[Bug rtl-optimization/87763] [9 Regression] aarch64 target testcases fail after r265398

["rearnsha at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87763

--- Comment #6 from Richard Earnshaw <rearnsha at gcc dot gnu.org> ---
(In reply to Wilco from comment #5)
> (In reply to Segher Boessenkool from comment #4)
> > (In reply to Wilco from comment #3)
> > > IRA costing doesn't consider the possibility of a simple move being
> > > removeable.
> > 
> > Not always, yeah (only if you have matching constraints, which are silly to
> > have for moves; IRA or LRA should do the work, not the machine description).
> 
> Would it take this into account in the costs? I believe it doesn't.
> 
> > > Neither does it consider that some variants are more expensive
> > > than others - use of '?' is the only way to mark more expensive variants.
> > 
> > ? and !, or ^ and $ if it only more expensive if it needs reloads.
> > 
> > > The example shown is fixed when I change the w=rZ variant to ?w=rZ in
> > > movsi_aarch64.
> > 
> > Is it always more expensive than the other alternatives?  Then it *should*
> > have "?" modifiers!
> 
> Today using '?' seems to be the only way to tell it to prefer certain
> alternatives. However this is annoying given there are already separate
> costs for moves between different register classes. So this means we end up
> hardcoding microarchitecture costs in the md files instead of using the
> existing cost hooks...

The '?' modifiers approach is not really suitable for any case where the cost
might depend on the micro-architecture.  There's only one cost that can be
applied this way, and if it doesn't suit all implementations, you're hosed.