[PATCH] libstdc++/8761 and libstdc++/7076

[Jerry Quinn]

Nathan Myers <ncm-nospam at cantrip dot org> wrote:
> On Fri, Jan 10, 2003 at 01:50:20AM -0500, Jerry Quinn wrote:
> > This patch provides specializations for integer formatting within libstdc++.
> > It nearly halves the time for the PR 8761 testcase from 21 seconds to 11
> > seconds.  The PR 7076 testcase reduces from 7 seconds to 5.6 with this patch.
 
> This is most excellent.
> 
> It seems to me we could get more improvement by eliminating the
> format strings entirely, and working directly off the iosbase
> members.  We could do even better by caching decisions when 
> format flags are set, rather than looking them all up each time
> operator<< is called.  Minimally, this might mean calling a
> streamlined decimal-integer formatter when we know none of the
> complications (hex etc.) is selected.

The next patch I'm going to do will use the _M_literals cached string
to clobber the format string for integer output and also avoid a call
to ctype.widen().  This may shave off another 1 second or so.
 
> We should get a lot of improvement by eliminating memory allocations
> from the main code path, and instead using local buffers, or a 
> persistent buffer in the iosbase object.  (I seem to recall there 
> were some malloc calls there, but haven't rechecked.)


Benjamin Kosnik <bkoz at redhat dot com> wrote:

> Yeah. I'm more in favor of a __locale_cache, then the formatting bits
> (which dupicates parts of __num_base.) 
 
Saving the formatting bits amortizes the cost of a virtual function
call such as ctype.do_widen().  Every bit helps.

Do any of you think it's worth trying to do the same exercise with
floating point output?  It's going to be significantly more complex, I
suspect, between all the different floating point formats that are
supported.

Jerry Quinn