Re: [PATCH] fix dir2 shortform structures on ARM old ABI

[Timothy Shimmin <tes@sgi.com>]

Josef 'Jeff' Sipek wrote:
> On Tue, Mar 18, 2008 at 10:35:32AM +1100, Timothy Shimmin wrote:
>> Eric Sandeen wrote:
>>> Josef 'Jeff' Sipek wrote:
>>>> Josef 'Jeff' Sipek, wondering exactly how passionate one can get about
>>>> structure member alignment :)
>>> Very.  ;)
>>> Tossing packed at all the ondisk stuctures bloats things badly on ia64.
>>> cvs/linux-2.6-xfs> wc -l before.dis
>>> 166688 before.dis
>>> cvs/linux-2.6-xfs> wc -l after.dis
>>> 182294 after.dis
>>> That's +15606 lines.
>>> http://digitalvampire.org/blog/index.php/2006/07/31/why-you-shouldnt-use-__attribute__packed/
>> Interesting.
>> So the problem there is that gcc is doing the wrong thing
>> on some arches (the example being ia64, sparc64).
> 
> Actually, it's not doing the wrong thing...
> 
> __attribute__((packed)) means:
> 
> 1) condense the members of the struct leaving NO padding bytes
> 
> 2) do NOT assume the entire structure is aligned on any boundary
>
Okay I only knew about (1) - cause that sounds more like "pack"ing to me.
So you can't assume alignment for the start of the variable
without aligned() if you use packed - Ok.

Thanks,
--Tim

> This means, that even if you have a member that'd be nicely aligned without
> the packed attribute (see below), the compiler will generate worst case
> alignment code.
> 
> struct foo {
> 	u64 a;
> } __attribute__((packed));
> 
> You can put struct foo anywhere in memory, and the code accessing ->a will
> _always_ work.
> 
> Using __attribute((packed,aligned(4))), tells it that the structure as a
> whole will be aligned on a 4-byte boundary, but there should be no padding
> bytes inserted.
> 
> Josef 'Jeff' Sipek.
>