Re: PING: vectorization, cft for toplevel bootstrap, PR/19210

[Daniel Jacobowitz <drow at false dot org>]

On Fri, May 13, 2005 at 04:33:07PM +0200, Paolo Bonzini wrote:
> 
> >>http://gcc.gnu.org/ml/gcc-patches/2005-05/msg00280.html
> >>CFT: hard-link tools in a combined tree
> >>   
> >>
> >Could you be more specific about what testing are you looking for?  I
> >don't see how cross builds would particularly stress this.  I also
> >don't see any way that this patch could possibly break.
> >
> Me too, but Richard Earnshaw reported that this broke arm-elf while 
> building libstdc++, and at the time I could reproduce it.  But now I cannot.
> 
> Just a combined tree build on any embedded target will do; the patch was 
> even already approved, but then reverted because of the breakage.

Conveniently I'm starting one right now.  However I can already see a
good reason for it to break.  Nothing else in configure-gcc assumes
that the in-tree assembler has already been built; for instance, we get
its version number out of the Makefile.  And the top level dependencies
do not enforce that gas be built before gcc is configured.  So
sometimes the $LN just fails.

It's a little touchy to reproduce (sometimes gcc/ spontaneously
reconfigures, thus hiding the problem).  But I did successfully
reproduce it by running "make configure-gcc" (which configured
everything, but did not build gas/ld/binutils), and then "make
all-gcc", which bombed using the wrong assembler.

-- 
Daniel Jacobowitz
CodeSourcery, LLC