1 package #hide from PAUSE
2 DBIx::Class::Storage::DBIHacks;
5 # This module contains code supporting a battery of special cases and tests for
6 # many corner cases pushing the envelope of what DBIC can do. When work on
7 # these utilities began in mid 2009 (51a296b402c) it wasn't immediately obvious
8 # that these pieces, despite their misleading on-first-sighe-flakiness, will
9 # become part of the generic query rewriting machinery of DBIC, allowing it to
10 # both generate and process queries representing incredibly complex sets with
11 # reasonable efficiency.
13 # Now (end of 2015), more than 6 years later the routines in this class have
14 # stabilized enough, and are meticulously covered with tests, to a point where
15 # an effort to formalize them into user-facing APIs might be worthwhile.
17 # An implementor working on publicizing and/or replacing the routines with a
18 # more modern SQL generation framework should keep in mind that pretty much all
19 # existing tests are constructed on the basis of real-world code used in
20 # production somewhere.
22 # Please hack on this responsibly ;)
28 use base 'DBIx::Class::Storage';
31 use List::Util 'first';
32 use Scalar::Util 'blessed';
33 use DBIx::Class::_Util qw(UNRESOLVABLE_CONDITION serialize dump_value);
34 use SQL::Abstract qw(is_plain_value is_literal_value);
35 use DBIx::Class::Carp;
39 # This code will remove non-selecting/non-restricting joins from
40 # {from} specs, aiding the RDBMS query optimizer
42 sub _prune_unused_joins {
43 my ($self, $attrs) = @_;
45 # only standard {from} specs are supported, and we could be disabled in general
46 return ($attrs->{from}, {}) unless (
47 ref $attrs->{from} eq 'ARRAY'
51 ref $attrs->{from}[0] eq 'HASH'
53 ref $attrs->{from}[1] eq 'ARRAY'
55 $self->_use_join_optimizer
59 $attrs->{_precalculated_aliastypes}
61 $self->_resolve_aliastypes_from_select_args($attrs)
64 my $new_aliastypes = { %$orig_aliastypes };
66 # we will be recreating this entirely
67 my @reclassify = 'joining';
69 # a grouped set will not be affected by amount of rows. Thus any
70 # purely multiplicator classifications can go
71 # (will be reintroduced below if needed by something else)
72 push @reclassify, qw(multiplying premultiplied)
73 if $attrs->{_force_prune_multiplying_joins} or $attrs->{group_by};
75 # nuke what will be recalculated
76 delete @{$new_aliastypes}{@reclassify};
78 my @newfrom = $attrs->{from}[0]; # FROM head is always present
80 # recalculate what we need once the multipliers are potentially gone
81 # ignore premultiplies, since they do not add any value to anything
83 for ( @{$new_aliastypes}{grep { $_ ne 'premultiplied' } keys %$new_aliastypes }) {
84 # add all requested aliases
85 $need_joins{$_} = 1 for keys %$_;
87 # add all their parents (as per joinpath which is an AoH { table => alias })
88 $need_joins{$_} = 1 for map { values %$_ } map { @{$_->{-parents}} } values %$_;
91 for my $j (@{$attrs->{from}}[1..$#{$attrs->{from}}]) {
92 push @newfrom, $j if (
93 (! defined $j->[0]{-alias}) # legacy crap
95 $need_joins{$j->[0]{-alias}}
99 # we have a new set of joiners - for everything we nuked pull the classification
100 # off the original stack
101 for my $ctype (@reclassify) {
102 $new_aliastypes->{$ctype} = { map
103 { $need_joins{$_} ? ( $_ => $orig_aliastypes->{$ctype}{$_} ) : () }
104 keys %{$orig_aliastypes->{$ctype}}
108 return ( \@newfrom, $new_aliastypes );
112 # This is the code producing joined subqueries like:
113 # SELECT me.*, other.* FROM ( SELECT me.* FROM ... ) JOIN other ON ...
115 sub _adjust_select_args_for_complex_prefetch {
116 my ($self, $attrs) = @_;
118 $self->throw_exception ('Complex prefetches are not supported on resultsets with a custom from attribute') unless (
119 ref $attrs->{from} eq 'ARRAY'
121 @{$attrs->{from}} > 1
123 ref $attrs->{from}[0] eq 'HASH'
125 ref $attrs->{from}[1] eq 'ARRAY'
128 my $root_alias = $attrs->{alias};
130 # generate inner/outer attribute lists, remove stuff that doesn't apply
131 my $outer_attrs = { %$attrs };
132 delete @{$outer_attrs}{qw(from bind rows offset group_by _grouped_by_distinct having)};
134 my $inner_attrs = { %$attrs, _simple_passthrough_construction => 1 };
135 delete @{$inner_attrs}{qw(for collapse select as)};
137 # there is no point of ordering the insides if there is no limit
138 delete $inner_attrs->{order_by} if (
139 delete $inner_attrs->{_order_is_artificial}
141 ! $inner_attrs->{rows}
144 # generate the inner/outer select lists
145 # for inside we consider only stuff *not* brought in by the prefetch
146 # on the outside we substitute any function for its alias
147 $outer_attrs->{select} = [ @{$attrs->{select}} ];
149 my ($root_node, $root_node_offset);
151 for my $i (0 .. $#{$inner_attrs->{from}}) {
152 my $node = $inner_attrs->{from}[$i];
153 my $h = (ref $node eq 'HASH') ? $node
154 : (ref $node eq 'ARRAY' and ref $node->[0] eq 'HASH') ? $node->[0]
158 if ( ($h->{-alias}||'') eq $root_alias and $h->{-rsrc} ) {
160 $root_node_offset = $i;
165 $self->throw_exception ('Complex prefetches are not supported on resultsets with a custom from attribute')
168 # use the heavy duty resolver to take care of aliased/nonaliased naming
169 my $colinfo = $self->_resolve_column_info($inner_attrs->{from});
170 my $selected_root_columns;
172 for my $i (0 .. $#{$outer_attrs->{select}}) {
173 my $sel = $outer_attrs->{select}->[$i];
176 $colinfo->{$sel} and $colinfo->{$sel}{-source_alias} ne $root_alias
179 if (ref $sel eq 'HASH' ) {
180 $sel->{-as} ||= $attrs->{as}[$i];
181 $outer_attrs->{select}->[$i] = join ('.', $root_alias, ($sel->{-as} || "inner_column_$i") );
183 elsif (! ref $sel and my $ci = $colinfo->{$sel}) {
184 $selected_root_columns->{$ci->{-colname}} = 1;
187 push @{$inner_attrs->{select}}, $sel;
189 push @{$inner_attrs->{as}}, $attrs->{as}[$i];
192 my $inner_aliastypes = $self->_resolve_aliastypes_from_select_args($inner_attrs);
194 # In the inner subq we will need to fetch *only* native columns which may
195 # be a part of an *outer* join condition, or an order_by (which needs to be
196 # preserved outside), or wheres. In other words everything but the inner
198 # We can not just fetch everything because a potential has_many restricting
199 # join collapse *will not work* on heavy data types.
201 # essentially a map of all non-selecting seen columns
202 # the sort is there for a nicer select list
206 { keys %{$_->{-seen_columns}||{}} }
208 { values %{$inner_aliastypes->{$_}} }
210 { $_ ne 'selecting' }
211 keys %$inner_aliastypes
213 my $ci = $colinfo->{$_} or next;
215 $ci->{-source_alias} eq $root_alias
217 ! $selected_root_columns->{$ci->{-colname}}++
219 # adding it to both to keep limits not supporting dark selectors happy
220 push @{$inner_attrs->{select}}, $ci->{-fq_colname};
221 push @{$inner_attrs->{as}}, $ci->{-fq_colname};
225 # construct the inner {from} and lock it in a subquery
226 # we need to prune first, because this will determine if we need a group_by below
227 # throw away all non-selecting, non-restricting multijoins
228 # (since we def. do not care about multiplication of the contents of the subquery)
229 my $inner_subq = do {
231 # must use it here regardless of user requests (vastly gentler on optimizer)
232 local $self->{_use_join_optimizer} = 1;
234 # throw away multijoins since we def. do not care about those inside the subquery
235 # $inner_aliastypes *will* be redefined at this point
236 ($inner_attrs->{from}, $inner_aliastypes ) = $self->_prune_unused_joins ({
238 _force_prune_multiplying_joins => 1,
239 _precalculated_aliastypes => $inner_aliastypes,
242 # uh-oh a multiplier (which is not us) left in, this is a problem for limits
243 # we will need to add a group_by to collapse the resultset for proper counts
245 grep { $_ ne $root_alias } keys %{ $inner_aliastypes->{multiplying} || {} }
247 # if there are user-supplied groups - assume user knows wtf they are up to
248 ( ! $inner_aliastypes->{grouping} or $inner_attrs->{_grouped_by_distinct} )
251 my $cur_sel = { map { $_ => 1 } @{$inner_attrs->{select}} };
253 # *possibly* supplement the main selection with pks if not already
254 # there, as they will have to be a part of the group_by to collapse
256 my $inner_select_with_extras;
257 my @pks = map { "$root_alias.$_" } $root_node->{-rsrc}->primary_columns
258 or $self->throw_exception( sprintf
259 'Unable to perform complex limited prefetch off %s without declared primary key',
260 $root_node->{-rsrc}->source_name,
263 push @{ $inner_select_with_extras ||= [ @{$inner_attrs->{select}} ] }, $col
264 unless $cur_sel->{$col}++;
267 ($inner_attrs->{group_by}, $inner_attrs->{order_by}) = $self->_group_over_selection({
269 $inner_select_with_extras ? ( select => $inner_select_with_extras ) : (),
270 _aliastypes => $inner_aliastypes,
274 # we already optimized $inner_attrs->{from} above
275 # and already local()ized
276 $self->{_use_join_optimizer} = 0;
278 # generate the subquery
279 $self->_select_args_to_query (
280 @{$inner_attrs}{qw(from select where)},
285 # Generate the outer from - this is relatively easy (really just replace
286 # the join slot with the subquery), with a major caveat - we can not
287 # join anything that is non-selecting (not part of the prefetch), but at
288 # the same time is a multi-type relationship, as it will explode the result.
290 # There are two possibilities here
291 # - either the join is non-restricting, in which case we simply throw it away
292 # - it is part of the restrictions, in which case we need to collapse the outer
293 # result by tackling yet another group_by to the outside of the query
295 # work on a shallow copy
296 my @orig_from = @{$attrs->{from}};
299 $outer_attrs->{from} = \ my @outer_from;
301 # we may not be the head
302 if ($root_node_offset) {
303 # first generate the outer_from, up to the substitution point
304 @outer_from = splice @orig_from, 0, $root_node_offset;
306 # substitute the subq at the right spot
309 -alias => $root_alias,
310 -rsrc => $root_node->{-rsrc},
311 $root_alias => $inner_subq,
313 # preserve attrs from what is now the head of the from after the splice
314 @{$orig_from[0]}[1 .. $#{$orig_from[0]}],
319 -alias => $root_alias,
320 -rsrc => $root_node->{-rsrc},
321 $root_alias => $inner_subq,
325 shift @orig_from; # what we just replaced above
327 # scan the *remaining* from spec against different attributes, and see which joins are needed
329 my $outer_aliastypes = $outer_attrs->{_aliastypes} =
330 $self->_resolve_aliastypes_from_select_args({ %$outer_attrs, from => \@orig_from });
333 my ($outer_select_chain, @outer_nonselecting_chains) = map { +{
334 map { $_ => 1 } map { values %$_} map { @{$_->{-parents}} } values %{ $outer_aliastypes->{$_} || {} }
335 } } qw/selecting restricting grouping ordering/;
337 # see what's left - throw away if not selecting/restricting
338 my $may_need_outer_group_by;
339 while (my $j = shift @orig_from) {
340 my $alias = $j->[0]{-alias};
343 $outer_select_chain->{$alias}
347 elsif (first { $_->{$alias} } @outer_nonselecting_chains ) {
348 push @outer_from, $j;
349 $may_need_outer_group_by ||= $outer_aliastypes->{multiplying}{$alias} ? 1 : 0;
353 # also throw in a synthetic group_by if a non-selecting multiplier,
354 # to guard against cross-join explosions
355 # the logic is somewhat fragile, but relies on the idea that if a user supplied
356 # a group by on their own - they know what they were doing
357 if ( $may_need_outer_group_by and $attrs->{_grouped_by_distinct} ) {
358 ($outer_attrs->{group_by}, $outer_attrs->{order_by}) = $self->_group_over_selection ({
360 from => \@outer_from,
364 # FIXME: The {where} ends up in both the inner and outer query, i.e. *twice*
366 # This is rather horrific, and while we currently *do* have enough
367 # introspection tooling available to attempt a stab at properly deciding
368 # whether or not to include the where condition on the outside, the
369 # machinery is still too slow to apply it here.
370 # Thus for the time being we do not attempt any sanitation of the where
371 # clause and just pass it through on both sides of the subquery. This *will*
372 # be addressed at a later stage, most likely after folding the SQL generator
373 # into SQLMaker proper
375 # OTOH it can be seen as a plus: <ash> (notes that this query would make a DBA cry ;)
380 # This is probably the ickiest, yet most relied upon part of the codebase:
381 # this is the place where we take arbitrary SQL input and break it into its
382 # constituent parts, making sure we know which *sources* are used in what
383 # *capacity* ( selecting / restricting / grouping / ordering / joining, etc )
384 # Although the method is pretty horrific, the worst thing that can happen is
385 # for a classification failure, which in turn will result in a vocal exception,
386 # and will lead to a relatively prompt fix.
387 # The code has been slowly improving and is covered with a formiddable battery
388 # of tests, so can be considered "reliably stable" at this point (Oct 2015).
390 # A note to implementors attempting to "replace" this - keep in mind that while
391 # there are multiple optimization avenues, the actual "scan literal elements"
392 # part *MAY NEVER BE REMOVED*, even if it is limited only ot the (future) AST
393 # nodes that are deemed opaque (i.e. contain literal expressions). The use of
394 # blackbox literals is at this point firmly a user-facing API, and is one of
395 # *the* reasons DBIC remains as flexible as it is. In other words, when working
396 # on this keep in mind that the following is widespread and *encouraged* way
397 # of using DBIC in the wild when push comes to shove:
400 # select => \[ $random, @stuff],
401 # from => \[ $random, @stuff ],
402 # where => \[ $random, @stuff ],
403 # group_by => \[ $random, @stuff ],
404 # order_by => \[ $random, @stuff ],
407 # Various incarnations of the above are reflected in many of the tests. If one
408 # gets to fail, you get to fix it. A "this is crazy, nobody does that" is not
409 # acceptable going forward.
411 sub _resolve_aliastypes_from_select_args {
412 my ( $self, $attrs ) = @_;
414 $self->throw_exception ('Unable to analyze custom {from}')
415 if ref $attrs->{from} ne 'ARRAY';
417 # what we will return
420 # see what aliases are there to work with
421 # and record who is a multiplier and who is premultiplied
423 for my $node (@{$attrs->{from}}) {
426 $j = $j->[0] if ref $j eq 'ARRAY';
427 my $al = $j->{-alias}
430 $alias_list->{$al} = $j;
432 $aliases_by_type->{multiplying}{$al} ||= { -parents => $j->{-join_path}||[] }
433 # not array == {from} head == can't be multiplying
434 if ref($node) eq 'ARRAY' and ! $j->{-is_single};
436 $aliases_by_type->{premultiplied}{$al} ||= { -parents => $j->{-join_path}||[] }
437 # parts of the path that are not us but are multiplying
438 if grep { $aliases_by_type->{multiplying}{$_} }
441 @{ $j->{-join_path}||[] }
444 # get a column to source/alias map (including unambiguous unqualified ones)
445 my $colinfo = $self->_resolve_column_info ($attrs->{from});
447 # set up a botched SQLA
448 my $sql_maker = $self->sql_maker;
450 # these are throw away results, do not pollute the bind stack
451 local $sql_maker->{where_bind};
452 local $sql_maker->{group_bind};
453 local $sql_maker->{having_bind};
454 local $sql_maker->{from_bind};
456 # we can't scan properly without any quoting (\b doesn't cut it
457 # everywhere), so unless there is proper quoting set - use our
458 # own weird impossible character.
459 # Also in the case of no quoting, we need to explicitly disable
460 # name_sep, otherwise sorry nasty legacy syntax like
461 # { 'count(foo.id)' => { '>' => 3 } } will stop working >:(
462 local $sql_maker->{quote_char} = $sql_maker->{quote_char};
463 local $sql_maker->{name_sep} = $sql_maker->{name_sep};
465 unless (defined $sql_maker->{quote_char} and length $sql_maker->{quote_char}) {
466 $sql_maker->{quote_char} = ["\x00", "\xFF"];
467 # if we don't unset it we screw up retarded but unfortunately working
468 # 'MAX(foo.bar)' => { '>', 3 }
469 $sql_maker->{name_sep} = '';
472 my ($lquote, $rquote, $sep) = map { quotemeta $_ } ($sql_maker->_quote_chars, $sql_maker->name_sep);
474 # generate sql chunks
477 ($sql_maker->_recurse_where ($attrs->{where}))[0],
478 $sql_maker->_parse_rs_attrs ({ having => $attrs->{having} }),
481 $sql_maker->_parse_rs_attrs ({ group_by => $attrs->{group_by} }),
484 $sql_maker->_recurse_from (
485 ref $attrs->{from}[0] eq 'ARRAY' ? $attrs->{from}[0][0] : $attrs->{from}[0],
486 @{$attrs->{from}}[1 .. $#{$attrs->{from}}],
490 # kill all selectors which look like a proper subquery
491 # this is a sucky heuristic *BUT* - if we get it wrong the query will simply
492 # fail to run, so we are relatively safe
494 { $_ !~ / \A \s* \( \s* SELECT \s+ .+? \s+ FROM \s+ .+? \) \s* \z /xsi }
496 { ($sql_maker->_recurse_fields($_))[0] }
501 ( my $sql = (ref $_ ? $_->[0] : $_) ) =~ s/ \s+ (?: ASC | DESC ) \s* \z //xi;
504 $sql_maker->_order_by_chunks( $attrs->{order_by} ),
508 # we will be bulk-scanning anyway - pieces will not matter in that case,
509 # thus join everything up
510 # throw away empty-string chunks, and make sure no binds snuck in
511 # note that we operate over @{$to_scan->{$type}}, hence the
512 # semi-mindbending ... map ... for values ...
513 ( $_ = join ' ', map {
515 ( ! defined $_ ) ? ()
516 : ( length ref $_ ) ? $self->throw_exception(
517 "Unexpected ref in scan-plan: " . dump_value $_
519 : ( $_ =~ /^\s*$/ ) ? ()
522 } @$_ ) for values %$to_scan;
524 # throw away empty to-scan's
526 length $to_scan->{$_}
528 delete $to_scan->{$_}
529 ) for keys %$to_scan;
533 # these will be used for matching in the loop below
534 my $all_aliases = join ' | ', map { quotemeta $_ } keys %$alias_list;
536 $lquote ( $all_aliases ) $rquote $sep (?: $lquote ([^$rquote]+) $rquote )?
538 \b ( $all_aliases ) \. ( [^\s\)\($rquote]+ )?
542 my $all_unq_columns = join ' | ',
546 # using a regex here shows up on profiles, boggle
547 { index( $_, '.') < 0 }
550 my $unq_col_re = $all_unq_columns
552 $lquote ( $all_unq_columns ) $rquote
554 (?: \A | \s ) ( $all_unq_columns ) (?: \s | \z )
560 # the actual scan, per type
561 for my $type (keys %$to_scan) {
564 # now loop through all fully qualified columns and get the corresponding
565 # alias (should work even if they are in scalarrefs)
567 # The regex captures in multiples of 4, with one of the two pairs being
568 # undef. There may be a *lot* of matches, hence the convoluted loop
569 my @matches = $to_scan->{$type} =~ /$fq_col_re/g;
571 while( $i < $#matches ) {
576 $aliases_by_type->{$type}{$matches[$i]} ||= { -parents => $alias_list->{$matches[$i]}{-join_path}||[] };
578 $aliases_by_type->{$type}{$matches[$i]}{-seen_columns}{"$matches[$i].$matches[$i+1]"} = "$matches[$i].$matches[$i+1]"
579 if defined $matches[$i+1];
588 # now loop through unqualified column names, and try to locate them within
589 # the chunks, if there are any unqualified columns in the 1st place
590 next unless $unq_col_re;
592 # The regex captures in multiples of 2, one of the two being undef
593 for ( $to_scan->{$type} =~ /$unq_col_re/g ) {
595 my $alias = $colinfo->{$_}{-source_alias} or next;
596 $aliases_by_type->{$type}{$alias} ||= { -parents => $alias_list->{$alias}{-join_path}||[] };
597 $aliases_by_type->{$type}{$alias}{-seen_columns}{"$alias.$_"} = $_
602 # Add any non-left joins to the restriction list (such joins are indeed restrictions)
606 ! $aliases_by_type->{restricting}{ $_->{-alias} }
611 $_->{-join_type} !~ /^left (?: \s+ outer)? $/xi
614 $aliases_by_type->{restricting}{ $_->{-alias} } = { -parents => $_->{-join_path}||[] }
615 ) for values %$alias_list;
620 keys %{$aliases_by_type->{$_}}
622 delete $aliases_by_type->{$_}
623 ) for keys %$aliases_by_type;
629 # This is the engine behind { distinct => 1 } and the general
630 # complex prefetch grouper
631 sub _group_over_selection {
632 my ($self, $attrs) = @_;
634 my $colinfos = $self->_resolve_column_info ($attrs->{from});
636 my (@group_by, %group_index);
638 # the logic is: if it is a { func => val } we assume an aggregate,
639 # otherwise if \'...' or \[...] we assume the user knows what is
640 # going on thus group over it
641 for (@{$attrs->{select}}) {
642 if (! ref($_) or ref ($_) ne 'HASH' ) {
645 if ($colinfos->{$_} and $_ !~ /\./ ) {
646 # add a fully qualified version as well
647 $group_index{"$colinfos->{$_}{-source_alias}.$_"}++;
652 my @order_by = $self->_extract_order_criteria($attrs->{order_by})
653 or return (\@group_by, $attrs->{order_by});
655 # add any order_by parts that are not already present in the group_by
656 # to maintain SQL cross-compatibility and general sanity
658 # also in case the original selection is *not* unique, or in case part
659 # of the ORDER BY refers to a multiplier - we will need to replace the
660 # skipped order_by elements with their MIN/MAX equivalents as to maintain
661 # the proper overall order without polluting the group criteria (and
662 # possibly changing the outcome entirely)
664 my ($leftovers, $sql_maker, @new_order_by, $order_chunks, $aliastypes);
666 my $group_already_unique = $self->_columns_comprise_identifying_set($colinfos, \@group_by);
668 for my $o_idx (0 .. $#order_by) {
670 # if the chunk is already a min/max function - there is nothing left to touch
671 next if $order_by[$o_idx][0] =~ /^ (?: min | max ) \s* \( .+ \) $/ix;
673 # only consider real columns (for functions the user got to do an explicit group_by)
676 @{$order_by[$o_idx]} != 1
678 # only declare an unknown *plain* identifier as "leftover" if we are called with
679 # aliastypes to examine. If there are none - we are still in _resolve_attrs, and
680 # can just assume the user knows what they want
681 ( ! ( $chunk_ci = $colinfos->{$order_by[$o_idx][0]} ) and $attrs->{_aliastypes} )
683 push @$leftovers, $order_by[$o_idx][0];
686 next unless $chunk_ci;
688 # no duplication of group criteria
689 next if $group_index{$chunk_ci->{-fq_colname}};
692 $attrs->{_aliastypes}
694 $self->_resolve_aliastypes_from_select_args({
695 from => $attrs->{from},
696 order_by => $attrs->{order_by},
698 ) if $group_already_unique;
700 # check that we are not ordering by a multiplier (if a check is requested at all)
702 $group_already_unique
704 ! $aliastypes->{multiplying}{$chunk_ci->{-source_alias}}
706 ! $aliastypes->{premultiplied}{$chunk_ci->{-source_alias}}
708 push @group_by, $chunk_ci->{-fq_colname};
709 $group_index{$chunk_ci->{-fq_colname}}++
712 # We need to order by external columns without adding them to the group
713 # (eiehter a non-unique selection, or a multi-external)
715 # This doesn't really make sense in SQL, however from DBICs point
716 # of view is rather valid (e.g. order the leftmost objects by whatever
717 # criteria and get the offset/rows many). There is a way around
718 # this however in SQL - we simply tae the direction of each piece
719 # of the external order and convert them to MIN(X) for ASC or MAX(X)
720 # for DESC, and group_by the root columns. The end result should be
721 # exactly what we expect
723 $sql_maker ||= $self->sql_maker;
725 map { ref $_ eq 'ARRAY' ? $_ : [ $_ ] } $sql_maker->_order_by_chunks($attrs->{order_by})
728 my ($chunk, $is_desc) = $sql_maker->_split_order_chunk($order_chunks->[$o_idx][0]);
730 # we reached that far - wrap any part of the order_by that "responded"
731 # to an ordering alias into a MIN/MAX
732 $new_order_by[$o_idx] = \[
733 sprintf( '%s( %s )%s',
734 ($is_desc ? 'MAX' : 'MIN'),
736 ($is_desc ? ' DESC' : ''),
738 @ {$order_chunks->[$o_idx]} [ 1 .. $#{$order_chunks->[$o_idx]} ]
743 $self->throw_exception ( sprintf
744 'Unable to programatically derive a required group_by from the supplied '
745 . 'order_by criteria. To proceed either add an explicit group_by, or '
746 . 'simplify your order_by to only include plain columns '
747 . '(supplied order_by: %s)',
748 join ', ', map { "'$_'" } @$leftovers,
751 # recreate the untouched order parts
753 $new_order_by[$_] ||= \ $order_chunks->[$_] for ( 0 .. $#$order_chunks );
758 (@new_order_by ? \@new_order_by : $attrs->{order_by} ), # same ref as original == unchanged
762 sub _resolve_ident_sources {
763 my ($self, $ident) = @_;
765 my $alias2source = {};
767 # the reason this is so contrived is that $ident may be a {from}
768 # structure, specifying multiple tables to join
769 if ( blessed $ident && $ident->isa("DBIx::Class::ResultSource") ) {
770 # this is compat mode for insert/update/delete which do not deal with aliases
771 $alias2source->{me} = $ident;
773 elsif (ref $ident eq 'ARRAY') {
777 if (ref $_ eq 'HASH') {
780 if (ref $_ eq 'ARRAY' and ref $_->[0] eq 'HASH') {
784 $alias2source->{$tabinfo->{-alias}} = $tabinfo->{-rsrc}
785 if ($tabinfo->{-rsrc});
789 return $alias2source;
792 # Takes $ident, \@column_names
794 # returns { $column_name => \%column_info, ... }
795 # also note: this adds -result_source => $rsrc to the column info
797 # If no columns_names are supplied returns info about *all* columns
799 sub _resolve_column_info {
800 my ($self, $ident, $colnames) = @_;
802 return {} if $colnames and ! @$colnames;
804 my $sources = $self->_resolve_ident_sources($ident);
806 $_ = { rsrc => $_, colinfos => $_->columns_info }
807 for values %$sources;
809 my (%seen_cols, @auto_colnames);
811 # compile a global list of column names, to be able to properly
812 # disambiguate unqualified column names (if at all possible)
813 for my $alias (keys %$sources) {
815 ++$seen_cols{$_}{$alias}
819 push @auto_colnames, "$alias.$_"
820 ) for keys %{ $sources->{$alias}{colinfos} };
825 ( grep { keys %{$seen_cols{$_}} == 1 } keys %seen_cols ),
830 my ($colname, $source_alias) = reverse split /\./, $_;
835 # if the column was seen exactly once - we know which rsrc it came from
839 keys %{$seen_cols{$colname}} == 1
841 ( %{$seen_cols{$colname}} )[0]
847 $self->throw_exception(
848 "No such column '$colname' on source " . $sources->{$assumed_alias}{rsrc}->source_name
849 ) unless $seen_cols{$colname}{$assumed_alias};
852 %{ $sources->{$assumed_alias}{colinfos}{$colname} },
853 -result_source => $sources->{$assumed_alias}{rsrc},
854 -source_alias => $assumed_alias,
855 -fq_colname => "$assumed_alias.$colname",
856 -colname => $colname,
859 $return{"$assumed_alias.$colname"} = $return{$_}
860 unless $source_alias;
866 # The DBIC relationship chaining implementation is pretty simple - every
867 # new related_relationship is pushed onto the {from} stack, and the {select}
868 # window simply slides further in. This means that when we count somewhere
869 # in the middle, we got to make sure that everything in the join chain is an
870 # actual inner join, otherwise the count will come back with unpredictable
871 # results (a resultset may be generated with _some_ rows regardless of if
872 # the relation which the $rs currently selects has rows or not). E.g.
873 # $artist_rs->cds->count - normally generates:
874 # SELECT COUNT( * ) FROM artist me LEFT JOIN cd cds ON cds.artist = me.artistid
875 # which actually returns the number of artists * (number of cds || 1)
877 # So what we do here is crawl {from}, determine if the current alias is at
878 # the top of the stack, and if not - make sure the chain is inner-joined down
881 sub _inner_join_to_node {
882 my ($self, $from, $alias) = @_;
884 my $switch_branch = $self->_find_join_path_to_node($from, $alias);
886 return $from unless @{$switch_branch||[]};
888 # So it looks like we will have to switch some stuff around.
889 # local() is useless here as we will be leaving the scope
890 # anyway, and deep cloning is just too fucking expensive
891 # So replace the first hashref in the node arrayref manually
892 my @new_from = ($from->[0]);
893 my $sw_idx = { map { (values %$_), 1 } @$switch_branch }; #there's one k/v per join-path
895 for my $j (@{$from}[1 .. $#$from]) {
896 my $jalias = $j->[0]{-alias};
898 if ($sw_idx->{$jalias}) {
899 my %attrs = %{$j->[0]};
900 delete $attrs{-join_type};
914 sub _find_join_path_to_node {
915 my ($self, $from, $target_alias) = @_;
917 # subqueries and other oddness are naturally not supported
921 ref $from->[0] ne 'HASH'
923 ! defined $from->[0]{-alias}
926 # no path - the head is the alias
927 return [] if $from->[0]{-alias} eq $target_alias;
929 for my $i (1 .. $#$from) {
930 return $from->[$i][0]{-join_path} if ( ($from->[$i][0]{-alias}||'') eq $target_alias );
933 # something else went quite wrong
937 sub _extract_order_criteria {
938 my ($self, $order_by, $sql_maker) = @_;
941 my ($sql_maker, $order_by, $orig_quote_chars) = @_;
943 return scalar $sql_maker->_order_by_chunks ($order_by)
946 my ($lq, $rq, $sep) = map { quotemeta($_) } (
947 ($orig_quote_chars ? @$orig_quote_chars : $sql_maker->_quote_chars),
952 for ($sql_maker->_order_by_chunks ($order_by) ) {
953 my $chunk = ref $_ ? [ @$_ ] : [ $_ ];
954 ($chunk->[0]) = $sql_maker->_split_order_chunk($chunk->[0]);
956 # order criteria may have come back pre-quoted (literals and whatnot)
957 # this is fragile, but the best we can currently do
958 $chunk->[0] =~ s/^ $lq (.+?) $rq $sep $lq (.+?) $rq $/"$1.$2"/xe
959 or $chunk->[0] =~ s/^ $lq (.+) $rq $/$1/x;
961 push @chunks, $chunk;
968 return $parser->($sql_maker, $order_by);
971 $sql_maker = $self->sql_maker;
973 # pass these in to deal with literals coming from
974 # the user or the deep guts of prefetch
975 my $orig_quote_chars = [$sql_maker->_quote_chars];
977 local $sql_maker->{quote_char};
978 return $parser->($sql_maker, $order_by, $orig_quote_chars);
982 sub _order_by_is_stable {
983 my ($self, $ident, $order_by, $where) = @_;
986 ( map { $_->[0] } $self->_extract_order_criteria($order_by) ),
987 ( $where ? keys %{ $self->_extract_fixed_condition_columns($where) } : () ),
990 my $colinfo = $self->_resolve_column_info($ident, \@cols);
992 return keys %$colinfo
993 ? $self->_columns_comprise_identifying_set( $colinfo, \@cols )
998 sub _columns_comprise_identifying_set {
999 my ($self, $colinfo, $columns) = @_;
1002 $cols_per_src -> {$_->{-source_alias}} -> {$_->{-colname}} = $_
1003 for grep { defined $_ } @{$colinfo}{@$columns};
1005 for (values %$cols_per_src) {
1006 my $src = (values %$_)[0]->{-result_source};
1007 return 1 if $src->_identifying_column_set($_);
1013 # this is almost similar to _order_by_is_stable, except it takes
1014 # a single rsrc, and will succeed only if the first portion of the order
1016 # returns that portion as a colinfo hashref on success
1017 sub _extract_colinfo_of_stable_main_source_order_by_portion {
1018 my ($self, $attrs) = @_;
1020 my $nodes = $self->_find_join_path_to_node($attrs->{from}, $attrs->{alias});
1022 return unless defined $nodes;
1026 ( $self->_extract_order_criteria($attrs->{order_by}) )
1028 return unless @ord_cols;
1030 my $valid_aliases = { map { $_ => 1 } (
1031 $attrs->{from}[0]{-alias},
1032 map { values %$_ } @$nodes,
1035 my $colinfos = $self->_resolve_column_info($attrs->{from});
1037 my ($colinfos_to_return, $seen_main_src_cols);
1039 for my $col (@ord_cols) {
1040 # if order criteria is unresolvable - there is nothing we can do
1041 my $colinfo = $colinfos->{$col} or last;
1043 # if we reached the end of the allowed aliases - also nothing we can do
1044 last unless $valid_aliases->{$colinfo->{-source_alias}};
1046 $colinfos_to_return->{$col} = $colinfo;
1048 $seen_main_src_cols->{$colinfo->{-colname}} = 1
1049 if $colinfo->{-source_alias} eq $attrs->{alias};
1052 # FIXME the condition may be singling out things on its own, so we
1053 # conceivable could come back wi "stable-ordered by nothing"
1054 # not confient enough in the parser yet, so punt for the time being
1055 return unless $seen_main_src_cols;
1057 my $main_src_fixed_cols_from_cond = [ $attrs->{where}
1061 ( $colinfos->{$_} and $colinfos->{$_}{-source_alias} eq $attrs->{alias} )
1062 ? $colinfos->{$_}{-colname}
1065 keys %{ $self->_extract_fixed_condition_columns($attrs->{where}) }
1070 return $attrs->{result_source}->_identifying_column_set([
1071 keys %$seen_main_src_cols,
1072 @$main_src_fixed_cols_from_cond,
1073 ]) ? $colinfos_to_return : ();
1076 # Attempts to flatten a passed in SQLA condition as much as possible towards
1077 # a plain hashref, *without* altering its semantics. Required by
1078 # create/populate being able to extract definitive conditions from preexisting
1079 # resultset {where} stacks
1081 # FIXME - while relatively robust, this is still imperfect, one of the first
1082 # things to tackle when we get access to a formalized AST. Note that this code
1083 # is covered by a *ridiculous* amount of tests, so starting with porting this
1084 # code would be a rather good exercise
1085 sub _collapse_cond {
1086 my ($self, $where, $where_is_anded_array) = @_;
1093 elsif ($where_is_anded_array or ref $where eq 'HASH') {
1097 my @pieces = $where_is_anded_array ? @$where : $where;
1099 my $chunk = shift @pieces;
1101 if (ref $chunk eq 'HASH') {
1102 for (sort keys %$chunk) {
1104 # Match SQLA 1.79 behavior
1105 unless( length $_ ) {
1106 is_literal_value($chunk->{$_})
1107 ? carp 'Hash-pairs consisting of an empty string with a literal are deprecated, use -and => [ $literal ] instead'
1108 : $self->throw_exception("Supplying an empty left hand side argument is not supported in hash-pairs")
1112 push @pairs, $_ => $chunk->{$_};
1115 elsif (ref $chunk eq 'ARRAY') {
1116 push @pairs, -or => $chunk
1119 elsif ( ! length ref $chunk) {
1121 # Match SQLA 1.79 behavior
1122 $self->throw_exception("Supplying an empty left hand side argument is not supported in array-pairs")
1123 if $where_is_anded_array and (! defined $chunk or ! length $chunk);
1125 push @pairs, $chunk, shift @pieces;
1128 push @pairs, '', $chunk;
1132 return unless @pairs;
1134 my @conds = $self->_collapse_cond_unroll_pairs(\@pairs)
1137 # Consolidate various @conds back into something more compact
1138 for my $c (@conds) {
1139 if (ref $c ne 'HASH') {
1140 push @{$fin->{-and}}, $c;
1143 for my $col (sort keys %$c) {
1145 # consolidate all -and nodes
1146 if ($col =~ /^\-and$/i) {
1147 push @{$fin->{-and}},
1148 ref $c->{$col} eq 'ARRAY' ? @{$c->{$col}}
1149 : ref $c->{$col} eq 'HASH' ? %{$c->{$col}}
1150 : { $col => $c->{$col} }
1153 elsif ($col =~ /^\-/) {
1154 push @{$fin->{-and}}, { $col => $c->{$col} };
1156 elsif (exists $fin->{$col}) {
1157 $fin->{$col} = [ -and => map {
1158 (ref $_ eq 'ARRAY' and ($_->[0]||'') =~ /^\-and$/i )
1162 } ($fin->{$col}, $c->{$col}) ];
1165 $fin->{$col} = $c->{$col};
1171 elsif (ref $where eq 'ARRAY') {
1172 # we are always at top-level here, it is safe to dump empty *standalone* pieces
1175 for (my $i = 0; $i <= $#$where; $i++ ) {
1177 # Match SQLA 1.79 behavior
1178 $self->throw_exception(
1179 "Supplying an empty left hand side argument is not supported in array-pairs"
1180 ) if (! defined $where->[$i] or ! length $where->[$i]);
1182 my $logic_mod = lc ( ($where->[$i] =~ /^(\-(?:and|or))$/i)[0] || '' );
1186 $self->throw_exception("Unsupported top-level op/arg pair: [ $logic_mod => $where->[$i] ]")
1187 unless ref $where->[$i] eq 'HASH' or ref $where->[$i] eq 'ARRAY';
1189 my $sub_elt = $self->_collapse_cond({ $logic_mod => $where->[$i] })
1192 my @keys = keys %$sub_elt;
1193 if ( @keys == 1 and $keys[0] !~ /^\-/ ) {
1194 $fin_idx->{ "COL_$keys[0]_" . serialize $sub_elt } = $sub_elt;
1197 $fin_idx->{ "SER_" . serialize $sub_elt } = $sub_elt;
1200 elsif (! length ref $where->[$i] ) {
1201 my $sub_elt = $self->_collapse_cond({ @{$where}[$i, $i+1] })
1204 $fin_idx->{ "COL_$where->[$i]_" . serialize $sub_elt } = $sub_elt;
1208 $fin_idx->{ "SER_" . serialize $where->[$i] } = $self->_collapse_cond( $where->[$i] ) || next;
1215 elsif ( keys %$fin_idx == 1 ) {
1216 $fin = (values %$fin_idx)[0];
1221 # at this point everything is at most one level deep - unroll if needed
1222 for (sort keys %$fin_idx) {
1223 if ( ref $fin_idx->{$_} eq 'HASH' and keys %{$fin_idx->{$_}} == 1 ) {
1224 my ($l, $r) = %{$fin_idx->{$_}};
1230 ( @$r == 1 and $l =~ /^\-and$/i )
1243 $l =~ /^\-(?:and|or)$/i
1253 push @or, $fin_idx->{$_};
1261 # not a hash not an array
1262 $fin = { -and => [ $where ] };
1265 # unroll single-element -and's
1271 my $and = delete $fin->{-and};
1274 # at this point we have @$and == 1
1276 ref $and->[0] eq 'HASH'
1278 ! grep { exists $fin->{$_} } keys %{$and->[0]}
1285 $fin->{-and} = $and;
1290 # compress same-column conds found in $fin
1291 for my $col ( grep { $_ !~ /^\-/ } keys %$fin ) {
1292 next unless ref $fin->{$col} eq 'ARRAY' and ($fin->{$col}[0]||'') =~ /^\-and$/i;
1293 my $val_bag = { map {
1294 (! defined $_ ) ? ( UNDEF => undef )
1295 : ( ! length ref $_ or is_plain_value $_ ) ? ( "VAL_$_" => $_ )
1296 : ( ( 'SER_' . serialize $_ ) => $_ )
1297 } @{$fin->{$col}}[1 .. $#{$fin->{$col}}] };
1299 if (keys %$val_bag == 1 ) {
1300 ($fin->{$col}) = values %$val_bag;
1303 $fin->{$col} = [ -and => map { $val_bag->{$_} } sort keys %$val_bag ];
1307 return keys %$fin ? $fin : ();
1310 sub _collapse_cond_unroll_pairs {
1311 my ($self, $pairs) = @_;
1316 my ($lhs, $rhs) = splice @$pairs, 0, 2;
1318 if (! length $lhs) {
1319 push @conds, $self->_collapse_cond($rhs);
1321 elsif ( $lhs =~ /^\-and$/i ) {
1322 push @conds, $self->_collapse_cond($rhs, (ref $rhs eq 'ARRAY'));
1324 elsif ( $lhs =~ /^\-or$/i ) {
1325 push @conds, $self->_collapse_cond(
1326 (ref $rhs eq 'HASH') ? [ map { $_ => $rhs->{$_} } sort keys %$rhs ] : $rhs
1330 if (ref $rhs eq 'HASH' and ! keys %$rhs) {
1331 # FIXME - SQLA seems to be doing... nothing...?
1333 # normalize top level -ident, for saner extract_fixed_condition_columns code
1334 elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{-ident}) {
1335 push @conds, { $lhs => { '=', $rhs } };
1337 elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{-value} and is_plain_value $rhs->{-value}) {
1338 push @conds, { $lhs => $rhs->{-value} };
1340 elsif (ref $rhs eq 'HASH' and keys %$rhs == 1 and exists $rhs->{'='}) {
1341 if ( length ref $rhs->{'='} and is_literal_value $rhs->{'='} ) {
1342 push @conds, { $lhs => $rhs };
1345 for my $p ($self->_collapse_cond_unroll_pairs([ $lhs => $rhs->{'='} ])) {
1347 # extra sanity check
1349 local $Data::Dumper::Deepcopy = 1;
1350 $self->throw_exception(
1351 "Internal error: unexpected collapse unroll:"
1352 . dump_value { in => { $lhs => $rhs }, out => $p }
1361 # the unroller recursion may return a '=' prepended value already
1362 ref $r eq 'HASH' and keys %$rhs == 1 and exists $rhs->{'='}
1367 : { $l => { '=' => $r } }
1372 elsif (ref $rhs eq 'ARRAY') {
1373 # some of these conditionals encounter multi-values - roll them out using
1374 # an unshift, which will cause extra looping in the while{} above
1376 push @conds, { $lhs => [] };
1378 elsif ( ($rhs->[0]||'') =~ /^\-(?:and|or)$/i ) {
1379 $self->throw_exception("Value modifier not followed by any values: $lhs => [ $rhs->[0] ] ")
1382 if( $rhs->[0] =~ /^\-and$/i ) {
1383 unshift @$pairs, map { $lhs => $_ } @{$rhs}[1..$#$rhs];
1385 # if not an AND then it's an OR
1387 unshift @$pairs, $lhs => $rhs->[1];
1390 push @conds, { $lhs => [ @{$rhs}[1..$#$rhs] ] };
1393 elsif (@$rhs == 1) {
1394 unshift @$pairs, $lhs => $rhs->[0];
1397 push @conds, { $lhs => $rhs };
1400 # unroll func + { -value => ... }
1404 ( my ($subop) = keys %$rhs ) == 1
1406 length ref ((values %$rhs)[0])
1408 my $vref = is_plain_value( (values %$rhs)[0] )
1410 push @conds, { $lhs => { $subop => $$vref } }
1413 push @conds, { $lhs => $rhs };
1421 # Analyzes a given condition and attempts to extract all columns
1422 # with a definitive fixed-condition criteria. Returns a hashref
1423 # of k/v pairs suitable to be passed to set_columns(), with a
1424 # MAJOR CAVEAT - multi-value (contradictory) equalities are still
1425 # represented as a reference to the UNRESOVABLE_CONDITION constant
1426 # The reason we do this is that some codepaths only care about the
1427 # codition being stable, as opposed to actually making sense
1429 # The normal mode is used to figure out if a resultset is constrained
1430 # to a column which is part of a unique constraint, which in turn
1431 # allows us to better predict how ordering will behave etc.
1433 # With the optional "consider_nulls" boolean argument, the function
1434 # is instead used to infer inambiguous values from conditions
1435 # (e.g. the inheritance of resultset conditions on new_result)
1437 sub _extract_fixed_condition_columns {
1438 my ($self, $where, $consider_nulls) = @_;
1439 my $where_hash = $self->_collapse_cond($_[1]);
1443 for $c (keys %$where_hash) {
1446 if (!defined ($v = $where_hash->{$c}) ) {
1447 $vals->{UNDEF} = $v if $consider_nulls
1454 if (exists $v->{-value}) {
1455 if (defined $v->{-value}) {
1456 $vals->{"VAL_$v->{-value}"} = $v->{-value}
1458 elsif( $consider_nulls ) {
1459 $vals->{UNDEF} = $v->{-value};
1462 # do not need to check for plain values - _collapse_cond did it for us
1464 length ref $v->{'='}
1467 ( ref $v->{'='} eq 'HASH' and keys %{$v->{'='}} == 1 and exists $v->{'='}{-ident} )
1469 is_literal_value($v->{'='})
1472 $vals->{ 'SER_' . serialize $v->{'='} } = $v->{'='};
1480 $vals->{"VAL_$v"} = $v;
1482 elsif (ref $v eq 'ARRAY' and ($v->[0]||'') eq '-and') {
1483 for ( @{$v}[1..$#$v] ) {
1484 my $subval = $self->_extract_fixed_condition_columns({ $c => $_ }, 'consider nulls'); # always fish nulls out on recursion
1485 next unless exists $subval->{$c}; # didn't find anything
1487 ! defined $subval->{$c} ? 'UNDEF'
1488 : ( ! length ref $subval->{$c} or is_plain_value $subval->{$c} ) ? "VAL_$subval->{$c}"
1489 : ( 'SER_' . serialize $subval->{$c} )
1494 if (keys %$vals == 1) {
1495 ($res->{$c}) = (values %$vals)
1496 unless !$consider_nulls and exists $vals->{UNDEF};
1498 elsif (keys %$vals > 1) {
1499 $res->{$c} = UNRESOLVABLE_CONDITION;