https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93444
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
Target Milestone|--- |8.4
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Isn't this kind-of PR39612? I realize that was store-motion and here's
invariant motion.
I also note that more aggressively "optimizing" things like
_7 = (int) y_21(D);
would have allowed propagation of the uninitialized use to a context where
it isn't used invariantly. That is, optimize the above to
_9(D);
introduce an uninit register of 'int' type and optimize (int) y_21(D) to that.
Goes away with useful warnings about the uninit use because _9(D) would be
unnamed (we could use sth named y as well but it would technically be a
different
y).
For the special case of uninitialized SSA names (but not loads from
uninitialized
memory as in PR39612) we can "easily" fix LIM here:
static class loop *
outermost_invariant_loop (tree def, class loop *loop)
{
...
def_stmt = SSA_NAME_DEF_STMT (def);
def_bb = gimple_bb (def_stmt);
if (!def_bb)
return superloop_at_depth (loop, 1);
return NULL if ssa_undefined_value_p (def, true) with a comment explaining
things.
