Re: [v3] Revised locale caches

[Jerry Quinn <jlquinn at optonline dot net>]

B. Kosnik writes:
 > Sorry about the delay.
 > 
 > You'll find enclosed something that I think is the way to go. It's
 > zero-alloc, as fast as mainline, per-facet, extensible for user-defined
 > facets, removes the known issues and problems with the current approach,
 > etc, etc etc.
 > 
 > This is somewhat like the libstdc++-v3.20030429-base-idea sources, in
 > that the facet is now made up of the cache, not the other way around.
 > Encapsulation seems like a better way to go with this.

I think this is generally fine.  A few issues, though:

1) If someone specializes numpunct<sometype>, there probably won't be
   a cache.  With the solution I had posted, we will still cache the
   values.  The only way for that to work is to have it outside the
   facet itself - or perhaps in locale::facet, where everyone is
   forced to pick it up.

   If you think that this issue isn't important enough to keep the
   caches external to the facets, then I won't argue the point
   further.

1) There should probably be an accessor to the underlying cache data,
   rather than relying on numpunct == numpunct_cache*

2) I'd rather see the __use_cache function in the tcc file as a
   template.  Except then we run afoul of #1 :-(

   Moving the __use_cache function specializations here buys a 2-3%
   speedup on the int output test.

3) I don't think we need the allocated flag.  The only time the data
   isn't allocated is for the posix locale, and the destructor will
   never run.

4) It's not safe to require the existence of ctype in setting up the
   numpunct cache.  Some uses will not require it, and shouldn't blow
   up as a result.  My posted solution suffers from a similar issue.

6) This can't be backported to 3.3 since it changes the ABI.  So, can
   we keep the version I posted alive for that?  I'm pretty confident
   I can make that work in 3.3 within the ABI.  Including a cache per
   facet.

It's too late tonight, but I'll do the minor changes above tomorrow
night.

 > Let me know what you think. I'd like to go through and do num_get
 > (painless now), as well as the other data facets (__timepunct,
 > __moneypuncts) in a similar style.

Can we get this in first and then do the other facets? :-)

Thanks,
Jerry