[Bug middle-end/54146] Very slow compile with attribute((flatten))

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54146

--- Comment #33 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-08-08 08:09:25 UTC ---
(In reply to comment #29)
> Created attachment 27957 [details]
> Do not traverse sibling loops
> 
> The idea here is to note that for a nested loop we know for sure that the loop
> exits into a common loop father must all need PHI nodes, because we have have
> loop pre-headers. I'm not sure what the effect of this is on irreducible loops,
> and I haven't tested this much yet. Comments sought...

Nice results.  A few comments - if you rely on preheaders then please assert
in rewrite_into_loop_closed_ssa that loops_state_satisfies_p
(LOOPS_HAVE_PREHEADERS).

Irreducible loops are "one" loop to cfgloop and to loop-closed SSA form, so
they should be fine.

What do you mean by "for a nested loop we know for sure that the loop
exits into a common loop father must all need PHI nodes"?  What happens for
nested loops that do not exit into a common loop father?  That is what's
"common" here?  I think what you want to say is that a loop cannot exit
to a child of a sibling of a father, thus:

  for (;;)
    for (;;)
      goto x;
  for (;;)
    x:

cannot happen in the sense that loop detection would not detect the loop
nest as written (but had a single irreducible loop)?

I think you should simply move compute_global_livein to its single use
and make it static.