On 03/31/2011 04:14 PM, Russell King wrote:

On Thu, Mar 31, 2011 at 04:09:32PM +0200, Jiri Slaby wrote:

quoted

On 03/31/2011 03:51 PM, Russell King wrote:

quoted

On Thu, Mar 31, 2011 at 03:48:20PM +0200, Jiri Slaby wrote:

quoted

How it can? Packed only tells the compiler to have one byte alignment
steps. And it should not matter here as all the members are 32-bit long.

It also tells the compiler that it may be misaligned, so to avoid
alignment faults it will use byte loads/stores.

No, it should not (if offsetof % sizeof == 0). Otherwise it's a bug in
the compiler.

Am I missing something?

$ cat t1.c
struct foo {
        unsigned long bar;
} __attribute__((packed));

unsigned long baz(struct foo *f)
{
        return f->bar;
}
$ arm-linux-gcc -O2 -S -o - t1.c
...
        .global baz
        .type   baz, %function
baz:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        ldrb    r3, [r0, #0]    @ zero_extendqisi2
        ldrb    r2, [r0, #1]    @ zero_extendqisi2
        ldrb    r1, [r0, #2]    @ zero_extendqisi2
        orr     r3, r3, r2, asl #8
        ldrb    r0, [r0, #3]    @ zero_extendqisi2
        orr     r3, r3, r1, asl #16
        orr     r0, r3, r0, asl #24
        mov     pc, lr
        .size   baz, .-baz
        .ident  "GCC: (GNU) 4.3.5"
        .section        .note.GNU-stack,"",%progbits

It does this because with the packed attribute, it can't make assumptions
about the alignment of 'f'.

I don't see why it couldn't like without the packed attribute. (The
manual says packed attribute for a struct is equivalent to specifying
packed to each member of that struct. It doesn't say anything about the
structure alignment itself.)

Ok, not that I'm fully convinced, but it doesn't matter.

So let's NOT mark it packed.

Anyway the info you sent was useful for me.

thanks,
-- 
js
suse labs