Re: [abc] Simplifying codegen for tracematches

Another optimisation that we discussed was the following.

Once all variables on a state are bound, its outgoing transitions
merely check (via addBindings and addNegativeBindings) what the
values are. It is thus not necessary for Disjunct.addBindings and
Disjunct.addNegativeBindings to construct a new Disjunct in the
result: it is always just "this" or falseD.

However, the situation is complicated by the use of weak and
strong references, and the full condition is therefore:
if we take a transition s->t, and both s and t have all
variables bound, and their NeedStrongRef sets are the same,
then we can use versions of
   addBindings and addNegativeBindings
that always return "this" or falseD.

We weren't sure whether this is too specific to jhotdraw.

On Mon, 1 Aug 2005, Pavel Avgustinov wrote:

> Following a discussion with Oege, we propose the following optimisation:
>
> Note that any temporary label is never copied to a different state (while it
> is temporary). All that happens is that we accumulate disjuncts in it, then
> make it the 'proper' label for the associated state. Only once it's become
> the 'proper' label can it happen that it is copied (with additional bindings)
> to another state.
>
> Now observe the operations that are performed on lab'(), the temporary labels
> of the states. The only operation that occurs is lab'().or(). Moreover, the
> constraint's or() method is only ever called on temporary constraints.
>
> These two observations justify the following conclusion: Disjunction
> operations on constraints can be destructive. Since the constraint isn't used
> for anything else until it is copied into the state's non-temporary label,
> after which its or() method isn't called again, modifying the constraint
> in-place for operations like c1.or(c2) (which would add all of c2's disjuncts
> to c1) is justified.
>
> I'm about to make this change to the Constraint class. add[Neg]Bindings will
> still behave functionally, but or() will modify the constraint in-place.
>
> - P
>
> On Tuesday 19 July 2005 20:25, Oege de Moor wrote:
> > optimisation:
> >
> > all initial states are always labelled with true,
> > so we should leave them alone: it is never necessary
> > to compute new constraints for these.
> >
> > Julian, can you amend the codegen accordingly?
> >
> > On Mon, 18 Jul 2005, Damien Sereni wrote:
> > > aspect Tracematch {
> > > // For each initial state sj (1<=j<=N)
> > > private Constraint lab(sj) = true();
> > >
> > >>>>> private Constraint lab'(sj) = false();
> >
> > delete this line
> >
> > > // For each non-initial state sj
> > > private Constraint lab(sj) = false();
> > > private Constraint lab'(sj) = false();
> > >
> > > private Constraint skip = true();
> > >
> > > // For each symbol ai (1<=i<=n)
> > > ai(y1, ..., yk) {
> > >
> > >>>> for each state s
> >
> > replace by: for each non-initial state s
> >
> > > for each state s' with s' -->ai s
> > > lab'(s) = or(lab'(s), and(lab(s'), eqs(vsi, [1, ...,
> > > yk])));
> > >
> > >>>> skip = and(skip(not(eqs(vsi,[y1,...,yk])))); // THIS IS NEW
> >
> > (typo) and(skip,not(eqs(vsi,[y1,...,yk])));
> >
> > > }
> > >
> > > some() {
> > >
> > >>>> for each state s
> >
> > replace by: for each non-initial state s
> >
> > > if s has a skip-loop then // THIS IS NEW
> > > lab'(s) = or(lab'(s), and(lab(s),skip));
> > > lab(s) = lab'(s);
> > > lab'(s) = false();
> > > skip = true(); // THIS IS NEW
> > > for each final state s
> > > for each solution v of lab(s)
> > > run the advice body with bindings v;
> > > }
> > > }
> > >
> > > I think that this works [modulo typos]; please check it to make sure!
> > >
> > > Damien
> > >
> > > On 18/07/05, Pavel Avgustinov <pavel.avgustinov@magdalen.oxford.ac.uk>
> wrote:
> > > > (This is mainly for Julian and Oege, as I think only they are following
> > > > the implementation efforts)
> > > >
> > > > The current interface to the generated classes is as follows (I have
> > > > added methods in TraceMatchCodeGen to get the respective class names
> > > > from the TraceMatch object -- I'll check those in when I've written
> > > > this up)
> > > >
> > > > public class Constraint$TraceMatchName {
> > > > // 'true' and 'false' constraints -- globally unique
> > > > public final Constraint$TraceMatchName trueC;
> > > > public final Constraint$TraceMatchName falseC;
> > > >
> > > > // disjunction of constraints
> > > > public Constraint$TraceMatchName or(Constraint$TraceMatchName o);
> > > >
> > > > // conjunction of constraints -- is in atm, but I don't think it's
> > > > needed
> > > > public Constraint$TraceMatchName and(Constraint$TraceMatchName o);
> > > >
> > > > // cloning of the constraint -- don't think it's needed outside of
> > > > this class
> > > > public Constraint$TraceMatchName copy();
> > > >
> > > > // This bit I'm not entirely sure about yet, but I think the
> > > > following is the best
> > > > // interface. For any declared symbol X which binds variables
> > > > var1..varn,
> > > > public Constraint$TraceMatchName addBindingsForSymbolX(int
> > > > StateNumber, Object varr1, Object var2, ..., Object varn);
> > > > // Here, StateNumber is the number of the 'target' state, i.e. the
> > > > one on the LHS of the
> > > > // update equation. It's how the code figures out what to do wrt
> > > > weak/strong refs.
> > > > }
> > > >
> > > > I *think* this is all you'll need to access from the top level. This
> > > > lacks the method for getting solutions from constraints, because I
> > > > haven't yet thought about what that should look like -- once you know
> > > > what you'd like, let me know and I'll provide the method.
> > > >
> > > > - P
> > >
> > > --
> > > Damien Sereni
> > > Oxford University Computing Laboratory
> > > Wolfson Building, Parks Road
> > > Oxford OX1 3QD
> > > UK
> > >
> > > dsereni@gmail.com / dsereni@comlab.ox.ac.uk
> > > http://web.comlab.ox.ac.uk/work/damien.sereni
>
Received on Mon Aug 1 15:46:05 2005