http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55860
Jeffrey A. Law <law at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |law at redhat dot com
--- Comment #2 from Jeffrey A. Law <law at redhat dot com> ---
My current work won't help this PR. edge equivalences are not well handled
for threading. While the edge containing the equivalence for jlk < n dominates
the true arm of the condition, it does not dominate the join point where the
true/else arms meet. So it's not going to be available in the expression hash
table when we try to thread the true arm.
dom_opt_leave_block does some work to recover edge equivalences, but only for
the edge from the block we're leaving to its successor blocks. The edge
equivalence is higher up in the chain.
In theory we could walk up the CFG and add more stuff to the tables, which
would probably help. Assuming that worked, I still doubt threading is going to
be successful as we have to thread through the backedge in the CFG and we're
currently very conservative with that, both in terms of recording jump threads
and in terms of which jump threads we'll ultimately realize.
In this particular case, threading from the true arm through the backedge will
create a new, inner loop which the threading code tries really hard not to do
because it doesn't know how to update all the various loop structures.
We have the same problem for the FSA optimization, but in that case I'm tempted
to just throw away all the loop information -- if we thread a backedge and
eliminate a switch, that's huge from a performance standpoint and we can
probably afford to just throw away the loop info at that point.
