3 The available [[ikiwiki/pagespec/sorting]] methods are currently hard-coded in
4 IkiWiki.pm, making it difficult to add any extra sorting mechanisms. I've
5 prepared a branch which adds 'sort' as a hook type and uses it to implement a
6 new `meta_title` sort type.
8 Someone could use this hook to make `\[[!inline sort=title]]` prefer the meta
9 title over the page name, but for compatibility, I'm not going to (I do wonder
10 whether it would be worth making sort=name an alias for the current sort=title,
11 and changing the meaning of sort=title in 4.0, though).
13 *[sort-hooks branch now withdrawn in favour of sort-package --s]*
15 I briefly tried to turn *all* the current sort types into hook functions, and
16 have some of them pre-registered, but decided that probably wasn't a good idea.
17 That earlier version of the branch is also available for comparison:
19 *[also withdrawn in favour of sort-package --s]*
21 >> I wonder if IkiWiki would benefit from the concept of a "sortspec", like a [[ikiwiki/PageSpec]] but dedicated to sorting lists of pages rather than defining lists of pages? Rather than defining a sort-hook, define a SortSpec class, and enable people to add their own sort methods as functions defined inside that class, similarly to the way they can add their own pagespec definitions. --[[KathrynAndersen]]
23 >>> [[!template id=gitbranch branch=smcv/ready/sort-package author="[[Simon_McVittie|smcv]]"]]
24 >>> I'd be inclined to think that's overkill, but it wasn't very hard to
25 >>> implement, and in a way is more elegant. I set it up so sort mechanisms
26 >>> share the `IkiWiki::PageSpec` package, but with a `cmp_` prefix. Gitweb:
27 >>> <http://git.pseudorandom.co.uk/smcv/ikiwiki.git?a=shortlog;h=refs/heads/sort-package>
29 >>>> I agree it seems more elegant, so I have focused on it.
31 >>>> I don't know about reusing `IkiWiki::PageSpec` for this.
34 >>>>> Fair enough, `IkiWiki::SortSpec::cmp_foo` would be just
35 >>>>> as easy, or `IkiWiki::Sorting::cmp_foo` if you don't like
36 >>>>> introducing "sort spec" in the API. I took a cue from
37 >>>>> [[ikiwiki/pagespec/sorting]] being a subpage of
38 >>>>> [[ikiwiki/pagespec]], and decided that yes, sorting is
39 >>>>> a bit like a pagespec :-) Which name would you prefer? --s
41 >>>>>> `SortSpec` --[[Joey]]
45 >>>> I would be inclined to drop the `check_` stuff. --[[Joey]]
47 >>>>> It basically exists to support `title_natural`, to avoid
48 >>>>> firing up the whole import mechanism on every `cmp`
49 >>>>> (although I suppose that could just be a call to a
50 >>>>> memoized helper function). It also lets sort specs that
51 >>>>> *must* have a parameter, like
52 >>>>> [[field|plugins/contrib/field/discussion]], fail early
53 >>>>> (again, not so valuable).
55 >>>>>> AFAIK, `use foo` has very low overhead when the module is already
56 >>>>>> loaded. There could be some evalation overhead in `eval q{use foo}`,
57 >>>>>> if so it would be worth addressing across the whole codebase.
60 >>>>>>> check_cmp_foo now dropped. --s
62 >>>>> The former function could be achieved at a small
63 >>>>> compatibility cost by putting `title_natural` in a new
64 >>>>> `sortnatural` plugin (that fails to load if you don't
65 >>>>> have `title_natural`), if you'd prefer - that's what would
66 >>>>> have happened if `title_natural` was written after this
67 >>>>> code had been merged, I suspect. Would you prefer this? --s
69 >>>>>> Yes! (Assuming it does not make sense to support
70 >>>>>> natural order sort of other keys than the title, at least..)
73 >>>>>>> Done. I added some NEWS.Debian for it, too. --s
75 >>>> Wouldn't it make sense to have `meta(title)` instead
76 >>>> of `meta_title`? --[[Joey]]
78 >>>>> Yes, you're right. I added parameters to support `field`,
79 >>>>> and didn't think about making `meta` use them too.
80 >>>>> However, `title` does need a special case to make it
81 >>>>> default to the basename instead of the empty string.
83 >>>>> Another special case for `title` is to use `titlesort`
84 >>>>> first (the name `titlesort` is derived from Ogg/FLAC
85 >>>>> tags, which can have `titlesort` and `artistsort`).
86 >>>>> I could easily extend that to other metas, though;
87 >>>>> in fact, for e.g. book lists it would be nice for
88 >>>>> `field(bookauthor)` to behave similarly, so you can
89 >>>>> display "Douglas Adams" but sort by "Adams, Douglas".
91 >>>>> `meta_title` is also meant to be a prototype of how
92 >>>>> `sort=title` could behave in 4.0 or something - sorting
93 >>>>> by page name (which usually sorts in approximately the
94 >>>>> same place as the meta-title, but occasionally not), while
95 >>>>> displaying meta-titles, does look quite odd. --s
97 >>>>>> Agreed. --[[Joey]]
99 >>>>>>> I've implemented meta(title). meta(author) also has the
100 >>>>>>> `sortas` special case; meta(updated) and meta(date)
101 >>>>>>> should also work how you'd expect them to (but they're
102 >>>>>>> earliest-first, unlike age). --s
104 >>>> As I read the regexp in `cmpspec_translate`, the "command"
105 >>>> is required to have params. They should be optional,
106 >>>> to match the documentation and because most sort methods
107 >>>> do not need parameters. --[[Joey]]
109 >>>>> No, `$2` is either `\w+\([^\)]*\)` or `[^\s]+` (with the
110 >>>>> latter causing an error later if it doesn't also match `\w+`).
111 >>>>> This branch doesn't add any parameterized sort methods,
112 >>>>> in fact, although I did provide one on
113 >>>>> [[field's_discussion_page|plugins/contrib/report/discussion]]. --s
115 >>>> I wonder if it would make sense to add some combining keywords, so
116 >>>> a sortspec reads like `sort="age then ascending title"`
117 >>>> In a way, this reduces the amount of syntax that needs to be learned.
118 >>>> I like the "then" (and it could allow other operations than
119 >>>> simple combination, if any others make sense). Not so sure about the
120 >>>> "ascending", which could be "reverse" instead, but "descending age" and
121 >>>> "ascending age" both seem useful to be able to explicitly specify.
124 >>>>> Perhaps. I do like the simplicity of [[KathrynAndersen]]'s syntax
125 >>>>> from [[plugins/contrib/report]] (which I copied verbatim, except for
126 >>>>> turning sort-by-`field` into a parameterized spec).
128 >>>>> If we're getting into English-like (or at least SQL-like) queries,
129 >>>>> it might make sense to change the signature of the hook function
130 >>>>> so it's a function to return a key, e.g.
131 >>>>> `sub key_age { return -%pagemtime{$_[0]) }`. Then we could sort like
134 >>>>> field(artistsort) or field(artist) or constant(Various Artists) then meta(titlesort) or meta(title) or title
136 >>>>> with "or" binding more closely than "then". Does this seem valuable?
137 >>>>> I think the implementation would be somewhat more difficult. and
138 >>>>> it's probably getting too complicated to be worthwhile, though?
139 >>>>> (The keys that actually benefit from this could just
140 >>>>> have smarter cmp functions, I think.)
142 >>>>> If the hooks return keys rather than cmp results, then we could even
143 >>>>> have "lowercase" as an adjective used like "ascending"... maybe.
144 >>>>> However, there are two types of adjective here: "lowercase"
145 >>>>> really applies to the keys, whereas "ascending" applies to the "cmp"
146 >>>>> result. Again, I think this is getting too complex, and could just
147 >>>>> be solved with smarter cmp functions.
149 >>>>>> I agree. (Also, I think returning keys may make it harder to write
150 >>>>>> smarter cmp functions.) --[[Joey]]
152 >>>>> Unfortunately, `sort="ascending mtime"` actually sorts by *descending*
153 >>>>> timestamp (but`sort=age` is fine, because `age` could be defined as
154 >>>>> now minus `ctime`). `sort=freshness` isn't right either, because
155 >>>>> "sort by freshness" seems as though it ought to mean freshest first,
156 >>>>> but "sort by ascending freshness" means put the least fresh first. If
157 >>>>> we have ascending and descending keywords which are optional, I don't
158 >>>>> think we really want different sort types to have different default
159 >>>>> directions - it seems clearer to have `ascending` always be a no-op,
160 >>>>> and `descending` always negate.
162 >>>>>> I think you've convinced me that ascending/descending impose too
163 >>>>>> much semantics on it, so "-" is better. --[[Joey]]
165 >>>>>>> I've kept the semantics from `report` as-is, then:
166 >>>>>>> e.g. `sort="age -title"`. --s
168 >>>>> Perhaps we could borrow from `meta updated` and use `update_age`?
169 >>>>> `updateage` would perhaps be a more normal IkiWiki style - but that
170 >>>>> makes me think that updateage is a quantity analagous to tonnage or
171 >>>>> voltage, with more or less recently updated pages being said to have
172 >>>>> more or less updateage. I don't know whether that's good or bad :-)
174 >>>>> I'm sure there's a much better word, but I can't see it. Do you have
175 >>>>> a better idea? --s
177 [Regarding the `meta title=foo sort=bar` special case]
179 > I feel it sould be clearer to call that "sortas", since "sort=" is used
180 > to specify a sort method in other directives. --[[Joey]]
185 I notice the implementation does not use the magic `$a` and `$b` globals.
186 That nasty perl optimisation is still worthwhile:
188 perl -e 'use warnings; use strict; use Benchmark; sub a { $a <=> $b } sub b ($$) { $_[0] <=> $_[1] }; my @list=reverse(1..9999); timethese(10000, {a => sub {my @f=sort a @list}, b => sub {my @f=sort b @list}, c => => sub {my @f=sort { b($a,$b) } @list}})'
189 Benchmark: timing 10000 iterations of a, b, c...
190 a: 80 wallclock secs (76.74 usr + 0.05 sys = 76.79 CPU) @ 130.23/s (n=10000)
191 b: 112 wallclock secs (106.14 usr + 0.20 sys = 106.34 CPU) @ 94.04/s (n=10000)
192 c: 330 wallclock secs (320.25 usr + 0.17 sys = 320.42 CPU) @ 31.21/s (n=10000)
194 Unfortunatly, I think that c is closest to the new implementation.
197 > Unfortunately, `$a` isn't always `$main::a` - it's `$Package::a` where
198 > `Package` is the call site of the sort call. This was a showstopper when
199 > `sort` was a hook implemented in many packages, but now that it's a
200 > `SortSpec`, I may be able to fix this by putting a `sort` wrapper in the
201 > `SortSpec` namespace, so it's like this:
206 > return sort $cmp @_;
209 > which would mean that the comparison used `$IkiWiki::SortSpec::a`.
212 >> I've now done this. On a wiki with many [[plugins/contrib/album]]s
213 >> (a full rebuild takes half an hour!), I tested a refresh after
214 >> `touch tags/*.mdwn` (my tag pages contain inlines of the form
215 >> `tagged(foo)` sorted by date, so they exercise sorting).
216 >> I also tried removing sorting from `pagespec_match_list`
217 >> altogether, as an upper bound for how fast we can possibly make it.
219 >> * `master` at branch point: 63.72user 0.29system
220 >> * `master` at branch point: 63.91user 0.37system
221 >> * my branch, with `@_`: 65.28user 0.29system
222 >> * my branch, with `@_`: 65.21user 0.28system
223 >> * my branch, with `$a`: 64.09user 0.28system
224 >> * my branch, with `$a`: 63.83user 0.36system
225 >> * not sorted at all: 58.99user 0.29system
226 >> * not sorted at all: 58.92user 0.29system
230 > I do notice that `pagespec_match_list` performs the sort before the
231 > filter by pagespec. Is this a deliberate design choice, or
232 > coincidence? I can see that when `limit` is used, this could be
233 > used to only run the pagespec match function until `limit` pages
234 > have been selected, but the cost is that every page in the wiki
235 > is sorted. Or, it might be useful to do the filtering first, then
236 > sort the sub-list thus produced, then finally apply the limit? --s
238 >> Yes, it was deliberate, pagespec matching can be expensive enough that
239 >> needing to sort a lot of pages seems likely to be less work. (I don't
240 >> remember what benchmarking was done though.) --[[Joey]]
242 >>> We discussed this on IRC and Joey pointed out that this also affects
243 >>> dependency calculation, so I'm not going to get into this now... --s
245 Joey pointed out on IRC that the `titlesort` feature duplicates all the
246 meta titles. I did that in order to sort by the unescaped version, but
247 I've now changed the branch to only store that if it makes a difference.
250 ## Documentation from sort-package branch
252 ### advanced sort orders (conditionally added to [[ikiwiki/pagespec/sorting]])
254 * `title_natural` - Orders by title, but numbers in the title are treated
255 as such, ("1 2 9 10 20" instead of "1 10 2 20 9")
256 * `meta(title)` - Order according to the `\[[!meta title="foo" sortas="bar"]]`
257 or `\[[!meta title="foo"]]` [[ikiwiki/directive]], or the page name if no
258 full title was set. `meta(author)`, `meta(date)`, `meta(updated)`, etc.
261 ### Multiple sort orders (added to [[ikiwiki/pagespec/sorting]])
263 In addition, you can combine several sort orders and/or reverse the order of
264 sorting, with a string like `age -title` (which would sort by age, then by
265 title in reverse order if two pages have the same age).
267 ### meta sortas parameter (added to [[ikiwiki/directive/meta]])
271 An optional `sort` parameter will be used preferentially when
272 [[ikiwiki/pagespec/sorting]] by `meta(title)`:
274 \[[!meta title="The Beatles" sort="Beatles, The"]]
276 \[[!meta title="David Bowie" sort="Bowie, David"]]
280 An optional `sortas` parameter will be used preferentially when
281 [[ikiwiki/pagespec/sorting]] by `meta(author)`:
283 \[[!meta author="Joey Hess" sortas="Hess, Joey"]]
285 ### Sorting plugins (added to [[plugins/write]])
287 Similarly, it's possible to write plugins that add new functions as
288 [[ikiwiki/pagespec/sorting]] methods. To achieve this, add a function to
289 the IkiWiki::SortSpec package named `cmp_foo`, which will be used when sorting
290 by `foo` or `foo(...)` is requested.
292 The names of pages to be compared are in the global variables `$a` and `$b`
293 in the IkiWiki::SortSpec package. The function should return the same thing
294 as Perl's `cmp` and `<=>` operators: negative if `$a` is less than `$b`,
295 positive if `$a` is greater, or zero if they are considered equal. It may
296 also raise an error using `error`, for instance if it needs a parameter but
299 The function will also be passed one or more parameters. The first is
300 `undef` if invoked as `foo`, or the parameter `"bar"` if invoked as `foo(bar)`;
301 it may also be passed additional, named parameters.