On 29 sept. 2012, at 19:54, m...@mikesolomon.org wrote:
>
> On 29 sept. 2012, at 19:53, "Keith OHara" <k-ohara5...@oco.net> wrote:
>
>> On Sat, 29 Sep 2012 10:30:32 -0700, m...@mikesolomon.org
>> <m...@mikesolomon.org> wrote:
>>
>>>
>>> The way you're using "tentative" is almost exactly how pure properties are
>>> used in LilyPond.
>>
>> Specifically, 'pure-height being the estimated vertical extent before
>> line-breaking, while 'height is its extent after line-breaking.
>>
>> If there are distinct properties to describe the position at different
>> stages, then each property can be evaluated just once (as HanWen suggested,
>> and as Mike agreed 100%).
More thinking. I'm not enthusiastic about stages - it is a top down approach
that locks us into certain points of evaluation. What if we decided to add or
get rid of a stage? Would we need to create things like unpure-pure-containers
for various stages? What qualifies as a stage? Could users make custom stages?
Furthermore, the idea of properties being pure in LilyPond is impossible to
police. The code almost guarantees that won't be pure, as once the dim_cache_
of a grob is filled, it gives that value instead of the result of a function
call (see Grob::pure_relative_y_coordinate, specifically line 350 of grob.cc as
of 9e605a1bb6644d89cf110f20cb6d46bb0339fca7). I like Keith's idea of
"tentative" and "permanent". The model I'm thinking of is:
--) LilyPond evaluates all callbacks as permanent unless tentative is
explicitly asked for.
--) Permanent properties are stashed in a cache that cannot be erased.
--) Tentative properties may (i.e. heights) or may not (i.e. color) be cached
depending on how we choose to optimize LilyPond. It is the responsibility of
the coder, if she wants an uncached property, to wipe the cache.
--) Most importantly, all notions of pure callbacks disappear. It is the job
of functions to police themselves. For example, a function Grob::has_full_x
can be created to determine if a grob has an X position with respect to the
system. It'd check to see if the dim_cache_[X_AXIS].offset_ of a Paper_column
pointing to a real location in memory. The answer will be false if
System::break_into_pieces has been called and true otherwise. Then, we could
do something like:
MAKE_SCHEME_CALLBACK_WITH_OPTARGS (Slur, height, 1, 2, "");
SCM
Slur::height (SCM smob, SCM begscm, SCM endscm)
{
Grob *me = unsmob_grob (smob);
if (me->has_full_x ())
return permanent_height (me);
int beg = robust_scm2int (begscm, 0);
int end = robust_scm2int (endscm, 0);
return tentative_height (me, beg, end);
}
This'd allow to entirely get rid of all the pure callback lists as well as
unpure-pure-containers.
--) Both permanent and tentative properties can result in calls to
set_property, set_object, translate_axis, and suicide. All setters should be
used sparingly and internally to cache values (like we do for minimum
translations, for example). We can police this via warnings and errors.
Cheers,
MS
_______________________________________________
lilypond-devel mailing list
lilypond-devel@gnu.org
https://lists.gnu.org/mailman/listinfo/lilypond-devel
