D32 Code generation

General Overview

The code generation activity has been undertaken at the University of Southampton. This has been a new line of work for DEPLOY that was not identified in the original Description of Work for the project. The development of the approach, and the tools to support, it involved a number of team members at Southampton; and also at other institutions. This work draws on our recent experience with technologies such as Shared Event Decomposition , and the EMF Framework for Event-B . There was collaboration at an early stage with Newcastle University, where we explored the commonalities between their flow plug-in and the flow control structures used in our approach. Collaboration with the University of York was also established since we chose to use their Epsilon model-to-model transformation technology.

Motivations

The decision was taken in 2009 to include code generation as a project goal . It had been recognised that support for generation of code from refined Event-B models would be an important factor in ensuring eventual deployment of the DEPLOY approach within their organisations. This was especially true for Bosch and Space Systems Finland (SSF). After receiving more detailed requirements from Bosch and SSF, it became clear we should focus our efforts on supporting the generation of code for typical real-time embedded control software.

Choices / Decisions

Strategic Overview

During the last year we have focussed on supporting the generation of code for typical real-time embedded control software. To this end we have evolved a multi-tasking approach which is conceptually similar to that of the Ada tasking model. Tasks are modelled by an extension to Event-B, called Tasking Machines. Tasking Machines are an extension of the existing Event-B Machine component. In implementations such as Ada, tasks share the resources and have mutually exclusive access to shared state through the use of a protection mechanism. An Event-B machine can also be viewed as an abstraction of a shared resource, and the mechanism protecting it. We use existing Event-B machines with minimal extensions (called Shared Machines) to represent shared resources.

For real-time control, periodic and one-shot activation is currently supported; and it is planned to support triggered tasks in the near future. Tasks have priorities to ensure appropriate responsiveness of the control software. For the DEPLOY project, it was regarded as sufficient to support construction of programs with a fixed number of tasks and a fixed number of shared variables – no dynamic creation of processes or objects has been accommodated.

Our main goal this year has been to devise an approach for, and provide tool support for, code generation (initially to Ada). In accord with the resources available during the year it was decided to limit the provision of tool support to that of a demonstrator tool. The tool is a proof-of-concept only, and lacks the productivity enhancements expected in a more mature tool. Nevertheless much insight has been gained in undertaking this work; it lays a foundation for future research, and will be useful since it will allow interested parties to explore the approach.

The Tasking Extension for Event-B

The following text can be read in conjunction with the slides from the Deploy Plenary Meeting - Zurich 2010.

Tasking Event-B can be viewed as an extension of the existing Event-B language. We use the existing approaches of refinement and decomposition to structure a development that is suitable for construction of a Tasking Development. At some point during the modelling phase parameters may have to be introduced to facilitate decomposition. This constitutes a natural part of the refinement process as it moves towards decomposition and on to the implementation level. During decomposition parameters form part of the interface that enables event synchronization. We make use of this interface and add information (see #Implementing Events) to facilitate code generation.

A Tasking Development is generated programmatically, at the direction of the user; the Tasking Development consists of a number of machines (and perhaps associated contexts). In our approach we make use of the Event-B EMF extension mechanism which allows addition of new constructs to a model. The tasking extension consists of the constructs in the following table.

Construct	Options
Machine Type	DeclaredTask, AutoTask, SharedMachine
Control	Sequence, Loop, Branch, EventSynch
Task Type	Periodic(n), Triggered, Repeating, OneShot
Priority	-
Event Type	Branch, Loop, ProcedureDef, ProcedureSynch
Parameter Type	ActualIn, ActualOut, FormalIn, FormalOut

The machines in the Tasking Development are extended with the constructs shown in the table, and may be viewed as keywords in a textual representation of the language. With extensions added, a Tasking Development can be translated to a common language model for mapping to implementation source code. There is also a translator that constructs new machines/contexts modelling the implementation, and these should refine/extend the existing elements of the Event-B project.

Tasking Machines

The following constructs relate only to Tasking Machines, and provide implementation details. Timing of periodic tasks is not modelled formally. Tasking Machines are related to the concept of an Ada task. These can be implemented in Ada using tasks, in C using the pthread library C, or in Java using threads.

• Tasking Machines may be characterised by the following types:
  • AutoTasks - Singleton Tasks.
  • Declared tasks - (Not currently used) A task template relating to an Ada tasktype declaration.
  • TaskType - Defines the scheduling, cycle and lifetime of a task. i.e. one-shot periodic or triggered.
  • Priority - An integer value is supplied, the task with the highest value priority takes precedence when being scheduled.

Shared Machines

A Shared Machine corresponds to the concept of a protected resource, such as a monitor. They may be implemented in Ada as a Protected Object, in C using mutex locking, or in Java as a monitor.

• Applied to the Shared Machine we have:
  • A SharedMachine keyword that identifies a machine as a Shared Machine.

Tasks and Events

Control Constructs

Each Tasking Machine has a task body which contains the flow control, or algorithmic, constructs.

• We have the following constructs available in the Tasking Machine body:
  • Sequence - for imposing an order on events.
  • Branch - choice between a number of mutually exclusive events.
  • Loop - event repetition while it's guard remains true.
  • Event Synchronisation - synchronization between an event in a Tasking Machine and an event in a Shared Machine. Synchronization corresponds to an subroutine call with atomic (with respect to an external viewer) updates. The updates in the protected resource are implemented by a procedure call to a protected object, and tasks do not share state. The synchronization construct also provides the means to specify parameter passing, both in and out of the task.
  • Event wrappers - The event synchronization construct is contained in an event wrapper. The wrapper may also contain a single event (we re-use the synchronization construct, but do not use it for synchronizing). The event may belong to the Tasking Machine, or to a Shared Machine that is visible to the task. Single events in a wrapper correspond to a subroutine call in an implementation.

Implementing Events

An event's role in the implementation is identified using the following extensions which are added to the event. Events used in task bodies are 'references' that make use of existing event definitions from the abstract development. The events are extended. to assist with translation, with a keyword indicating their role in the implementation.

• Event implementation.
  • Branch - In essence a task's event is split in the implementation; guards are mapped to branch conditions and actions are mapped to the branch body. If the branch refers to a Shared Machine event (procedureDef) then this is mapped to a simple procedure call.
  • Loop - The task's event guard maps to the loop condition and actions to to loop body. If the loop refers to a Shared Machine event then it is mapped to a simple procedure call.
  • ProcedureSych - This usually indicates to the translator that the event maps to a subroutine, but an event in a task may not require a subroutine implementation if its role is simply to provide parameters for a procedure call.
  • ProcedureDef - Identifies an event that maps to a (potentially blocking) subroutine definition. Event guards are implemented as a conditional wait; in Ada this is an entry barrier, and in C may use a pthread condition variable .

In an implementation, when an subroutine is defined, its formal parameters are replaced by actual parameter values at run-time. To assist the code generator we extend the Event-B parameters. We identify formal and actual parameters in the implementation, and add the following keywords to the event parameters, as follows:

• Event parameter types
  • FormalIn or FormalOut - event parameters are extended with the ParameterType construct. Extension with formal parameters indicates a mapping to formal parameters in the implementation.
  • ActualIn or ActualOut - Extension with an actual parameter indicates a mapping to an actual parameter in the implementation.

Other Technical Issues

Translation Technology

In order to provide a structured extensible code generation tool it was decided to use a multi-stage translation approach. The Event-B EMF model provided by the Event-B EMF Framework is extended to accommodate the tasking constructs as described above. The Tasking model is then translated to an intermediate model, the Common Language Model. The Common Language Meta-model is an abstraction of some useful generic programming constructs such as sequence, loop, branch and subroutine call/definition and so on. The translation of the Common Language Model to programme source code is then a relatively small step. The main translation activity takes place in the step between Tasking and Common Language models.

The decision was made to use Epsilon to facilitate model to model translation for this stage. It was felt that an extensible, easily maintainable solution was required for this. Various model-to-model technologies (Java code, ATL, Epsilon) were appraised and it was judged that the Epsilon tool best matched our requirements. It proved to be a good choice initially for the specification of translations, especially in simpler areas of the project where the correspondence between models were simple. However the lack of debugging facilities, and productivity enhancements that are found in more mature tools, somewhat hindered rapid development as the project increased in complexity.

Implementation - Source Code

Early in the current phase of work we identified the possibility of translating the Common Language Model to EMF models of programming languages such as Ada and C, in addition to producing textual source. While the EMF route still remains an option, it was decided that we would produce a PrettyPrinter for the Ada code. This allows a user to cut and paste the Ada source code from the PrettyPrinter window to an Ada editor, and was the optimal route to code for this phase of the code generation activity in DEPLOY.

Editing the Tasking Model

The editor for the Tasking Development is based on a EMF tree-editor. The tree editor provides a facility for adding the extensions to Event-B constructs. The readability of the editor is enhanced by a PrettyPrinter, which provides a textual version of the Tasking Development, which is easier to read. It is envisaged that the textual notation will be fully integrated as a Camille extension when the facility/resources become available.

The Tool Deliverable

The demonstrator tool was released on 30 November 2010, and is available as an update site, or bundled Rodin package from:

https://sourceforge.net/projects/codegenerationd/files 

Sources are available from:

https://codegenerationd.svn.sourceforge.net/svnroot/codegenerationd

The tool is based on a build of Rodin 1.3.1 (not Rodin 2.0.0 due to dependency conflicts).

• The Code Generation tool consists of,
  • a Tasking Development Generator.
  • a Tasking Development Editor (Based on an EMF Tree Editor).
  • a translator, from Tasking Development to Common Language Model (IL1).
  • a translator, from the Tasking Development to Event-B model of the implementation.
  • a pretty-printer for the Tasking Development.
  • a pretty-printer for Common Language Model, which generates Ada Source Code.

Available Documentation

Technical Background

Much insight was gained during the work on code generation reported in the thesis Providing Concurrent Implementations for Event-B Developments

Tooling issues were reported in a paper Tool Support for Event-B Code Generation which was presented at Workshop on Tool Building in Formal Methods, http://abzconference.org/

There are technical notes available , that give more precise details of the approach and the mapping between Event-B and the common language meta-model, and its corresponding Event-B model.

For Users

There is an overview at Tasking Event-B Overview for D32

There is a wiki page at Code_Generation_Activity

There is a tutorial at Code_Generation_Tutorial

Planning

During 2011 we plan to develop the code generation tools further, taking on board any feedback from interested parties. The tool support should advance to the prototype stage, with improvements in the tool's usability in terms of features and user experience.