Weird handling of r+r (pre-inc) addresses

Ulrich Weigand Ulrich.Weigand at de.ibm.com
Fri Mar 22 08:01:49 PDT 2013


Hal Finkel <hfinkel at anl.gov> wrote on 22.03.2013 07:47:36:

> These patches LGTM, regarding this last point:
>
> I think that in cases where the returned base/offset will fail these
> checks, we should check the swapped ordering and return that when
> possible. This will mean repeating the checks in the target code,
> but they're fairly small (pasted here for the convenience of list
readers):

With this change, we now actually get somewhat *more* pre-inc
accesses than before ...

I've now checked in the patches (including this change).

Thanks,
Ulrich




More information about the llvm-commits mailing list