https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102646
--- Comment #3 from Aldy Hernandez ---
Most if not all the performance changes I've seen so far have been,
not due to the jump threader changes, but to the restrictions we've
put into place for jump threadable paths. Before, we used to thread
paths that destroyed loop form. We are much more cautious now. In
theory, the vectorizer should be able to do an even better job with
loops preserved longer.
That being said, this has been a bit of a moving target, with the
thread validity model changing a few times in the past month.
A good exercise would be to compare the old and new threaders, with
the vectorizer kept constant (whether -fno-tree-vectorize or
-ftree-vectorize), but without the loop threading restrictions. That
is, with something like this patch:
diff --git a/gcc/tree-ssa-threadupdate.c b/gcc/tree-ssa-threadupdate.c
index 32ce1e3af40..1a49cb61ca3 100644
--- a/gcc/tree-ssa-threadupdate.c
+++ b/gcc/tree-ssa-threadupdate.c
@@ -2853,9 +2853,6 @@ jt_path_registry::register_jump_thread
(vec *path)
return false;
}
- if (cancel_invalid_paths (*path))
-return false;
-
if (dump_file && (dump_flags & TDF_DETAILS))
dump_jump_thread_path (dump_file, *path, true);
On Mon, Oct 11, 2021 at 4:59 PM hubicka at kam dot mff.cuni.cz
wrote:
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102646
>
> --- Comment #2 from hubicka at kam dot mff.cuni.cz ---
> > I think most of the regressions are fixed, we get even better numbers now.
> Because we enabled vectorization. I would say they should still
> reproduce with -fno-tree-vectorize, right?
>
> Honza
>
> --
> You are receiving this mail because:
> You are on the CC list for the bug.
>
