In section 11.5 "Operations and Well-formedness Rules" for the five standard PrimitiveTypes Boolean, Integer, Real, String, and UnlimitedNatural, we have specifications for logical comparison operators <, >, <=, >= but no specification for the equality operator.

It seems to me that if these operators must be explicitly implemented by tools, they must also implement a method for equality which should be specified. In C++, this is of course necessary. The operation String::equalsIgnoreCase(), for example, depends on it directly:

"Queries whether s and self are equivalent under case-insensitive collation.
post: result = (self.toUpperCase() = s.toUpperCase())"

Issue 12948 "Making OclAny denote any object" was resolved to allow Collection to conform to OclAny.

A.2.6 still states "A simple set inclusion semantics for subtype relationships as described in the next sub section would not be possible due to cyclic domain definitions if OclAny
were the supertype of Set(OclAny)."

A premise for Annex A has therefore been violated requiring rework to consider at least a less simple set inclusion semantics.

According to section 9.3.49, there are two forms of comments, line comments (-- this is a comment) and paragraph comments (/* this is a comment */):

It is possible to include comments anywhere in a text composed according to the above concrete syntax. There will be no
mapping of any comments to the abstract syntax. Comments are simply skipped when the text is being parsed. There are
two forms of comments, a line comment, and a paragraph comment. The line comment starts with the string ‘--’ and ends
with the next newline. The paragraph comment starts with the string ‘/’ and ends with the string ‘/.’ Paragraph
comments may be nested.

In section 7.4.12, only the syntax for line comments is described:

Comments in OCL are written following two successive dashes (minus signs). Everything immediately following the two
dashes up to and including the end of line is part of the comment.

For example:
– this is a comment

The two possibilities should be mentioned in 7.4.12.

Fig 10.2 presents some classes that to the best of my knowledge have never been used in an OCL implementation. The LocalSnapshot is a possible way of reconciling the need for two distinct system states when evaluating @pre or oclIsNew. It is however perhaps no more than some draft thoughts. Some much more promising work has evolved as 'Filmstrip's in conjunction with the USE tool. A Filmstrip supports OCL expressions that apply across time.

The LocalSnapshot and associated support are absolutely not a normative part of an OCL implementation.

OCL25-80 comment 1 identifies that the xxxValue classes are misleading and could be trivially rephrased as "xxx value semantics".

Similarly the xxxEval classes are unhelpful and could be triially rephrased as "xxx evaluation semantics".

Simple or efficient implementations of OCL have no need to reify each evaluation as an instance, the execution can just return the appropriate result. Only an implementation with rigorous tracing for debugging or incremental execution need reify the execution and the implementation should be free to choose its own implementation details.

There is absolutely no need for a normative existence of xxxEval classes. Only the semantics is normative.

Figure 10.5 StringValue only appears here. No explanation was given before the diagram. Add definition or explanation of StringValue before this diagram

Special Types violate UML Generalization Semantics The definition of OclAny as a general of all classes in the UML model and of OclVoid and OclInvalid as specializations of all classes in the UML model are in violation of the semantics of generalization. Classifiers in the UML may only specialize other classifiers of the same or a conformant metaclass. From the description of Classifier in the UML: [3] A classifier may only specialize classifiers of a valid type. self.parents()->forAll(c | self.maySpecializeType(c)) [8] The query maySpecializeType() determines whether this classifier may have a generalization relationship to classifiers of the specified type. By default a classifier may specialize classifiers of the same or a more general type. It is intended to be redefined by classifiers that have different specialization constraints. Thus, it is not valid for OclAny (an instance of the AnyType metaclass) to be the general of any other class. Likewise, OclVoid and OclInvalid may not specialize any other class at all. This could be corrected in OclVoid and OclInvalid by redefining the maySpecializeType() query. For OclAny, the solution is not so straight-forward, as I don't see that OCL can redefine maySpecializeType() on behalf of the UML metaclasses; according to Section 7.4.4, it is not permitted to attempt to define an additional operation that clashes with an intrinsic operation of the context classifier.

7.4.6 states: If the actual type of the object, at evaluation time, is not
a subtype of the type to which it is re-typed, then the result of oclAsType is invalid.

7.5.8 states: self.oclAsType(A).p1 – accesses the p1 property defined in A - this is incompatible with the no-class-change-to-A so A::p1 is not available for null.

But the normative 11.3.2 states: oclAsType(type : Classifier) : T
Evaluates to self.

The OCL 2.4 contribution to 11.3.2 was a typo. It should be "evaluates to invalid".

Section 8.3.8 neglects to specify what it means for allInstances to be invoked for a Classifier that does not have a finite number of instances.

The Real/Integer/UnlimitedNatural/String PrimitiveTypes, CollectionType, TupleType, DataType, ,,, overloads could reasonably return invalid or an empty Set. Just need the specification to say which by specifying each of the overloads in Section 8 / 11.

Overloading Classifier::allInstances() with a crash (invalid) seems unkind, so suggest the empty set to avoid users needing to exclude many metatypes in a many-classifiers loop..

The flawed resolution of issue 8937 introduced Section 7.5.10 in which it is suggested that :: is an operator that can be contrasted with the . object navigation operator.

It isn't. It is a name qualification operator allowing e.g. A::B::C::d to clearly locate a feature with respect to any prolematic name clashes.

If :: is an invocable operator, which of the :: in A::B::C::d are to be executed, and how?

As a qualification operator the referredOperation/referredProperty field in an OperationCallExp/PropertyCallExp is assigned at compile-time to the unambiguous resolution of the qualified reference. No extra execution for the qualification.

The flawed resolution of issue 8937 introduced Section 7.5.10 and the Employee::id() example that invokes the static Employee::uniqueID() to initialize each Employee instance with a presumably distinct unique id.

This can only work if uniqueID() returns a distinct value from each call. It is therefore not a query and not side-effect-free. It cannot be invoked from OCL.

Unless the execution semantics introduces some new magic static-time between load-time and run-time in which a side-effecting static semantics is defined and supported, this is impossible. No idea how this could be made deterministic on a multi-processor execution engine.

Rather we need to be declarative and do everything at once, rather than one at a time imperatively.

context Employee
static def: allEmployees : Sequence(Employee)
= Employee.allInstances()->asSequence()
def: id : String
= 'ID' + allEmployees->indexOf(self).toString()

The quadratic cost of indexOf can be avoided if we have maps and co-iterators:

context Employee
static def: employee2index : Map(Employee,Integer)
= Employee.allInstances()->collectByValue(key with value | key)
def: id : String
= 'ID' + employee2index->at(self).toString()

An optimized execution engine could recognize the idiom and optimize away the single use static Map.

toString() is a familiar debug facility that OCL currently provides ony for logical / numeric usage. This is unfriendly.

Better to provide toString() for all values including null and invalid, with a carefully designed formatting to ensure that a monotonic forward/reverse round trip is possible.

(OCL25-75 vaguely suggests toString()).

The Isabelle evolution for Annex A highlight two errors in the OCL 2.4 clarifications.

OclVoid::oclIsKindOf/oclIsTypeOf should return true/false rather than invalid.

? oclIsKindof true for all types except OclIsInvalid.

There seems to be some confusion in ptc-13-08-13 regarding the relation of OclAny to other OCL types. Sections 8.2 and 11
state that it is a supertype of all other OCL types, including collection and tuple types, but Annex A excludes collection/tuple types from inclusion in OclAny.

The new OCL for UML 2.5.1 seems not to have been checked by any tool; shame on you. There are two syntax errors:

OpaqueExpression-only_in_or_return_parameters-specification makes use of ParameterDirectionKind::in but 'in' is a reserved word and so it must be escaped as in other similar constraints. Change to ParameterDirectionKind::'in'

Lifeline-interaction_uses_share_lifeline-_specification has a Bag rather than Boolean argument for the final implies. Perhaps the select(...) should be followed by a ->notEmpty() or the select(...) could be an exists(...)

in UML the rules for distinguishable names are:

• Have different Names
• Same Names but different Types
  metaclasses are different
  Neither is a (direct or indirect) generalization/specialization of the other
  (e.g. Operations distinguished by signature)

so we can have a class Person and a DataType Person... each could have an association to it... so the association end implicit name is the same... and would need to be differentiated by its Type... this is not currently supported in OCL

A Collection containing both 'a', and

must be typed as a
Collection of OclAny since that is the only common type.

This is unhelpful for tree structures.

Suggest introduction of a pseudo-collection type Tree(T) to which
both T and Collection(Tree(T)) conform.

A String content tree then then be declared as a Tree(String).

12: Why no applicability to Essential MOF models?

Although EMOF::Operation has no precondition feature, there is no problem in an operationContextDecl imposing a truth upon the operation.

Similarly, if Initial value and Derived values are defined as Constraints in the same way as PreConditions, then the propertyContextDecl can constrain EMOF by imposing a truth upon the Property.default : String even though String is not an Expression

Sections 12.5.1, 12.6.1, 12.7.1 specify the "invariant", "precondition", "postcondition" spellings.

Sections 12.8, 12.9, 12.10, 12.11 do not specify their corresponding spellings.

Suggest: "initial", "derivation", "body", "guard" as the conventional adjective used to qualify "constraint".

Languages such as Groovy have found it helpful to intrioduce the safe navigation operator ?: so that navigation over null yields null rather than a problem (invalid in OCL).

In OCL it could be a pure syntax sugar:

(x?.y) = (if x == null then null else x.y endif)

or perhaps an additional PropertyCallExp operator

Issue 12953 made changes to the CollectionLiteralExpCS grammar and these appear in OCL 2.3 but not in OCL 2.4.

It would appear that some major editorial snafu occurred while reconstructing a FrameMaker base for OCL 2.4 after the OCL 2.3.1 corruptions.

Application of OCL 2.2, 2.3, 2,3.1, 2.4 changes need to be checked to understand the snafu.

Typo in sentence "If the constraint is shown in a diagram, with the proper stereotype and the dashed lines to connect it to its contextual
element, there is no need for an explicit context declaration in the test of the constraint."

test should be text

The sortedBy iteration provides an elegant solution to a sort problem in which the sort metric is a projection of the sorted object. Thus sortedBy(p|p.name) or just sortedBy(name) is short and avoids the opportunities for typos from a more conventional exposition involving comparison of two objects. The natural solution may well be an efficient one for large collections with non-trivial metrics.

However the sortedBy solution is unfamiliar and so confusing for newcomers and unsuitable for multi-key sorting for which an artificial compound single key may need to be constructed.

One solution might be to provide a more conventional iterator such as sort(p1, p2 | comparison-expression) allowing a two key sort:

sort(p1, p2 | let diff1 = p1.key1.compareTo(p2.key1) in
if diff1 <> 0 the diff 1 else p1.key2.compareTo(p2.key2) endif)

However this has poor readability and ample opportunities for typos.

Alternatively sortedBy with a Tuple-valued metric might support multiple keys as:

sortedBy(Tuple

)

(The alphabetical order of the Tuple part names determines the priority.)

(Since sortedBy is declaratively clear and compact, inefficient small/trivial implementations can be optimized to their sort() equivalents.)

The restriction to sort the OrderedSet with the lowest value coming first is very restrictive. Sometimes it is beneficial to sort in reverse order. Think about a statement that would allow a > sort order.

The following problems are common to all definitions of sortedBy in 11.9:

Invalid algorithm:

• When the current element is not the first one, but is the
  greatest one so far, for example 2 in
  Sequence {1,2}

  ->sortedBy(x|x),
  then
  result->select (item | body (item) > body (iterator))
  is empty and calling ->first on it would result in invalid.

Inconsistent use of < and >:

• The text descriptions mention using < operation on the elements,
  but the algorithm uses > operator.

Typos:

• The accumulator initialization incorrectly uses : instead of =.
• The source for the iterate expression is missing.
• The closing parenthesis for the iterate expression is missing.
• The operations append and insertAt are called by . operator.

SOLUTION:

• Introduce a new local variable upperBounds, containing elements from the
  accumulator that are greater than the current element.
• Swap body (iterator) and body (item).
• Fix typos.

Corrected algorithm for Set (11.9.2) and OrderedSet (11.9.5):

****************************************
source->sortedBy(iterator | body) =
source->iterate( iterator ; result : OrderedSet(T) = OrderedSet {} |
let upperBounds : OrderedSet(T) =
result->select (item | body (iterator) < body (item) )
in
if upperBounds->isEmpty() then
result->append(iterator)
else
let position : Integer = result->indexOf ( upperBounds->first() )
in
result->insertAt(position, iterator)
endif
)
****************************************

Corrected algorithm for Sequence (11.9.4) and Bag (11.9.3):

****************************************
source->sortedBy(iterator | body) =
source->iterate( iterator ; result : Sequence(T) = Sequence {} |
let upperBounds : Sequence(T) =
result->select (item | body (iterator) < body (item) )
in
if upperBounds->isEmpty() then
result->append(iterator)
else
let position : Integer = result->indexOf ( upperBounds->first() )
in
result->insertAt(position, iterator)
endif
)
****************************************

FURTHER SUGGESTIONS:

• The presented algorithm works only under assumption that indexOf(obj:T)
  on Sequence(T) returns the first occurence. This is not mentioned in the
  definition of indexOf. If indexOf can return any index, for example:
  Sequence {2, 2}

  ->indexOf(2) = 2,
  then the above algorithm could insert the element at the wrong place,
  for example:
  Sequence

  {2, 2, 1}

  ->sortedBy(x|x) = Sequence

  {2, 1, 2}
• The algorithm is stable for Sequence and OrderedSet, this fact could be
  mentioned in the description, to make it clear that successive
  sortedBy iterations can be used to sort by multiple criteria.

OCL supports a '*' value for UnlimitedNatural in order to accommodate the
full range of UML multiplicities.

UnlimitedNatural conforms to Integer and Real, so that any UnlimitedNatural
conversion must perform a run-time check in order to convert '*' to invalid.
This conversion cannot be replicated in the reverse direction.

Suggest that '*' be aligned with the conventional IEEE math notion of
infinity, so that

• and -* are valid values for Integer and Real. UnlimitedNatural is then a
  simple restriction of Integer which is a simple restriction of Real.

Consider: my general purpose library provides:

mylib::Aggregate<T>

My subsystems exploit this by drawing:

mysub1::Aggregate<T -> String>
mysub2::Aggregate<T -> String>

and my application integrates the subsystems by drawing

myapp::Aggregate<T -> String>

Since each of the three Aggregates of String have different packages, oops, it would seem that they are different types and so cannot be passed interchangeably between subsystems.

The above arises from the natural encoding of a graphical representation in which Aggregate<String> is drawn once per diagram/package and referenced many times. The package owns and so appropriates the shared type.

In contrast, textual representations, in the absence of a typedef, elaborate Aggregate<String> for each reference. Each reference 'owns' the shared type, which is not appropriated by a prevailing package.

(This has been a major problem for attempts to provide UML-aligned XMI serialization for OCL since e.g. there is obviously only one Set(String) but who owns it, who persists it, and how is it referenced? Eclipse OCL has prototyped a solution whereby a shared orphanage package is responsible for owning the singletons, but everyone persists their singletons. This incurs unpleasant costs in creating local orphanages during a serialization and destroying them again during loading. Struggling with this finally made me see that there is a fundamental UML limitation.)

In similar scenarios where there is an ownership/reference dilemma, UML provides e.g. TemplateParameter::parameteredElement/ownedParameteredElement. Surely there should be a TypedElement::type/ownedType so that a TypedElement::ownedType can introduce a package-less type that is consequently shared by all similar introductions? Bloat could be avoided by one TypedElement::type referencing another TypedElement::ownedType. Alternatively a Package::sharedElements could contain local copies that are logically shared globally. (This could also solve the problem of everyone wanting their own Boolean PrimitiveType.)

The pseudocode for OrderedSet::insertAt assumes that the inserted element is not already contained.

Presumably insertion of an existing element does not increase the OrderedSet size.

Design choice. Is the old element retained or is the old element removed and the new element inserted? The latter might preserve the postcondition that the element at the index is the inserted object, but what if the removed element was earlier so that everything moved along first?

Surely the only predictable behavior is that the old element is retained?

Although it is clear that using a constructor to write an object to the system being modelled breaks the property of being side-effect free, it would be useful to return new objects as a result of the query functionality of OCL. It is possible to create a new Tuple for return from a function. For example: def: newDate1() : TupleType

= Tuple

; Rather than write this object to the model under query, it would only be returned as a query result so, under these circumstances, would not break the property of being side-effect free. This feature would be extremely useful to the HL7 GELLO project, which works with data models defined in absense of defined methods or constructors. If this feature were to only apply to classes marked in some way to guarantee they have no side-effects from construction then that would remain useful.

"asSequence" and "asOrderedSet" are specified as returning an unknown ordering. How deterministic is this? repeatable with the same tool version, the same tool, the same CPU, the same number of CPUs, any OCL tool?

"any" is specified as indeterminate.

Non-determinism is a very bad language characteristic.

Surely "any" should be the earliest element in asSequence()?

If a tool avoids anarchic sets, perhaps keeping a list and set for each 'set', the resulting behavior can be deterministic. But is specifying deterministic set evaluation algorithms appropriate for the OCL specification?

Suggest the specification should just require that a re-execution with:

• the same tool version
• the same Operating System etc
• the same single CPU
• the same command line options
  shall give the same result (unknown but deterministic).

OCL is an executable specification language that uses "invalid" as its exception handling mechanism for untoward execution, which clearly includes deterministic internal program control faults:

• divide by zero
• null-navigation
• ordered collection index out of bounds
  Does this behavior also include non-deterministic external problems:
• network/disk access failure
• stack overflow
• tooling malfunction

Without an answer to this question, it i impossible to know whether e.g. "aSequence->includes->last()" can be optimized to "x" or not. Is it necessary to fully evaluate aSequence, perhaps run out of memory, in order to incur an external issue that influences an evaluation?

The % (modulo/remainder) operator is realtively standard operator in many langages. It is missing from the OCL Standard Library. There is only Integer::mod().

QVTo 1,3 8.4.4 4 implies that there is a % modulo operator even though there isn't. See QVT14-44.

Suggest add Real::%

UML/OCL hardwire three capabilities of an Operation.

"body" is very useful
"precondition", "postcondition" are sometimes useful
anything else is unsupported.

Some users want frame conditions.

Code generation/analysis wants other forms of meta-evaluation.

Rather than expand the hardwired options, allow an Operation to have an extensible set of named capabilities.Some names such as "body", "precondition", "postcondition" can be standardized. Others could be reserved in anticipation.

OCL supports feature overloading and attempts to comply with UML. OCL does not define feature overloading itself. UML leaves feature overloading as a semantic variation point.

The OCL behaviour is therefore undefined. However a variety of behaviours particularly those involving null and invalid only make sense if operations are selected at run-time according to the dynamic type.

The example from an earilier pending issue:

Sequence(OclAny)

>count((-3)) = 2

wherein

UnlimitedNatural 3 widens to Integer 3 for successful comparison with Integer 3
Real 3.0 is successfully compared to Integer 3 widened to Real
String '3' widens to OclAny and fails to compare to Integer 3 widened to OclAny.

is difficult to realise if static type resolution is applicable in which case all pair-wise value comparisons would use OclAny::= rather than Real::=.

Suggest:

OCL define that features are selected dynamically at run-time.

A UML InstanceSpecification introduces classification within the M1 level. OCL is normally applicable to inter level typing. M0/M1 or M1/M2.

Does anInstanceSpecification.oclIsKindOf(XXX) use a same-level or meta XXX?

Does anInstanceSpecification.YYY use the YYY feature of the InstanceSpecification metaclass or any one of the anInstanceSpecification.classifier?

Is a new navigation operator needed to distinguish the two cases?

Practical evaluation of stereotypes requires OCL to reify the Stereotype instances. These instances should be accessible by Stereotype.allInstances().

e.g. if an ElementExtension is the Stereotype instance and Element.extensions are the applied stereotypes.

Stereotype.allInstances() is equivalent to

Element.allInstances().extensions->selectByKind(MyStereotype)

The last sentence of the first paragraph on page 37 says "All metaclasses defined as part of the OCL abstract syntax are shown with a light gray background.", but subsequent abstract syntaxs show OLC defined metaclasses in yellow

asOrderedSet() : OrderedSet(T)
Redefines the Collection operation. An OrderedSet that contains all the elements from self, in undefined order.

If asOrderedSet() is invoked more than once on the same Set, then presumably the subsequent calls have a defined order; the order previously used.

When used by QVTo the lack of determinism of Set ordering is a significant debugging nuisance. Suggest that there is a repeatable but unspecified underlying order determined perhaps by object creation order.

object.assoc[qualifier] returns a particular element from a qualified association.

With Map already suggested/prototyped for OCL, let object.assoc[] return a Map<qualifier,element> that avoids the loss of information from object.assoc that returns Collection<element>

null-in-Collections was added by the OCL 2.0 FTF in Issue 5972 in the unanalyzed hope that it would be useful.

https://ocl2015.lri.fr/OCL_2015_paper_0921_1400.pdf examines the null safety of OCL nad concludes that an ability to declare null-free collections is essential.

Since almost all collections in practical OCL are null-free, suggest that the OCL default should be for null-free collections, with an option to have null-full collections when really needed.

Qualifiers, written in brackets after the path name of a feature call,
can express two different things.

• qualifying use: A qualifier is used to give the qualifing value of
  a qualified association (chapter 7.5.7).
• navigational use: A qualifier is used to refine the navigation to
  association classes. While this navigational use is necessary
  only with recursive associations, it is legal for every navigation
  to an association class (chapter 7.5.5).

There is no way to distinguish these two sorts of qualifiers. There are
even expressions where both uses of the qualifiers would be necessary at
once, but this problem is restricted to such models that contain a
recursive, qualified association that has an association class.

Example where navigational and qualifing use cannot be distinguished:

There are two classes "Bank" and "Person", with a association between
them qualified by the account number (an Integer). The association end
at the class Person is named "customers".
An additional class "Company" has an attribute "customers" of type
Integer.

Now consider the subexpression "bank.person[customers]" in the context
of Company. "bank" clearly is a navigational expression. But "customers"
could either mean the attribute of Company, since Company is the context
of the expression (that is qualifying use as defined in 7.5.7); or
"customer" could mean the name of the association end (navigational use
as defined in 7.5.5). In the first case, the result type would be
Person, in the second case Set(Person).

37. – [2] The parameters of the referredOperation become attributes of the instance
– of OclMessageType context OclMessageType
inv: referredOperation->size() = 1 implies
self.feature = referredOperation.Parameter->collect(p |
p.asAttribute().oclAsType(Feature) ).asOrderedSet()
==> should be:
context OclMessageType
inv: referredOperation->size() = 1 implies
self.feature = referredOperation.Parameter->collect(p | p.asAttribute()
).asOrderedSet()

Issue 15836 suggests that a negative CollectionRanbge should be invalid even after considering the Sequence

->forAll idiom.

The idiom clearly requires that a negative range gives an empty range.

Description: At some places, template parameter T appears in operation signatures, e.g., oclAsType(typename:OclType) : T (e.g., Sect. 6.2.1). At other places, this is denoted by "instance of OclType" or <<the return type of the invoked operation>>. It would be more meaningful when these informal return type descriptions are replaced by "OclAny". An additional constraint about the actual return type should be given when necessary.

OCL 2 doesn't really define what a collection is. In essence,
a particular UML construct is arbitrarily designated as
the OCL collection, and a series of properties are assigned
to it

This question arises in 2 different ways:

• can you sub-class one of the concrete descendents in
  a UML diagram - by referring to the OCL package - and thereby
  add functionality to your own collection types

• are all parameterised classifiers collections? if you
  define a parameterised class, how does an OCL user or
  environment know whether it's a collection or not?

perhaps a stereotype should be intrroduced to allow for
unambiguous resolution of these issues

Synthesized attributes, [D], [F], [G] The meaning of “and” at the end of the line is not clear and seems inconsistent with other parts. Are these logical operations? Add text to clarify this

Issue 12953 considered the problem of the element type of an empty collection and rejected OclVoid (Option 1) because Set{}->including(1) would fail.

This rejection is appropriate when the collection is modified as in Java, but is inappropriate for OCL where a new Collection is created. There is therefore no problem in deciding that the most derived common element type of Set{} (OclVoid) and 1 (Integer) is Integer causing Set(Integer)::including(Integer) to be invoked.