CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS and bcm2835_defconfig

[Måns Rullgård]

Stephen Warren <swarren at wwwdotorg.org> writes:

> On 03/12/2014 05:03 AM, Russell King - ARM Linux wrote:
>> On Tue, Mar 11, 2014 at 10:35:24PM -0600, Stephen Warren wrote:
>>> Hmm. I thought that ARMv6 didn't have efficient unaligned accesses, but
>>> I suppose it must; ARM selects HAVE_EFFICIENT_UNALIGNED_ACCESS for
>>> v6/v6k/v7. BUILTIN_BSWAP seems innocuous. Can you verify which of those
>>> two options actually causes problems for you?
>> 
>> What do you think is different between v6 and v7 for unaligned accesses?
>
> I don't know the details of unaligned accesses;

In ARMv6 the SCTLR.U bit selects the alignment configuration.  When this
bit is cleared, the behaviour is (mostly) compatible with ARMv5.  When
set, the new model for unaligned accesses is used.  In ARMv7 this bit is
always set (read-only).

The SCTLR.A bit selects whether unaligned accesses are permitted (clear)
or trapped (set).

Linux on ARMv6 and up sets the U bit and clears the A bit so v6 and v7
behave identically.

> I'm just remembering (probably incorrectly and/or misinterpreting)
> some discussion in U-Boot about ARMv7 being able to do unaligned
> accesses OK, yet earlier generations not being able to (although I
> don't recall which), and hence the toolchain options are set up to
> disallow them, and the ARMv7(?)  config bit that enables them is left
> disabled, so such accesses cause an exception, so it's easy to catch
> the problems that must happen on older HW even when running on newer
> HW, even though such access could be allows on ARMv7.

That discussion was full of nonsense as well as being irrelevant to the
kernel.  In the kernel, the SCTLR.A bit is cleared, enabling unaligned
accesses on ARMv6 and later.

-- 
Måns Rullgård
mans at mansr.com