Arc Forumnew | comments | leaders | submitlogin
3 points by shader 1380 days ago | link | parent

Thanks for sharing! I always like ideas that integrate previously distinct concepts into a single recursive hierarchy.

I am curious what makes it slow? Is it just an implementation detail, or something fundamental to the algorithms / concepts?



3 points by rocketnia 1378 days ago | link

There are a number of factors going into what makes it slow (and what might make it faster). I'm particularly hopeful about points 4 and 5 here, basically the potential to design data structures which pay substantially better attention to avoiding redundant memory use.

By the way, thanks for asking! I'm reluctant to talk about this in depth in the documentation because I don't want to get people's hopes up; I don't know if these optimizations will actually help, and even if they do, I don't know how soon I'm going to be able to work on building them. Still, it's something I've been thinking about.

1. Contracts on the whole

Punctaffy does a lot of redundant contract checking. I have a compile-time constant that turns off contract checking[1], and turning it off gives a time reduction in the unit tests of about 70%, reducing a time like 1h20m to something more like 20m. That's still pretty slow, but since this is a quick fix for most of the issue, it's very tempting to just publish a contractless variation of the library for people who want to live on the edge.

2. Whether the contracts trust the library

Currently, all the contracts are written as though they're for documentation, where they describe the inputs and the outputs. This constrains not just that the library is being used with the correct inputs but that it's producing the correct outputs. Unless the library is in the process of being debugged, these contracts can be turned off.

(Incidentally, I do have a compile-time constant that turns on some far more pervasive contract-checking within Punctaffy to help isolate bugs.[2] I'll probably want it to control these output-verifying contracts as well.)

3. The contract `hypernest/c`

One of the most fixable aspects here is that my `hypernest/c` contract checks a hypernest's entire structure every time, verifying that all the different branches zip together. If I verified this at the time a hypernest was constructed, the `hypernest/c` contract could just take it for granted that every hypernest was well-formed.

4. Avoidable higher-dimensional redundancy in the data structure

Of course, even without contracts, 20 minutes is still a pretty long time to wait for some simple tests to compile. I don't want to imagine the time it would take to compile a project that made extensive use of Punctaffy. So what's the remaining issue?

Well, one clue is a previous implementation of hypersnippets I wrote before I refactored it and polished it up. This old implementation represented hypertees not as trees that corresponded to the nesting structure, but as plain old lists of brackets. Every operation on these was implemented in terms of hyperstacks, and while this almost imperative style worked, it didn't give me confidence in the design. This old implementation isn't hooked up to all the tests, but it's hooked up to some tests that correspond to ones that take about 3 minutes to run on the polished-up implementation. On the old list-of-brackets implementation, they take about 7 seconds.

I think there's a reason for that. When I represent a hole in a hypertee, the shape of the hole itself is a hypertee, and the syntax beyond each of the holes of that hole is a hypertee that fits there. A hypertee fits in a hole if its low-degree holes match up. That means that in the tree representation, I have some redundancy: Certain holes are in multiple places that have to match up. Once we're dealing with, say, degree-3 hypertees, which can have degree-2 hypertees for holes, which can have degree-1 hypertees for holes, which have a degree-0 hypertee for a hole, the duplication compounds on itself. The data structure is using too much space, and traversing that space is taking too much time.

I think switching back to using lists of brackets and traversing them with hyperstacks every time will do a lot to help here.

But I have other ideas...

5. Avoidable copying of the data structure

Most snippets could be views, carrying an index into some other snippet's list of brackets rather than carrying a whole new list of brackets of their own. In particular, since Punctaffy's main use is for parsing hyperbracketed code, most snippets will probably be views over a programmer's hand-written code.

6. The contract `snippet-sys-unlabeled-shape/c`

There's also another opportunity that might pay off a little. Several of Punctaffy's operations expect values of the form "snippet with holes that contain only trivial values," using the `snippet-sys-unlabeled-shape/c` contract combinator to express this. It would probably be easy for each snippet to carry some precomputed information saying what its least degree of nontrivial-value-carrying hole is (if any). That would save a traversal every time this contract was checked.

This idea gets into territory that makes some more noticeable compromises to conceptual simplicity for the sake of performance. Now a snippet system would have a new dedicated method for computing this particular information. While that would help people implement efficient snippet systems, it might intimidate people who find snippet systems to be complicated enough already.

It's not that much more complicated, so I suspect it's worth it. But if it turns out this optimization doesn't pay off very well, or if the other techniques already bring Punctaffy's performance to a good enough level, it might not turn out to be a great tradeoff.

---

[1] `debugging-with-contracts-suppressed` at https://github.com/lathe/punctaffy-for-racket/blob/399657556...

[2] `debugging-with-contracts` at https://github.com/lathe/punctaffy-for-racket/blob/399657556...

-----

2 points by shader 1360 days ago | link

I want to think more about this and continue the conversation, but I'm worried the reply window will close first.

Since I presume your input space is relatively small (the AST of a program, which usually only has a few thousand nodes), it sounds like you have some sort of state-space explosion. Your comment about recursive matching of hypertees sounds like the biggest problem. Just a shot in the dark (having not studied what you're doing yet), but is there any chance you could use partial-order reduction, memoization, backtracking, etc. to reduce the state-space explosion?

I could be wrong, but most of the other optimizations sounded like they address constant factors, like contract checking. But then I don't know much about how contracts work; I guess the verification logic could be rather involved itself.

If the window closes, maybe we could continue at #arc-language:matrix.org

-----

3 points by rocketnia 1354 days ago | link

I'm sorry, I really appreciate it, but right now I have other things I need to focus on. I hope we can talk about Punctaffy in the future.

-----