => $j->{-join_path}||[] } if ( - (not $j->{-join_type}) + ( + $_->{-alias} + and + ! $aliases_by_type->{restricting}{ $_->{-alias} } + and + ( + not $_->{-join_type} or - ($j->{-join_type} !~ /^left (?: \s+ outer)? $/xi) - ); - } + $_->{-join_type} !~ /^left (?: \s+ outer)? $/xi + ) + and + $aliases_by_type->{restricting}{ $_->{-alias} } = { -parents => $_->{-join_path}||[] } + ) for values %$alias_list; + + + # final cleanup + ( + keys %{$aliases_by_type->{$_}} + or + delete $aliases_by_type->{$_} + ) for keys %$aliases_by_type; - for (keys %$aliases_by_type) { - delete $aliases_by_type->{$_} unless keys %{$aliases_by_type->{$_}}; - } - return $aliases_by_type; + $aliases_by_type; } -# This is the engine behind { distinct => 1 } +# This is the engine behind { distinct => 1 } and the general +# complex prefetch grouper sub _group_over_selection { my ($self, $attrs) = @_; @@ -588,36 +649,114 @@ sub _group_over_selection { } } - # add any order_by parts *from the main source* that are not already - # present in the group_by - # we need to be careful not to add any named functions/aggregates - # i.e. order_by => [ ... { count => 'foo' } ... ] - my @leftovers; - for ($self->_extract_order_criteria($attrs->{order_by})) { + my @order_by = $self->_extract_order_criteria($attrs->{order_by}) + or return (\@group_by, $attrs->{order_by}); + + # add any order_by parts that are not already present in the group_by + # to maintain SQL cross-compatibility and general sanity + # + # also in case the original selection is *not* unique, or in case part + # of the ORDER BY refers to a multiplier - we will need to replace the + # skipped order_by elements with their MIN/MAX equivalents as to maintain + # the proper overall order without polluting the group criteria (and + # possibly changing the outcome entirely) + + my ($leftovers, $sql_maker, @new_order_by, $order_chunks, $aliastypes); + + my $group_already_unique = $self->_columns_comprise_identifying_set($colinfos, \@group_by); + + for my $o_idx (0 .. $#order_by) { + + # if the chunk is already a min/max function - there is nothing left to touch + next if $order_by[$o_idx][0] =~ /^ (?: min | max ) \s* \( .+ \) $/ix; + # only consider real columns (for functions the user got to do an explicit group_by) - if (@$_ != 1) { - push @leftovers, $_; - next; + my $chunk_ci; + if ( + @{$order_by[$o_idx]} != 1 + or + # only declare an unknown *plain* identifier as "leftover" if we are called with + # aliastypes to examine. If there are none - we are still in _resolve_attrs, and + # can just assume the user knows what they want + ( ! ( $chunk_ci = $colinfos->{$order_by[$o_idx][0]} ) and $attrs->{_aliastypes} ) + ) { + push @$leftovers, $order_by[$o_idx][0]; } - my $chunk = $_->[0]; - if ( - !$colinfos->{$chunk} + next unless $chunk_ci; + + # no duplication of group criteria + next if $group_index{$chunk_ci->{-fq_colname}}; + + $aliastypes ||= ( + $attrs->{_aliastypes} or - $colinfos->{$chunk}{-source_alias} ne $attrs->{alias} + $self->_resolve_aliastypes_from_select_args({ + from => $attrs->{from}, + order_by => $attrs->{order_by}, + }) + ) if $group_already_unique; + + # check that we are not ordering by a multiplier (if a check is requested at all) + if ( + $group_already_unique + and + ! $aliastypes->{multiplying}{$chunk_ci->{-source_alias}} + and + ! $aliastypes->{premultiplied}{$chunk_ci->{-source_alias}} ) { - push @leftovers, $_; - next; + push @group_by, $chunk_ci->{-fq_colname}; + $group_index{$chunk_ci->{-fq_colname}}++ } + else { + # We need to order by external columns without adding them to the group + # (eiehter a non-unique selection, or a multi-external) + # + # This doesn't really make sense in SQL, however from DBICs point + # of view is rather valid (e.g. order the leftmost objects by whatever + # criteria and get the offset/rows many). There is a way around + # this however in SQL - we simply tae the direction of each piece + # of the external order and convert them to MIN(X) for ASC or MAX(X) + # for DESC, and group_by the root columns. The end result should be + # exactly what we expect + # + $sql_maker ||= $self->sql_maker; + $order_chunks ||= [ + map { ref $_ eq 'ARRAY' ? $_ : [ $_ ] } $sql_maker->_order_by_chunks($attrs->{order_by}) + ]; + + my ($chunk, $is_desc) = $sql_maker->_split_order_chunk($order_chunks->[$o_idx][0]); + + # we reached that far - wrap any part of the order_by that "responded" + # to an ordering alias into a MIN/MAX + $new_order_by[$o_idx] = \[ + sprintf( '%s( %s )%s', + ($is_desc ? 'MAX' : 'MIN'), + $chunk, + ($is_desc ? ' DESC' : ''), + ), + @ {$order_chunks->[$o_idx]} [ 1 .. $#{$order_chunks->[$o_idx]} ] + ]; + } + } - $chunk = $colinfos->{$chunk}{-fq_colname}; - push @group_by, $chunk unless $group_index{$chunk}++; + $self->throw_exception ( sprintf + 'Unable to programatically derive a required group_by from the supplied ' + . 'order_by criteria. To proceed either add an explicit group_by, or ' + . 'simplify your order_by to only include plain columns ' + . '(supplied order_by: %s)', + join ', ', map { "'$_'" } @$leftovers, + ) if $leftovers; + + # recreate the untouched order parts + if (@new_order_by) { + $new_order_by[$_] ||= \ $order_chunks->[$_] for ( 0 .. $#$order_chunks ); } - return wantarray - ? (\@group_by, (@leftovers ? \@leftovers : undef) ) - : \@group_by - ; + return ( + \@group_by, + (@new_order_by ? \@new_order_by : $attrs->{order_by} ), # same ref as original == unchanged + ); } sub _resolve_ident_sources { @@ -659,53 +798,66 @@ sub _resolve_ident_sources { # for all sources sub _resolve_column_info { my ($self, $ident, $colnames) = @_; - my $alias2src = $self->_resolve_ident_sources($ident); + + return {} if $colnames and ! @$colnames; + + my $sources = $self->_resolve_ident_sources($ident); + + $_ = { rsrc => $_, colinfos => $_->columns_info } + for values %$sources; my (%seen_cols, @auto_colnames); # compile a global list of column names, to be able to properly # disambiguate unqualified column names (if at all possible) - for my $alias (keys %$alias2src) { - my $rsrc = $alias2src->{$alias}; - for my $colname ($rsrc->columns) { - push @{$seen_cols{$colname}}, $alias; - push @auto_colnames, "$alias.$colname" unless $colnames; - } + for my $alias (keys %$sources) { + ( + ++$seen_cols{$_}{$alias} + and + ! $colnames + and + push @auto_colnames, "$alias.$_" + ) for keys %{ $sources->{$alias}{colinfos} }; } $colnames ||= [ @auto_colnames, - grep { @{$seen_cols{$_}} == 1 } (keys %seen_cols), + ( grep { keys %{$seen_cols{$_}} == 1 } keys %seen_cols ), ]; - my (%return, $colinfos); - foreach my $col (@$colnames) { - my ($source_alias, $colname) = $col =~ m/^ (?: ([^\.]+) \. )? (.+) $/x; - - # if the column was seen exactly once - we know which rsrc it came from - $source_alias ||= $seen_cols{$colname}[0] - if ($seen_cols{$colname} and @{$seen_cols{$colname}} == 1); + my %return; + for (@$colnames) { + my ($colname, $source_alias) = reverse split /\./, $_; - next unless $source_alias; + my $assumed_alias = + $source_alias + || + # if the column was seen exactly once - we know which rsrc it came from + ( + $seen_cols{$colname} + and + keys %{$seen_cols{$colname}} == 1 + and + ( %{$seen_cols{$colname}} )[0] + ) + || + next + ; - my $rsrc = $alias2src->{$source_alias} - or next; + $self->throw_exception( + "No such column '$colname' on source " . $sources->{$assumed_alias}{rsrc}->source_name + ) unless $seen_cols{$colname}{$assumed_alias}; - $return{$col} = { - %{ - ( $colinfos->{$source_alias} ||= $rsrc->columns_info )->{$colname} - || - $self->throw_exception( - "No such column '$colname' on source " . $rsrc->source_name - ); - }, - -result_source => $rsrc, - -source_alias => $source_alias, - -fq_colname => $col eq $colname ? "$source_alias.$col" : $col, + $return{$_} = { + %{ $sources->{$assumed_alias}{colinfos}{$colname} }, + -result_source => $sources->{$assumed_alias}{rsrc}, + -source_alias => $assumed_alias, + -fq_colname => "$assumed_alias.$colname", -colname => $colname, }; - $return{"$source_alias.$colname"} = $return{$col} if $col eq $colname; + $return{"$assumed_alias.$colname"} = $return{$_} + unless $source_alias; } return \%return; @@ -729,31 +881,9 @@ sub _resolve_column_info { sub _inner_join_to_node { my ($self, $from, $alias) = @_; - # subqueries and other oddness are naturally not supported - return $from if ( - ref $from ne 'ARRAY' - || - @$from <= 1 - || - ref $from->[0] ne 'HASH' - || - ! $from->[0]{-alias} - || - $from->[0]{-alias} eq $alias # this last bit means $alias is the head of $from - nothing to do - ); - - # find the current $alias in the $from structure - my $switch_branch; - JOINSCAN: - for my $j (@{$from}[1 .. $#$from]) { - if ($j->[0]{-alias} eq $alias) { - $switch_branch = $j->[0]{-join_path}; - last JOINSCAN; - } - } + my $switch_branch = $self->_find_join_path_to_node($from, $alias); - # something else went quite wrong - return $from unless $switch_branch; + return $from unless @{$switch_branch||[]}; # So it looks like we will have to switch some stuff around. # local() is useless here as we will be leaving the scope @@ -781,6 +911,29 @@ sub _inner_join_to_node { return \@new_from; } +sub _find_join_path_to_node { + my ($self, $from, $target_alias) = @_; + + # subqueries and other oddness are naturally not supported + return undef if ( + ref $from ne 'ARRAY' + || + ref $from->[0] ne 'HASH' + || + ! defined $from->[0]{-alias} + ); + + # no path - the head is the alias + return [] if $from->[0]{-alias} eq $target_alias; + + for my $i (1 .. $#$from) { + return $from->[$i][0]{-join_path} if ( ($from->[$i][0]{-alias}||'') eq $target_alias ); + } + + # something else went quite wrong + return undef; +} + sub _extract_order_criteria { my ($self, $order_by, $sql_maker) = @_; @@ -829,130 +982,526 @@ sub _extract_order_criteria { sub _order_by_is_stable { my ($self, $ident, $order_by, $where) = @_; - my $colinfo = $self->_resolve_column_info($ident, [ - (map { $_->[0] } $self->_extract_order_criteria($order_by)), - $where ? @{$self->_extract_fixed_condition_columns($where)} :(), - ]); + my @cols = ( + ( map { $_->[0] } $self->_extract_order_criteria($order_by) ), + ( $where ? keys %{ $self->_extract_fixed_condition_columns($where) } : () ), + ) or return 0; - return undef unless keys %$colinfo; + my $colinfo = $self->_resolve_column_info($ident, \@cols); + + return keys %$colinfo + ? $self->_columns_comprise_identifying_set( $colinfo, \@cols ) + : 0 + ; +} + +sub _columns_comprise_identifying_set { + my ($self, $colinfo, $columns) = @_; my $cols_per_src; - $cols_per_src->{$_->{-source_alias}}{$_->{-colname}} = $_ for values %$colinfo; + $cols_per_src -> {$_->{-source_alias}} -> {$_->{-colname}} = $_ + for grep { defined $_ } @{$colinfo}{@$columns}; for (values %$cols_per_src) { my $src = (values %$_)[0]->{-result_source}; return 1 if $src->_identifying_column_set($_); } - return undef; + return 0; } -# this is almost identical to the above, except it accepts only +# this is almost similar to _order_by_is_stable, except it takes # a single rsrc, and will succeed only if the first portion of the order # by is stable. # returns that portion as a colinfo hashref on success -sub _main_source_order_by_portion_is_stable { - my ($self, $main_rsrc, $order_by, $where) = @_; +sub _extract_colinfo_of_stable_main_source_order_by_portion { + my ($self, $attrs) = @_; - die "Huh... I expect a blessed result_source..." - if ref($main_rsrc) eq 'ARRAY'; + my $nodes = $self->_find_join_path_to_node($attrs->{from}, $attrs->{alias}); + + return unless defined $nodes; my @ord_cols = map { $_->[0] } - ( $self->_extract_order_criteria($order_by) ) + ( $self->_extract_order_criteria($attrs->{order_by}) ) ; return unless @ord_cols; - my $colinfos = $self->_resolve_column_info($main_rsrc); + my $valid_aliases = { map { $_ => 1 } ( + $attrs->{from}[0]{-alias}, + map { values %$_ } @$nodes, + ) }; + + my $colinfos = $self->_resolve_column_info($attrs->{from}); + + my ($colinfos_to_return, $seen_main_src_cols); + + for my $col (@ord_cols) { + # if order criteria is unresolvable - there is nothing we can do + my $colinfo = $colinfos->{$col} or last; + + # if we reached the end of the allowed aliases - also nothing we can do + last unless $valid_aliases->{$colinfo->{-source_alias}}; + + $colinfos_to_return->{$col} = $colinfo; + + $seen_main_src_cols->{$colinfo->{-colname}} = 1 + if $colinfo->{-source_alias} eq $attrs->{alias}; + } + + # FIXME the condition may be singling out things on its own, so we + # conceivable could come back wi "stable-ordered by nothing" + # not confient enough in the parser yet, so punt for the time being + return unless $seen_main_src_cols; + + my $main_src_fixed_cols_from_cond = [ $attrs->{where} + ? ( + map + { + ( $colinfos->{$_} and $colinfos->{$_}{-source_alias} eq $attrs->{alias} ) + ? $colinfos->{$_}{-colname} + : () + } + keys %{ $self->_extract_fixed_condition_columns($attrs->{where}) } + ) + : () + ]; + + return $attrs->{result_source}->_identifying_column_set([ + keys %$seen_main_src_cols, + @$main_src_fixed_cols_from_cond, + ]) ? $colinfos_to_return : (); +} + +# Attempts to flatten a passed in SQLA condition as much as possible towards +# a plain hashref, *without* altering its semantics. Required by +# create/populate being able to extract definitive conditions from preexisting +# resultset {where} stacks +# +# FIXME - while relatively robust, this is still imperfect, one of the first +# things to tackle when we get access to a formalized AST. Note that this code +# is covered by a *ridiculous* amount of tests, so starting with porting this +# code would be a rather good exercise +sub _collapse_cond { + my ($self, $where, $where_is_anded_array) = @_; + + my $fin; + + if (! $where) { + return; + } + elsif ($where_is_anded_array or ref $where eq 'HASH') { + + my @pairs; + + my @pieces = $where_is_anded_array ? @$where : $where; + while (@pieces) { + my $chunk = shift @pieces; + + if (ref $chunk eq 'HASH') { + for (sort keys %$chunk) { + + # Match SQLA 1.79 behavior + unless( length $_ ) { + is_literal_value($chunk->{$_}) + ? carp 'Hash-pairs consisting of an empty string with a literal are deprecated, use -and => [ $literal ] instead' + : $self->throw_exception("Supplying an empty left hand side argument is not supported in hash-pairs") + ; + } + + push @pairs, $_ => $chunk->{$_}; + } + } + elsif (ref $chunk eq 'ARRAY') { + push @pairs, -or => $chunk + if @$chunk; + } + elsif ( ! length ref $chunk) { + + # Match SQLA 1.79 behavior + $self->throw_exception("Supplying an empty left hand side argument is not supported in array-pairs") + if $where_is_anded_array and (! defined $chunk or ! length $chunk); + + push @pairs, $chunk, shift @pieces; + } + else { + push @pairs, '', $chunk; + } + } + + return unless @pairs; + + my @conds = $self->_collapse_cond_unroll_pairs(\@pairs) + or return; + + # Consolidate various @conds back into something more compact + for my $c (@conds) { + if (ref $c ne 'HASH') { + push @{$fin->{-and}}, $c; + } + else { + for my $col (sort keys %$c) { + + # consolidate all -and nodes + if ($col =~ /^\-and$/i) { + push @{$fin->{-and}}, + ref $c->{$col} eq 'ARRAY' ? @{$c->{$col}} + : ref $c->{$col} eq 'HASH' ? %{$c->{$col}} + : { $col => $c->{$col} } + ; + } + elsif ($col =~ /^\-/) { + push @{$fin->{-and}}, { $col => $c->{$col} }; + } + elsif (exists $fin->{$col}) { + $fin->{$col} = [ -and => map { + (ref $_ eq 'ARRAY' and ($_->[0]||'') =~ /^\-and$/i ) + ? @{$_}[1..$#$_] + : $_ + ; + } ($fin->{$col}, $c->{$col}) ]; + } + else { + $fin->{$col} = $c->{$col}; + } + } + } + } + } + elsif (ref $where eq 'ARRAY') { + # we are always at top-level here, it is safe to dump empty *standalone* pieces + my $fin_idx; + + for (my $i = 0; $i <= $#$where; $i++ ) { + + # Match SQLA 1.79 behavior + $self->throw_exception( + "Supplying an empty left hand side argument is not supported in array-pairs" + ) if (! defined $where->[$i] or ! length $where->[$i]); + + my $logic_mod = lc ( ($where->[$i] =~ /^(\-(?:and|or))$/i)[0] || '' ); + + if ($logic_mod) { + $i++; + $self->throw_exception("Unsupported top-level op/arg pair: [ $logic_mod => $where->[$i] ]") + unless ref $where->[$i] eq 'HASH' or ref $where->[$i] eq 'ARRAY'; + + my $sub_elt = $self->_collapse_cond({ $logic_mod => $where->[$i] }) + or next; + + my @keys = keys %$sub_elt; + if ( @keys == 1 and $keys[0] !~ /^\-/ ) { + $fin_idx->{ "COL_$keys[0]_" . serialize $sub_elt } = $sub_elt; + } + else { + $fin_idx->{ "SER_" . serialize $sub_elt } = $sub_elt; + } + } + elsif (! length ref $where->[$i] ) { + my $sub_elt = $self->_collapse_cond({ @{$where}[$i, $i+1] }) + or next; + + $fin_idx->{ "COL_$where->[$i]_" . serialize $sub_elt } = $sub_elt; + $i++; + } + else { + $fin_idx->{ "SER_" . serialize $where->[$i] } = $self->_collapse_cond( $where->[$i] ) || next; + } + } + + if (! $fin_idx) { + return; + } + elsif ( keys %$fin_idx == 1 ) { + $fin = (values %$fin_idx)[0]; + } + else { + my @or; + + # at this point everything is at most one level deep - unroll if needed + for (sort keys %$fin_idx) { + if ( ref $fin_idx->{$_} eq 'HASH' and keys %{$fin_idx->{$_}} == 1 ) { + my ($l, $r) = %{$fin_idx->{$_}}; + + if ( + ref $r eq 'ARRAY' + and + ( + ( @$r == 1 and $l =~ /^\-and$/i ) + or + $l =~ /^\-or$/i + ) + ) { + push @or, @$r + } + + elsif ( + ref $r eq 'HASH' + and + keys %$r == 1 + and + $l =~ /^\-(?:and|or)$/i + ) { + push @or, %$r; + } + + else { + push @or, $l, $r; + } + } + else { + push @or, $fin_idx->{$_}; + } + } + + $fin->{-or} = \@or; + } + } + else { + # not a hash not an array + $fin = { -and => [ $where ] }; + } + + # unroll single-element -and's + while ( + $fin->{-and} + and + @{$fin->{-and}} < 2 + ) { + my $and = delete $fin->{-and}; + last if @$and == 0; - for (0 .. $#ord_cols) { + # at this point we have @$and == 1 if ( - ! $colinfos->{$ord_cols[$_]} - or - $colinfos->{$ord_cols[$_]}{-result_source} != $main_rsrc + ref $and->[0] eq 'HASH' + and + ! grep { exists $fin->{$_} } keys %{$and->[0]} ) { - $#ord_cols = $_ - 1; + $fin = { + %$fin, %{$and->[0]} + }; + } + else { + $fin->{-and} = $and; last; } } - # we just truncated it above - return unless @ord_cols; + # compress same-column conds found in $fin + for my $col ( grep { $_ !~ /^\-/ } keys %$fin ) { + next unless ref $fin->{$col} eq 'ARRAY' and ($fin->{$col}[0]||'') =~ /^\-and$/i; + my $val_bag = { map { + (! defined $_ ) ? ( UNDEF => undef ) + : ( ! length ref $_ or is_plain_value $_ ) ? ( "VAL_$_" => $_ ) + : ( ( 'SER_' . serialize $_ ) => $_ ) + } @{$fin->{$col}}[1 .. $#{$fin->{$col}}] }; + + if (keys %$val_bag == 1 ) { + ($fin->{$col}) = values %$val_bag; + } + else { + $fin->{$col} = [ -and => map { $val_bag->{$_} } sort keys %$val_bag ]; + } + } - my $order_portion_ci = { map { - $colinfos->{$_}{-colname} => $colinfos->{$_}, - $colinfos->{$_}{-fq_colname} => $colinfos->{$_}, - } @ord_cols }; + return keys %$fin ? $fin : (); +} - # since all we check here are the start of the order_by belonging to the - # top level $rsrc, a present identifying set will mean that the resultset - # is ordered by its leftmost table in a stable manner - # - # RV of _identifying_column_set contains unqualified names only - my $unqualified_idset = $main_rsrc->_identifying_column_set({ - ( $where ? %{ - $self->_resolve_column_info( - $main_rsrc, $self->_extract_fixed_condition_columns($where) - ) - } : () ), - %$order_portion_ci - }) or return; +sub _collapse_cond_unroll_pairs { + my ($self, $pairs) = @_; - my $ret_info; - my %unqualified_idcols_from_order = map { - $order_portion_ci->{$_} ? ( $_ => $order_portion_ci->{$_} ) : () - } @$unqualified_idset; + my @conds; - # extra optimization - cut the order_by at the end of the identifying set - # (just in case the user was stupid and overlooked the obvious) - for my $i (0 .. $#ord_cols) { - my $col = $ord_cols[$i]; - my $unqualified_colname = $order_portion_ci->{$col}{-colname}; - $ret_info->{$col} = { %{$order_portion_ci->{$col}}, -idx_in_order_subset => $i }; - delete $unqualified_idcols_from_order{$ret_info->{$col}{-colname}}; + while (@$pairs) { + my ($lhs, $rhs) = splice @$pairs, 0, 2; - # we didn't reach the end of the identifying portion yet - return $ret_info unless keys %unqualified_idcols_from_order; + if (! length $lhs) { + push @conds, $self->_collapse_cond($rhs); + } + elsif ( $lhs =~ /^\-and$/i ) { + push @conds, $self->_collapse_cond($rhs, (ref $rhs eq 'ARRAY')); + } + elsif ( $lhs =~ /^\-or$/i ) { + push @conds, $self->_collapse_cond( + (ref $rhs eq 'HASH') ? [ map { $_ => $rhs->{$_} } sort keys %$rhs ] : $rhs + ); + } + else { + if (ref $rhs eq 'HASH' and ! keys %$rhs) { + # FIXME - SQLA seems to be doing... nothing...? + } + # normalize top level -ident, for saner extract_fixed_condition_columns code + elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{-ident}) { + push @conds, { $lhs => { '=', $rhs } }; + } + elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{-value} and is_plain_value $rhs->{-value}) { + push @conds, { $lhs => $rhs->{-value} }; + } + elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{'='}) { + if ( length ref $rhs->{'='} and is_literal_value $rhs->{'='} ) { + push @conds, { $lhs => $rhs }; + } + else { + for my $p ($self->_collapse_cond_unroll_pairs([ $lhs => $rhs->{'='} ])) { + + # extra sanity check + if (keys %$p > 1) { + require Data::Dumper::Concise; + local $Data::Dumper::Deepcopy = 1; + $self->throw_exception( + "Internal error: unexpected collapse unroll:" + . Data::Dumper::Concise::Dumper { in => { $lhs => $rhs }, out => $p } + ); + } + + my ($l, $r) = %$p; + + push @conds, ( + ! length ref $r + or + # the unroller recursion may return a '=' prepended value already + ref $r eq 'HASH' and keys %$rhs == 1 and exists $rhs->{'='} + or + is_plain_value($r) + ) + ? { $l => $r } + : { $l => { '=' => $r } } + ; + } + } + } + elsif (ref $rhs eq 'ARRAY') { + # some of these conditionals encounter multi-values - roll them out using + # an unshift, which will cause extra looping in the while{} above + if (! @$rhs ) { + push @conds, { $lhs => [] }; + } + elsif ( ($rhs->[0]||'') =~ /^\-(?:and|or)$/i ) { + $self->throw_exception("Value modifier not followed by any values: $lhs => [ $rhs->[0] ] ") + if @$rhs == 1; + + if( $rhs->[0] =~ /^\-and$/i ) { + unshift @$pairs, map { $lhs => $_ } @{$rhs}[1..$#$rhs]; + } + # if not an AND then it's an OR + elsif(@$rhs == 2) { + unshift @$pairs, $lhs => $rhs->[1]; + } + else { + push @conds, { $lhs => [ @{$rhs}[1..$#$rhs] ] }; + } + } + elsif (@$rhs == 1) { + unshift @$pairs, $lhs => $rhs->[0]; + } + else { + push @conds, { $lhs => $rhs }; + } + } + # unroll func + { -value => ... } + elsif ( + ref $rhs eq 'HASH' + and + ( my ($subop) = keys %$rhs ) == 1 + and + length ref ((values %$rhs)[0]) + and + my $vref = is_plain_value( (values %$rhs)[0] ) + ) { + push @conds, { $lhs => { $subop => $$vref } } + } + else { + push @conds, { $lhs => $rhs }; + } + } } - die 'How did we get here...'; + return @conds; } -# returns an arrayref of column names which *definitely* have som -# sort of non-nullable equality requested in the given condition -# specification. This is used to figure out if a resultset is -# constrained to a column which is part of a unique constraint, -# which in turn allows us to better predict how ordering will behave -# etc. +# Analyzes a given condition and attempts to extract all columns +# with a definitive fixed-condition criteria. Returns a hashref +# of k/v pairs suitable to be passed to set_columns(), with a +# MAJOR CAVEAT - multi-value (contradictory) equalities are still +# represented as a reference to the UNRESOVABLE_CONDITION constant +# The reason we do this is that some codepaths only care about the +# codition being stable, as opposed to actually making sense +# +# The normal mode is used to figure out if a resultset is constrained +# to a column which is part of a unique constraint, which in turn +# allows us to better predict how ordering will behave etc. +# +# With the optional "consider_nulls" boolean argument, the function +# is instead used to infer inambiguous values from conditions +# (e.g. the inheritance of resultset conditions on new_result) # -# this is a rudimentary, incomplete, and error-prone extractor -# however this is OK - it is conservative, and if we can not find -# something that is in fact there - the stack will recover gracefully -# Also - DQ and the mst it rode in on will save us all RSN!!! sub _extract_fixed_condition_columns { - my ($self, $where, $nested) = @_; + my ($self, $where, $consider_nulls) = @_; + my $where_hash = $self->_collapse_cond($_[1]); - return unless ref $where eq 'HASH'; + my $res = {}; + my ($c, $v); + for $c (keys %$where_hash) { + my $vals; - my @cols; - for my $lhs (keys %$where) { - if ($lhs =~ /^\-and$/i) { - push @cols, ref $where->{$lhs} eq 'ARRAY' - ? ( map { $self->_extract_fixed_condition_columns($_, 1) } @{$where->{$lhs}} ) - : $self->_extract_fixed_condition_columns($where->{$lhs}, 1) - ; + if (!defined ($v = $where_hash->{$c}) ) { + $vals->{UNDEF} = $v if $consider_nulls + } + elsif ( + ref $v eq 'HASH' + and + keys %$v == 1 + ) { + if (exists $v->{-value}) { + if (defined $v->{-value}) { + $vals->{"VAL_$v->{-value}"} = $v->{-value} + } + elsif( $consider_nulls ) { + $vals->{UNDEF} = $v->{-value}; + } + } + # do not need to check for plain values - _collapse_cond did it for us + elsif( + length ref $v->{'='} + and + ( + ( ref $v->{'='} eq 'HASH' and keys %{$v->{'='}} == 1 and exists $v->{'='}{-ident} ) + or + is_literal_value($v->{'='}) + ) + ) { + $vals->{ 'SER_' . serialize $v->{'='} } = $v->{'='}; + } + } + elsif ( + ! length ref $v + or + is_plain_value ($v) + ) { + $vals->{"VAL_$v"} = $v; + } + elsif (ref $v eq 'ARRAY' and ($v->[0]||'') eq '-and') { + for ( @{$v}[1..$#$v] ) { + my $subval = $self->_extract_fixed_condition_columns({ $c => $_ }, 'consider nulls'); # always fish nulls out on recursion + next unless exists $subval->{$c}; # didn't find anything + $vals->{ + ! defined $subval->{$c} ? 'UNDEF' + : ( ! length ref $subval->{$c} or is_plain_value $subval->{$c} ) ? "VAL_$subval->{$c}" + : ( 'SER_' . serialize $subval->{$c} ) + } = $subval->{$c}; + } } - elsif ($lhs !~ /^\-/) { - my $val = $where->{$lhs}; - push @cols, $lhs if (defined $val and ( - ! ref $val - or - (ref $val eq 'HASH' and keys %$val == 1 and defined $val->{'='}) - )); + if (keys %$vals == 1) { + ($res->{$c}) = (values %$vals) + unless !$consider_nulls and exists $vals->{UNDEF}; + } + elsif (keys %$vals > 1) { + $res->{$c} = UNRESOLVABLE_CONDITION; } } - return $nested ? @cols : \@cols; + + $res; } 1;