52 releases

0.2.25 Oct 23, 2024
0.2.22 Aug 25, 2024
0.2.20 May 7, 2024
0.2.17 Mar 25, 2024
0.1.9 Mar 27, 2022

#2 in Value formatting

Download history 1323586/week @ 2024-08-21 1284752/week @ 2024-08-28 1435748/week @ 2024-09-04 1371128/week @ 2024-09-11 1461892/week @ 2024-09-18 1597844/week @ 2024-09-25 1931968/week @ 2024-10-02 1931155/week @ 2024-10-09 1930018/week @ 2024-10-16 1482867/week @ 2024-10-23 1359124/week @ 2024-10-30 1349082/week @ 2024-11-06 1469081/week @ 2024-11-13 1411748/week @ 2024-11-20 1201826/week @ 2024-11-27 1211659/week @ 2024-12-04

5,546,824 downloads per month
Used in 4,911 crates (300 directly)

MIT/Apache

245KB
6K SLoC

prettyplease::unparse

github crates.io docs.rs build status

A minimal syn syntax tree pretty-printer.


Overview

This is a pretty-printer to turn a syn syntax tree into a String of well-formatted source code. In contrast to rustfmt, this library is intended to be suitable for arbitrary generated code.

Rustfmt prioritizes high-quality output that is impeccable enough that you'd be comfortable spending your career staring at its output — but that means some heavyweight algorithms, and it has a tendency to bail out on code that is hard to format (for example rustfmt#3697, and there are dozens more issues like it). That's not necessarily a big deal for human-generated code because when code gets highly nested, the human will naturally be inclined to refactor into more easily formattable code. But for generated code, having the formatter just give up leaves it totally unreadable.

This library is designed using the simplest possible algorithm and data structures that can deliver about 95% of the quality of rustfmt-formatted output. In my experience testing real-world code, approximately 97-98% of output lines come out identical between rustfmt's formatting and this crate's. The rest have slightly different linebreak decisions, but still clearly follow the dominant modern Rust style.

The tradeoffs made by this crate are a good fit for generated code that you will not spend your career staring at. For example, the output of bindgen, or the output of cargo-expand. In those cases it's more important that the whole thing be formattable without the formatter giving up, than that it be flawless.


Feature matrix

Here are a few superficial comparisons of this crate against the AST pretty-printer built into rustc, and rustfmt. The sections below go into more detail comparing the output of each of these libraries.

prettyplease rustc rustfmt
non-pathological behavior on big or generated code 💚
idiomatic modern formatting ("locally indistinguishable from rustfmt") 💚 💚
throughput 60 MB/s 39 MB/s 2.8 MB/s
number of dependencies 3 72 66
compile time including dependencies 2.4 sec 23.1 sec 29.8 sec
buildable using a stable Rust compiler 💚
published to crates.io 💚
extensively configurable output 💚
intended to accommodate hand-maintained source code 💚

Comparison to rustfmt

If you weren't told which output file is which, it would be practically impossible to tell — except for line 435 in the rustfmt output, which is more than 1000 characters long because rustfmt just gave up formatting that part of the file:

            match segments[5] {
                0 => write!(f, "::{}", ipv4),
                0xffff => write!(f, "::ffff:{}", ipv4),
                _ => unreachable!(),
            }
        } else { # [derive (Copy , Clone , Default)] struct Span { start : usize , len : usize , } let zeroes = { let mut longest = Span :: default () ; let mut current = Span :: default () ; for (i , & segment) in segments . iter () . enumerate () { if segment == 0 { if current . len == 0 { current . start = i ; } current . len += 1 ; if current . len > longest . len { longest = current ; } } else { current = Span :: default () ; } } longest } ; # [doc = " Write a colon-separated part of the address"] # [inline] fn fmt_subslice (f : & mut fmt :: Formatter < '_ > , chunk : & [u16]) -> fmt :: Result { if let Some ((first , tail)) = chunk . split_first () { write ! (f , "{:x}" , first) ? ; for segment in tail { f . write_char (':') ? ; write ! (f , "{:x}" , segment) ? ; } } Ok (()) } if zeroes . len > 1 { fmt_subslice (f , & segments [.. zeroes . start]) ? ; f . write_str ("::") ? ; fmt_subslice (f , & segments [zeroes . start + zeroes . len ..]) } else { fmt_subslice (f , & segments) } }
    } else {
        const IPV6_BUF_LEN: usize = (4 * 8) + 7;
        let mut buf = [0u8; IPV6_BUF_LEN];
        let mut buf_slice = &mut buf[..];

This is a pretty typical manifestation of rustfmt bailing out in generated code — a chunk of the input ends up on one line. The other manifestation is that you're working on some code, running rustfmt on save like a conscientious developer, but after a while notice it isn't doing anything. You introduce an intentional formatting issue, like a stray indent or semicolon, and run rustfmt to check your suspicion. Nope, it doesn't get cleaned up — rustfmt is just not formatting the part of the file you are working on.

The prettyplease library is designed to have no pathological cases that force a bail out; the entire input you give it will get formatted in some "good enough" form.

Separately, rustfmt can be problematic to integrate into projects. It's written using rustc's internal syntax tree, so it can't be built by a stable compiler. Its releases are not regularly published to crates.io, so in Cargo builds you'd need to depend on it as a git dependency, which precludes publishing your crate to crates.io also. You can shell out to a rustfmt binary, but that'll be whatever rustfmt version is installed on each developer's system (if any), which can lead to spurious diffs in checked-in generated code formatted by different versions. In contrast prettyplease is designed to be easy to pull in as a library, and compiles fast.


Comparison to rustc_ast_pretty

This is the pretty-printer that gets used when rustc prints source code, such as rustc -Zunpretty=expanded. It's used also by the standard library's stringify! when stringifying an interpolated macro_rules AST fragment, like an $:expr, and transitively by dbg! and many macros in the ecosystem.

Rustc's formatting is mostly okay, but does not hew closely to the dominant contemporary style of Rust formatting. Some things wouldn't ever be written on one line, like this match expression, and certainly not with a comma in front of the closing brace:

fn eq(&self, other: &IpAddr) -> bool {
    match other { IpAddr::V4(v4) => self == v4, IpAddr::V6(_) => false, }
}

Some places use non-multiple-of-4 indentation, which is definitely not the norm:

pub const fn to_ipv6_mapped(&self) -> Ipv6Addr {
    let [a, b, c, d] = self.octets();
    Ipv6Addr{inner:
                 c::in6_addr{s6_addr:
                                 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF,
                                  0xFF, a, b, c, d],},}
}

And although there isn't an egregious example of it in the link because the input code is pretty tame, in general rustc_ast_pretty has pathological behavior on generated code. It has a tendency to use excessive horizontal indentation and rapidly run out of width:

::std::io::_print(::core::fmt::Arguments::new_v1(&[""],
                                                 &match (&msg,) {
                                                      _args =>
                                                      [::core::fmt::ArgumentV1::new(_args.0,
                                                                                    ::core::fmt::Display::fmt)],
                                                  }));

The snippets above are clearly different from modern rustfmt style. In contrast, prettyplease is designed to have output that is practically indistinguishable from rustfmt-formatted code.


Example

// [dependencies]
// prettyplease = "0.2"
// syn = { version = "2", default-features = false, features = ["full", "parsing"] }

const INPUT: &str = stringify! {
    use crate::{
          lazy::{Lazy, SyncLazy, SyncOnceCell}, panic,
        sync::{ atomic::{AtomicUsize, Ordering::SeqCst},
            mpsc::channel, Mutex, },
      thread,
    };
    impl<T, U> Into<U> for T where U: From<T> {
        fn into(self) -> U { U::from(self) }
    }
};

fn main() {
    let syntax_tree = syn::parse_file(INPUT).unwrap();
    let formatted = prettyplease::unparse(&syntax_tree);
    print!("{}", formatted);
}

Algorithm notes

The approach and terminology used in the implementation are derived from Derek C. Oppen, "Pretty Printing" (1979), on which rustc_ast_pretty is also based, and from rustc_ast_pretty's implementation written by Graydon Hoare in 2011 (and modernized over the years by dozens of volunteer maintainers).

The paper describes two language-agnostic interacting procedures Scan() and Print(). Language-specific code decomposes an input data structure into a stream of string and break tokens, and begin and end tokens for grouping. Each beginend range may be identified as either "consistent breaking" or "inconsistent breaking". If a group is consistently breaking, then if the whole contents do not fit on the line, every break token in the group will receive a linebreak. This is appropriate, for example, for Rust struct literals, or arguments of a function call. If a group is inconsistently breaking, then the string tokens in the group are greedily placed on the line until out of space, and linebroken only at those break tokens for which the next string would not fit. For example, this is appropriate for the contents of a braced use statement in Rust.

Scan's job is to efficiently accumulate sizing information about groups and breaks. For every begin token we compute the distance to the matched end token, and for every break we compute the distance to the next break. The algorithm uses a ringbuffer to hold tokens whose size is not yet ascertained. The maximum size of the ringbuffer is bounded by the target line length and does not grow indefinitely, regardless of deep nesting in the input stream. That's because once a group is sufficiently big, the precise size can no longer make a difference to linebreak decisions and we can effectively treat it as "infinity".

Print's job is to use the sizing information to efficiently assign a "broken" or "not broken" status to every begin token. At that point the output is easily constructed by concatenating string tokens and breaking at break tokens contained within a broken group.

Leveraging these primitives (i.e. cleverly placing the all-or-nothing consistent breaks and greedy inconsistent breaks) to yield rustfmt-compatible formatting for all of Rust's syntax tree nodes is a fun challenge.

Here is a visualization of some Rust tokens fed into the pretty printing algorithm. Consistently breaking beginend pairs are represented by «», inconsistently breaking by , break by ·, and the rest of the non-whitespace are string.

use crate::«{·
‹    lazy::«{·‹Lazy,· SyncLazy,· SyncOnceCell›·}»,·
    panic,·
    sync::«{·
‹        atomic::«{·‹AtomicUsize,· Ordering::SeqCst›·}»,·
        mpsc::channel,· Mutex›,·
    }»,·
    thread›,·
}»;·
«‹«impl<«·T‹›,· U‹›·»>» Into<«·U·»>· for T›·
where·
    U:‹ From<«·T·»>›,·
{·
«    fn into(·«·self·») -> U {·
‹        U::from(«·self·»)›·
»    }·
»}·

The algorithm described in the paper is not quite sufficient for producing well-formatted Rust code that is locally indistinguishable from rustfmt's style. The reason is that in the paper, the complete non-whitespace contents are assumed to be independent of linebreak decisions, with Scan and Print being only in control of the whitespace (spaces and line breaks). In Rust as idiomatically formatted by rustfmt, that is not the case. Trailing commas are one example; the punctuation is only known after the broken vs non-broken status of the surrounding group is known:

let _ = Struct { x: 0, y: true };

let _ = Struct {
    x: xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx,
    y: yyyyyyyyyyyyyyyyyyyyyyyyyyyyyy,   //<- trailing comma if the expression wrapped
};

The formatting of match expressions is another case; we want small arms on the same line as the pattern, and big arms wrapped in a brace. The presence of the brace punctuation, comma, and semicolon are all dependent on whether the arm fits on the line:

match total_nanos.checked_add(entry.nanos as u64) {
    Some(n) => tmp = n,   //<- small arm, inline with comma
    None => {
        total_secs = total_secs
            .checked_add(total_nanos / NANOS_PER_SEC as u64)
            .expect("overflow in iter::sum over durations");
    }   //<- big arm, needs brace added, and also semicolon^
}

The printing algorithm implementation in this crate accommodates all of these situations with conditional punctuation tokens whose selection can be deferred and populated after it's known that the group is or is not broken.


License

Licensed under either of Apache License, Version 2.0 or 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.

Dependencies

~225–650KB
~16K SLoC