[Bug middle-end/39326] Segmentation fault with -O1, out of memory with -O2

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39326

--- Comment #52 from Richard Biener <rguenth at gcc dot gnu.org> 2013-03-26 10:09:52 UTC ---
(In reply to comment #51)
> >     (struct mem_ref): Replace mem member with ao_ref typed member.
> 
> RTL gcse (-O2) suffers from the same slowness in its dependence tests.  Caching
> ao_ref instead of just mem and alias-set in RTL mem-attrs would improve
> this tremendously.
>
> A quick try reveals that gengtype is not at all happy with including
> something in rtl.h though :/

Doing that (caching the ao_ref) doesn't help as much as it did on the
GIMPLE level.  The most time-consuming part of canon_true_dependence
(which consumes 56% of compile-time) is find_base_term (37% of that 56%)
followed by memrefs_conflict_p and only then (after caching ao_ref)
rtx_refs_may_alias_p (18%).  The find_base_term result is another thing
that could be easily cached when doing dependence checks that repeatedly
use one or another operand.  For this testcase the most expensive caller
is compute_transp.

There is some obvious way to do less find_base_term calls in alias.c
itself.  I'm going to cleanup things there a bit.