Build 32 bit binaries on amd64

[Warner Losh]

On Aug 21, 2012, at 10:34 AM, Nathan Whitehorn wrote:

> On 08/21/12 08:44, John Baldwin wrote:
>> On Tuesday, August 21, 2012 4:49:30 am Konstantin Belousov wrote:
>>> On Mon, Aug 20, 2012 at 08:32:41PM -0600, Dan McGregor wrote:
>>>> Hi.
>>>> 
>>>> I've been working on porting compiler-rt/clang's support for address
>>>> sanitization (asan) to FreeBSD.  So far I have it building and it
>>>> appears to work properly, however the build system expects to be able
>>>> to build 32 bit binaries on amd64.
>>>> 
>>>> amd64 doesn't include i386's machine/foo headers.  The included patch
>>>> is my proposed solution:
>>>> 
>>>> Add i386 headers to /usr/include/i386, and in machine/foo.h, check if
>>>> it's a 32 bit build and include the appropriate header from i386.
>>>> 
>>>> For example machine/ucontext.h will include i386/ucontext.h if
>>>> compiled with -m32.
>>>> 
>>>> Thoughts?
>>>> 
>>>> If anyone's curious about the compiler_rt port, I have it at
>>>> github.com/dannomac/compiler-rt on the branch named freebsd.
>>> There was a work by Tijl Coosemans in the similar, but somewhat less hacky
>>> direction. The headers are moved into sys/x86/include and unified as much
>>> as possible, while machine/ counterpart includes corresponding header
>>> from x86/include.
>>> 
>>> I even lost track of how much more headers is left to convert. In fact,
>>> not all headers are equal, some are only useful for kernel or base system.
>>> Also, parts of the critically important headers do not live in machine/
>>> at all, e.g. the headers from libm.
>>> 
>>> The work seems to be stale, do you want to cooperate with Tijl or continue ?
>> I think we should probably follow Tijl's model since we are on that path.
>> I originally preferred the /usr/include/i386 approach, but have come around
>> to Tjil's approach instead.
>> 
> 
> I just wanted to add that the unified 32/64 header route is where we went on PowerPC (and MIPS?) and it works very well for -m32.

I don't know if mips has full -m32 support yet, but we do have one set of headers there and that works for all mips modes.  We also have one set of headers for ARM too, but that's all ISA related, not 32 vs 64 related... yet.

The original i386 headers were copied to amd64 so they could be cleansed of all the historical baggage from i386.  This turned out to be a good approach in many parts of the code, but not such a good idea for the headers.  I'm glad to see we're finally cleaning this up.

Warner