[aarch64-port-dev ] RFR: AArch64 -- patch to make volatile reads and stores use ldar/stlr

[Andrew Dinn]

On 21/04/15 19:18, Aleksey Shipilev wrote:
> There is also a pending idea for x86: try to replace volatile
> stores with "xchg", instead of "mov + lock addl". It would be nice
> to make sure the way AArch64 does the replacement would help the
> future x86 code.

I don't know for sure if this is going to be susceptible to the same
sort of implementation as I have provided but I think it is very likely.

As I understand it the x86 code generates the lock addl instruction in
response to a MemBarVolatile node. So, a predicate something like the
one I provided could be used to identify that the MemBarVolatile
belongs to a volatile load or store and hence that the lock addl can
be elided.

Similarly, I suspect the rules which generate the mov for normal
stores could probably be duplicated and a lower cost variant could use
a predicate similar to my needs_releasing_store() and plant an xchg
instead of a mov if the predicate decided the store was volatile.

regards,


Andrew Dinn
-----------
Senior Principal Software Engineer
Red Hat UK Ltd
Registered in UK and Wales under Company Registration No. 3798903
Directors: Michael Cunningham (USA), Matt Parson (USA), Charlie Peters
(USA), Michael O'Neill (Ireland)