[PATCH] USB: ehci: use packed,aligned(4) instead of removing the packed attribute

[Alexander Holler]

Am 16.06.2011 19:55, schrieb Arnd Bergmann:
> On Thursday 16 June 2011, Alan Stern wrote:
>>
>> On Thu, 16 Jun 2011, Alexander Holler wrote:
>>
>>> In commit 139540170d9d9b7ead3caaf540f161756b356d56 the attribute
>>> packed is removed from the structs which are used to access the EHCI-registers.
>>>
>>> This is done to circumvent a problem with gcc 4.6, which might access members of
>>> packed structs on a byte by byte basis. But using packed, aligned(4) fixes that
>>> too and is imho the better solution. Otherwise (without packed) the compiler would be free
>>> to choose whatever alignment he thinks fits best, which might be e.g. 8-byte on 64-bit machines.
>>
>> Is that really true?
>
> No.
>
>> I thought the compiler was not allowed to insert
>> padding if the natural alignment of the data types didn't require any.
>
> It's architecture dependent. The alignment of the structure is the maximum alignment
> of its members, so it gets to be 8 bytes if there is a 64 bit member in the struct
> on most architectures, but 4 bytes on x86.

Hmm, sorry, but that sentence just says something about the alignment of 
the structure itself and nothing about the alignment of it's members or 
do I understand something wrong?

I've had a look at c99 again, and in addition to the two points in c99 I 
mentioned in the mail before (6.7.2.1 13 and 6.7.2.1. 15), I've only 
found the following on that topic:

6.7.2.1 12  Each non-bit-field member of a structure or union object is 
aligned in an implementationdefined manner appropriate to its type.

And, under "J.1 Unspecified behaviour":

Many aspects of the representations of types (6.2.6).

I even haven't found anything which says something about the alignment 
of a structure itself. But I'm no compiler expert and I look only seldom 
at c99 and usually try to avoid such aspects as the one we are talking 
about. ;)

For me that means that I understand that when packed(,aligned(4)) is 
used, it's pretty sure, that there is no padding inbetween the members 
of e.g. struct ehci_regs. But without I'm unsure, so I would avoid that.

That aligned(4) is necessary (for ARM) is only a workaround because of 
the implementation of readl(), at least that is how I understood the 
discussion. But that is discussed elsewhere and don't want to take part 
in that discussion (and can't).

Regards,

Alexander