>>> Samuel Thibault <samuel.thibault@xxxxxxxxxxxxx> 14.03.08 14:59 >>>
>Jan Beulich, le Fri 14 Mar 2008 13:51:26 +0000, a écrit :
>> >memory-clobber would be needed for the atomic ops, but a dummy memory
>> >operand would suffice for non-atomic ops. I used memory clobber everywhere
>>
>> Atomic ops imply a barrier (otherwise the compiler can defeat the
>> purpose of the atomic operation). The non-atomic ones don't need a
>> dummy operand, but one that precisely describes the place in memory
>> that changes.
>
>Mmm, won't that unnecessarily make the compiler put code to compute that
>place?
Not generally - if both the 'used' and the 'fake' operands are apart just
by a constant offset, the same base/index expression will be used by
the compiler, just with different offsets. If the bit position is not a
constant that might well be true, though (and I ought to check) - since
in that case the compiler can't predict which part of the array changes*,
splitting the two cases with a __builtin_constant_p() should be possible,
though. The constant section then would need two operands (one of
them fake), while the variable section would need just one.
* In theory, the compiler could even then detect certain cases of
accesses to the same field. If we're afraid of that, the we'd need to
use precise operand descriptions here or, as Keir suggested, just
live with using a memory barrier instead of precise operand constraints.
Jan
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home