On Thu, 11 Aug 2022, Aldy Hernandez wrote:
> On Thu, Aug 11, 2022 at 3:59 PM Andrew MacLeod <amacl...@redhat.com> wrote:
> >
> >
> > On 8/11/22 07:42, Richard Biener wrote:
> > > This avoids going BBs outside of the path when adding def chains
> > > to the set of imports. It also syncs the code with
> > > range_def_chain::get_def_chain to not miss out on some imports
> > > this function would identify.
> > >
> > > Bootstrap / regtest pending on x86_64-unknown-linux-gnu.
> > >
> > > The question still stands on what the path_range_query::compute_ranges
> > > actually needs in its m_imports - at least I don't easily see how
> > > the range-folds will use the path range cache or be path sensitive
> > > at all.
> >
> > All the range folding code is in gimple_range_fold.{h,cc}, and its
> > driven by the mystical FUR_source classes. fur_source stands for
> > Fold_Using_Range source, and its basically just an API class which all
> > the folding routines use to make queries. it is used by all the fold
> > routines to ask any questions about valueizing relations, ssa name,
> > etc.. but abstracts the actual source of the information. Its the
> > distillation from previous incarnations where I use to pass an edge, a
> > stmt and other stuff to each routine that it might need, and decided to
> > abstract since it was very unwieldy. The base class requires only a
> > range_query which is then used for all queries.
>
> Note that not only is ranger and path_query a range_query, so is
> vr_values from legacy land. It all shares the same API. And the
> simplify_using_ranges class takes a range_query, so it can work with
> legacy or ranger, or even (untested) the path_query class.
>
> >
> > Then I derive fur_stmt which is instantiated additionally with the stmt
> > you wish to fold at, and it will perform queries using that stmt as the
> > context source.. Any requests for ranges/relations/etc will occur as
> > if that stmt location is the source. If folding a particular stmt, you
> > use that stmt as the fur_stmt source. This is also how I do
> > recalculations.. when we see
> > bb4:
> > a_32 = f_16 + 10
> > <...>
> > bb88:
> > if (f_16 < 20)
> > b_8 = a_32 + 8
> > and there is sufficient reason to think that a_32 would have a different
> > value , we can invoke a re-fold of a_32's defintion stmt at the use
> > point in b_8.. using that stmt as the fur_source. Ranger will take into
> > account the range of f_16 being [0,19] at that spot, and recalculate
> > a_32 as [10,29]. Its expensive to do this at every use point, so we
> > only do it if we think there is a good reason at this point.
> >
> > The point is that the fur_source mechanism is how we provide a context,
> > and that class talkes care of the details of what the source actually is.
> >
> > There are other fur_sources.. fur_edge allows all the same questions to
> > be answered, but using an edge as the source. Meaning we can calculate
> > an arbitrary stmt/expressions as if it occurs on an edge.
> >
> > There are also a couple of specialized fur_sources.. there is an
> > internal one in ranger which communicates some other information called
> > fur_depend which acts like range_of_stmt, but with additional
> > functionality to register dependencies in GORI as they are seen.
>
> This is a really good explanation. I think you should save it and
> included it in the documentation when you/we get around to writing it
> ;-).
>
> >
> > Aldy overloads the fur_depend class (called jt_fur_source-- Im not sure
> > the origination of the name) to work with the values in the path_query
> > class. You will note that the path_range_query class inherits from a
> > range_query, so it supports all the range_of_expr, range_of_stmt, and
> > range_on_edge aspect of rangers API.
>
> The name comes from "jump thread" fur_source. I should probably
> rename that to path_fur_source. Note that even though the full
> range_query API is available in path_range_query, only range_of_expr
> and range_of_stmt are supported (or tested). As I mention in the
> comment for the class:
>
> // This class is a basic block path solver. Given a set of BBs
> // indicating a path through the CFG, range_of_expr and range_of_stmt
> // will calculate the range of an SSA or STMT as if the BBs in the
> // path would have been executed in order.
>
> So using range_on_edge would probably give unexpected results, using
> stuff in the cache as it would appear at the end of the path, or some
> such. We could definitely harden this class and make it work solidly
> across the entire API, but we've had no uses so far for anything but
> range_of_expr and range_of_stmt-- and even those are only supported
> for a range as it would appear at the end of the path. So if you call
> range_of_expr with a statement anywhere but the end of the path,
> you're asking for trouble.
>
> >
> > I believe all attempts are first made to pick up the value from the path
> > cache, and failing that, a query is made from ranger at the start of the
> > path. So as the walk thru the path happens, and edges are taken/chosen,
> > all the context information local to the path should be placed into the
> > path cache, and used in any future queries using the path_range_query.
> > Ranger will only be invoked if there is no entry in the path query, and
> > it would provide the range as it would appear at entry to the path.
> >
> > Thats my high level understanding of how the path_query class provides
> > context.
>
> That's actually a really good explanation of how it all works (or at
> least how it's supposed to work ;-)). Thanks.
Yes, thanks - that was helpful (and the general back threader stuff
confirms what I reverse engineered).
> >
> > So I think to answer the other question, the m_imports list Is probably
> > the list of ssa-names that may have relevant context information which
> > the path_query would provide ranges during the walk instead of ranger?
> > I think Aldy pre-calculates all those during the walk, and then uses
> > this pre-filled cache to answer questions at the thread exit? Thats my
> > guess
>
> Yeah, though I should probably sit down and do some testing, making
> sure we're not adding more imports than we need to. Richi has found a
> whole bunch of imports that ended up in the list that were ultimately
> not needed. My original comment that adding more imports to the
> bitmap had no penalty should be nuked-- it obviously has a performance
> effect, especially for pathological cases.
>
> Regarding the name "imports". I think it's confusing all of us,
> because GORI has a very clear definition of what imports are. OTOH,
> the path solver starts with the imports from GORI land, but ends up
> adding more SSA names that would be useful to solve, to provide
> context as Andrew says. Maybe, we should call the "imports" in the
> path solver something else to avoid confusion. Interesting names?
> Must-be-solved? Context-SSA? Suggestions?
I understand them to be path exit dependences, the 'interesting names'
thing is already used. So maybe
s/compute_imports/compute_exit_dependences/, this name clash did
confuse me a bit. Though we do stop at the path entry and have
the "path imports" in the list of dependences as well (but not
the dependences of those imports).
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
query start (const vec<basic_block> &);
and make range_of_* and friends members of a new 'query' class
instantiated by path_range_query. I ran into this when trying
to axe the linear array walks for the .contains() query on the
path where I need a convenient way to "clenanup" after a path
query is done.
Richard.
