[Bug rtl-optimization/56181] [4.8 Regression] ICE in verify_loop_structure, at cfgloop.c:1581 with -ftracer

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56181

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-06 14:55:25 UTC ---
(In reply to comment #9)
> Created attachment 29369 [details]
> updated patch
> 
> This works better.  Probably flow_loops_find should not handle loop removal
> though, and the IRA use is plain ugly (but it was before).

What I like about this idea is consolidating of fix_loop_structure and
flow_loops_find (and possibly parts of loop_optimizer_init).  What I like
less is that we allow even more leeway to passes to fixup after them.
That is, at least we should ICE when a loop marked for removal by a pass
is then re-discovered with exactly the same header.  And we should probably
dump info somewhere when we incrementally discover new loops.

Separately fixing the IRA oddness would be nice (use
loop_optimizer_init (AVOID_CFG_MANIPULATIONS) and current_loops instead of
ira_loops).