From: Peter Rabbitson Date: Sat, 3 Mar 2012 11:02:01 +0000 (+0100) Subject: Cleanup of stale constructor codepath comments X-Git-Tag: v0.08240~25 X-Git-Url: http://git.shadowcat.co.uk/gitweb/gitweb.cgi?p=dbsrgits%2FDBIx-Class.git;a=commitdiff_plain;h=2d0b795a54a018d5c9cf2593cf83045962cd9b93;hp=4e9fc3f33df616fb7340d05e304ff985b9cce9cb Cleanup of stale constructor codepath comments --- diff --git a/lib/DBIx/Class/ResultSet.pm b/lib/DBIx/Class/ResultSet.pm index 2c0207e..a9c4ab7 100644 --- a/lib/DBIx/Class/ResultSet.pm +++ b/lib/DBIx/Class/ResultSet.pm @@ -1237,14 +1237,22 @@ sub next { return shift @{$self->{stashed_objects}}; } -# takes a single DBI-row of data and coinstructs as many objects -# as the resultset attributes call for. -# This can be a bit of an action at a distance - it takes as an argument -# the *current* cursor-row (already taken off the $sth), but if -# collapsing is requested it will keep advancing the cursor either -# until the current row-object is assembled (the collapser was able to -# order the result sensibly) OR until the cursor is exhausted (an -# unordered collapsing resultset effectively triggers ->all) +# Constructs as many objects as it can in one pass while respecting +# cursor laziness. Several modes of operation: +# +# * Always builds everything present in @{$self->{stashed_rows}} +# * If called with $fetch_all true - pulls everything off the cursor and +# builds all objects in one pass +# * If $self->_resolved_attrs->{collapse} is true, checks the order_by +# and if the resultset is ordered properly by the left side: +# * Fetches stuff off the cursor until the "master object" changes, +# and saves the last extra row (if any) in @{$self->{stashed_rows}} +# OR +# * Just fetches, and collapses/constructs everything as if $fetch_all +# was requested (there is no other way to collapse except for an +# eager cursor) +# * If no collapse is requested - just get the next row, construct and +# return sub _construct_objects { my ($self, $fetch_all) = @_; @@ -1257,10 +1265,11 @@ sub _construct_objects { # a suprising amount actually my $rows = (delete $self->{stashed_rows}) || []; if ($fetch_all) { - # FIXME - we can do better, cursor->next/all (well diff. methods) should return a ref + # FIXME SUBOPTIMAL - we can do better, cursor->next/all (well diff. methods) should return a ref $rows = [ @$rows, $cursor->all ]; } elsif (!$attrs->{collapse}) { + # FIXME SUBOPTIMAL - we can do better, cursor->next/all (well diff. methods) should return a ref push @$rows, do { my @r = $cursor->next; @r ? \@r : () } unless @$rows; } @@ -1286,8 +1295,8 @@ sub _construct_objects { } # since all we check here are the start of the order_by belonging to the - # top level $rsrc, the order stability check will fail unless the whole - # thing is ordered as we need it + # top level $rsrc, a present identifying set will mean that the resultset + # is ordered by its leftmost table in a tsable manner (@ord_cols and $rsrc->_identifying_column_set({ map { $colinfos->{$_}{-colname} => $colinfos->{$_} } @ord_cols @@ -1298,6 +1307,7 @@ sub _construct_objects { push @$rows, do { my @r = $cursor->next; @r ? \@r : () }; } # instead of looping over ->next, use ->all in stealth mode + # FIXME - encapsulation breach, got to be a better way elsif (! $cursor->{done}) { push @$rows, $cursor->all; $cursor->{done} = 1; @@ -1342,8 +1352,9 @@ sub _construct_objects { selection => $attrs->{select}, collapse => $attrs->{collapse}, }) or die $@)->($rows, $fetch_all ? () : ( - sub { my @r = $cursor->next or return; \@r }, - ($self->{stashed_rows} = []), + # FIXME SUBOPTIMAL - we can do better, cursor->next/all (well diff. methods) should return a ref + sub { my @r = $cursor->next or return; \@r }, # how the collapser gets more rows + ($self->{stashed_rows} = []), # where does it stuff excess )); # modify $rows in-place, shrinking/extending as necessary $_ = $inflator->($res_class, $rsrc, @$_) for @$rows; diff --git a/lib/DBIx/Class/ResultSource/RowParser.pm b/lib/DBIx/Class/ResultSource/RowParser.pm index d71ca30..550c9e5 100644 --- a/lib/DBIx/Class/ResultSource/RowParser.pm +++ b/lib/DBIx/Class/ResultSource/RowParser.pm @@ -63,11 +63,12 @@ sub _resolve_prefetch { } } -# Takes a selection list and generates a collapse-map representing +# Takes an arrayref selection list and generates a collapse-map representing # row-object fold-points. Every relationship is assigned a set of unique, # non-nullable columns (which may *not even be* from the same resultset) # and the collapser will use this information to correctly distinguish -# data of individual to-be-row-objects. +# data of individual to-be-row-objects. See t/resultset/rowparser_internals.t +# for extensive RV examples sub _resolve_collapse { my ($self, $as, $as_fq_idx, $rel_chain, $parent_info, $node_idx_ref) = @_; @@ -270,54 +271,28 @@ sub _resolve_collapse { } # Takes an arrayref of {as} dbic column aliases and the collapse and select -# attributes from the same $rs (the slector requirement is a temporary -# workaround), and returns a coderef capable of: -# my $me_pref_clps = $coderef->([$rs->cursor->next]) -# Where the $me_pref_clps arrayref is the future argument to -# ::ResultSet::_collapse_result. -# -# $me_pref_clps->[0] is always returned (even if as an empty hash with no -# rowdata), however branches of related data in $me_pref_clps->[1] may be -# pruned short of what was originally requested based on {as}, depending -# on: -# -# * If collapse is requested, a definitive collapse map is calculated for -# every relationship "fold-point", consisting of a set of values (which -# may not even be contained in the future 'me' of said relationship -# (for example a cd.artist_id defines the related inner-joined artist)). -# Thus a definedness check is carried on all collapse-condition values -# and if at least one is undef it is assumed that we are dealing with a -# NULLed right-side of a left-join, so we don't return a related data -# container at all, which implies no related objects -# -# * If we are not collapsing, there is no constraint on having a selector -# uniquely identifying all possible objects, and the user might have very -# well requested a column that just *happens* to be all NULLs. What we do -# in this case is fallback to the old behavior (which is a potential FIXME) -# by always returning a data container, but only filling it with columns -# IFF at least one of them is defined. This way we do not get an object -# with a bunch of has_column_loaded to undef, but at the same time do not -# further relationships based off this "null" object (e.g. in case the user -# deliberately skipped link-table values). I am pretty sure there are some -# tests that codify this behavior, need to find the exact testname. +# attributes from the same $rs (the selector requirement is a temporary +# workaround... I hope), and returns a coderef capable of: +# my $me_pref_clps = $coderef->([$rs->cursor->next/all]) +# Where the $me_pref_clps arrayref is the future argument to inflate_result() # # For an example of this coderef in action (and to see its guts) look at -# t/prefetch/_internals.t +# t/resultset/rowparser_internals.t # -# This is a huge performance win, as we call the same code for -# every row returned from the db, thus avoiding repeated method -# lookups when traversing relationships +# This is a huge performance win, as we call the same code for # every row +# returned from the db, thus avoiding repeated method lookups when traversing +# relationships # # Also since the coderef is completely stateless (the returned structure is # always fresh on every new invocation) this is a very good opportunity for # memoization if further speed improvements are needed # -# The way we construct this coderef is somewhat fugly, although I am not -# sure if the string eval is *that* bad of an idea. The alternative is to -# have a *very* large number of anon coderefs calling each other in a twisty -# maze, whereas the current result is a nice, smooth, single-pass function. +# The way we construct this coderef is somewhat fugly, although the result is +# really worth it. The final coderef does not perform any kind of recursion - +# the entire nested structure constructor is rolled out into a single scope. +# # In any case - the output of this thing is meticulously micro-tested, so -# any sort of rewrite should be relatively easy +# any sort of adjustment/rewrite should be relatively easy (fsvo relatively) # sub _mk_row_parser { my ($self, $args) = @_; @@ -327,8 +302,30 @@ sub _mk_row_parser { ( 0 .. $#{$args->{inflate_map}} ) }; - my ($parser_src); - if ($args->{collapse}) { + my $parser_src; + + # the non-collapsing assembler is easy + # FIXME SUBOPTIMAL there could be a yet faster way to do things here, but + # need to try an actual implementation and benchmark it: + # + # First setup the nested data structure you want for each row + # Then call bind_col() to alias the row fields into the right place in + # the data structure, then to fetch the data do: + # push @rows, dclone($row_data_struct) while ($sth->fetchrow); + # + if (!$args->{collapse}) { + $parser_src = sprintf('$_ = %s for @{$_[0]}', __visit_infmap_simple( + $inflate_index, + { rsrc => $self }, # need the $rsrc to sanity-check inflation map once + )); + + # change the quoted placeholders to unquoted alias-references + $parser_src =~ s/ \' \xFF__VALPOS__(\d+)__\xFF \' /"\$_->[$1]"/gex; + } + + # the collapsing parser is more complicated - it needs to keep a lot of state + # + else { my $collapse_map = $self->_resolve_collapse ( # FIXME @@ -337,7 +334,8 @@ sub _mk_row_parser { # alias random crap to existing column names anyway, but still - just in # case # FIXME !!!! - this does not yet deal with unbalanced selectors correctly - # (it is now trivial as the attrs specify where things go out of sync) + # (it is now trivial as the attrs specify where things go out of sync + # needs MOAR tests) { map { ref $args->{selection}[$inflate_index->{$_}] ? () : ( $_ => $inflate_index->{$_} ) } keys %$inflate_index @@ -389,24 +387,16 @@ sub _mk_row_parser { ### END STRING EVAL EOS + # !!! note - different var than the one above # change the quoted placeholders to unquoted alias-references $parser_src =~ s/ \' \xFF__VALPOS__(\d+)__\xFF \' /"\$cur_row->[$1]"/gex; $parser_src =~ s/ \' \xFF__IDVALPOS__(\d+)__\xFF \' /"\$cur_row_ids[$1]"/gex; } - else { - $parser_src = sprintf('$_ = %s for @{$_[0]}', __visit_infmap_simple( - $inflate_index, { rsrc => $self }), # need the $rsrc to determine left-ness - ); - - # change the quoted placeholders to unquoted alias-references - # !!! note - different var than the one above - $parser_src =~ s/ \' \xFF__VALPOS__(\d+)__\xFF \' /"\$_->[$1]"/gex; - } - $parser_src; } +# the simple non-collapsing nested structure recursor sub __visit_infmap_simple { my ($val_idx, $args) = @_; @@ -423,18 +413,18 @@ sub __visit_infmap_simple { my @relperl; for my $rel (sort keys %$rel_cols) { - my $rel_rsrc = __get_related_source($args->{rsrc}, $rel, $rel_cols->{$rel}); - + # DISABLEPRUNE #my $optional = $args->{is_optional}; #$optional ||= ($args->{rsrc}->relationship_info($rel)->{attrs}{join_type} || '') =~ /^left/i; push @relperl, join ' => ', perlstring($rel), __visit_infmap_simple($rel_cols->{$rel}, { - non_top => 1, + rsrc => __get_related_source($args->{rsrc}, $rel, $rel_cols->{$rel}), + # DISABLEPRUNE + #non_top => 1, #is_optional => $optional, - rsrc => $rel_rsrc, }); - # FIXME SUBOPTIMAL - disabled to satisfy t/resultset/inflate_result_api.t + # FIXME SUBOPTIMAL DISABLEPRUNE - disabled to satisfy t/resultset/inflate_result_api.t #if ($optional and my @branch_null_checks = map # { "(! defined '\xFF__VALPOS__${_}__\xFF')" } # sort { $a <=> $b } values %{$rel_cols->{$rel}} @@ -458,6 +448,7 @@ sub __visit_infmap_simple { ); } +# the collapsing nested structure recursor sub __visit_infmap_collapse { my ($val_idx, $collapse_map, $parent_info) = @_; @@ -512,6 +503,7 @@ sub __visit_infmap_collapse { ); } + # DISABLEPRUNE #my $known_defined = { %{ $parent_info->{known_defined} || {} } }; #$known_defined->{$_}++ for @{$collapse_map->{-node_id}}; @@ -524,10 +516,11 @@ sub __visit_infmap_collapse { node_idx => $collapse_map->{-node_index}, sequenced_node_id => $sequenced_node_id, relname => $rel, + # DISABLEPRUNE #known_defined => $known_defined, }); - # FIXME SUBOPTIMAL - disabled to satisfy t/resultset/inflate_result_api.t + # FIXME SUBOPTIMAL DISABLEPRUNE - disabled to satisfy t/resultset/inflate_result_api.t #if ($collapse_map->{$rel}{-is_optional} and my @null_checks = map # { "(! defined '\xFF__IDVALPOS__${_}__\xFF')" } # sort { $a <=> $b } grep @@ -575,7 +568,7 @@ sub __visit_dump { ($dumper_obj ||= do { require Data::Dumper; Data::Dumper->new([]) - ->Useperl (1) + ->Useperl (0) ->Purity (1) ->Pad ('') ->Useqq (0) diff --git a/lib/DBIx/Class/Row.pm b/lib/DBIx/Class/Row.pm index c59f70a..51b5325 100644 --- a/lib/DBIx/Class/Row.pm +++ b/lib/DBIx/Class/Row.pm @@ -1140,15 +1140,9 @@ sub inflate_result { foreach my $pre (keys %{$prefetch||{}}) { my @pre_vals; - if (! @{$prefetch->{$pre}}) { - # nothing, empty @pre_vals is put in the caches - } - elsif (ref $prefetch->{$pre}[0] eq 'ARRAY') { - @pre_vals = @{$prefetch->{$pre}}; - } - else { - @pre_vals = $prefetch->{$pre}; - } + @pre_vals = (ref $prefetch->{$pre}[0] eq 'ARRAY') + ? @{$prefetch->{$pre}} : $prefetch->{$pre} + if @{$prefetch->{$pre}}; my $pre_source = $source->related_source($pre); @@ -1160,7 +1154,8 @@ sub inflate_result { # FIXME SUBOPTIMAL - the new row parsers can very well optimize # this away entirely, and *never* return such empty rows. - # For now we maintain inflate_result API backcompat + # For now we maintain inflate_result API backcompat, see + # t/resultset/inflate_result_api.t next unless first { defined $_ } values %{$me_pref->[0]}; push @pre_objects, $pre_source->result_class->inflate_result(