[PATCH 1/2] Seperate -funroll-loops for GIMPLE unroller and RTL unroller

[David Edelsohn]

On Mon, May 25, 2020 at 1:58 PM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On May 25, 2020 7:40:00 PM GMT+02:00, Segher Boessenkool <segher@kernel.crashing.org> wrote:
> >On Mon, May 25, 2020 at 02:14:02PM +0200, Richard Biener wrote:
> >> On Mon, May 25, 2020 at 1:10 PM guojiufu <guojiufu@linux.ibm.com>
> >wrote:
> >> Since a new flag is not needed to fix the regression please avoid
> >> adding -fcomplete-unroll-loops.
> >>
> >> For -frtl-unroll-loops you should be able to use
> >
> >Erm.  That *is* a new command-line option (the internal flags I do not
> >care about so much: new implementation details are *good*).  And a new
> >name that is a mistake in my opinion, for many reasons (users do not
> >know and should not have to care about "rtl"; the name is not
> >descriptive; it is useless churn, it is not the same name as we have
> >had for decades now; it is adding a new option for a future where we
> >will do most unrolling at gimple level, a future we do not know will
> >ever exist, and we do not know what that will look like anyway; it is
> >an extra level of indirection (in the name)).
> >
> >We should not have an -frtl-unroll-loops if we do not have a
> >-ftree-unroll-loops (or whatever).
> >
> >Unrolling early is not a good idea in general (the problems with the
> >very trivial complete unrolling case just underline that).  But we
> >*should* know which code we expect to unroll later, for better costing.
> >Adding names like "rtl-unroll-loops" only stands in the way of getting
> >a better design here.
>
> You folks made ppc specific hacks instead of a better design. Those now stand in the way as well. But sure, simply do not expose the flag to the users, use
> Var(flag_rtl_unroll_loops). My other points still stand.
>
> Feel free to ignore the regression part on the branch and come up with a great design. But don't expect to backport that then.

I completely agree.

This path is digging a deeper and deeper hole.

- David