svn commit: r204309 - in head/sys: amd64/amd64 amd64/isa conf i386/bios i386/cpufreq i386/i386 i386/isa i386/xen isa modules/bios/smbios modules/bios/vpd modules/cpufreq pc98/pc98 x86 x86/bios x86/...

[M. Warner Losh]

In message: <20100301091756.GA73361 at server.vk2pj.dyndns.org>
            Peter Jeremy <peterjeremy at acm.org> writes:
: On 2010-Feb-28 16:07:32 -0700, "M. Warner Losh" <imp at bsdimp.com> wrote:
: >In message: <4B8AF55C.6040103 at freebsd.org>
: >            Nathan Whitehorn <nwhitehorn at FreeBSD.org> writes:
: >: M. Warner Losh wrote:
: >: > In message: <20100228195041.GA68446 at server.vk2pj.dyndns.org>
: >: >             Peter Jeremy <peterjeremy at acm.org> writes:
: >: > : On 2010-Feb-26 14:57:36 +0100, Attilio Rao <attilio at freebsd.org>
: >: > wrote:
: >: > : >For the future, however, probabilly we would need to do something
: >: > like
: >: > : >pc98 already does wrt i386 (i386/include/ pc98/include/
: >: > amd64/include/
: >: > : >just have files wrappers to the generic one under x86/include/ when
: >: > : >necessary).
: >: > : : Not that something like this is also critical for getting
: >: > i386/amd64
: >: > : cross compilation to work - one of the blocking issues is that the
: >: > : correct machine/ includes are not available in cross-compilatio
: >: > mode.
: >: >
: >: > Huh?
: >: >
: >: > I cross build all the time, and the machine/ includes are always
: >: > right, with or without this patch.  Maybe you could explain what you
: >: > mean here...
: >: >   
: >: I'm guessing what was meant is that one benefit of a merged tree (like
: >: we will have with powerpc) is that you can do cc -m32 on a 64-bit host
: >: to get a 32-bit binary, which is less than trivial now.
: 
: Yep.
: 
: >True enough.  But we don't have anywhere close to a merged tree on
: >x86, and won't for the foreseeable future.  I don't really want to see
: >an include32, when all the cross building tools exist in the tree
: >already (see XDEV in Makefile.inc1)...
: 
: I agree that we don't want an include32 - all that we need is the
: machine-dependent directory (i386, amd64, ...).

No.  We need a merged tree to make it actually work, otherwise
machine/foo.h won't work right.  simply having both i386/foo.h and
amd64/foo.h is not sufficient.

: >  If you are cross building, you
: >really need a full cross building environment.  The -m32 hacks we have
: >now are, as you describe, insufficient to the task...
: 
: For building i386 executables in an amd64 world, we already have
: almost everything needed - the toolchain is present, the libraries are
: present, most of the headers are present.  The only parts of the
: environment that aren't installed in default are the machine-dependent
: headers (and some glue to include the appropriate architecture).
: 
: I'm not saying that installing the i386 headers will magically make
: '-m32' work but it's a necessary prerequisite.

We need merged support for both out of one tree, and we're nowhere
near that.  The toolchain also is only mostly present, last time I
looked there were small bits and pieces missing here and there.

Also, -m32 is a bit of a hack, since it isn't as useful as you'd like
once you start to compile ports.  Too many of them don't like it when
you pass in a cc that has args with it, or they don't religiously use
$CC, which means you get some 64-bit and some 32-bit goo.  We have the
same problem with MIPS...

Warner