Re: Arch Hurd progress

[Allan McRae]

On 13/04/10 18:20, Samuel Thibault wrote:
> Allan McRae, le Tue 13 Apr 2010 10:04:25 +1000, a écrit :
> > The major stumbling block there was gnumach.  When compiling for i686
> > it requires -O1 to be used.  -O0 results in a build error and -O2
> > results in the boot process stalling very, very early on.  There are
> > no such issues for i[45]86.
>
> Mmm, doesn't building with native i686 support and -O2 enable
> vectorization, and thus the use of mmx/sse?  This is not supported in
> kernel space for context switch performance reasons (like on Linux). Try
> with -mno-mmx -mno-sse -mno-sse2 -mno-3dnow.
>
> We probably need to enable them by default in the configure script.

Thanks for the hint.  We will give this a try and report back the results.

> > Once that was out of the way, the biggest issue we had was getting gcc
> > to work natively.
>
> What were the issues?  Just cross-compiling, or are patches needed? (in
> addition to what we have in Debian)

Getting the C++ cross-compiler working was the main issue.  That is 
needed to build gmp with C++ support which is require for mpfr.  That 
was giving us headaches so we tried adjusting from the crosshurd script 
style approach to something more akin to CLFS.  That resulted in 
non-booting systems so was a step back.

In the end I think it was part of a patch from Debian that solved the 
C++ issue (which we had overlooked due to other parts in it that we did 
not want to use).  This has since appeared on the gcc-patches list as 
pointed out on this list a couple of days ago.  Then we needed a lot of 
configure flags to gcc to pick up the cross built libraries.  Perhaps we 
made it harder for ourselves by enabling graphite support...

Allan