{% set rfcid = "RFC-0055" %} {% include "docs/contribute/governance/rfcs/_common/_rfc_header.md" %}
Note: Formerly known as FTP-009.
Documenting FIDL.
This RFC was amended by:
Not only is good documentation an important part of scaling a team, documenting our APIs is an important part of defining a stable API. The Fuchsia API is primarily implemented in FIDL and a large amount of documentation is being left in comments that is hard to surface. Even worse, it is not uncommon for people to look at generated bindings to figure out how to use an interface. This proposal is the first step of a comprehensive documentation strategy for the FIDL language and it's interfaces.
There are two proposed FIDL source language changes. A standard Doc attribute
and syntactic sugar to improve the ergonomics of writing documentation.
Arbitrary attributes are an already supported component of the FIDL language.
Standardizing on the Doc attribute is for the tooling that generates the
formated documentation. Using an attribute as the basis of the documentation
tooling also allows for additional options to be added to the formatted output
without breaking changes.
[Discoverable, Doc = "Primary Bluetooth control service to access bluetooth"]
interface Control {
...
[Doc = "Sets the public Bluetooth |name| for this device"]
10: SetName(string? name) -> (fuchsia.bluetooth.Status status);
}
Currently, there can be only one Doc attribute per language element. This
makes it so all text must be put into the Attribute braces, which could lead to
excessively long lines.
To address the poor ergonomics of using attributes, a layer of syntactic sugar is proposed.
This involves a small change to the FIDL language specification. Comments are currently ignored during the lexing of FIDL. This FTP does not address adding general comments to the AST, only documentation comments.
Attributes are the primary way that the FIDL language expresses the idea of metadata attached to structures. Having documentation comments as just a special case of this simplifies the consumption of the metadata in the IR.
The suggested modification of the grammar is in the appendix of the FTP, but mostly involves adding an additional rule and minor rule re-ordering.
documentation-comment = "///", STRING-LITERAL, "\n"
interface Control {
/// Sent when an adapter with the given |identifier| has been
/// removed from the system.
10102: -> OnAdapterRemoved(string identifier);
}
This would de-sugar to:
[Doc="Sent when an adapter with the given |identifier| has been\n removed from the system\n"]
Doc comments are primarily freeform text. Any specific style of formatting is up to the author, team, or future style guides. The only primitive added is an identifier marker, currently proposed as pipes (|) encasing the local identifier. Unqualified identifiers are scoped to the members beneath the object that the attribute is attached too. A fully qualified identifier can be used (Ex: |fuchsia.bluetooth.ErrorCode|) to refer to objects outside of the current scope.
Eventually, fidldoc documentation generation should fail if any of the identifiers are missing, but the attribute will still be included and passed into language bindings. This will prevent documentation rot. Adding the identifiers to the IR or as part of the parsing step is intentionally avoided due to complicating these steps. Extracting the identifiers belongs in the documentation tooling (fidldoc). Documentation generation should be added as a mandatory part of a standard debug build and the overall build should fail if the documentation is not successfully generated.
A standard tool, called fidldoc, should be added to the tools directory. Fidldoc will generate markdown after consuming the FIDL JSON IR. Markdown is the current format that we are using with the other first-class language's documentation tooling.
The Wire format is unaffected by these changes. How language bindings chose to surface docstrings, or if they surface them, is left as an implementation detail for their respective communities or potentially as additional FTPs.
The style guide should be amended to prefer /// over the doc attribute, but
otherwise is left alone.
Triple comments are a relatively common way of denoting documentation comments and should not be a large barrier to understanding the fidl language. Example using triple comments should be added to the existing documentation as well as an explanation of how to use the attribute annotations.
The primary way that people will consume this feature is in the generated output.
Backwards Compatibility This feature is already backwards compatible with all recent prior fidlc compilers. While the new functionality will not exist for the triple comment syntactic sugar, they do not break earlier compilers.
Document attribute comments will work without any language changes.
No performance changes are expected except for a small increase in the JSON IR size. We will also generate documentation at compile time, which will slow down the build a small amount.
n/a
n/a
General agreement of the approach and the specific syntax used are required for adoption. The syntax is easily modified (and bikeshedded) and does alter the core ideas of the proposal.
Potential alternatives with regards to fidldoc is that the compiler generates the documentation itself. It may also be worth using the existing backend generator approach for this. Output format of the generated documentation may also be up for discussion.
Another alternative is to represent the documentation comments as first-class citizens in the AST. While there aren't any real drawbacks to this strategy, you lose some of the extensibility benefits of modeling it as an attribute. Someday we may want to add additional information for our documentation tooling and the attribute style makes this possible without breaking changes. For example, we may want to allow specifying the markdown language of the comment. This would then keep all of the information for generating the documentation within the same output (the attributes). It also enforces a nice regularity where the doc comments and attributes, which have similar placement constraints, are parsed in the same way.
Most languages have documentation tooling. This draws from prior art of dartdoc, rustdoc, and javadoc (mostly in what not to do)