https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
vries at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #35983|0 |1
is obsolete| |
--- Comment #7 from vries at gcc dot gnu.org ---
Created attachment 35986
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35986&action=edit
Updated tentative patch
I found that always doing graphite before parloops resulted in failures to
parallelize reduction testcases.
I've split things up now:
- first we do parloopsred, a parloops variant in which we only handle
reductions
- then we do graphite
- then we do the normal parloops
This seems to combine the best of graphite and parloops.
The only gotcha is that I had to disable pass_iv_canon when
tree_parallelize_loops > 1. It seems to interfere with graphite. I did not
observe any failures to parallelize due to not running pass_iv_canon.
