On Tue, Dec 30, 2014 at 04:34:42PM -0600, Dennis Ferguson wrote:
>
> On 30 Dec, 2014, at 12:52 , David Laight <da...@l8s.co.uk> wrote:
> > Is that the correct fix?
> > Unless the rdpr actually accesses memory (don't think it does) then
> > then problem is probably a missing 'volatile' instead.
> >
> > Certainly the way those asm functions are defined looks to be
> > rather more obfuscated than necessary.
>
> Or maybe just get rid of the __constfunc? I think that, plus
> the (void) argument list to the inline, is explicitly telling the
> compiler it should feel free to move the call absolutely anywhere
> it feels like placing it.
The relevant optimisations happen after the inline, so the function
definition shouldn't affect things.
I'm guessing that __constfunc just means that the compiler can merge
multiple calls (and cache the result), not at all sure that is
valid with a "memory" clobber.
> I'm a bit surprised the "memory" clobber by itself changed anything
> at all since I thought a "memory" clobber without a 'volatile' is
> supposed to refer only to memory-based input and output arguments
> to the asm(), of which there are none.
The compiler doesn't know which args to the asm reference memory
unless you tell it with an appropriate 'clobber' - either the
global one, or a ranged one based on one if the arguments.
In this case I suspect removing __constfunc and making the
asm volatile will force correct sequencing.
David
--
David Laight: da...@l8s.co.uk
