[cacao] RFC: CACAO PR157 ARM memory barrier patch

[Andrew Haley]

On 03/11/2011 04:27 PM, Robert Lougher wrote:
> On 11 March 2011 15:55, Andrew Haley <aph at redhat.com> wrote:
>> On 03/11/2011 03:48 PM, Robert Lougher wrote:
>>>
>>> This is how I define MBARRIER for ARM in JamVM:
>>>
>>> #ifdef __ARM_ARCH_7A__
>>> #define MBARRIER() __asm__ __volatile__ ("dmb" ::: "memory")
>>> #else
>>> #define MBARRIER() __asm__ __volatile__ ("" ::: "memory")
>>> #endif
>>
>> But that's wrong for GNU/Linux binaries, surely.
> 
> Ubuntu defaults to ARMv7 and Thumb2, so this is OK for Ubuntu.  But
> yes, I believe Debian builds binaries for ARMv4t.  I can understand
> this in the past, as older binaries would run with no problem on later
> chips.  But ARMv7 needs proper memory barriers, so I don't think the
> Debian policy makes any sense here.  Using the kernel helper like this
> just seems nasty to me (defining a function pointer via __kernel_dmb
> (*(__kernel_dmb_t *) 0xffff0fa0)).

In what way is it nasty?  You get the correct behaviour with very little
overhead.

> My opinion is that if you build for ARMv4t, don't expect it to run on ARMv7.

I don't get it.  Why not just do it right for all cases?

Andrew.