Date: Wed, 17 Nov 2010 17:52:40 -0600 From: Nathan Whitehorn <nwhitehorn@freebsd.org> To: Warner Losh <imp@bsdimp.com> Cc: Tijl Coosemans <tijl@coosemans.org>, Dimitry Andric <dim@freebsd.org>, Garrett Cooper <gcooper@freebsd.org>, freebsd-arch@freebsd.org, Warner Losh <imp@freebsd.org> Subject: Re: Support for cc -m32 Message-ID: <C96B2549-4BAE-4800-925D-57EA7DFD7733@freebsd.org> In-Reply-To: <4CE46602.9000303@bsdimp.com> References: <201007291718.12687.tijl@coosemans.org> <4CE417B3.3030102@bsdimp.com> <201011172058.05683.tijl@coosemans.org> <201011171718.37798.jhb@freebsd.org> <4CE46602.9000303@bsdimp.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Nov 17, 2010, at 5:32 PM, Warner Losh wrote:
> On 11/17/2010 15:18, John Baldwin wrote:
>> On Wednesday, November 17, 2010 2:57:51 pm Tijl Coosemans wrote:
>>> cc-m32-3.diff:
>>> Modify amd64 headers to include i386 headers when compiling 32
>>> bit code.
>>>
>>> All amd64 headers follow the following format:
>>>
>>> #ifndef _AMD64_HEADER_H_
>>> #define _AMD64_HEADER_H_
>>>
>>> #ifdef __i386__
>>> #include<i386/header.h>
>>> #else
>>>
>>> /* Amd64 declarations go here. */
>>>
>>> #endif /* __i386__ */
>>> #endif /* !_AMD64_HEADER_H_ */
>> I find this to be really ugly, and error prone (since it is a
>> manual process).
>> I'd prefer something that autogenerated headers in /usr/include/
>> machine that
>> #include the appropriate version similar to what Warner suggested.
>>
>> However, one issue with that approach (and this one) are headers
>> that only
>> exist for one platform. The end result would be that that header
>> would now
>> exist for both platforms (in that if you do 'if [ -r
>> /usr/include/machine/foo.h ]' it will be true). We can make it
>> #error or
>> otherwise fail (by including a non-existing file for example), but
>> if there
>> was some way to have cc -m32 "magically" substitute "i386/" for
>> "machine",
>> that is what I would most prefer. (This has problems too in that
>> #include
>> <machine/foo.h> would work with -m32 even though /usr/include/
>> machine/foo.h
>> doesn't exist, but /usr/include/i386/foo.h does.
> "magically" converting machine -> i386 requires cpp hacking.
>
> However, the if [] test is beyond the scope of the API that we
> support. Scripts that use -m32 will have to cope with other issues.
>
> We could 'solve' this by having an /usr/include32, but even that
> still isn't complete.
>
> I contend that the least bad solution is to auto generate the
> machine directory from the sys/{i386,amd64}/include. If we do that,
> we could implement -m64 on i386 too, but that needs a lot more
> infrastructure.
The other way of solving this, which continues to work very well on
powerpc64, is to have the machine/ stuff be identical for the two
platforms (which, as far as I can tell, really are the same platform,
but with a different ABI) and to use appropriate #ifdefs to select the
right things. I would imagine, based on the continued exodus of these
headers to x86/ anyway, that the differences are not enormously large.
They certainly were not for PPC.
This could be done either with piece-by-piece modifications of the
header files, as was done for PPC, or (perhaps automatically) install
some ugly stub headers that look like
#ifdef __LP64__
#include <amd64/stuff.h>
#else
#include <i386/stuff.h>
#endif
-Nathan
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?C96B2549-4BAE-4800-925D-57EA7DFD7733>