IDL based codegen

What is a matter IDL

A matter IDL is a text-file that aims to be a concise representation of data structures, cluster definitions and bindings. It is intended for human readability (has to be clear and concise, supports comments) as well as machine parsing (well defined syntax, not free form).

The actual grammar is defined as a Lark parsing grammar, however it may be easier to understand with an example:

/* C++ comments are supported */
// so are C-style ones


// Types such as enums and structs can be defined globally
// An enum has a name and must have an underlying type (a sized integer defining
// storage size and how it gets transmitted over the wire)
enum StatusCode : ENUM8 {
  kBusy = 1;                  // every enum MUST have an integer value
  kPAKEParameterError = 0x02; // throughout, either HEX or integer can be used
  kWindowNotOpen = 3;
}

// Structures just have a name
struct LabelStruct {
    CHAR_STRING<16> label = 0;  // structure fields have a type. Some types can be sized (strings, octet strings)
    CHAR_STRING<16> value = 1;  // every structure field has an index. this is the tag when encoding over the wire
}

// Server clusters are clusters that are expected to be exposed as servers
//
// Clusters may have optional things (e.g. optional attributes). A server
// cluster only defines attributes actually exposed by the server.
//
// Every cluster has an identifier that is defined by the matter spec (31 in this case)
server cluster AccessControl = 31 {

  // Enums and structs can be defined globally or be cluster specific.
  // IDL generation rules will take into account scoping (i.e. pick local defined
  // name first, things defined in one cluster are not visible in another).
  enum AuthMode : ENUM8 {
    kPase = 1;
    kCase = 2;
    kGroup = 3;
  }

  // structures may be fabric scoped by tagging them as 'fabric_scoped'
  // in a fabric scoped structure, fields may be 'fabric_sensitive'
  fabric_scoped struct AccessControlEntry {
    fabric_idx fabricIndex = 0;
    fabric_sensitive Privilege privilege = 1;
    fabric_sensitive AuthMode authMode = 2;
    nullable fabric_sensitive INT64U subjects[] = 3;  // fields in structures may be lists and
    nullable fabric_sensitive Target targets[] = 4;   // they may have attributes: nullable
  }

  // request structures are regular structures that are used
  // as command inputs. Special tagging to make the use clear.
  request struct AnnounceOTAProviderRequest {}
  request struct ConnectNetworkRequest {
    OCTET_STRING networkID = 0;
    INT64U breadcrumb = 1;
  }

  // Response structures are used for command outputs
  // Responses are encoded as a command and use a unique ID for encoding
  response struct ConnectNetworkResponse = 123 {
    CHAR_STRING debugText = 1;
    INT32S errorValue = 2;
  }

  // events can be specified with a type (critical/info) and may contain data
  critical event StartUp = 0 {
       INT32U softwareVersion = 0;
  }

  // no-data events are supported
  info event Leave = 2 {
  }

  // events default to 'view' privilege however this can be modified
  info event access(read: manage) RestrictedEvent = 3 {
  }

  attribute AccessControlEntry acl[] = 0;    // attributes are read-write by default
  attribute ExtensionEntry extension[] = 1;  // and require a (spec defined) number

  // Access control privileges on attributes default to:
  //
  // access(read: view, write: operate)
  //
  // These defaults can be modified to any of view/operate/manage/administer roles.
  attribute access(read: manage, write: administer) int32u customAcl = 3;

  // attributes may be read-only as well
  readonly attribute int16u clusterRevision = 65533;

  // Commands have spec-defined numbers which are used for over-the-wire
  // invocation.
  //
  // Commands have input and output data types, generally encoded as structures.
  command ConnectNetwork(ConnectNetworkRequest): ConnectNetworkResponse = 0;

  // An output is always available even if just for ok/failure, however
  // IDLs specifically do not define a structure for DefaultSuccess.
  //
  // which is considered an internal type.
  command AnnounceOTAProvider(AnnounceOTAProviderRequest): DefaultSuccess = 1;

  // Some commands may take no inputs at all
  command On(): DefaultSuccess = 2;
  command Off(): DefaultSuccess = 3;

  // command invocation default to "operate" privilege, however these
  // can be modified as well
  command access(invoke: administer) Off(): DefaultSuccess = 4;

  // command invocation can require timed invoke usage
  timed command access(invoke: administer) RevokeCommissioning(): DefaultSuccess = 2;

  // commands may be fabric scoped
  fabric command ViewGroup(ViewGroupRequest): ViewGroupResponse = 1;

  // commands may have multiple attributes
  fabric timed command RequiresTimedInvoke(): DefaultSuccess = 7;
}

// A client cluster represents something that is used by an app
// but no server is exposed.
//
// A code generation may generate either combination of client and server
// clusters for a given cluster name.
//
// Even if both client and server cluster are declared within an IDL, their
// content may differ. For example
//    - server cluster contains only attributes that the server exposes. As such
//      some optional attributes may be missing.
//
//    - client cluster contains attributes that the app wants to access as a
//      client. So an optional attribute may be presented because the
//      underlying application can make use of that attribute.
client cluster OtaSoftwareUpdateProvider = 41 {
    ///.... content removed: it is very similar to a server cluster
}


// On every endpoint number (non-dynamic)
// a series of clusters can be exposed
endpoint 0 {
  // A binding cluster is a CLIENT cluster that can be bound to for the
  // application to make use.
  //
  // As an example, a light switch can be bound to a light bulb or
  // a cluster can be bound to a OTA provider to use for updates.
  binding cluster OtaSoftwareUpdateProvider;

  // A server cluster is a server that gets exposed to the world.
  //
  // As an example, a light bulb may expose a OnOff cluster.
  server  cluster OtaSoftwareUpdateRequestor {

    // Each endpoint server cluster instantiations will have individual
    // attributes chosen for storage/defaults
    //
    // If no storage default is given, the value is initialized with 0/false/empty
    //
    // Defaults are currently only supported for primitive types (i.e. not
    // list/struct/array, but supports strings)

    ram attribute zeroInit;                    // initialized with 0.
    ram attribute stringDefault default="abc"; // Strings can have defaults.
    ram attribute boolDefault   default=true;  // bools can have defaults.
    ram attribute inRam default=123;           // stored in RAM, lost on reboot.
    persist attribute persist;                 // persisted in NVM across reboot.
    callback attribute usesCallback;           // the zap/ember 'EXTERNAL' callback.
  }
}

Parsing of IDLs

IDL parsing is done within the matter_idl python package (this is the current directory of this README). Most of the heavy lifting is done by the lark using matter_grammar.lark, which is then turned into an AST:

Code generation

Code generators are defined in generators and their purpose is to convert the parsed AST into one or several output files. In most cases the output will be split per cluster so that large CPP files are not generated (faster and more parallel compilation).

Code generator base functionality

Generators use Jinja2 as a templating language. The general CodeGenerator class in generators/init.py provides the ability to output files based on jinja templates.

In order to build working jinja2 templates, some further processing of the AST data is required. Some facilities for lookup namespacing (e.g. search for named data types within cluster first then globally) as well interpretation of data types into more concrete types is provided by generators/types.py.

Implementing generators

Beyond default AST processing, each generator is expected to add language-specific filters to create templates that work. This includes:

  • add any additional filters and transforms for data
  • add any additional type processing that is language specific

See the java code generator in generators/java as an example of codegen.

Testing generators

Tests of generation are based on checking that a given input matches an expected output. The tests/available_tests describe for each input and generator the expected output.

Intent for tests is to be focused and still easy to see deltas. Input IDLs are expected to be small and focusing on a specific functionality. Keep in mind that the test outputs are expected to be human-reviwed when codegen logic changes.

These generator tests are run by test_generators.py.