Re: ctype optimization work

[Jerry Quinn <jlquinn at optonline dot net>]

Paolo Carlini writes:
 > Jerry Quinn wrote:
 > 
 > >Any thoughts?
 > >  
 > >
 > Well, mine is *really* down-to-earth: as my preliminary tests show
 > 
 >     http://gcc.gnu.org/ml/libstdc++/2003-11/msg00351.html
 > 
 > dealing with the wchar_t case first with a trivial table prepared at 
 > construction time, doesn't remove the virtual call overhead and doesn't 
 > provide all the advantages of a dynamic cache, but for sure promise to 
 > speed up noticeably many applications, in particular when the 
 > time_get/time_put facets are involved:
 > 
 >     http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13341
 > 
 > I mean to just implement this baseline thing, against which we can judge 
 > more sophisticated, but tricky alternatives.

I just looked closer at this.  One problem is that you can't build the
table in the ctype<wchar_t> constructor.  There's no way to delay the
table construction until a derived class constructor finishes
executing.  If you construct the table too soon, that may break a
derived do_narrow function.

This means filling the cache on demand, within the do_narrow
function.  However, it seems the cost of wctob is enough to justify
even this.

Also, in that code, you would have to also test for __wc >= 0, since
wchar_t is probably signed.

I thought it would be a cool trick to initialize the table in
use_facet, but that doesn't cover using the facet outside of the
locale :-(

Jerry