Hi Richard,
From: Richard Sandiford <[email protected]>
> > @@ -3719,6 +3722,7 @@ static unsigned int rest_of_handle_sched (void)
> > { #ifdef INSN_SCHEDULING
> > + first_moveable_pseudo = last_moveable_pseudo;
> > if (flag_selective_scheduling
> > && ! maybe_skip_selective_scheduling ())
> > run_selective_scheduling ();
>
> I think instead we should zero out both variables at the end of IRA.
> There are other places besides the scheduler that call into the IRA code, so
> tackling the problem there seems more general.
If you zero first_moveable_pseudo and last_moveable_pseudo after IRA then
they'll be zero for the second scheduler pass, which uses them.
In fact, I've just realised that the GCSE and move_loop_invariants passes
also use them (they both call ira_set_pseudo_classes which calls
find_costs_and_classes which uses these variables).
They need to be zeroed or set equal to each other before the first pass that
uses them (move_loop_invariants), but kept alive until after the last pass
that uses them (GCSE). So if there's a function that sets things up right
before the RTL passes start then I think that's a good location candidate.
> > +/* We have plenty of spare registers, so check nothing has been
> > +spilled. */
> > +/* { dg-final { scan-assembler-not "str" } } */
>
> The testcase looks good, but it's probably better to make that “\tstr\t”.
> The problem with plain “str” is that assembly output can often include
> pathnames and version strings, and it's possible that one of those could
> contain “str”.
Good catch, I'll keep that tip in mind for future!
Thanks,
Omar
