[U-Boot] [PATCH] disk:efi: avoid unaligned access on efi partition

[Måns Rullgård]

Albert ARIBAUD <albert.u.boot at aribaud.net> writes:

> Hi Måns,
>
> On Mon, 14 Oct 2013 14:05:13 +0100, Måns Rullgård <mans at mansr.com>
> wrote:
>
>> Albert ARIBAUD <albert.u.boot at aribaud.net> writes:
>> 
>> >> > Please do not advise using native unaligned accesses on code that is
>> >> > not strictly used by ARMv6+ architectures: the present code, for
>> >> > instance, might be run on pre-ARMv6 or non-ARM platforms, and thus,
>> >> > should never assume ability to perform unaligned accesses natively.
>> >> 
>> >> I'm advising no such thing.  I said two things:
>> >> 
>> >> 1.  Declaring a struct with the 'packed' attribute makes gcc
>> >>     automatically generate correct code for all targets.  _IF_ the
>> >>     selected target supports unaligned ldr/str, these might get used.
>> >> 
>> >> 2.  If your target is ARMv6 or later _AND_ you enable strict alignment
>> >>     checking in the system control register, you _MUST_ build with the
>> >>     -mno-unaligned-access flag.
>> >
>> > Then I apologize; I had read "Note that on ARMv6 and later ldr/str
>> > support unaligned addresses unless this is explicitly disabled in
>> > the system control register" as a suggestion to use that capability.
>> 
>> If building for ARMv6 or later, I do suggest allowing unaligned
>> accesses.  The moment you add -march=armv6 (or equivalent), you allow
>> for a number of things not supported by older versions, so why not
>> unaligned memory accesses?
>
> doc/README.arm-unaligned-accesses probably has the answer to your
> question, especially from line 21 onward. If not, I'll be happy to
> provide further clarification.

That is about as backwards as it can get.  By adding -munaligned-access
you are telling gcc that unaligned accesses are supported and welcome.
With this setting enabled, gcc can and will generate unaligned accesses
just about anywhere.  This setting is NOT compatible with the SCTRL.A
bit being set (strict hardware alignment checking).

You cannot enable strict alignment checking in hardware, tell the
compiler unaligned accesses are OK, and then expect not to have
problems.  This is no more a valid combination than allowing
floating-point instructions when the target has no FPU.

-- 
Måns Rullgård
mans at mansr.com