On Tue, Aug 16, 2022, 14:26 Richard Biener <rguent...@suse.de> wrote:
> On Tue, 16 Aug 2022, Aldy Hernandez wrote:
>
> > On Tue, Aug 16, 2022 at 1:38 PM Richard Biener <rguent...@suse.de>
> wrote:
> > >
> > > On Tue, 16 Aug 2022, Aldy Hernandez wrote:
> > >
> > > > On Mon, Aug 15, 2022 at 11:53 AM Richard Biener <rguent...@suse.de>
> wrote:
> > > > >
> > > > > The remaining issue I have with the path_range_query is that
> > > > > we re-use the same instance in the back threader but the
> > > > > class doesn't provide any way to "restart", aka give m_path
> > > > > a lifetime. The "start a new path" API seems to essentially
> > > > > be compute_ranges (), but there's no convenient way to end.
> > > > > It might be more appropriate to re-instantiate the
> path_range_query,
> > > > > though that comes at a cost. Or abstract an actual query, like
> > > > > adding a
> > > >
> > > > Yes, compute_ranges() is the way to start a new path. It resets exit
> > > > dependencies, the path, relations, etc. I think it would be clearer
> > > > to name it set_path (or reset_path if we want to share nomenclature
> > > > with the path_oracle).
> > > >
> > > > Instantiating a new path_range_query per path is fine, as long as you
> > > > allocate the ranger it uses yourself, instead of letting
> > > > path_range_query allocate it. Instantiating a new ranger does have a
> > > > cost, and it's best to let path_range_query re-use a ranger from path
> > > > to path. This is why path_range_query is (class) global in the
> > > > backwards threader. Andrew mentioned last year making the ranger
> > > > start-up 0-cost, but it still leaves the internal caching the ranger
> > > > will do from path to path (well, the stuff outside the current path,
> > > > cause the stuff inside the path is irrelevant since it'll get
> > > > recalculated).
> > > >
> > > > However, why can't you use compute_ranges (or whatever we rename it
> to ;-))??
> > >
> > > I've added
> > >
> > > auto_bb_flag m_on_path;
> > >
> > > to the path query and at set_path time set m_on_path on each BB so
> > > the m_path->contains () linear walks go away. But I need to clear
> > > the flag for which I would need something like finish_path (),
> > > doing it just at the point we deallocate the path query object
> > > or when we set the next path via compute_ranges doesn't look right
> > > (and in fact it doesn't work out-of-the-box without adjusting the
> > > lifetime of the path query object).
> > >
> > > So a more incremental thing would be to add such finish_path ()
> > > or to make the whole path query object single-shot, thus remove
> > > compute_ranges and instead use the CTOR for this.
> > >
> > > Probably not too important (for short paths).
> >
> > On a high level, I wonder if this matters since we don't allow long
> > paths for other performance reasons you've already tackled. But OTOH,
> > I've always been a little uncomfortable with contains_p linear search,
> > so if you think this makes a difference, go right ahead :).
> >
> > I'm fine with either the finish_path() or the single-shot thing you
> > speak of. Although making path query inmutable makes things cleaner
> > in the long run. I like it! My guess is that the non-ranger
> > instantiation penalty would be minimal. I'd even remove the default
> > (auto-allocated) ranger from path_range_query, to make it obvious that
> > you need to manage that yourself and avoid folks shooting themselves
> > in the foot.
>
> We currently have
>
> path_range_query::path_range_query (bool resolve, gimple_ranger *ranger)
> : m_cache (new ssa_global_cache),
> m_has_cache_entry (BITMAP_ALLOC (NULL)),
> m_resolve (resolve),
> m_alloced_ranger (!ranger)
> {
> if (m_alloced_ranger)
> m_ranger = new gimple_ranger;
> else
> m_ranger = ranger;
>
> m_oracle = new path_oracle (m_ranger->oracle ());
>
> if (m_resolve && flag_checking)
> verify_marked_backedges (cfun);
> }
>
> so at least verify_marked_backedges will explode, I suppose we
> want to hoist that somehow ...
>
Good point.
> then we allocate the path_oracle - that one does have a
> reset_path () function at least. It's allocation looks
> quite harmless, but we should only need it when m_resolve?
>
Yes.
> > Wanna have a go at it? If you'd rather not, I can work on it.
>
> If you have cycles go ahead - I'm fiddling with other parts of
> the threader right now.
>
Sure, I'll take a stab at it. Thanks for the other stuff you're doing on
the threader.
Aldy
> Richard.
>
>
