[Bug target/70048] [6 Regression][AArch64] Inefficient local array addressing

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70048

--- Comment #8 from amker at gcc dot gnu.org ---
(In reply to Richard Henderson from comment #6)
> Created attachment 37886 [details]
> proposed patch
> 
> I agree -- at minimum virtual and eliminable frame registers ought to be
> special-cased.  If we separate the constants too far, we'll never be able
> to fold the constant plus the adjustment back together.
> 
> If the statement in #c4 is taken at face value -- that r233136 was applied
> to simplify frame-based array accesses...   Well, I simply don't believe
> that.
> 
> I can see how the patch would aid reduction of access to members of a
> structure that are in an array which in turn is *not* on the stack.  But
> for the average stack-based access I can't see except that it would hurt.

Hi Richard, my comment was about when to legitimize address expression in the
form of "base + reg << scaling + offset", the gimple passes could be improved
to catch CSE opportunities of "reg << scaling" so that base + offset can be
moved out of memory reference.  it's beneficial since "base_offset" is likely
invariant in loop context, especially when the base is sp/fp related.   IMHO,
this is the same transform as your propose patch does.  Or maybe I
mis-understood something important?

Thanks.