https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71366
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so it happens that we optimize away the backedge of the loop father when
canonicalizing IVs in the inner loop which appears as
<>
# d_9 = PHI<2(7)>
<header>
# d_3 = PHI<d_9(6), d_11(5)>
if (d_3 >= 0)
;
else
-> loop latch of outer loop
There is a missed constant propagation from the CCP pass immediately before
cunrolli as well.
So we peel the loop based on undefinedness of a[0][d] in the last iteration.
So we never take the d<0 exit to the latch of the outer loop which in turn
prunes this edge.
In fact it is removed as pointless in remove_redundant_iv_tests (optimized to
1 != 0) and then removed by to_remove edges marked by
gimple_duplicate_loop_to_header_edge via remove_path which in turn adjusts
stuff to eventually unloop outer loops (sth the pass doesn't really expect).
In fact we carefully avoid doing this kind of stuff to the loop we unroll
itself.
