Greg Schafer wrote:
Hmmm, thinking about this some more, there might actually be another
option and that's the one already raised by Alex ie: don't bootstrap GCC
pass1 (possibly move the bootstrap to pass2 as suggested by Jeremy). As it
currently stands, the new tools start being used in combo with the host
glibc during stage2 of the bootstrap. We could avoid this by dropping the
bootstrap of pass1. That leaves the only exposed areas as kernel headers
installation and glibc configure, but if they cause problems, we might be
able to get away some PATH twiddling. There might be some merit here..
I'll chew on it for a while and maybe try some tests.
Alright, I've done some testing on this. (Greg, have you been able to
look at anything related on your end?) Let me just say first of all,
that the more I think about it, the saner it seems to save bootstrapping
GCC until pass 2. Bootstrapping in pass 1 only tells us that GCC can
compile itself fully and correctly using our new ld (in stages 2 and 3)
and the host's libc. That's well and good, but we don't necessarily need
that. What we _do_ need, is for a minimal GCC to be able to flawlessly
build a new Glibc. On broken(?) hosts, that is a subtle but important
difference.
The host that Alexander suggested, multilib Debian Lenny for x86_64 with
32-bit userspace makes a good testcase. On this system we can't
bootstrap GCC right away, even if using HJL Binutils. But, if we forgo
bootstrapping and disable building of libmudflap, we can build what
appears to be a solid Glibc. Note that I still need to set up the
testcase that Greg has mentioned on the diy-linux-dev list so that I can
run a full testsuite on the temporary Glibc. However, I was able to
bootstrap GCC pass 2 on the aforementioned broken host. Here's what i did:
* Install Debian testing (lenny) for amd64
* apt-get install debootstrap
* debootstrap --arch i386 --variant=buildd --exclude=mawk \
--include=gawk,gcc-4.1-multilib,m4,bison,flex,bzip2,texinfo \
lenny /mnt/testbed http://debian.osuosl.org/debian
* chroot /mnt/testbed /bin/bash --login
* mount -t proc proc /proc
* mount -t sysfs sysfs /sys
* Build up to Chapter 5 ncurses from the instructions here:
http://www.linuxfromscratch.org/~jhuntwork/lfs-JH/
(This currently has changes that have not been committed to
the branch.)
Logs of my build to chapter 5 ncurses are here:
http://www.linuxfromscratch.org/~jhuntwork/build-logs-20070820/
(The gcc-pass1-with-libmudflap.log used the same configure commands
but without '--disable-libmudflap'.)
The diff of my working copy for the relevant changes:
http://www.linuxfromscratch.org/~jhuntwork/lfs-jh-20070820.diff
As usual, any insights or suggestions you may have are welcome.
--
JH
--
http://linuxfromscratch.org/mailman/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page
