QVTd Relation Overriding

The QVT specification has a relatively clear specification that QVTr relation may override, but no clues as to how this is mapped to QVTc that has no overrides concrete syntax, even though it inherits an overrides capability from QVTb.

This working document considers how overriding might be supported. It takes over from

• https://bugs.eclipse.org/bugs/show_bug.cgi?id=508267 [qvts] Introduce MicroMapping Alternator/Switch
• https://bugs.eclipse.org/bugs/show_bug.cgi?id=515327 [qvtc] Support overrides / refines

QVTc Choices

QVTc refinement syntax is an additive capability and therefore not suitable for implementing the QVTr replacement semantics of QVTr overrides.

We could add overrides syntax and semantics to QVTc, which then passes the buck to QVTi. But since we already need an 'alternator' to dispatch multiple similar but differently predicated mappings efficiently, we may be able to kill two birds with one stone.

QVTc is relatively simple. It solves the mapping sequencing problem by inter-trace class references/dependencies. Can it solve the overrides problem as-is? Let's go with this.

Consideration here led to QVT14-51 advocating the addition of overrides to the QVTc concrete syntax.

QVTr2QVTc Overrides 'Rewrite'

There is a simple rewrite that reifies overrides in QVTc without requiring any new capabilities, although it may stress some existing ones.

Given a transformation comprising Mx mappings/relations, Gx guard functions, and Bx bottom actions:

MA: when { GA(); } BA();
MB overrides MA: when { GB(); } BB();
MC overrides MB: when { GC(); } BC();
MD overrides MA: when { GD(); } BD();

we can rewrite the polymorphs (relations that participate in an overrides hierarchy) as siblings (relations that share a similar structural interface), refined guard functions and dispatchers:

MAsibling: when { GAsibling(); } BA();
MBsibling: when { GBsibling(); } BB();
MCsibling: when { GCsibling(); } BC();
MDsibling: when { GDsibling(); } BD();

GAsibling() = !GBsibling() && !GDsibling() && GA();
GBsibling() = !GCsibling() && GB();
GCsibling() = GC();
GDsibling() = GD();

MAdispatcher: when { GAdispatcher(); } where { MAsibling(); MBsibling(); MCsibling(); MDsibling(); }
MBdispatcher: when { GBdispatcher(); } where { MBsibling(); MCsibling(); }
MCdispatcher: when { GCdispatcher(); } where { MCsibling(); }
MDdispatcher: when { GDdispatcher); } where { MDsibling(); }

GAdispatcher() = GAsibling() || GBsibling() || GCsibling() || GDsibling();
GBdispatcher() = GBsibling() || GCsibling();
GCdispatcher() = GCsibling();
GDdispatcher() = GDsibling();

• introduce Gxsibling() equal to Gx() anded with the Gxsibling()es of all directly overriding relations.
• introduce Mxsibling() equal to Mx() using Gxsibling() to replace the override.
• replace Mx by an Mxdispatcher to all the new siblings

Mxdispatcher and Gxdispatcher can be omitted if there is no direct invocation of Mx.

(The aggregating Gxdispatcher() operations are required to ensure that Mxdispatcher dispatch fails if no override matches.)

Rewrite reification in QVTr2QVTc

We need to synthesize Gx() as a Function/Operation invoked from the Mapping Guard rather than being inlined in the Mapping Guard. All type checks must use explicit oclIsKindOf's since we cannot exploit 'mismatch is predicate fail' semantics within Function/Operations.

Whether Function/Operation should be used to exploit/bypass uniqueness optimizations is unclear. Probably doesn't matter. If QVTi understands what's happening, QVTi can optimize merged invocations behind the uniqueness of the polymorphic dispatcher.

Should relations that are not overridden also have their predicates structured as guard functions?

• pro: a simplification reducing QVTr2QVTc complexity
• pro: sibling mappings are also exposed to QVTi dispatch optimization
• con: increases stress on QVTs deep operation analysis - needs to work anyway
• con: further deviation from RelToCore

2 strong pro's, 2 weak con's => always create guard functions.

A distinct guard function may allow a failed guard to have a distinct and more optimum trace class for incremental update.

In order to avoid recomputing all the intermediate guard/bottom pattern variables the Mxdispatcher's trace class must provide properties to pass context between Gxsibling, Mxdispatcher and Mxsibling. For the degenerate case of no dispatcher the sibling's trace class should serve the same function.

Rewrite reification in QVTc, QVTu, QVTm, QVTs

There should be no change, but deep operation dependency analysis will need to work for QVTs.

But, what about partitioning? Is it necessary to partition guard functions?

Rewrite reification in QVTi / run-time

Nothing necessary, but significant optimizations available.

The dispatch of one or Micromappings from a Connection currently invokes each relevant Micromapping. No change since each synthetic override has full exclusions in its own predicate. However there may be many available Micromappings each with a guard function. These can be hoisted into a dispatcher and CSE applied to dramatically reduce recomputation. Since the failed guard functions are invoked by the dispatcher, the failures do not need to be traced, incremental redispatch from the Connection will redetermine what needs to be done correctly. These optimizations apply to siblings as well as polymorphs.

If we add a QVTi syntax to support the decision tree, the optimized dispatch can work in interpreted as well as code generated execution.

The CSE will need to work on the OCL expressions and so we need to generalize CSE support from CG-only. [1]

It may be helpful to structure the guard functions in disjunctive normal form within a let-variable CSE hierarchy to facilitate partitioning.

Add QVTc Overrides

The alternative to a QVTr2QVTc synthesis is to just propagate the 'overrides' semantic till much later.

Rather than worry about how to partition functions, perhaps the Partitioner could partition the Gx() of each Mx() without ever needing to create the artificial Gx(). We have the flexibility to introduce QVTs/QVTi constructs to directly support the reified semantic rather than contort into standard QVTc whose idioms must be dismantled later.

Both approaches have the same semantic to resolve:

'QVTr2QVTc rewrite' adds unpleasant complexity to the QVTr2QVTc to express predicates overtly as functions so that another layer of functions defines the overrides. QVTs then has to struggle to make sense of the functions.

'Overrides partitioning' just passes 'overrides' through QVTc, then uses the Partitioner to perform the rewrites where rewrites are already performed. The predicates are partitioned as preliminary micromappings and not-success dependencies lock-out the unwanted executions.

In both cases QVTi/Connection synthesis optimizes the treatment of siblings. For the 'QVTr2QVTc rewrite' idiomatic functions need recognition, for the 'partitioner' shared state is possible.

QVTr2QVTc synthesis

The QVTc synthesis is still not easy.

For a polymorphic invocation of the Mxx with a Txx trace, we require a calling signature Sxx containing the RelationCallExp arguments and a Txx for each polymorph. Each Txx is potentially independent since it may have properties with distinct types. Sxx is similarly independent.

The caller creates an Sxx which non-top Txxes depend on. The QVTs partitioner reifies the inter Txx dependencies.

This is fairly easy for a polymorphic where call, since the realized caller can realize an Sxx in addition to realizing its own trace. It might be beneficial to aggregate each Sxx into the caller's trace. Perhaps slightly awkward/bloated for multiple calls of the same relation; but probably an implementation detail that will become clear.

This is not easy for a polymorphic when call, since the caller's trace has yet to be realized and no additional Sxx can be realized either. Promoting a nest of when calls to the root caller solves the nested calls, but the root caller might be a top relation, so again no one can create the Sxx to initiate the dependency. A solution is to split the the caller

MC : when (GC()) BC()

as

MC1 : when (GC1()) where{ MC2() }
MC2 : when (GC2()) BC()

MC1, GC1() verifies all the non-top-when-call predicates and realizes the Sxx for each non-top-when-call. MC2, GC2() use the Sxx created by MC1 to provide the non-top-when-call objects to correctly guard BC().

Any monomorphic or polymorphic non-top-when-call requires the invoking mapping to be two-part. For simplicity, let the first-part trace aggregate each non-top-when-call Sxx. The second-part depends on the Sxx success. An optimization for most cases: the first part trace can inherit each Sxx, more so if Sxx property names are long and so no clashing.

Any polymorphic top-when-call is a lookup so just predicates the Sxx result. A distinct Sxx result must be maintained by each polymorph for each possible top-when invocation. No. A single Sxx result for (each) root polymorphism is sufficient. The derived Sxx results can be deduced from the root. ? cycle hazard: if we wait for the full root polymorphism tree to resolve when we only needed a branch, do we introduce a deadlock hazard? Perhaps the wait for derived Sxx can accurately test for only the required branch.

Any monomorphic top-when-call is a lookup so just predicates the Txx result.

top when/where calls

top relations do not need invocation, so what does it mean to invoke them?

For top-when calls, the call is just a possibly polymorphic lookup.

Perhaps for fixed point execution a top-where call might provoke an execution on some internally created objects that would not otherwise occur. But the internally created objects must be in the input domain, so execution occurs. The option to declare that a domain uses another domain allows output into another domain to contribute to the input. top-where calls therefore seem to be a semantic error.

Signature class

An Sxx signature class instance is required for every non-top when or where call. It provides the critical predicate that communicates the RelationCallExp arguments to the non-top relation.

An Sxx signature class instance is also required for each distinct polymorphic invocation of top when calls.

There is one property per invoked Relation parameter. For when-calls the enforced domain parameters may be 'output's.

For when calls there is a success status that may participate in trivial or complex predicates. The status is three-valued

• true=success
• false=fail
• null=not-ready.

A distinct failure status is required for each polymorph so that the chain of failures can eventually enable a partially overridden relation to succeed.

What about multiple successes? If a polymorphic when-call invokes multiple non-orthogonal overrides, there can be multiple results! The transformation is a statement of final truth, each of the multiple results is a truth, so we must treat a polymorphic when-call as an iteration over the potentially many results. This requires a distinct

Connection dispatching

Must accommodate an iteration over multiple polymorphic results.