On Fri, 22 Nov 2019, Jeff Law wrote:
> On 11/21/19 1:00 AM, Richard Biener wrote:
> > On Wed, 20 Nov 2019, Jeff Law wrote:
> >
> > Sure, but we're not being called with this but with RTL
> > expanded via cselib usually:
> >
> > (plus:SI (plus:SI (and:SI (symbol_ref:SI ("*.LANCHOR0") [flags 0x182])
> > (const_int 3 [0x3]))
> > (value:SI 8:8 @0x3089c18/0x30f5e40))
> > (const_int -8 [0xfffffffffffffff8]))
> Ugh. I'm suspect all the ANCHOR stuff came in well after the original
> code to try and optimize the low-order bit-logical masking for addressing.
The PR also shows an example w/o the ANCHOR.
> Without knowing something more, I'd claim that can't actually be a pointer.
>
> >
> > so there are no REGs involved on the "subexpressions" we could
> > check REG_POINTER on ... and in fact the symbol_ref _is_
> > a pointer! It's just not something we should consider the
> > base of the memory reference. We're talking about sth like
> >
> > *(&a + ((uintptr_t)&b - (uintptr_t)&a))
> >
> > where it's nearly impossible to tell (on RTL, if the address
> > parts are not split to separate insns and thus REG_POINTER is
> > applicable) what object we are looking at for aliasing purposes.
> Right. And in that case we have to make the most conservative
> assumptions.
Right.
>
> >
> > The appropriate way here is to not to too clever things with
> > the address but instead rely on MEM_EXPR here - unfortunately
> > that is non-existent for spill slots which is where most of
> > the regressions with removing the code appear.
> That's what doesn't make much sense to me. We ought to be able to
> analyze stack slots reasonably well. But the RTL DSE bits have changed
> a lot since Christian's implementation -- it actually used to be easy
> to follow.
So alias analysis of stack slots vs. non-stack slots relies on
finding some non-stack "base" in find_base_term, that's basically
saying "yeah, this MEM cannot alias [this] stack slot". But we
have to get those other MEMs out of the way to DSE stores to spill
slots.
It looks like MEM_EXPR (mem) == get_spill_slot_decl (false) should
handle this somewhat but I guess that spill slot decl is aliased
(marked as address taken) and thus that MEM_EXPR doesn't help us
very much, plus DSE is happily throwing away MEM_EXPRs in its
meta it throws to the alias oracle.
Richard.
