Re: 7+ days of dogfood

On Mon, 2013-02-11 at 09:15 -0500, Glen Barber wrote:
> On Mon, Feb 11, 2013 at 02:56:47PM +0100, Fabian Keil wrote:
> > >                                  WITHOUT_GCC=yes in src.conf is
> > > worthwhile.  WITHOUT_GDB=yes is probably also sensible, as the gdb in
> > > base is so old that it doesn't understand most of the DWARF that clang
> > > uses.  We should have lldb ready for import in a few months, but until
> > > then using gdb from ports is more sensible if you plan on actually doing
> > > any debugging.
> > 
> > So far I didn't consider not building gdb, but I agree that it's
> > not too useful when compiling with clang and am already using
> > gdb751 for debugging anyway.
> > 
> > My impression was that the base gdb compiles rather quickly
> > (compared to more recent versions) and that it thus wouldn't
> > matter, but I'll give it a try.
> > 
> > Thanks for the suggestion.
> > 
> 
> You might also want to try MALLOC_PRODUCTION=1 in make.conf.  This took
> my buildworld/buildkernel times from 35/15 minutes to 8/5 minutes,
> respectively.
> 
> Glen

Actually, anyone who is routinely setting MALLOC_PRODUCTION to run
-current, please considering trying again without it.

One particular sanity check enabled without MALLOC_PRODUCTION was
responsible for virtually all of the performance degradation, and the
recent import of a new jemalloc reworked that code to eliminate the
problem (it was needlessly validating that freshly mmap'd memory was
zeroed; now it only does that validation if it internally recycles a
block, and I think it never even does that on freebsd).

-- Ian