Re: More ARM binutils fuckage

[Nicolas Pitre <nico at cam dot org>]

On Wed, 6 Dec 2006, Daniel Jacobowitz wrote:

> I don't think MV's kernel is any different in this area, but I'm
> not authoritative.

I am.

> > You pass your old-ABI compiler the option -mabi=aapcs-linux, which works
> > fine with my gcc 4.1 old-ABI toolchain and is exactly what mainline 2.6
> > does.
> 
> I don't recommend doing this.  The two compilers (...gnu-gcc
> -mabi=aapcs-linux and ...gnueabi-gcc) do not have exactly the same
> configuration; I don't know for sure what might be different between
> them, but I do know we only expect EABI compliance from the EABI
> compilers.  -mabi=aapcs-linux versus -mabi=aapcs was mainly for
> interoperation between arm-none-eabi-gcc and
> arm-none-linux-gnueabi-gcc.

In fact, the kernel doesn't care much about EABI full compliance.  The 
only requirement on the kernel to be able to run EABI userspace is to 
have the same structure member alignment and function argument layout as 
the "true" EABI compliance so the kernel and user space understand each 
other across syscall boundaries.  And using -mabi=aapcs worked just fine 
for that so far regardless of the gcc config.

If that behavior was not to hold true then gcc should not accept the 
conflicting -mabi argument.  But that would be unfortunate if it 
happened.


Nicolas

--
For unsubscribe information see http://sourceware.org/lists.html#faq