On Wed, Dec 1, 2021 at 5:00 PM Aldy Hernandez <al...@redhat.com> wrote:
>
> On Wed, Dec 1, 2021 at 2:36 PM Richard Biener
> <richard.guent...@gmail.com> wrote:
> >
> > On Tue, Nov 30, 2021 at 4:48 PM Aldy Hernandez via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > We are currently restricting loop crossing paths in the generic copier
> > > used by the back threader, but we should be able to handle them after
> > > loop_done has completed.
> >
> > But this isn't a cost thing but a correctness (as in, can we update
> > loops properly)
> > thing. We are passing 'loop' down to copy_bbs and very much rely on
> > seeing either BBs of that loop or copying complete subloops of it.
> >
> > You will likely see excess loop fixup caused by this or ICEs in case the
> > caller of this function will not always set LOOPS_NEED_FIXUP
> > (not that copy_bbs handling of loops is perfect).
>
> Ughh, fair enough.
>
> In which case at least I've commented in the PR what the problem is
> here. Perhaps we could benefit from a BB copier that could handle
> loops, which is way above my circle of competence ;-).
Well, more high level APIs exercising the copier are able to handle loops
but it really depends on the high-level transform what you need to do ;)
So for "loop-crossing" threadings it depends on the actual details what
the "loop transform" is that you are doing. If you just rely on
LOOPS_NEED_FIXUP then for example a #pragma GCC unroll 0
might be lost and that is - counter to PROP_after_loop_opts - also
effective to prevent the RTL unroller from unrolling a loop.
What's probably OK is for example changing the code to only
prohibit copying loop headers:
+ /* We do not handle subloops, i.e. all the blocks must belong to the
+ same loop. */
+ if (region[i]->loop_father != loop
&& region[i]->loop_father->header != region[i])
+ return false;
that should allow threadings covering an exit edge for example. It's
reasonable (with some extra setup here!) to also handle the case
of entering and exiting a loop (the "subloop" the comment mentions),
but that requires us to do the loop structure copying/setup before
copying the BBs (maybe look at gimple_duplicate_sese_tail).
Note the above adjustment to handle exits does not necessarily result in correct
->loop_father assignment but LOOPS_NEED_FIXUP should be able to
deal with it without using loop annotations. These kind of threads should also
be OK when loop opts are not finished yet.
Richard.
> Aldy
>
