festive: A Constrained Size Output Virtual Machine

An attempt at defining a small VM to create a stack machine for size-constrained clients and servers.

Original motivation was to create a simple templating renderer for USSD clients, combined with an agnostic data-retrieval reference that may conceal any level of complexity.

Opcodes

The VM defines the following opcode symbols:

• BACK - Return to the previous execution frame (will fail if at top frame). It leaves to the state of the execution layer to define what "previous" means.
• CATCH <symbol> <signal> - Jump to symbol if signal is set (see signal below).
• CROAK <signal> - Clear state and restart execution from top if signal is set (see signal below).
• LOAD <symbol> <size> - Execute the code symbol symbol and cache the data, constrained to the given size. Can be exposed with MAP within scope,
• RELOAD <symbol> - Execute a code symbol already loaded by LOAD and cache the data, constrained to the previously given size for the same symbol.
• MAP <symbol> - Expose a code symbol previously loaded by LOAD to the rendering client. Roughly corresponds to the global directive in Python.
• MOVE <symbol> - Create a new execution frame, invalidating all previous MAP calls. More detailed: After a MOVE call, a BACK call will return to the same execution frame, with the same symbols available, but all MAP calls will have to be repeated.

External code

LOAD is used to execute code symbols in the host environment. It is loaded with a size constraint, and returned data violating this constraint should generate an error.

Any symbol successfully loaded with LOAD will be associated with the call stack frame it is loaded. The symbol will be available in the same frame and frames below it. Once the frame goes out of scope (e.g. BACK is called in that frame) the symbols should be freed as soon as possible. At this point they are not available to the abandoned scope.

Loaded symbols are not automatically exposed to the rendering client. To expose symbols ot the rendering client the MAP opcode must be used.

The associated content of loaded symbols may be refreshed using the RELOAD opcode. RELOAD only works within the same constraints as MAP. However, updated content must be available even if a MAP precedes a RELOAD within the same frame.

External symbol optimizations

Only LOAD and RELOAD should trigger external code side-effects.

In an effort to prevent leaks from unnecessary external code executions, the following constraints are assumed:

• An explicit MAP must exist in the scope of any LOAD.
• All symbols declared in MAP must be used for all template renderings of a specific node.

Any code compiler or checked should generate an error on any orphaned LOAD or MAP symbols as described above.

Rendering

The fixed-size output is generated using a templating language, and a combination of one or more max size properties, and an optional sink property that will attempt to consume all remaining capacity of the rendered template.

For example, in this example

• maxOutputSize is 256 bytes long.
• template is 120 bytes long.
• param one has max size 10 but uses 5.
• param two has max size 20 but uses 12.
• param three is a sink.

The renderer may use up to 256 - 120 - 5 - 12 = 119 bytes from the sink when rendering the output.

Multipage support

Multipage outputs, like listings, are handled using the sink output constraints:

• first calculate what the rendered display size is when all symbol results that are not sinks are resolved.
• split and cache the list data within its semantic context, given the sink limitation after rendering.
• provide a next and previous menu item to browse the prepared pagination of the list data.

Languages support

Language for rendering is determined at the top-level state.

Lookups dependent on language are prefixed by either ISO 639-1 or ISO 639-3 language codes, followed by :.

Default language means records returned without prefix if no language is set. Default language should be settable at the top-level.

Node names must be defined using 7-bit ASCII.

Virtual machine interface layout

This is the version 0 of the VM. That translates to highly experimental.

Currently the following rules apply for encoding in version 0:

• A code instruction is a big-endian 2-byte value. See vm/opcodes.go for valid opcode values.
• symbol value is encoded as one byte of string length, after which the byte-value of the string follows.
• size value is encoded as one byte of numeric length, after which the big-endian byte-value of the integer follows.
• signal value is encoded as one byte of byte length, after which a byte-array representing the defined signal follows.

Reference implementation

This repository provides a golang reference implementation for the festive concept.

In this reference implementation some constraints apply

Structure

TODO: state will be separated into cache and session.

• vm: Defines instructions, and applies transformations according to the instructions.
• state: Holds the code cache, contents cache aswell as error tates from code execution.
• resource: Retrieves data and bytecode from external symbols, and retrieves and renders templates.
• engine: Outermost interface. Orchestrates execution of bytecode against input.

Template rendering

Template rendering is done using the text/template faciilty in the golang standard library.

It expects all replacement symbols to be available at time of rendering, and has no tolerance for missing ones.

Assembly language

TBD

An assmebly language will be defined to generate the routing and execution bytecodes for each menu node.