[Monotone-devel] Re: Re: monotone on alpha, possibly only Debian (was Re: ikiwiki monotone support)

[Pavel Cahyna]

On Mon, Jan 14, 2008 at 01:04:24PM -0500, Zack Weinberg wrote:
> On Jan 14, 2008 3:54 AM, Pavel Cahyna <address@hidden> wrote:
> > > I'm seeing mysterious testsuite failures when the entire binary is
> > > compiled with -O2, and fewer, different mysterious testsuite failures
> > > when the entire binary is compiled with -O1.  See
> >
> > I had to disable optimizations; -O1 did not work, either.
> 
> Just to be clear: you get *reproducible* testsuite failures when the
> entire binary is compiled with -O1, and none when it is compiled with
> -O0?

I haven't run any testsuite. The binary does not work when compiled with
-O1. It works without optimizations and also with
CXXFLAGS=-mcpu=21164a -g3 CFLAGS=-O2 -mcpu=21164a -g3 MTN_CXXFLAGS=-O2

which ensures that only C++ 3rd party libraries are built without
optimizations and everything else is built with -O2. (I was not able to
restrict the scope of disabling optimizations further to include only
botan.)

> If so, that is either a compiler bug indeed, or a place where gcc has
> gotten cleverer with its optimizations and broken an improper
> assumption in the source code.  Either way, the next step, if you're
> up for it, would be to find out which object file has the bad machine
> code in it (start from an -O0 build, and one at a time, swap out .o
> files for ones compiled at -O1, relink and retest; don't neglect
> lib3rdparty.a).

lib3rdparty.a is where the error must be. Unfortunately due to extreme
slowness of monotone compilation I don't think I'll have the patience to
investigate further in the near future. And unfortunately I've deleted the
object files for the compilations to free disk space.

Pavel