[Bug tree-optimization/68549] [6 Regression] ICE: in verify_loop_structure, at cfgloop.c:1669

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68549

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-11-26
                 CC|                            |vries at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's a different cause though as it is from rtl_loop_init checking for
the case LOOPS_NEED_FIXUP is not set.  Thus a pass preceeding isn't
updating loop->latch.

I think that was by design (LOOPS_MAY_HAVE_MULTIPLE_LATCHES was supposed
to be "->latch doesn't matter, only keep header correct" which is what
we require from passes not having initialized loops).

Of course that's a somewhat conflicting requirement and thus the cleanest
approach would be to split the latch flags even more.  LOOPS_HAVE_LATCHES
(otherwise ->latch cannot be relied on), LOOPS_MAY_HAVE_MULTIPLE_LATCHES
(->latch may be NULL and do not disambiguate such loops),
~LOOPS_MAY_HAVE_MULTIPLE_LATCHES (disambiguate, all loops have ->latch
non-NULL), LOOPS_HAVE_SIMPLE_LATCHES (latches are forwarders).

The current state is really broken though in case we have
LOOPS_MAY_HAVE_MULTIPLE_LATCHES users that rely on ->latch reflecting
reality.  The flag is not explicitely requested by any pass though
so a simple option would be to revert the checking bits for now
and revisit the whole issue in stage1.