fieldx v0.1.8

Procedural macro for constructing structs with lazily initialized fields, builder pattern, and serde support with a focus on declarative syntax.

Let's start with an example:

use fieldx::fxstruct;

#[fxstruct(lazy)]
struct Foo {
    count: usize,
    foo:   String,
    #[fieldx(lazy(off), get)]
    order: RefCell<Vec<&'static str>>,
}

impl Foo {
    fn build_count(&self) -> usize {
        self.order.borrow_mut().push("Building count.");
        12
    }

    fn build_foo(&self) -> String {
        self.order.borrow_mut().push("Building foo.");
        format!("foo is using count: {}", self.count())
    }
}

let foo = Foo::new();
assert_eq!(foo.order().borrow().len(), 0);
assert_eq!(foo.foo(), "foo is using count: 12");
assert_eq!(foo.foo(), "foo is using count: 12");
assert_eq!(foo.order().borrow().len(), 2);
assert_eq!(foo.order().borrow()[0], "Building foo.");
assert_eq!(foo.order().borrow()[1], "Building count.");

What happens here is:

• a struct with all fields been lazy by default
• laziness is explicitly disabled for field order
• methods build_count and build_foo return initial values for corresponding fields

At run-time we first ensure that the order vector is empty meaning none of the build_ methods was called. Then we read from foo using its accessor method. Then we make sure that each build_ method was invoked only once.

As one can notice, a minimal amount of handcraft is needed here as most of boilerplate is handled by the macro, which provides even basic new associated function.

Also notice that we don't need to remember the order of initialization of fields. Builder of foo is using count without worrying if it's been initialized yet or not because it will always be.

Basic

The module provides two attributes: fxstruct, and fieldx. The first is responsible for configuring structs, the second for adjusting field parameters.

The macro can only be used with named structures, no union types, nor enums are supported. When applied, it rewrites the type it is applied to according to the parameters provided. Here is a list of most notable changes and additions:

• field types may be be wrapped into container types (see The Inner Workings)

  In the above example foo and count become [OnceCell<String>][OnceCell] and OnceCell<usize>, whereas order remains unchanged.

• a partial implementation of Foo is added with helper and special methods and associated functions (Field Or Method in this section)

  I.e. this is where accessor methods and new live.

• depending on parameters, an implicit implementation of the [Default] trait may be be added

• if requested, builder struct and builder() associated function will be implemented

• also, if requested, a shadow struct for correct serde support will be there too

Field Or Method?

Normally it is recommended to use module-generated helper methods to access, modify, or otherwise interact with struct fields. Use of the methods provides both better code readability and, sometimes, better functionality. Like, for example, marking a field as #[fieldx(get(clone)) would always be returning a plain cloned instance of the field value.

But when there is a need to work with a field directly (for example, to implement own accessor with additional functionality) fieldx provides own container types that are aimed at providing necessary API. See [FXProxySync] and [FXProxyAsync] container types.

Sync, Async, And Plain Structs

Note: "Async" is considered to be a synonym to "sync" since both require concurrency safety. Even the code generated for sync and async cases is mostly identical.

If a thread-safe struct is needed then fxstruct must take the sync argument: #[fxstruct(sync, ...)]. When instructed so, the macro will do its best to provide concurrency safety at the field level. It means that:

• lazy builder methods are guaranteed to be invoked once and only once per each initialization, be it single- or multi-threaded application
• access to field is lock-protected for lazy or optional fields implicitly

In less strict cases it is possible to mark individual fields as sync.

Plain non-mutable accessors normally return a reference to their field. Accessors of sync structs, unless directed to use [clone][Clone] or [copy][Copy], or used with a non-protected field, return some kind of lock-guard object.

Wrapper types for sync struct fields are non-std and provided with the module.

Protected And Unprotected Fields Of Sync Structs

For a fieldx sync struct to be Sync+Sent all of its fields are expected to be lock-protected (or, sometimes we could just say "protected"). But "expected" doesn't mean "has to be". Unless defaults, specified with fxstruct attribute (i.e. with struct-level arguments) tell otherwise, fields not marked with fieldx attribute with corresponding arguments will remain unprotected. I.e.:

#[fxstruct(sync)]
struct Foo {
    #[fieldx(lazy)]
    foo: String, // protected
    #[fieldx(get_mut)]
    bar: String, // unprotected
}

Of course, whether the struct remains thread-safe would then depend on the safety of unprotected fields.

Optional Fields

Optional in this context has the same meaning, as in the [Option] type. Sure thing, one can simply declare a field using the core type (and, as a matter of fact, this is what fieldx is using internally anyway). What's the advantages of using fieldx then?

First of all, manual declaration may mean additional boilerplate code to implement an accessor, among other things. With fieldx most of it can be hidden under a single declaration:

#[fxstruct]
struct Foo {
    #[fieldx(predicate, clearer, get, set(into))]
    description: String,
}

let mut obj = Foo::new();
assert!( !obj.has_description() );
obj.set_description("foo");
assert!( obj.has_description() );
assert_eq!( obj.description(), &Some(String::from("foo")) );
obj.clear_description();
assert!( !obj.has_description() );

<digression_mode> Besides, aesthetically, to some has_description is more appealing than obj.description().is_some(). </digression_mode>

Next, optional fields of sync structs are lock-protected by default. This can be changed with explicit lock(off), but one has to be aware that then sync status of the struct will depend the safety of the field.

And the last note to be made is that if at some point it would prove to be useful to convert a field into a lazy then refactoring could be reduced to simply adding corresponding argument the fieldx attribute and implementing a new builder for it.

Laziness Protocol

Though being very simple concept, laziness has its own peculiarities. The basics, as shown above, are such that when we declare a field as lazy the macro wraps it into some kind of proxy container type ([OnceCell] for plain fields). The first read¹ from an uninitialized field will result in the lazy builder method to be invoked and the value it returns to be stored in the field.

Here come the caveats:

1. A builder is expected to be infallible. This requirement comes from the fact that when we call field's accessor we expect a value of field's type to be returned. Since Rust requires errors to be handled semi-in-place (contrary to exceptions in many other languages) there is no way for us to overcome this limitation. The builder could panic, but this is rarely a good option.

   For cases when it is important to have controllable error handling, one could give the field a [Result] type. Then obj.field()? could be a way to take care of errors. But this approach has its own complications, especially for sync fields.

2. Field builder methods cannot mutate their objects. This limitation also comes from the fact that a typical accessor method doesn't need and must not use mutable &self. Of course, it is always possible to use internal mutability, as in the first example here.

Field Interior Mutability

Marking fields with inner_mut flag is a shortcut for using [RefCell] wrapper. This effectively turns such fields to be plain ones.

#[fxstruct]
struct Foo {
    #[fieldx(inner_mut, get, get_mut, set, default(String::from("initial")))]
    modifiable: String,
}

let foo = Foo::new();
let old = foo.set_modifiable(String::from("manual"));
assert_eq!(old, String::from("initial"));
assert_eq!(*foo.modifiable(), String::from("manual"));
*foo.modifiable_mut() = String::from("via mutable accessor");
assert_eq!(*foo.modifiable(), String::from("via mutable accessor"));

Note that this pattern is only useful when the field must not be neither optional nor lock-protected in sync-declared structs.

Builder Pattern

IMPORTANT! First of all, it is necessary to mention unintended terminological ambiguity here. The terms build and builder are used for different, though identical in nature, processes. As mentioned in the previous section, the lazy builders are methods that return initial values for associated fields. The struct builder in this section is an object that collects initial values from user and then is able to create the final instance of the original struct. This ambiguity has some history spanning back to the times when Perl's Moo module was one of the author's primary tools. Then it was borrowed by Raku AttrX::Mooish and, finally, automatically made its way into fieldx which, initially, didn't implement the builder pattern.

The default new method generated by fxstruct macro accepts no arguments and simply creates a bare-bones object initialized from type defaults. Submitting custom values for struct fields is better be done by using the builder pattern:

#[fxstruct(builder)]
struct Foo {
    #[fieldx(lazy)]
    description: String,
    count: usize,
}

impl Foo {
    fn build_description(&self) -> String {
        format!("this is item #{}", self.count)
    }
}

let obj = Foo::builder()
            .count(42)
            .build()
            .expect("Foo builder failure");
assert_eq!( obj.description(), &String::from("this is item #42") );

let obj = Foo::builder()
            .count(13)
            .description(String::from("count is ignored"))
            .build()
            .expect("Foo builder failure");
// Since the `description` is given a value the `count` field is not used
assert_eq!( obj.description(), &String::from("count is ignored") );

Since the only fieldx-related failure that may happen when building a new object instance is a required field not given a value, the build() method would return FieldXError if this happens.

Crate Features

The following featues are supported by this crate:

Usage

Most arguments of both fxstruct and fieldx can take either of the two forms: a keyword (arg), or a "function" (arg(subarg)).

Also, most of the arguments are shared by both fxstruct and fieldx. But their meaning and the way their arguments are interpreted could be slightly different for each attribute. For example, if an argument takes a literal string sub-argument it is likely to be a method name when associated with fieldx; but for fxstruct it would define common prefix for method names.

There is also a commonality between most of the arguments: they can be temporarily (say, for testing purposes) or permanently turned off by using off sub-argument with them. See lazy(off) in the above example.

Do We Need The Default Trait?

Unless explicit default argument is used with the fxstruct attribute, fieldx tries to avoid implementing the Default trait unless really required. Here is the conditions which determine if the implementation is needed:

1. Method new is generated by the procedural macro.

   This is, actually, the default behavior which is disabled with no_new argument of the fxstruct attribute.

2. A field is given a default value.

3. The struct is sync and has a lazy field.

Why get/get_mut and reader/writer For Sync Structs?

It may be confusing at first as to why there are, basically, two different kinds of accessors for sync structs. But there are reasons for it.

First of all, let's take into account these important factors:

• fields, that are protected, cannot provide their values directly; lock-guards are required for this
• lazy fields are expected to always get some value when read from

Let's focus on a case of lazy fields. They have all properties of lock-protected and optional fields, so we loose nothing in the context of the get/get_mut and reader/writer differences.

get vs reader

A bare bones get accessor helper is the same thing, as the reader helper². But, as soon as a user decides that they want copy or clone accessor behavior, reader becomes the only means of reaching out to field's lock-guard:

#[fxstruct(sync)]
struct Foo {
    #[fieldx(get(copy), reader, lazy)]
    bar: u32
}
impl Foo {
    fn build_bar(&self) -> u32 { 1234 }
    fn do_something(&self) -> u32 {
        // We need to protect the field value until we're done using it.
        let bar_guard = self.read_bar();
        let outcome = *bar_guard * 2;
        outcome
    }
}
let foo = Foo::new();
assert_eq!(foo.do_something(), 2468);

get_mut vs writer

This case if significantly different. Despite both helpers are responsible for mutating fields, the get_mut helper remains an accessor in first place, whereas the writer is not. In the context of lazy fields it means that get_mut guarantees the field to be initialized first. Then we can mutate its value.

writer, instead, provides direct and immediate access to the field's container. It allows to store a value into it without the builder method to be involved. Since building a lazy field can be expensive, it could be helpful to avoid it in cases when we don't actually need it³.

Basically, the guard returned by the writer helper can only do two things: store an entire value into the field, and clear the field.

#[fxstruct(sync)]
struct Foo {
    #[fieldx(get_mut, get(copy), writer, lazy)]
    bar: u32
}
impl Foo {
    fn build_bar(&self) -> u32 {
        eprintln!("Building bar");
        1234
    }
    fn do_something1(&self) {
        eprintln!("Using writer.");
        let mut bar_guard = self.write_bar();
        bar_guard.store(42);
    }
    fn do_something2(&self) {
        eprintln!("Using get_mut.");
        let mut bar_guard = self.bar_mut();
        *bar_guard = 12;
    }
}

let foo = Foo::new();
foo.do_something1();
assert_eq!(foo.bar(), 42);

let foo = Foo::new();
foo.do_something2();
assert_eq!(foo.bar(), 12);

This example is expected to output something like this:

Using writer.
Using get_mut.
Building bar

As you can see, use of the bar_mut accessor results in the build_bar method invoked.

The Inner Workings

As it was mentioned in the Basics section, fieldx rewrites structures with fxstruct applied. The following table reveals the final types of fields. T in the table represents the original field type, as specified by the user; O is the original struct type.

Apparently, skipped fields retain their original type. Sure enough, if such a field is of non-Send or non-Sync type the entire struct would be missing these traits despite all the efforts from the fxstruct macro.

There is also a difference in how the initialization of lazy fields is implemented. For plain fields this is done directly in their accessor methods. Sync structs delegate this functionality to the [FXProxySync] type.

Traits

fieldx additionally implement traits FXStructNonSync and FXStructSync for corresponding kind of structs. Both traits are empty and only used to distinguish structs from non-fieldx ones and from each other. For both of them FXStruct is a super-trait.

Sync Primitives

The functionality of sync structs are backed by primitives provided by the parking_lot crate.

Support Of De-/Serialization With serde

Transparently de-/serializing container types is a non-trivial task. Luckily, serde allows us to use special parameters from and into to perform indirect marshalling via a shadow struct. The way this functionality implemented by serde (and it is for a good reason) requires our original struct to implement the [Clone] trait. fxstruct doesn't automatically add a #[derive(Clone)] because implementing the trait might require manual work from the user.

Normally one doesn't need to interfere with the marshalling process. But if such a need emerges then the following implementation details might be helpful to know about:

• shadow struct mirror-fields of lazy and optional originals are [Option]-wrapped
• the struct may be given a custom name using string literal sub-argument of the serde argument
• a shadow field may share its attributes with the original if they are listed in forward_attrs sub-argument of the serde argument
• forward_attrs is always applied to the fields, no matter if it is used with struct- or field-level serde argument
• if you need custom attributes applied to the shadow struct, use the attributes*-family of serde sub-arguments
• same is about non-shared field-level custom attributes: they are to be declared with field-level attributes* of serde

License

Licensed under the BSD 3-Clause License.