Preinterpet - The code generation toolkit

This crate provides the preinterpret! macro, which works as a simple pre-processor to the token stream. It takes inspiration from and effectively combines the quote, paste and syn crates, to empower code generation authors and declarative macro writers, bringing:

• Heightened readability - quote-like variable definition and substitution make it easier to work with code generation code.
• Heightened expressivity - a toolkit of simple commands reduce boilerplate, and mitigate the need to build custom procedural macros in some cases.
• Heightened simplicity - helping developers avoid the confusing corners [1, 2, 3, 4] of declarative macro land.

The preinterpret! macro can be used inside the output of a declarative macro, or by itself, functioning as a mini code generation tool all of its own.

[dependencies]
preinterpret = "0.2"

User Guide

Preinterpret works with its own very simple language, with two main syntax elements:

• Commands: [!command_name! ...input token stream...] take an input token stream and output a token stream. There are a number of commands which cover a toolkit of useful functions.
• Variables: [!set! #var_name = token stream...] defines a variable, and #var_name substitutes the variable into another command or the output.

Commands can be nested intuitively. The input of all commands (except [!raw! ...]) are first interpreted before the command itself executes.

Declarative macro example

The following artificial example demonstrates how preinterpret can be integrate into declarative macros, and covers use of variables, idents and case conversion:

macro_rules! create_my_type {
    (
        $(#[$attributes:meta])*
        $vis:vis struct $type_name:ident {
            $($field_name:ident: $inner_type:ident),* $(,)?
        }
    ) => {preinterpret::preinterpret! {
        [!set! #type_name = [!ident! My $type_name]]
        
        $(#[$attributes])*
        $vis struct #type_name {
            $($field_name: $inner_type,)*
        }

        impl #type_name {
            $(
                fn [!ident_snake! my_ $inner_type](&self) -> &$inner_type {
                    &self.$field_name
                }

                fn [!ident_snake! my_ $inner_type _mut](&mut self) -> &mut $inner_type {
                    &mut self.$field_name
                }
            )*
        }
    }}
}
create_my_type! {
    struct Struct {
        field0: String,
        field1: u64,
    }
}
assert_eq!(MyStruct { field0: "Hello".into(), field1: 21 }.my_string(), "Hello")

Quick background on token streams and macros

To properly understand how preinterpret works, we need to take a very brief detour into the language of macros.

In Rust, the input and output to a macro is a TokenStream. A TokenStream is simply an iterator of TokenTrees at a particular nesting level. A token tree is one of four things:

• A Group - typically (..), [..] or {..}. It consists of a matched pair of brackets "[Delimiters]` and an internal token stream. There is technically a confusing fourth type of group, with transparent brackets; used to encapsulate declarative macro substitutions. This is purposefully ignored/flattened in pre-interpret.
• An Ident - An unquoted string, used to identitied something named. Think MyStruct, or do_work or my_module.
• A Punct - A single piece of punctuation. Think ! or :.
• A Literal - This includes string literals "my string", char literals 'x' and numeric literals 23 / 51u64. Note that true/false are technically idents.

When you return output from a macro, you are outputting back a token stream, which the compiler will interpret.

Preinterpret commands take token streams as input, and return token streams as output.

Migration from paste

If migrating from paste, the main difference is that you need to specify what kind of concatenated thing you want to create. Paste tried to work this out magically from context, but sometimes got it wrong.

In other words, you typically want to replace [< ... >] with [!ident! ...], and sometimes [!string! ...] or [!literal! ...]:

• To create type and function names, use [!ident! My #preinterpret_type_name $macro_type_name], [!ident_camel! ...], [!ident_snake! ...] or [!ident_upper_snake! ...]
• For doc macros or concatenated strings, use [!string! "My type is: " #type_name]
• If you're creating literals of some kind by concatenating parts together, use [!literal! 32 u32]

For example:

preinterpret::preinterpret! {
    [!set! #type_name = [!ident! HelloWorld]]

    struct #type_name;

    #[doc = [!string! "This type is called [`" #type_name "`]"]]
    impl #type_name {
        fn [!ident_snake! say_ #type_name]() -> &'static str {
            [!string! "It's time to say: " [!title! #type_name] "!"]
        }
    }
}
assert_eq!(HelloWorld::say_hello_world(), "It's time to say: Hello World!")

Command List

Special commands

• [!set! #foo = Hello] followed by [!set! #foo = #bar(World)] sets the variable #foo to the token stream Hello and #bar to the token stream Hello(World), and outputs no tokens. Using #foo or #bar later on will output the current value in the corresponding variable.
• [!raw! abc #abc [!ident! test]] outputs its contents as-is, without any interpretation, giving the token stream abc #abc [!ident! test].
• [!ignore! $foo] ignores all of its content and outputs no tokens. It is useful to make a declarative macro loop over a meta-variable without outputting it into the resulting stream.

Concatenate and convert commands

Each of these commands functions in three steps:

• Apply the interpreter to the token stream, which recursively executes preinterpret commands.
• Convert each token of the resulting stream into a string, and concatenate these together. String and char literals are unquoted, and this process recurses into groups.
• Apply some command-specific conversion.

The following commands output idents:

• [!ident! X Y "Z"] outputs the ident XYZ
• [!ident_camel! my hello_world] outputs MyHelloWorld
• [!ident_snake! my_ HelloWorld] outputs my_hello_world
• [!ident_upper_snake! my_ const Name] outputs MY_CONST_NAME

The !literal! command outputs any kind of literal, for example:

• [!literal! 31 u 32] outputs the integer literal 31u32
• [!literal! '"' hello '"'] outputs the string literal "hello"

The following commands output strings, without dropping non-alphanumeric characters:

• [!string! X Y " " Z (Hello World)] outputs "XY Z(HelloWorld)"
• [!upper! foo_bar] outputs "FOO_BAR"
• [!lower! FooBar] outputs "foobar"
• [!capitalize! fooBar] outputs "FooBar"
• [!decapitalize! FooBar] outputs "fooBar"
• [!insert_spaces! fooBar] outputs "foo Bar"

The following commands output strings, whilst also dropping non-alphanumeric characters:

• [!snake! FooBar] and [!lower_snake! FooBar] are equivalent and output "foo_bar"
• [!upper_snake! FooBar] outputs "FOO_BAR"
• [!camel! foo_bar] and [!upper_camel! foo_bar] are equivalent and output "FooBar". This filters out non-alphanumeric characters.
• [!lower_camel! foo_bar] outputs "fooBar"
• [!kebab! fooBar] outputs "foo-bar"

[!NOTE]

These string conversion methods are designed to work intuitively across a wide class of input strings, by creating word boundaries when going from non-alphanumeric to alphanumeric, lowercase to uppercase, or uppercase to uppercase if the next character is lowercase.

The case-conversion commands which drop non-alphanumeric characters can potentially break up grapheme clusters and can cause unintuitive behaviour when used with complex unicode strings.

A wide ranging set of tests covering behaviour are in tests/string.rs.

Motivation

Readability

The preinterpret syntax is intended to be immediately intuitive even for people not familiar with the crate. And it enables developers to make more readable macros:

• Developers can name clear concepts in their macro output, and re-use them by name, decreasing code duplication.
• Developers can use variables to subdivide logic inside the macro, without having to resort to creating lots of small, functional helper macros.

These ideas are demonstrated with the following simple example:

macro_rules! impl_marker_traits {
    {
        impl [
            // The marker traits to implement
            $($trait:ident),* $(,)?
        ] for $type_name:ident
        $(
            // Arbitrary (non-const) type generics
            < $( $lt:tt $( : $clt:tt $(+ $dlt:tt )* )? $( = $deflt:tt)? ),+ >
        )?
    } => {preinterpret::preinterpret!{
        [!set! #impl_generics = $(< $( $lt $( : $clt $(+ $dlt )* )? ),+ >)?]
        [!set! #type_generics = $(< $( $lt ),+ >)?]
        [!set! #my_type = $type_name #type_generics]

        $(
            // Output each marker trait for the type
            impl #impl_generics $trait for #my_type {}
        )*
    }}
}
trait MarkerTrait1 {}
trait MarkerTrait2 {}
struct MyType<T: Clone>(T);
impl_marker_traits! {
    impl [MarkerTrait1, MarkerTrait2] for MyType<T: Clone>
};

Expressivity

Preinterpret provides a suite of simple, composable commands to convert token streams, literals and idents. The full list is documented in the Command List section.

For example:

macro_rules! create_struct_and_getters {
    (
        $name:ident { $($field:ident),* $(,)? }
    ) => {preinterpret::preinterpret!{
        // Define a struct with the given fields
        pub struct $name {
            $(
                $field: String,
            )*
        }

        impl $name {
            $(
                // Define get_X for each field X
                pub fn [!ident! get_ $field](&self) -> &str {
                    &self.$field
                }
            )*
        }
    }}
}
create_struct_and_getters! {
  MyStruct { hello, world }
}

Variable assignment works intuitively with the * + ? expansion operators, allowing basic procedural logic, such as creation of loop counts and indices before meta-variables are stabilized.

For example:

macro_rules! count_idents {
    {
        $($item: ident),*
    } => {preinterpret::preinterpret!{
        [!set! #current_index = 0usize]
        $(
            [!ignore! $item] // Loop over the items, but don't output them
            [!set! #current_index = #current_index + 1]
        )*
        [!set! #count = #current_index]
        #count
    }}
}

To quickly explain how this works, imagine we evaluate count_idents!(a, b, c). As count_idents! is the most outer macro, it runs first, and expands into the following token stream:

let count = preinterpret::preinterpret!{
  [!set! #current_index = 0usize]
  [!ignore! a]
  [!set! #current_index = #current_index + 1]
  [!ignore! = b]
  [!set! #current_index = #current_index + 1]
  [!ignore! = c]
  [!set! #current_index = #current_index + 1]
  [!set! #count = #current_index]
  #count
};

Now the preinterpret! macro runs, resulting in #count equal to the token stream 0usize + 1 + 1 + 1. This will be improved in future releases by adding support for mathematical operations on integer literals.

Simplicity

Using preinterpret partially mitigates some common areas of confusion when writing declarative macros.

Cartesian metavariable expansion errors

Sometimes you wish to output some loop over one meta-variable, whilst inside the loop of a non-parent meta-variable - in other words, you expect to create a cartesian product across these variables. But the macro evaluator only supports zipping of meta-variables of the same length, and gives an unhelpful error message.

The classical wisdom is to output an internal macro_rules! definition to handle the inner output of the cartesian product as per this stack overflow post, but this isn't very intuitive.

Standard use of preinterpret avoids this problem entirely, as demonstrated by the first readability example. If written out natively without preinterpret, the iteration of the generics in #impl_generics and #my_type wouldn't be compatible with the iteration over $trait.

Eager macro confusion

User-defined macros are not eager - they take a token stream in, and return a token stream; and further macros can then execute in this token stream.

But confusingly, some compiler built-in macros in the standard library (such as format_args!, concat!, concat_idents! and include!) don't work like this - they actually inspect their arguments, evaluate any macros inside eagerly, before then operating on the outputted tokens.

Don't get me wrong - it's useful that you can nest concat! calls and include! calls - but the fact that these macros use the same syntax as "normal" macros but use different resolution behaviour can cause confusion to developers first learning about macros.

Preinterpet commands also typically interpret their arguments eagerly and recursively, but it tries to be less confusing by:

• Having a clear name (Preinterpet) which suggests eager pre-processing.
• Using a different syntax [!command! ...] to macros to avoid confusion.
• Taking on the functionality of the concat! and concat_idents! macros so they don't have to be used alongside other macros.

The recursive macro paradigm shift

To do anything particularly advanced with declarative macros, you end up needing to conjure up various functional macro helpers to partially apply or re-order grammars. This is quite a paradigm-shift from most rust code.

In quite a few cases, preinterpret can allow developers to avoid writing these recursive helper macros entirely.

Limitations with paste support

The widely used paste crate takes the approach of magically hiding the token types from the developer, by attempting to work out whether a pasted value should be an ident, string or literal.

This works 95% of the time, but in other cases such as in attributes, it can cause developer friction. This proved to be one of the motivating use cases for developing preinterpret.

Preinterpret is more explicit about types, and doesn't have these issues:

macro_rules! impl_new_type {
    {
        $vis:vis $my_type:ident($my_inner_type:ty)
    } => {preinterpret::preinterpret!{
        #[xyz(as_type = [!string! $my_inner_type])]
        $vis struct $my_type($my_inner_type);
    }}
}

Future Extension Possibilities

Add github docs page / rust book

Add a github docs page / rust book at this repository, to allow us to build out a suite of examples, like serde or the little book of macros.

Destructuring / Parsing Syntax, and Declarative Macros 2.0

I have a vision for having preinterpret effectively replace the use of declarative macros in the Rust ecosystem, by:

• Enabling writing intuitive, procedural code, which feels a lot like normal rust
• Exposing the power of syn into this language, and preparing people to write procedural macros

This would avoid pretty much all of the main the complexities of declarative macros:

• The entirely lacking compile errors and auto-complete when it can't match tokens.
• What the metavariable types mean as :ty, :tt, :vis and how to use them all... And the fact that, once matched, all but tt are opaque for future matching.
• Having to learn a new paradigm of inverted thinking, which is pretty alien to rust.
• The macro_rules! declaration itself - I can never remember which brackets to use...

The idea is that we create two new tools:

• The parse command which can be built up of compasable parse-helpers (mostly wrapping syn calls), intuitively handling lots of common patterns seen in code
• Add control flow (for, match and the like) which runs lazily and declaratively, over the token stream primitive

In more detail:

• [!parse! (<PARSE_DESTRUCTURING>) = (<INPUT>)] is a more general [!set!] which acts like a let <XX> = <YY> else { panic!() }. It takes a ()-wrapped parse destructuring on the left and a token stream as input on the right. Any #x in the parse definition acts as a binding rather than as a substitution. Parse operations look like [<] Iterations are _not_ supported, but [!optional] This will handled commas intelligently, and accept intelligent parse-helpers like:
  • [<fields> { hello: #a, world?: #b }] - which can parse #x in any order, cope with trailing commas, and permit fields on the RHS not on the LHS
  • [<subfields> { hello: #a, world?: #b }] - which can parse fields in any order, cope with trailing commas, and permit fields on the RHS not on the LHS
  • [<item> { #ident, #impl_generics, ... }] - which calls syn's parse item on the token
  • [<ident> ...], [<literal> ...] and the like to parse idents / literals etc directly from the token stream (rather than token streams).
  • More tailored examples, such as [<generics> { impl: #x, type: #y, where: #z }] which uses syn to parse the generics, and then uses subfields on the result.
  • Possibly [<group> #x] to parse a group with no brackets, to avoid parser ambguity in some cases
  • Any complex logic (loops, matching), is delayed lazily until execution logic time - making it much more intuitive.
• [!for! (<PARSE_DESTRUCTURING>) in (<INPUT>) { ... }] which operates like the rust for loop, and uses a parse destructuring on the left, and has support for optional commas between values
• [!match! (<INPUT>) => { (<CASE1>) => { ... }, (<CASE2>) => { ... }, (#fallback) => { ... } }] which operates like a rust match expression, and can replace the function of the branches of declarative macro inputs.
• [!macro_rules! name!(<PARSE_DESTRUCTURING>) = { ... }] which can define a declarative macro, but just parses its inputs as a token stream, and uses preinterpret for its heavy lifting.

And then we can end up with syntax like the following:

// =================================================
// Hypothetical future syntax - not yet implemented!
// =================================================

// A simple macro can just take a token stream as input
preinterpret::preinterpret! {
    [!macro_rules! my_macro!(#input) {
        [!for! (#trait for #type) in (#input) {
            impl #trait for #type
        }]
    }]
}

// Or can parse input - although loops are kept as a token stream and delegated
// to explicit lazy iteration, allowing a more procedural code style,
// and clearer compiler errors.
preinterpret::preinterpret! {
    [!macro_rules! multi_impl_super_duper!(
        #type_list,
        ImplOptions [!fields! {
            hello: #hello,
            world?: #world (default "Default")
        }]
    ) = {
        [!for! (
            #type [!generics! { impl: #impl_generics, type: #type_generics }]
        ) in (#type_list) {
            impl<#impl_generics> SuperDuper for #type #type_generics {
                type Hello = #hello;
                type World = #world;
            }
        }]
    }]
}

preinterpret::preinterpret! {
    [!set! #input =
        MyTrait for MyType,
        MyTrait for MyType2,
    ]

    [!for! (#trait for #type) in (#input) {
        impl #trait for #type
    }]
}

Possible extension: Integer commands

Each of these commands functions in three steps:

• Apply the interpreter to the token stream, which recursively executes preinterpret commands.
• Iterate over each token (recursing into groups), expecting each to be an integer literal.
• Apply some command-specific mapping to this stream of integer literals, and output a single integer literal without its type suffix. The suffix can be added back manually if required with a wrapper such as [!literal! [!add! 1 2] u64].

Integer commands under consideration are:

• [!add! 5u64 9 32] outputs 46. It takes any number of integers and outputs their sum. The calculation operates in u128 space.
• [!sub! 64u32 1u32] outputs 63. It takes two integers and outputs their difference. The calculation operates in i128 space.
• [!mod! $length 2] outputs 0 if $length is even, else 1. It takes two integers a and b, and outputs a mod b.

We also support the following assignment commands:

• [!increment! #i] is shorthand for [!set! #i = [!add! #i 1]] and outputs no tokens.

Even better - we could even support calculator-style expression interpretation:

• [!usize! (5 + 10) / mod(4, 2)] outputs 7usize

Possible extension: User-defined commands

• [!define! [!my_command! <PARSE_DESTRUCTURING>] { <OUTPUT> }]

Possible extension: Boolean commands

Each of these commands functions in three steps:

• Apply the interpreter to the token stream, which recursively executes preinterpret commands.
• Expects to read exactly two token trees (unless otherwise specified)
• Apply some command-specific comparison, and outputs the boolean literal true or false.

Comparison commands under consideration are:

• [!eq! #foo #bar] outputs true if #foo and #bar are exactly the same token tree, via structural equality. For example:
  • [!eq! (3 4) (3 4)] outputs true because the token stream ignores spacing.
  • [!eq! 1u64 1] outputs false because these are different literals.
• [!lt! #foo #bar] outputs true if #foo is an integer literal and less than #bar
• [!gt! #foo #bar] outputs true if #foo is an integer literal and greater than #bar
• [!lte! #foo #bar] outputs true if #foo is an integer literal and less than or equal to #bar
• [!gte! #foo #bar] outputs true if #foo is an integer literal and greater than or equal to #bar
• [!not! #foo] expects a single boolean literal, and outputs the negation of #foo
• [!str_contains! "needle" [!string! haystack]] expects two string literals, and outputs true if the first string is a substring of the second string.

Possible extension: Token stream commands

• [!skip! 4 from [#stream]] reads and drops the first 4 token trees from the stream, and outputs the rest
• [!ungroup! (#stream)] outputs #stream. It expects to receive a single group (i.e. wrapped in brackets), and unwraps it.

Possible extension: Control flow commands

If statement

[!if! #cond then { #a } else { #b }] outputs #a if #cond is true, else #b if #cond is false.

The if command works as follows:

• It starts by only interpreting its first token tree, and expects to see a single true or false literal.
• It then expects to reads an unintepreted then ident, following by a single { .. } group, whose contents get interpreted and output only if the condition was true.
• It optionally also reads an else ident and a by a single { .. } group, whose contents get interpreted and output only if the condition was false.

For loop

• [!for! #token_tree in [#stream] { ... }]

Goto and label

• [!label! loop_start] - defines a label which can be returned to. Effectively, it takes a clones of the remaining token stream after the label in the interpreter.
• [!goto! loop_start] - jumps to the last execution of [!label! loop_start]. It unrolls the preinterpret stack (dropping all unwritten token streams) until it finds a stackframe in which the interpreter has the defined label, and continues the token stream from there.

// Hypothetical future syntax - not yet implemented!
preinterpret::preinterpret!{
    [!set! #i = 0]
    [!label! loop]
    const [!ident! AB #i]: u8 = 0;
    [!increment! #i]
    [!if! [!lte! #i 100] then { [!goto! loop] }]
}

Possible extension: Eager expansion of macros

When eager expansion of macros returning literals is stabilized, it would be nice to include a command to do that, which could be used to include code, for example: [!expand_literal_macros! include!("my-poem.txt")].

Possible extension: Explicit parsing feature to enable syn

The heavy syn library is (in basic preinterpret) only needed for literal parsing, and error conversion into compile errors.

We could add a parsing feature to speed up compile times a lot for stacks which don't need the parsing functionality.

License

Licensed under either of the Apache License, Version 2.0 or the MIT license at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this crate by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.