Re: [PATCH] Fix CFLAGS overwrite by Makefile

From: Arnaud Lacombe <lacombar_at_gmail.com>
Date: Wed, 25 May 2011 16:10:17 -0400
Hi,

On Wed, May 25, 2011 at 3:44 PM, John Baldwin <jhb_at_freebsd.org> wrote:
> On Wednesday, May 25, 2011 1:03:10 pm Arnaud Lacombe wrote:
>> Hi,
>>
>> On Wed, May 25, 2011 at 12:28 PM, John Baldwin <jhb_at_freebsd.org> wrote:
>> > On Wednesday, May 25, 2011 11:34:29 am Arnaud Lacombe wrote:
>> >> Hi,
>> >>
>> >> On Wed, May 25, 2011 at 9:43 AM, John Baldwin <jhb_at_freebsd.org> wrote:
>> >> >> The original trouble I met, is that building for an i586 target in a
>> >> >> 32bits jail, on top of an amd64 system[0] (I do not have control over
>> >> >> that setup) produces incorrect binaries. The current fix I've got is
>> >> >> to define MACHINE_ARCH=i386 and CPUTYPE=i586. This enforces
>> >> >> `-march=i586' to be passed to the compiler, for all except the
>> >> >> bootloader (because it overwrites CFLAGS). With this, binaries
>> >> >> produced works fine (ie. /bin/sh no longer SIGILL when bringing up the
>> >> >> system). So I suspect that gcc default to i686 in this setup and
>> >> >> corrupt all the binaries, thus the attached patch.
>> >> >
>> >> > Wait.  You must have something wrong in your jail if you can't do a
>> > buildworld
>> >> > with CPUTYPE set to none and have it do the right thing.  You need to
> find
>> >> > your root problem.  Forcing CPUCFLAGS for the boot code is a band-aid,
>> > it's
>> >> > not the right solution to your problem.
>> >> >
>> >> Unless error of my part, I never mentioned it was using `buildworld',
>> >> which it is not. The system uses bare calls to make(1) in the
>> >> sys/boot/ directory. As the jail is 32bits, it was expected not to be
>> >> an issue, but the jail compiler uses /lib/libstand.a to link the
>> >> loader, and it obviously contains i686-only instructions, which
>> >> trigger a reset of an i586-only CPU.
>> >>
>> >> The more broad issue with the setup is that gcc within that
>> >> environment, without being told -march=i586, produces i686
>> >> instructions which are incompatible with the target CPU.
>> >
>> > Huh?  GCC does not generate i686 instructions by default on FreeBSD/i386.
>  It
>> > generates i486 instructions but that is all.
>> something is odd somewhere.
>>
>> > Are you sure you aren't running
>> > the 64-bit gcc (which will generate i686 instructions by default)?
>> >
>> yes.
>>
>> # which gcc
>> /usr/bin/gcc
>>
>> # file /usr/bin/gcc
>> /usr/bin/gcc: ELF 32-bit LSB executable, Intel 80386, version 1
>> (FreeBSD), for FreeBSD 7.1, statically linked, FreeBSD-style, stripped
>>
>> # gcc -v
>> Using built-in specs.
>> Target: i386-undermydesk-freebsd
>> Configured with: FreeBSD/i386 system compiler
>> Thread model: posix
>> gcc version 4.2.1 20070719  [FreeBSD]
>>
>> # uname -a
>> FreeBSD build 7.2-RELEASE FreeBSD 7.2-RELEASE #0: Fri May  1 07:18:07
>> UTC 2009     root_at_driscoll.cse.buffalo.edu:/usr/obj/usr/src/sys/GENERIC
>>  amd64
>   ^^^^^
>
> I think this is probably going to confuse make and some other things as well.
>
This is what I meant when I said "canadian setup". HOST=amd64,
BUIILD=i386 and TARGET=i586.

I'm now trying to track down the original instruction triggering the
SIGILL, but it is in a library and that section of the memory does not
seem to be included in the core. Moreover I do not think I have any
way on a broken system to get the address at which libraries get
loaded (understand that ldd(1) is dynamically linked, and as the libc
the likely culprit, rendering ldd(1) useless).

 - Arnaud
Received on Wed May 25 2011 - 18:10:18 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:14 UTC