tags and all that...

[Sven Wischnowsky <wischnow@xxxxxxxxxxxxxxxxxxxxxxx>]

(So, was heisst denn `Grundsatzdiskussion' auf Englisch? ;-)


I've been thinking a lot about this file-patterns stuff we were
discussing. The biggest problem I had with the things suggested was
that there suddenly were two loops or loop-like whatevers: the tags
and the patterns. If we would make _files use only one tag, say
`files' functions like `_next_tags' would be in trouble. So I thought
about several ways how _files could be re-organised. For example we
could make the file-patterns style be inteded to be set globally,
containing tags and patterns for them. _files would then offer all
these tags or, if the style is not set, the tags it uses now with the
meaning they have now. That would use tag-order and friends as usual.

But somehow that didn't feel right. If file-patterns is needed only in 
certain contexts, one would 1) set it for those contexts and 2) set
tag-order for the same contexts, effectively repeating the tag names
and if set for a sufficiently specific context, file-patterns could
already give the ordering, as suggested by Bart. So I was back to the
beginning. Normally I try to generalise the problem in such cases (I
at least I wished I would do that always). Being able to specify ones
own patterns for files is just a special case. Similar things are
already possible with all other types of matches, for example with the 
ignored-patterns style. That made me think about tags some more.

They are kinda schizophrenic currently, aren't they? (And I'm only
talking about tags used when adding matches here...)

I mean, from a completion function's point of view, tags name `ways to
generate matches'. Nothing more. From a user's point of view, they
name a set of matches (trying to avoid the word `group' here, which is 
already used). So, if I may call the `ways to ...' `methods', then we
have the connection that completion functions know about one or
methods to generate matches and, currently, each method can generate
one set of matches. Several styles allow us to say, for example, how
the matches in the sets should be displayed (list-packed, list-colors, 
etc), which matches should not be in them (ignored-patterns) and so
on. Coming back to the _files example, we here have the case that we
have a function knowing about only one method to create matches, but
that can generate different sets of matches. And suddenly the tags
named only the sets, not the method (or the method has multiple names, 
which is equally ugly).

So I would like to first make this cleaner, by really separating
methods from sets of matches. I.e. functions would not call _tags to
say which sets they can generate, they would call, say, _methods to
say *how* they can generate matches. These methods would have
names. Users could say, which sets of matches they want, they would
have names, too. To make things a bit more confusing, let me use `tag' 
for the name of such a set. So users could say something like:

  zstyle ... tag-order foo=files bar=files

Which means that he wants to use the method `files' twice, generating
two sets of matches, named with the tags `foo' and `bar'. Then he can
use all the known styles to say, how the matches in the sets are
generated, how they are listed and so on. E.g.:

  zstyle '...:foo' file-patterns '*.ps'

Of course, we would have to add some default behaviour. And we should
try to make it look like what we have now (that's why I used those
styles in the examples). For tag-order that's relatively simple, I
think: if a value of tag-order does not contain an `=' (or whatever
token we choose):

  zstyle ... tag-order foo

it means, that `foo' is used both to name the method and the set
created. _files is probably the most complicated case: it looks at the 
options it gets (-/, -g, ...) and from that effectively builds a
default `table' mapping tags to methods and patterns (globbed-files to 
whatever was given with -g, directories to `*(-/)', all-files to `*').

But maybe we should make things more method-centered: leave tag-order
unchanged and add a new style looked up for the methods:

  zstyle '...:functions' method-tags non-comp-functions functions
  zstyle '...:non-comp-functions' ignored-patterns '_*'

So, whenever a functions offers to use the method `functions', we try
two generate two sets of matches: `non-comp-functions' and
`functions'. The tag-order style would then say in which order the
tags are tried (and we are back to having to specify tags more than
once).

On the completion function side this would of course mean that we
would have to change all places where _wanted, _tags, etc are used. In 
places where we now don't need a loop we would have to add one,
looping over the tags (sets). Where we now have a loop we would need
two nested loop, the outer one looping over the combinations of tags
and methods and the inner one looping over the methods, calling them
probably more than once, for different tags. Or the outer one looping
over the combinations and multiple inner loops looping over the
tags. That means that we can hide these tag-loops from completion
function writer by putting them into _wanted, _requested and so on
(and giveng them what is now after the `&&' as arguments which will be 
eval'ed).


Maybe I'm completely crazy now, maybe all this isn't that big a
problem to make is attempt something like this separation of methods
and tags. Two things that make me think that it may be worth it are:
Andrej (and my) question if it would be possible to have the matcher
style behave more like matcher-list. With the splitting one could
easily call a method more than once, use different tags and set the
matcher style for these tags. The second thing is my unhappiness with
all this alternate-set business. I haven't though too much about that
yet, but somehow I think there is the potential to get rid of the
alternate set altogether (leaving it only be used by compctl). That
would make me happy.


Actually, the biggest problem I see with all this is to find a good
(convenient) way to configure it. I don't like either of the ways I
showed above very much.


Opinions? Comments?


Bye
 Sven


--
Sven Wischnowsky                         wischnow@xxxxxxxxxxxxxxxxxxxxxxx