[Bug tree-optimization/103458] [12 Regression] ICE in verify_loop_structure, at cfgloop.c:1736 (error: loop with header 4 not in loop tree)

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103458

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the issue is that we have


   bb3:<-\ <--\
    |\   /    |
    | --      |
    |         |
   bb4:       |
    .ABNORMAL_DISPATCHER

at the start of CDDCE and the abnormal predecessor makes us not consider BB3
a loop header.  But then we have an opportunity to create a forwarder
(we avoid creating them for loop headers but this isn't one ...) which
shows the opportunity to represent this CFG as an outer loop with an inner
cycle with abnormal edge.  CFG cleanup doesn't undo this forwarder because
there are virtual PHIs in it.

ethread is the pass that introduces this problematic header while this is
all inside an irreducible region.

In principle handling the abnormal edge in this case would be OK but in
general abnormal latches pose a problem since with several loop opts we
are facing the need to split/redirect those so relaxing this condition
doesn't look feasible (but it creates nicer CFG for this particular testcase).