Index by Regex?

[burntsushi]

I’m honestly not a huge fan. It’s a nice trick, but it encourages task failure whenever the regex fails to match.

If you really want the indexing, I see no reason why this has to be in the regex crate. A wrapper type is a not a burden here, and it could even be less typing. (Just create your own constructor that calls Regex:new.)

Also, I think we should be moving toward stabilizing the regex crate. It’s rather popular, so we’ll risk making it de facto stable if we wait too long.

[kstep]

Yes, I get your POV. You are basically right about panics. In Ruby when regex “indexing” fails, it just returns nil, so it’s OK to have such sugar, but in Rust it’s not that simple. Now I think this feature doesn’t fit Rust ideology after all. And I don’t think a crate with such feature will be used by any significant number of users, so it just doesn’t worth efforts to create it.

[ArtemGr]

In Ruby when regex “indexing” fails, it just returns nil, so it’s OK to have such sugar, but in Rust it’s not that simple.

In Rust that would be Option<something>.

Now I think this feature doesn’t fit Rust ideology after all.

Right now Rust is in it’s “be minimalistic before 1.0” stage, everyone’s holding they breath waiting for Rust to become stable, but I’ve heard voices that Rust will start to grow some extra weight later.

Writing

if let Some (match) = a_string[regex!(r"...")] {}

seems useful to me, it’s easier to remember than the names of the methods regex crate uses (I keep hitting the docs to check them).

I’d vouch for making a crate even if there would be a few users. At least the gist will be easy to find that way.

[kstep]

You can’t do return Option<something> from index. See the signatures (simplified a little):

trait Index<Idx> {
  type Output;
  fn index(&'a self, index: &Idx) -> &'a Output;
}

You have to return a reference to something, no owned data can be returned from index(). And the reference inherently must to be something from self (due to lifetime restrictions).

[burntsushi]

In Rust that would be Option<something>.

Right, but when you’re indexing, the regex match has to be unwrapped which will cause task failure if the match fails. I suppose an alternative to this is to return a zero length slice if the regex doesn’t match. (This will conflate zero-length matches with non-matches, but that’s probably OK to gloss over for a convenience such as indexing.)

seems useful to me, it’s easier to remember than the names of the methods regex crate uses (I keep hitting the docs to check them).

I agree it’s a neat trick, I’m just trying to say that this can be done entirely outside of the regex crate with almost no downsides. (Wrapper types are usually inconvenient because you have to re-implement all of the behavior of the underlying type, but in this case, all you really need it for is construction.)

Also, more importantly, I’m pretty sure that your if let Some(match) = a_string[regex!(r"...")] {} construct is not possible. (Whoops, @kstep beat me to it.)

[kstep]

I suppose an alternative to this is to return a zero length slice if the regex doesn’t match.

Hmm, that’s not that elegant, but it may work. Not sure about ergonomics of such implementation, but I think it worth to make a PoC crate and give it a try in real world.

[ArtemGr]

Ahh. Too bad one can’t get an Option from an index. That makes it less useful than I thought.

[kstep]

I just published simple PoC crate regindex (repository).

[ArtemGr]

Apparently one can’t make it to work as transparent as in Ruby (e.g. without the ReIdx wrapper) without it being a part of the regex crate?

[burntsushi]

My suggestions to make this more ergonomic outside of the regex crate seem to have gotten lost in the noise. I filed an issue with suggestions: https://github.com/kstep/regindex/issues/1

[ArtemGr]

If we’re going so far as to define a new macro, rei!, then why bother and not define matches! ("hello", r"el"), returning an Option to boot. So far I feel this attempt falls short of the Ruby version.

[burntsushi]

I’m here to help you make the best with the tools that we have. I’m not a fan of indexing at all (I can’t imagine when I would use it in real code because it encourages task failure when the regex doesn’t match), but I was trying to work within the parameters of the OP. Certainly, some other macro that doesn’t use indexing at all might be more convenient, although I’m not convinced that it would belong in the regex crate proper.

[ArtemGr]

Sure. And thank you very much. I’m just venting some of my concerns, like not being able to properly implement the indexing operator outside of the regex crate. @kstep, it’s due to the http://www.reddit.com/r/rust/comments/2sg60s/only_traits_defined_in_the_current_crate_can_be/, right?

[kstep]

Yes, that’s the whole point of wrapper newtype in Rust: to work around “trait or type must be in current crate” rule.

[kstep]

I more or less finished the experimental version of regindex (github). Let’s see what happens next

[ArtemGr]

I’m not a fan of indexing at all (I can’t imagine when I would use it in real code because it encourages task failure when the regex doesn’t match)

I have to point out that it doesn’t use the task failure, it just returns an empty slice.

Also, after extending the syntax a bit, this feels quite useful to me:

let foo = &uri[ri! (regex! (r"^/path/(\w+)/$$"), 1)];

Arguably it’s intuitive, because taking a slice from a &str is obvious here. One doesn’t need to guess at the types, errors and how to handle them, we just get a slice, either an empty one or with the matched group.

If regex! dependency wouldn’t have been a problem, I’d even made it into

let foo = &uri[ri! (r"^/path/(\w+)/$$", 1)];

Yes, that’s the whole point of wrapper newtype in Rust: to work around “trait or type must be in current crate” rule.

I really wish Rust’ve choosed the other way 'round the orphans, allowing one to extend the types freely as in Google Go and Scala (and Haskell?).

[burntsushi]

Go uses structural subtyping, which is a completely different approach to polymorphism.

If you have a chance to talk to an experienced Haskell programmer, ask them about orphan instances and they’ll be eager to share a war story. More seriously though, the orphan rules has been the subject of a lot of attention. @nikomatsakis has a great write up on it: http://smallcultfollowing.com/babysteps/blog/2015/01/14/little-orphan-impls/

[ArtemGr]

I did some Haskell myself, a long time ago, and speaking of war stories orphan rules might be the least of one’s worries. Write up you mention only explores the one side of the coin and we see that it’s a troublesome side. There might be more truth on the other side, at least it hasn’t been experimentally explored that much in Rust, I take it? Scala choosed the other side (the Dark Side, woo! ) and it nailed the extensibility problem square with it’s implicit classes, IMHO, although the Rust story might be different.

[kstep]

I write production Scala code now, and what I want to say about implicit bounds, while they are very versatile and often useful, they are also often really confusing. I don’t like the idea of IDE to be a must have tool in order to program in some language, because without its type hints it’s often very difficult to say what’s really happens in your program. It often takes much time and effort just to make out where your implicit instances come from.

Upd. I actually like both Scala and Rust, but I think Rust approach in regard to orphan instances rules (while not being as versatile as Scala’s) is more right, in a sense it’s more obvious for a programmer, more explicit, and thus less error prone and easier to comprehend.

[system]

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.