Re: query on build and test times

["Kaveh R. Ghazi" <ghazi at caip dot rutgers dot edu>]

 > Mainline build and test times have quadrupled in the last two months. 
 > 
 > I used to be able to do a make-check-target-libstdc++-v3 and have it
 > complete in about 14 minutes. Then, PCH broke, and it went up to 31
 > minutes. Now, it's close to an hour.
 > 
 > Anybody know what's up? Known issue? This is a 1.8G P4, which is not a
 > slouch when it comes to compiling, or hasn't been in the past. Sorry
 > if I'm missing something obvious.
 > 
 > -benjamin

I think at least part of the slowdown (but certainly not all of it) is
due to the fact that we now disable the physmem opts when checking is
enabled.

The reason for this change is that the physmem opts cause GCC to
collect much less often on large memory systems.  This can mask
collection errors and deceive you into thinking your patch is "safe".
Recent bootstrap failures occured in part because the errors in
question only occured on small memory systems where collecting
happened more often.  Since lack of collecting masked the problem it
allowed the failures to get into circulation and caused a lot of
difficulty debugging it.  See this thread:
http://gcc.gnu.org/ml/gcc-patches/2003-09/msg00147.html

I also believe that the physmem opts mask incremental performance
degredation on small memory systems because developers do timing tests
on their large memory systems and mistakenly think their patch didn't
cause any slowdown in compile-times.  By using the small memory model
on mainline we will notice the impact.  I suspect we've already
suffered some slowdown on memory constrained systems because of this.

The unfortunate side-effect is that bootstraps with mainline sources
take longer until we branch and turn of checking.  Perhaps that's
incentive for some to speed up GCC even more, or at least get PCH
working again...

		--Kaveh
--
Kaveh R. Ghazi			ghazi@caip.rutgers.edu