Re: need for packed attribute

[Mikael Pettersson <mikpe@csd.uu.se>]

Russell King writes:
 > > As fas as I can tell, the AAPCS document (v2.03 7th Oct 2005) requires
 > > that a simple "struct foo { unsigned char c; };" should have both size
 > > and alignment equal to 1, but gcc makes them both 4. Do you have any
 > > information about why gcc is doing this on ARM/Linux? Is there an accurate
 > > ABI document for ARM/Linux somewhere?
 > 
 > That's the new EABI, which is a major change to the existing ABI which
 > the kernel and all of userspace is currently built using.
 > 
 > The old ABI has it's roots in 1993 when the kernel and userland was
 > initially built using an ANSI C compiler, and the work being done to
 > port GCC was to make it compliant with that version of the ABI.  This
 > ABI is documented only in dead-tree form.
 > 
 > Due to lack of manpower on the Linux side (iow, more or less just me)
 > this became the ABI of the early ARM Linux a.out toolchain.  At that
 > time, I did not consider this to be a problem - it wasn't a problem
 > as far as the kernel was concerned.
 > 
 > When ELF came along, other folk worked on the toolchain, but they stuck
 > with that ABI - you could not transition between the a.out ABI to the
 > ELF ABI without breaking the kernel - structure layouts would change.
 > 
 > Hence, this is the existing ABI we have.  Changing the padding or
 > alignment of structures changes the kernel ABI, making it incompatible
 > with current userland.

OK, thanks for this info. It means that GCC is the definitive authority
on calling conventions and data layouts, not the AAPCS; I wasn't aware of
that before.

(My interest in this issue comes from working on a port of a functional
programming language's JIT compiler and runtime system to XScale.)

/Mikael