http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34723
Jeffrey A. Law <law at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Blocks|19794 |
--- Comment #5 from Jeffrey A. Law <law at redhat dot com> ---
So "useless" really means threading from outside to an empty latch means the
entire loop is being peeled.
Reviewing the discussions between Zdenek, Dorit and myself, it was Zdenek's
contention that threading should never peel loops, leaving that to another
pass. The code to not thread to an empty latch block is a heuristic to avoid
peeling.
What I don't get is why we don't use changes in the loop structure to identify
cases where we're peeling a loop rather than threading to an exit edge (which
we certainly want to allow). Perhaps there wasn't enough infrastructure for
that in place in 2005..
Regardless of that, it looks like Zdenek's goal was to avoid peeling in DOM and
instead leave that to other passes which can make better estimates about
whether or not peeling is useful. I'm not 100% convinced that's the best thing
to do, but I don't think it's something we can/should revisit during stage3.
I'm removing this from the jump threading tracker as after reviewing the old
discussion, the threader is roughly doing what it's supposed to do.
Not sure how to best address the PR since this is a case where peeling is
actually quite cheap -- we don't have a good way to estimate when the
duplicated code from peeling for the most part just optimizes away, as it does
in this testcase.
