Instead of copying cached data just re-plant the fields to the new cache.
The new types-model to contain only the newly reported types. Types reported on earlier invocations are retained
in previously created models.
Refactor whe way TypecheckingResultsCache is updated:
all mutatiting methods extracted to Update subclass, which is
only available to the updater.
Make TypecheckingState a regular container for (unit) caches.
Make TypecheckingService responsible for managing caches.
Extract TypeModuleOwner and TypeModelHandler classes.
Get rid of QueryResultsCache, merge into TypecheckingResultsCache.
TypeModuleOwner to manage "repository" cache.
When running typechecking with trace=on too much memory is taken up
by the traces, the new option enables to see the evaluation results
without collecting the traces.
Optimization: Use repo/unit caches for storing various data in Caching strategy
Fix: introduce workaround for nodes coming from generator's transient model --
avoid clearing cache for non-resolvable origin nodes.
Fix: incrProgSpec not passed correctly downstream, incremental mode working only with trace on.
Refactoring: make TypesCollector a (private) singleton. TBD.
Rename SessionData -> CachedData.
Make use of MutableLogical's feature to track usages count.
A logical with 1 or less usages can't be contained in
another term, so it's safe to skip unification in this case.
During template application memo can be used to store results
of a computation in the session. The memo will be automatically
cleared on an event that invalidates the caches.
IncrProgramSpec influencing program eval is built from two sources:
internal analysis results & typechecking options.
Providing spec builder instead of spec allows to combine these
sources later at program pipeline creation in TypecheckingService.
Commit accomodates to addition of IncrementalSpec.Enabled.Unsafe
Commit also simplifies ProgramAnalysis -related classes.
Renames BasicProgramAnalysis to PrincipalsAnalysis.
The flag "recursive" is passed down to ProgramFactory to control whether
the input nodes are to be walked recursively.
This flag is set to false when calculating a node's type in isolation.
TypecheckingCache maintains two maps of root-to-cache: tracking and non-tracking.
Tracking cache is used in "watching" mode, caching mode doesn't require tracking.
Always merge the results of typechecking, avoid replacing the cached data.
Reimplement calcTypeLocally() to rely on cache for storing the resulting type.
Simplify cache API.
Fix TypecheckingTestHelper.
TypecheckingService has become too overloaded with responsibilities, and
so Strategy class is introduced to extract strategies to.
Cache holder can also be non-tracking, which comes in handy when
implementing purely caching typechecking, not anticipating any change.
TypecheckingService and provider component are explicitly dependent
on SRepository.
The repository instance comes from the first opened project, which
is a dirty hack and is temporary until Typechecking Facade API
is fixed to provide appropriate instance.
Operations such as 'error' and 'warning' are available on input
node in templates and serve to report issues found during
template application, in contrast with those found during evaluation.
Drop recording of "suppressed exception" by CoderulesHelper.
Report failed program evaluation as info message.
Ensure traces are available for failed evaluations.
Fixes IncrClosureInference.changeFunctionalInterfaceParam test.
Temporary rules built as macro applications weren't computing hash
due to origin!=null condition. Also macro inputs weren't processed.
Stable hash for TypeVarSubstitution is defined because it's used in macro args.
In general, any classes used as constraint/predicate/macro args
must adhere to stable hashCode() contract: hash should change
if there're important changes that may influence typechecking result.
Otherwise there may be unnecessary incremental recomputations.
Relevant for MPSCR-62
DataFormPrinter is hardcoded to support _term and _subterm
hacks and present the terms in a readable way.
Memoizer is searched for an implementation of
NodePresentable, which helps with visualization of SNode
and SNodeReference.
Term constructor creates all features with default values where
not overridden. The default values for "value" and "child"
features are wildcards for now (TBD), the default value for
list feature is an empty list.
The rationale is that a rule's head defining term pattern with
a logical variable capturing list feature will not be matched
with a free logical variable. Thus, to avoid having to always
explicitly specify all term features, a term constructor that
is not a pattern must use the default values for non-overridden
features.
Match constraint is provided for use cases where the intention
is to pattern-match a term capturing features into logical
variables.
A term constructor in the right-hand side of match constraint
is a pattern, and all its non-overridden features are
to be defined as wildcards.
It's possible to activate constraints in macros, so they must be also handled.
Also restrict a bit principal rules analysis: consider only rules with template input.
It makes definition of principal constraint clearer:
principal constraint is a constraint that has at least one arg
of type node<> and passes to it ruleOrigin in at least one rule.
Such constraints are most likely meant to define relations
of nodes to terms, e.g. `typeOf` -- which is what we're interested in.
At the same time semantics of values inside terms is less defined.
TermFeatures are also untyped, and so complete analysis of
ValueFeatures would require full code traversal of coderules program
to find out whether a value feature is used to carry smth of `node` type.
Relevant to MPSCR-62.
Motivation is that implicit (non-overriden) list features must unify
with list of any length. Consider an example:
`classifier(cls: node-ptr/A/) = classifier(cls: node-ptr/A/ param: [B,C])`
With previous approach it would fail, which seems counter-intuitive to me.
So, if empty list (e.g. of type parameters) is meant, then it must be
specified explicitly.
There's one thing to remember, though: such wildcard list features
won't match with MetaLogicalArrays in rule heads.
(e.g. see dpromote_* rules and consider what happens when
a classifier with unbound `parameter` feature comes in)
Avoides the need to exactly duplicate features of ancestors.
Opens possibility to match on super-terms
with partially provided (overriden) features.
For example, suppose there's `typeParameterized(child: param)`
term which is extended by `classifier` and `methodSignature`.
Then it becomes possible to write a common rule for both terms,
matching on `typeParameterized(param=[Smth])`.
Treat a language and its sidekick (if any) as a single vertex.
Add all extended languages to the pool before building the graph.
Drop dubious invariant check.
Fix the second pass in findAspectCliques ingnoring the fact
that there might be leafs without aspects.
Fedor Isakov