Subject: desugaring issue
From: Vassily Litvinov (vass@cs.washington.edu)
Date: Wed Jan 24 2001 - 18:06:38 PST
I'd like to reconsider the way the typechecker pre-processes type/object
declarations with "isa"/"subtypes" clause (this is solely a syntactic
issue). Problems encountered recently by Craig and Todd brought me back
to an ages-old problem with the Vortex typechecker.
Consider, for example, this declaration:
forall 'T, 'S>=T:
object i_vector[T] subtypes i_vector[S];
This is a syntactic sugar for two declarations:
... object i_vector[T];
... extend i_vector[T] subtypes i_vector[S];
Now the question is: what should appear in place of ... in each
declaration? A straightforward answer is to replicate the "forall"
clause for each declaration:
forall 'T, 'S>=T: object i_vector[T];
forall 'T, 'S>=T: extend i_vector[T] subtypes i_vector[S];
The problem, quite obvioussly, is that i_vector now has two type
parameters (one explicit (T) and one implicit (S)), so that upon each
instantiation the typechecker will be trying to find both T and S such
that S>=T. Nothing is going to go wrong because of that, but it just
doesn't sound right to me for i_vector to have an extra type parameter
that the programmer did not intend.
Why, you will ask, bother replicating the "forall" clause? Quite
obviously, i_vector[T] has only the T type parameter and nothing else.
Well, it happens so in the case of i_vector. But consider Todd's example:
forall `G <= AnalysisGraph[E, P],
`E <= IREdge, `P <= AnalysisPriority[P]:
object ScopeAnalysis[G]
isa Analysis[GenericScopeAnalysisInfo,G];
Semantically it is essential that all of `G, `E, and `P, together
with all the constraints, should be part of both the "object" and
"isa" desugared declarations.
It turns out that the current Vortex typechecker does poor job in both
the i_vector and ScopeAnalysis examples, apparently because I didn't
get right the desugaring rules. (Moreover, I think the rules are
undocumented.)
Therefore I solicit proposals for desugaring rules which will handle
both cases above and also other cases throughout Cecil stdlib/Vortex
code where we had to move the "isa" part into a separate declaration
(grep for "extend "). I would favor intuitiveness of the rule
over sofistication. For example, if the programmer knows that the
"forall" part will be replicated, then the i_vector declaration would
be written simply as:
object i_vector[T] subtypes i_vector[`S>=T];
Tied to this is the issue of constraints associated with types (e.g.,
the bounds on `G, `E, `P are associated with type ScopeAnalysis) and
how to desugar those. But that's yet another story.
v
This archive was generated by hypermail 2b25 : Wed Jan 24 2001 - 18:06:43 PST