Re: [RFC-PATCH 1/5] unaligned: introduce common header

["Will Newton" <will.newton@gmail.com>]

On Mon, Nov 10, 2008 at 6:51 PM, Harvey Harrison
<harvey.harrison@gmail.com> wrote:
> On Mon, 2008-11-10 at 18:35 +0000, Will Newton wrote:
>> On Mon, Nov 10, 2008 at 4:51 PM, Harvey Harrison
>> <harvey.harrison@gmail.com> wrote:
>> >
>> > In this particular case, packed isn't right as you know big_data is
>> > aligned (as long as you can guarantee the struct alignment), so you'd
>> > probably want:
>> >
>> > struct foo {
>> >        u64 big_data;
>> >        u8 small_data;
>> >        u32 medium_data __attribute__((__packed__));
>> > }
>> >
>> > But that's not what we're talking about in the kernel's case.
>>
>> Perhaps that would be a neater way of expressing what is required in
>> my simple example, but it's fairly common to use packed on the whole
>> struct which could be because a field that is "packed" by default on
>> one architecture might not be on another. You could mark every field
>> as packed but few people seem to do that and as far as I am aware
>> there is no documented difference between packing all members and the
>> whole struct. The gcc documentation for packed is pretty short:
>
> Actually it's documented that putting attribute(packed) on the struct
> is equivalent to putting attribute(packed) on _every_ member of the
> struct.
>
>>   The packed attribute specifies that a variable or structure field
>> should have the smallest
>>   possible alignment—one byte for a variable, and one bit for a field,
>> unless you specify a
>>   larger value with the aligned attribute.
>>
>> I'd love to know if the pointer alignment behaviour is widespread and
>> then maybe write a patch for the gcc manual.
>
> Well, it's kind of the whole point of __packed isn't it?  Otherwise the
> struct members get naturally (or some arch-dependent value) aligned,
> which the compiler can rely on unless you say __packed.
>
> So in my example above, the compiler _knows_ how it has aligned
> big_data and small_data and can use whatever access is most efficient,
> but it can't make any assumptions about medium_data, so access through
> a pointer _must_ be done unaligned.
>
> struct foo *bar;
> bar->medium_data; // compiler must do this unaligned

Agreed, but the packed struct unaligned access code does this:

struct __una_u16 { u16 x __attribute__((packed)); };

On an architecture with struct alignment > 1 this will not work. Which
is why I believe that the memmove code must stay for those
architectures.