Skip to content
This repository has been archived by the owner. It is now read-only.

Language ergonomic/learnability improvements #17

Open
aturon opened this issue Jan 31, 2017 · 61 comments
Open

Language ergonomic/learnability improvements #17

aturon opened this issue Jan 31, 2017 · 61 comments

Comments

@aturon
Copy link
Member

@aturon aturon commented Jan 31, 2017

Point of contact

@nikomatsakis @aturon @withoutboats

Overview

As part of the effort to lower Rust's learning curve, we are interested in pursuing a number of language improvements that will make Rust both nicer to use and easier to learn. The goal is to enhance productivity without sacrificing Rust's other core values. See the README for more details.

A kind of manifesto for this work was written up on the main blog.

Status

There are a lot of ideas in flight! The lang team has been focused on writing RFCs for the biggest changes; there are lots of opportunities to help either there, or with more narrow RFCs. If you are interested in the ideas listed below, please contact @aturon or others on the team to hook you up with a mentor for RFC writing!

The potential improvements are organized by broad areas of Rust:

Ownership/borrowing

RFCs

Sketches

Traits

RFCs

Sketches

Module and privacy system

Landed

RFCs

Sketches

Error handling

RFCs

Primitive types/core traits

Sketches

Type annotations/verbosity

RFCs

Sketches

FFI

Landed

  • Unions
    • FCP period for stabilization is complete!

Other

RFCs

  • Default struct field values
  • Coroutines, supporting nice async/await syntax
    • RFC postponed, pending some open questions; see the summary and related internals thread
    • There's now a working implementation and async/await notation built on top; we are considering landing that experimentally in the compiler.

More ideas

Ideas that didn't make it

@aturon
Copy link
Member Author

@aturon aturon commented Feb 1, 2017

Update: added the trait aliases RFC.

@aturon
Copy link
Member Author

@aturon aturon commented Feb 1, 2017

@aturon
Copy link
Member Author

@aturon aturon commented Feb 1, 2017

Update: added Delegation of impls

@aturon
Copy link
Member Author

@aturon aturon commented Feb 1, 2017

Update: added Add else match

@RUSshy
Copy link

@RUSshy RUSshy commented Feb 6, 2017

Please add class sometimes it's just faster and cleaner to use a class

@Ixrec
Copy link

@Ixrec Ixrec commented Feb 7, 2017

@RUSshy Could you be more specific about what "adding class" means to you? Most languages with "classes" have dozens of distinguishable features associated with their classes, and Rust already has analogues for many of them.

@RUSshy
Copy link

@RUSshy RUSshy commented Feb 7, 2017

pub trait Drawable
{
    fn draw(&self);
}

pub trait Updatable
{
    fn update(&self);
}

pub struct Entity
{
    pub id: i32,
    pub x: f64,
    pub y: f64,
    pub z: f64
}

impl Entity
{
    pub fn new(id: i32) -> Self
    {
        Entity{id: id, x: 0.0, y: 0.0, z:0.0}
    }
}

impl Drawable for Entity
{
    fn draw(&self)
    {
        println!("Draw: {}", self.id)
    }
}

impl Updatable for Entity
{
    fn update(&self)
    {
        println!("Update: {}", self.id)
    }
}

to this:

pub class Entity : Drawable, Updatable
{
    pub id: i32;
    pub x: f64;
    pub y: f64;
    pub z: f64;

    pub fn new(id: i32) -> Self
    {
        Entity{id: id, x: 0.0, y: 0.0, z:0.0}
    }

    fn update(&self)
    {
        println!("Update: {}", self.id)
    }

    fn draw(&self)
    {
        println!("Draw: {}", self.id)
    }
}

I love rust, i find it modern and really nice to use, but i think having class to merge struct and impl can help something with readability, because i find myself repeating and write duplicate code

@bestouff
Copy link

@bestouff bestouff commented Feb 7, 2017

@RUSshy why wouldn't this work:

pub struct Entity : Drawable, Updatable
{
...
}

?

(Otherwise to me, the two most useful items are the ones in "Early stage thoughts", i.e. non-lexical lifetimes and module system improvements)

@Nemikolh
Copy link

@Nemikolh Nemikolh commented Feb 7, 2017

@RUSshy It would give wrong impression to beginners in my opinion. One could believe that all the trait being implemented by a type would appear next to the class definition itself. But this is obviously not true (see Sync, Send or Into for instance).

Besides how would you deal with where clauses? What about bounds on type that apply to only specific trait impl but not others?

Would the code really be more readable at this point?

It seems, that you would definitely end up splitting your class into multiple impls as your code evolves.
Anyway, you can actually play with this class idea by writing a macro today. It would not look exactly as you have written it but would still save you a few braces. :)

@Ixrec
Copy link

@Ixrec Ixrec commented Feb 7, 2017

i think having class to merge struct and impl can help something with readability, because i find myself repeating and write duplicate code

I don't see any duplicate/repeated code in your example, just three impl blocks merged into the struct definition. I do see the appeal of that sugar unifying "obviously related" blocks of code, but it is just sugar, and it doesn't appear to save very much typing or confusion compared to some of the other proposed sugars such as delegation, and with that particular keyword there's definitely a risk of misleading novices. I agree with Nemikolh that experimenting with a "class macro" would make more sense for now.

Though we should probably move this discussion to https://users.rust-lang.org/ or https://internals.rust-lang.org/ since it's at most a very tiny piece of the puzzle this issue is about.

@withoutboats
Copy link

@withoutboats withoutboats commented Feb 7, 2017

My impression is that thread is for a high-level coordination of the work toward this roadmap goal. Brainstorming specific language feature ideas should be at either https://internals.rust-lang.org or an issue on the RFC repo https://github.com/rust-lang/rfcs/issues

@aturon
Copy link
Member Author

@aturon aturon commented Feb 10, 2017

Update: adding reference to this thread on internals which discusses some additional ideas.

@aturon
Copy link
Member Author

@aturon aturon commented Feb 10, 2017

@emberian
Copy link
Member

@emberian emberian commented Mar 3, 2017

One idea that would make working with capn proto and similar systems so so much nicer would be something like attributes as exist in C# or Python, where one can use field access syntax but underneath call functions to get/set the field. The current state of the art in Rust is explicit getters and setters, which leads to terrible ergonomics for code interacting with data at the edges of the process boundary, to the point of Java-level verbosity. In the past (pre-1.0) I know there was some discussion and rejection of this, but I'd like to bring it up again.

@aturon
Copy link
Member Author

@aturon aturon commented Mar 3, 2017

Updated with a link to the new blog post and several in-progress RFCs. If there are pre-RFC internals threads that should be cited, please leave comments and I'll add them!

@F001
Copy link

@F001 F001 commented Mar 4, 2017

The "Delegation of impls" and "Delegation of implementation" both point to rust-lang/rfcs#1806. They are duplicate.

@crumblingstatue
Copy link

@crumblingstatue crumblingstatue commented Mar 4, 2017

I see default struct field values is listed here.

Are there any learnability improvements that it provides, or are the benefits purely ergonomic?

In the latter case, I find it odd that it would be tracked here. There are lots of ergonomic improvements (e.g. default/keyword args) that are currently postponed, and we couldn't possibly fit them all into this year's roadmap.

@aturon
Copy link
Member Author

@aturon aturon commented Mar 6, 2017

@crumblingstatue I went through the open lang RFCs looking for any that touched on ergonomics, just to have them tracked in a central place.

I'm open to keeping this tracking issue more focused, but I'm not sure what the best "cutoff" line should be.

@LegNeato
Copy link

@LegNeato LegNeato commented Mar 7, 2017

What's the process for nominating an RFC for this? Sorry if I missed it.

As a Rust n00b I'd like to nominate rust-lang/rfcs#1078 🍻

@withoutboats
Copy link

@withoutboats withoutboats commented Mar 7, 2017

That issue has an RFC which will probably be merged soon :-) rust-lang/rfcs#1869

@hgrecco
Copy link

@hgrecco hgrecco commented Mar 9, 2017

I find this example from the book quite annoying:

enum Message {
    Quit,
    ChangeColor(i32, i32, i32),
    Move { x: i32, y: i32 },
    Write(String),
}

fn process_message(msg: Message) {
    match msg {
        Message::Quit => quit(),
        Message::ChangeColor(r, g, b) => change_color(r, g, b),
        Message::Move { x: x, y: y } => move_cursor(x, y),
        Message::Write(s) => println!("{}", s),
    };
}

If you are matching an Enum variant, cannot Message:: (in this case) be implicit?

fn process_message(msg: Message) {
    match msg {
        Quit => quit(),
        ChangeColor(r, g, b) => change_color(r, g, b),
        Move { x: x, y: y } => move_cursor(x, y),
        Write(s) => println!("{}", s),
    };
}

I have seen ways to mitigate this (such as use Message::*) but I think some implicitness here will be nicer.

@Manishearth
Copy link
Member

@Manishearth Manishearth commented Mar 9, 2017

Please file bugs on the Rust or rfcs repo, this is for tracking efforts to improve readability and not discussing the specifics themselves.

(In that specific case, it used to be the opposite, actually, but then everyone namespaced enums as QuitMessage, etc.)

@aturon
Copy link
Member Author

@aturon aturon commented Mar 9, 2017

@Manishearth @hgrecco to clarify, I'd actually recommend opening threads on https://internals.rust-lang.org/ with ideas, and then posting a link here. I am using the comment thread/headers to track ideas floating around.

@Manishearth
Copy link
Member

@Manishearth Manishearth commented Apr 15, 2017

how can we have hierarchies of traits without forcing the user to acknowledge the complexity all the way down?

This brushes on a very related issue with trait usability. We have a lot of "drill down" traits, like Eq, Serialize, Hash, etc. Most implementations are a #[derive()] away (with stable derive we may see an explosion of more such traits like this). Some are special. It's often a problem that a crate you use doesn't derive a particular trait (especially for non-std traits. this problem is assuaged by the automatic features RFC, but not fixed) and you need to go add it. Eq and Ord are the worst offenders here, because folks seem to often implement the partial ones but not the total ones.

I'm not really proposing a proper solution here, just mentioning that these traits have a class of problems associated with them. The class of problems comes from the fact that these traits are usually structurally-derived but sometimes need manual impls, and currently the structural-derive needs explicit opt in (which makes sense, really).

Throwing a half solution into the mix: in the past we have discussed things like deferred derive where derive(Trait) defers codegen until the crate that actually needs it is compiled. Similar mechanisms could be used to drill down to other objects, for example you could do something like #[derive(Hash)] pub struct Foo {x: u8, #[inline_derive(Hash)] y: somecrate::SomeType}, where SomeType doesn't impl Hash, but the compiler will try to structural derive Hash as deep as it needs to, if it can, and inline that code. Serde already has a (hacky) solution for this where you can specify a custom serializer function in an attribute which will be used instead. That feature has other, less hacky use cases, but when used to solve this problem you still have to do some annoying manual writing of derives.

This "solution" if used too much would have codesize bloat problems, and there's also the whole interaction with safety (you don't want someone to force-derive on an unsafe abstraction), but it's something that can be built upon.

@djc
Copy link

@djc djc commented Apr 16, 2017

It seems like @Manishearth's comment is really concerned with what I call deriving-at-a-distance. However, if derive(X) today is really syntax sugar for writing out an impl X for T, should Rust not provide you with the option to derive X for T for use in the local crate even if T's origin crate doesn't derive X? I was thinking about this recently in the context of Askama templates and wanting to implement a template for some context in another crate (mostly to allow plugin templates as dylibs).

@Manishearth
Copy link
Member

@Manishearth Manishearth commented Apr 16, 2017

should Rust not provide you with the option to derive X for T for use in the local crate even if T's origin crate doesn't derive X?

Well, you have coherence problems there.

Unless you're the one defining the deriving stuff. Crates like serde/heapsize may have to do manual tinkering to implement these traits since they themselves can't apply the deriving syntax to things from std.

My solution to the drill down issue was to ignore the part of generating impls at a distance, and just make the codegen accessible. Impls come with a whole host of other problems (coherence) 😄

Anyway, this is kind of off topic I guess.

@eddyb
Copy link
Member

@eddyb eddyb commented Apr 16, 2017

@Manishearth Crate introspection is possible after loading their metadata, so perhaps a feature along the lines of #[derive(Serialize, Deserialize)] use std::ops::Range; could be added.

@Manishearth
Copy link
Member

@Manishearth Manishearth commented Apr 16, 2017

Yep. We have discussed this before, I think 😄

I understand that it's feasible, it's just designing the exact API (and figuring out if we actually need it) that needs doing.

@burdges
Copy link

@burdges burdges commented Apr 16, 2017

Is there any problem with using specialization to reduce the boilerplate around PartialEq and PartialOrd?

impl<T> PartialEq<Self> for T where T: Ord {
    default fn eq(&self, other: &Self) -> bool { Ord::cmp(self, other) == Ordering::Equal }
}
impl<T> PartialOrd<Self> for T where T: Ord {
    default fn partial_cmp(&self, other: &Self) -> Option<Ordering> { Ok( Ord::cmp(self,other) ) }
}

I suppose impl<T> Eq for T where T: Ord { } might work too, but maybe you want Eq to serve as a reminder.

@nikomatsakis
Copy link

@nikomatsakis nikomatsakis commented Apr 19, 2017

@eddyb

Crate introspection is possible after loading their metadata, so perhaps a feature along the lines of #[derive(Serialize, Deserialize)] use std::ops::Range; could be added.

Can you clarify what you envision this syntax as doing? It seems to me that the interesting (and hard) problem here is deciding what to do about coherence. Ah, would this be something we would do in the crate that defines Serialize?

@eddyb
Copy link
Member

@eddyb eddyb commented Apr 19, 2017

@nikomatsakis Yupp, giving the derives an approximation of the original AST, from crate metadata.

@shepmaster
Copy link
Member

@shepmaster shepmaster commented Apr 21, 2017

Is there a more detailed location to add thoughts about "Allow owned values where references are expected" ?

@aturon
Copy link
Member Author

@aturon aturon commented Apr 24, 2017

@shepmaster There's not currently, but you're welcome to make an internals thread on the topic (and I'll link it here)!

FWIW, the strawman proposal is to automatically coerce via AsRef or AsMut, which will have the effect of allowing owned where borrowed is expected. This automatic coercion is not like full-blown implicit coercions, because the types involved (going from a reference to a reference) strongly limit what you can do, much like with Deref and deref coercions.

We already use this pattern in std, but it'd be nice to have it apply uniformly and not have to muck up signatures with it.

@aturon
Copy link
Member Author

@aturon aturon commented Apr 29, 2017

@shepmaster
Copy link
Member

@shepmaster shepmaster commented Apr 29, 2017

you're welcome to make an internals thread on the topic

@aturon I've created a thread as requested. We'll see where it goes ;-)

@aturon
Copy link
Member Author

@aturon aturon commented Jun 7, 2017

Updated with the latest status info.

@le-jzr
Copy link

@le-jzr le-jzr commented Jun 9, 2017

Any news regarding PartialEq/PartialOrd? I find @burdges's solution elegant and nobody suggested any problems.

It would also solve the weirdness where current bound Ord: PartialOrd<Self> means "for a type to be Ord, it must be PartialOrd first", whereas the true mathematical relation is that Ord is automatically PartialOrd<Self>, Eq and PartialEq<Self> as well, by virtue of Ord subsuming all those properties, so encoding the relationship in a trait bound actually makes no sense.

@sfackler
Copy link
Member

@sfackler sfackler commented Jun 9, 2017

@le-jzr I'd love to have automatic implementations like that - it'd also be possible to implement Read automatically for BufRead types. We may need to wait until after specialization stabilizes to be able to add them, though, since I believe you need to enable the feature to specialize an implementation.

@glaebhoerl
Copy link

@glaebhoerl glaebhoerl commented Jun 12, 2017

I think this RFC ("Allow comparisons between integers of different types") is also relevant here. Much like default binding modes for match, it takes a case where the compiler today issues errors and forces people to manually write boilerplate and makes it just do the obviously right thing instead. (Except in this case, the manual boilerplate is also highly error-prone.)

@suhr
Copy link

@suhr suhr commented Jun 23, 2017

use inside trait definitions.

Local use statements are an extremely cool rust feature. But unfortunately, there're some places where you can't use them. One of such places are trait definitions.

Long story short, I want to be able to write something like this:

trait Foo {
    use std::iter::IntoIterator;
    // ...
}
@crumblingstatue
Copy link

@crumblingstatue crumblingstatue commented Jun 23, 2017

@suhr: That's rust-lang/rfcs#1976, and it's about to get postponed.

@burdges
Copy link

@burdges burdges commented Jun 23, 2017

It turned out trickier than folks thought. I think you can mostly work around it using submodules judiciously.

mod UseNo1 {
use std::iter::IntoIterator;
pub(super) trait Foo { ... }
} // UseNo1
pub use UseNo1::*;
@suhr
Copy link

@suhr suhr commented Jun 23, 2017

it's about to get postponed

Well, that's unfortunate.

I think you can mostly work around it using submodules judiciously.

It's an ugly and awkward kludge that is not worth it.

@dobkeratops
Copy link

@dobkeratops dobkeratops commented Jul 25, 2017

I wonder if this would get a different reaction in the context of this thread.
RFC: rust-lang/rfcs#2063
link to text:
https://github.com/dobkeratops/rfcs/blob/infer-function-signatures-from-trait-declaration-into-'impl's/text/infer%20function%20signatures%20from%20trait%20declaration%20into%20impls.md

alternating between languages, I continue to find the need to micromanage traits to be irritating (i.e. the single function trait case), but strangely in haskell the typeclasses don't feel as bad. It's because the there, defining and referring to the 'type class' saves you writing details out again. Initially I found it odd, but their function signatures being split are actually clearer.

Hence this PR.

People are complaining 'you read it more than you write it' but it's the repetition that is annoying: if I'm specifying the types , I would rather it figures out the trait from them. If I've specified the trait, it makes sense to use that to infer the types. Users are going to read the trait declaration itself more surely, which has all the types, and you need to refer back to that to implement anyway, surely. Aren't you just taking up more screenspace in a useless way displaying it again. The trait defs are very easy to grep for ( i have that all working nicely in emacs), and there's rustdoc too.

Something else to mention is this mental flip of arguments in the single-function-trait case (e.g. used for overloading)..

I wish you could write the impl the other way round so that it's impl Type:Trait<..> , so the types appear in the same order as the function definition
e.g.

    // "I want to implement LHS.add(RHS)->Output"
    impl LHS:Add<RHS> { type Output=..;    // trait: "LHS,add, RHS, Output.."
     //  |          |           |
        fn add(&self,rhs){... }       //function: "add LHS,RHS,Output"
    }

wouldn't that be so much bettter? the mental flip between the intent "LHS,add,RHS" and expressing the trait makes the trait so much more annoying.

Again the 'multi-parameter-typeclass' case in haskell is more pleasant, possibly because you get to choose the order, so in the case of 'a single function class' you can write the class-types in the same order as the function types. (there's extra symmetry from the fact they have no 'self', but with better ordering I don't think special-Self would be a pain)

Combined with eliding the signature types, that might be enough for me to achieve peace with this system.

@burdges
Copy link

@burdges burdges commented Jul 25, 2017

I believe mutability, references, lifetimes, etc. change that calculation somewhat from Haskell, so maybe wait until folks become more comfortable with lifetime elision for impls rust-lang/rust#15872 ?

At least one can then deduce if type elision ever confuses the reader familiar with lifetime elision. If so, type elision could be restricted to cases where lifetime elision works. And maybe it should be restricted to not doing muts anyways.

@dobkeratops
Copy link

@dobkeratops dobkeratops commented Jul 26, 2017

most of the time operators are just reference inputs, and the lifetime elision kicks in;
Thinking in C++, I'm used to 2 main cases which are both quite simple:

  • 'accessors' ('self assumption' which Rust successfully allows elision for - this is great) ,
  • and 'no returned references'/'independent result' (harder to give this a catchy name, but basically 'there are no returned references to worry about' - any pointers are fully owned and managed by the result; I appreciate the need to mark up the difference, however I still suspect there might be an easier way to do this: often there's still a gap between my intent ("no escaped references", and the nature of the markup'(ok, show it which references in the inputs correspond to each other, specifically')

So basically the times you actually need to write lifetimes, you're doing something rarer (IMO).

Eliding information in the obvious cases will give you more 'cognitive budget' to spend on the unusual.

Also the information is still there and will surface if you hit compile.

I argue people reading code are more likely to be reading the trait, which does have the full types.

also I argue haskell's split is actually helping readability, regardless of how much type information there is

Conversely if I must write out the types then I'd prefer it to do the opposite: figure out the trait from the types.

This reminds me of years of people in C++ claiming "writing out the iterator types is good for you..", delaying the deployment of 'range based for', 'auto' ..

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Projects
None yet
Linked pull requests

Successfully merging a pull request may close this issue.

None yet