D45 Code Generation

Overview

The code generation plug-in allows the generation of code for typical real-time embedded control software from refined Event-B models. Such feature would be an important factor in ensuring eventual deployment of the DEPLOY approach within industrial organisations.

Motivations

DEPLOY industrial partners are interested in the formal development of multi-tasking, embedded control systems. During the project, work has been undertaken to investigate automatic generation, from Event-B models, for these type of systems. Initially, we chose to translate to the Ada programming language, and use it as a basis for the abstractions used in our approach. The first version of the code generator supported translation to Ada, and the current version also has limited support for C.
We described our previous code generation feature as a demonstrator tool; chiefly a tool designed as a proof of concept, used by us to validate the approach. In this sense, the tool as it stands now, is the first prototype intended to be used by developers.

Choices / Decisions

We described our previous code generation feature as a demonstrator tool; chiefly a tool designed as a proof of concept, used by us to validate the approach. In this sense, the tool as it stands now, is the first prototype intended to be used by developers. However, we can use the demonstrator as a baseline, and describe the new features as follows:

• Tasking Event-B is now integrated with the Event-B explorer. It uses the extensibility mechanism of Event-B EMF (In the previous version it was a separate model).
• Tasking Event-B is now integrated with the Event-B model editors. Tasking Event-B features can now be edited in the same place as the other Event-B features.
• We have the ability to translate to Java and C, in addition to Ada source code; and the source code is placed in appropriate files within the project.
• We use theories to define translations of the Event-B mathematical language (Theories for Ada and C are supplied).
• We use the theory plug-in as a mechanism for defining new data types , and the translations to target data types.
• The Tasking Event-B to Event-B translator is fully integrated. The previous tool generated a copy of the project, but this is no longer the case.
• The translator is extensible.
• The composed machine component is used to store event 'synchronizations'.
• Minimal use is made of the EMF tree editor in Rose.

Tasking Event-B and Editing

A text-based task body editor was added, to minimize the amount of editing required with the EMF tree editor. The task body editor is associated with a parser-builder; after the text is entered in the editor the EMF representation is generated (by clicking a button) that is, assuming parsing is successful. If the parser detects an error, information about the parse error is displayed in an adjoining text box. When specifying events in the task body, there is no longer a need to specify two events involved in a synchronization. The code generator automatically finds the corresponding event of a synchronization, based on the event name, and using the composed machine component. Composed machines are used to store event 'synchronizations', and are generated automatically during the decomposition process. This reduces the amount of typing in the TaskBody editor, since we no longer need to specify both local and remote (synchronizing) events. The new feature also overcomes the 'problem' that we previously experienced, with duplicate event names in a development, and event selection, when specifying the task body. The EMF tree editor in Rose is now only used minimally; to add annotations for Tasking, Shared and Environ Machines; typing annotations, and parameter direction information; and sensing/actuating annotations, where necessary. Further work is under way to integrate the code generation feature with the new Rodin editor.

Allowing Extensibility

The code generation approach is now extensible; in that, new target language constructs can be added using the Eclipse extension mechanism. The translation of target's mathematical language is now specified in the theory plug-in. This improves clarity since the the translation from source to target is achieved by specifying pattern matching rules. The theory plug-in is used to specify new data-types, and how they are implemented. Translated code is deposited in a directory in the appropriate files. An Ada project file is generated for use with AdaCore's GPS workbench. Eventually this could be enabled/disabled in a preferences dialog box. The Tasking Event-B to Event-B translator is now properly integrated. Control variable updates to the Event-B model are made in a similar way to the equivalent updates in the state-machine plug-in. The additional elements are added to the Event-B model and marked as 'generated'. This prevents users from manually modifying them, and allows them to be removed through a menu choice.

Targeting new Languages

Making the step from Event-B to code is a process that can be aided through automatic code generation. The code generation plug-in for Rodin is a new tool for translating Event-B models to concurrent programmes. However users of such a tool will likely require a diverse range of target languages and target platforms, for which we do not currently provide translations.Some of these languages may be subtly different to existing languages and only have modest differences between the translation rules, for example C and C++, whilst others may have more fundamental differences. As the translation from Event-B to executable code is non-trivial and to reduce the likelihood of error, we want to generalize as much of the translation as possible so that existing translation rules are re-used. Therefore significant effort is needed to ensure that such a translation tool is extensible to allow additional languages to be included with relative ease. Here we concentrate on translation from a previously defined Common Language Meta-model, the intermediary language IL1, which Event-B translates to directly. IL1 is an EMF metamodel representation of generic properties and functionality found in many programming languages. It has representations for key structural concepts such as variables, subroutines, function calls and parameters. The translation of predicates and expressions contained within the code are handled by a new extension to the theory plug-in, which allows translation rules to be developed for specific target languages within the Rodin environment. The generic nature of the intermediary language is designed to allow for a wide range of different target languages. Developers of new target languages are required to write translators in Java for the conversion from the EMF representation to the code of their target language. To do this we provide a central translation manager, that takes an IL1 model and automatically calls the appropriate translators for each element of the model, whilst also providing the link to the predicate and expression translators provided by the new theory plug-in. The developer registers their translators for the target language through an extension point, where currently there are 15 light-weight translators required for a new target language. To aid the developer, we provide abstract translators for each required element in the IL1 model that has to be translated. These translators perform the majority of the translation automatically, meaning that in most cases all the developer is required to do is format strings into the appropriate structure for their target language. For example in an branch statement, the developer would be required to write a method stating how a branch is defined and structured in their language, using a set of previously translated guard conditions and actions. Importantly, the flexibility remains for the developer to re-write any of the translations if the ones provided are not suitable.

Using Abstract Translators

To test our approach, we have built translators for C, Ada and Java using the same underlying abstract translators. Additionally we consider the case where a new language may be required that has only modest differences to an existing language. A good example of this is to consider the case where a different library may want to be used from one used in an existing translation. For instance in C, concurrency can be achieved through different mechanisms such as OpenMP or Pthreads. In this case it may be that all but the mechanism for handling a subroutine call are the same, meaning that the majority of the translation can occur using common translators, with separate translators for the different methods of handling a subroutine call. To allow for this we allow the developer to assign a core and specialisation language to each translator they build. In cases where a translator for the specialisation language does not exist, the translator will automatically defer to the default core language translator, if one exists. This means that default translators for a particular core language can be written for the majority of the translation, with specialisations being provided where differences occur. The core and specialisation of the language is also reflected in the theory translator, meaning that language theories are only required for the core languages, rather than for each individual specialization.

Available Documentation

TODO By Andy

Status

We released a new version of the code generator on 22-03-2012. We have made changes to the methodology, user interface, and tooling. The code generators have been completely re-written. The translators are now implemented using Java only. In our previous work we attempted to make use of the latest model-to-model transformation technology, available in the Epsilon tool set, but we decided to revert to Java since Epsilon lacked the debugging and productivity features of the Eclipse Java editor. We have also updated the documentation, including the Tasking Event-B Overview, and Tutorial materials.