On Fri, Feb 08, 2013 at 09:06:52AM -0800, H. Peter Anvin wrote:
> On 02/08/2013 07:28 AM, Russell King - ARM Linux wrote:
> > 
> > Whether that's safe for x86 or not, I don't know, but my suspicions are
> > that it's unsafe on x86 as it's possible to refer to the various bytes/
> > half-words of eax separately.
> > 
> > So, I came to the conclusion that if x86 remains a problem, there's
> > little point supporting it on ARM.
> > 
> 
> It is possible to access bytes separately, but gcc generally doesn't.
> However, whether or not that can be relied upon safely is a tricky question.
> 
> It *is* also worth nothing that the x86 ABI does allow two words to be
> returned in registers from a normal C function, simply by returning a
> structure.  That doesn't solve the problem at hand, though, which is how
> to make a type-neutral macro which can handle doublewords...

I just tried to compare a couple of options:

1)
	unsigned long __val_gu;
	unsigned long long __val_gu8;

2)
	register typeof(x) __val_gu asm("eax");

	I was still using a test app based on my orignal patch which
	used eax and edx for the value.

3)
	typeof(x) __val_gu;

While options 2 and 3 get rid of the warnings, they unfortunately
produce different results as well.

I've attached my test app in case you want to see it.

One other note is that if I include the memtest() call in the test,
option 1 and 2 produce the same results, but w/o the memtest() the
results differ. I didn't look into it any further.

-- 
Ville Syrjälä
Intel OTC