This is the mail archive of the
mailing list for the GCC project.
Re: Sharing stack slots
- To: kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)
- Subject: Re: Sharing stack slots
- From: Geoff Keating <geoffk at geoffk dot org>
- Date: 10 Oct 2001 12:59:33 -0700
- CC: gcc at gcc dot gnu dot org
- References: <10110092041.AA16529@vlsi1.ultra.nyu.edu>
firstname.lastname@example.org (Richard Kenner) writes:
> We have a lot of code to do it, but I'm wondering if we really should.
> There's a kludge in expr.c which says:
> /* If the address of the structure varies, then it might be on
> the stack. And, stack slots may be shared across scopes.
> So, two different structures, of different types, can end up
> at the same location. We will give the structures alias set
> zero; here we must be careful not to give non-zero alias sets
> to their fields. */
> That's a real efficiency hit.
> Moreover, one of the ideas behind the MEM tracking I'm doing is to record
> which decl a MEM is for with the idea that if two MEMs are in different
> decls, we they can't alias one another. But, of course, they can if they
> are in different scopes and have overlapped each other. But this is
> an important optimization, and I'm told particularly important on IA64.
> On the other hand, overlapping large objects that are in blocks of
> different scope are also quite important, especially in code
> that's generated by some program (and somewhat in Ada expanded code).
> We've been playing around with this compromise for a while, but I think we
> need to deal with it in a less ad-hoc manner now.
> Any thoughts?
- We can't stop sharing stack slots. That would cause an explosion in
the amount of stack space used.
- If we knew which decls mapped to which stack slots, then you could
still perform the optimisation, because it'd be possible to determine
which decls correspond to overlapping MEMs.
- With that information, it might also be possible to improve on the
alias set given to a structure on the stack...
- Geoffrey Keating <email@example.com>