[chrismorgan]

Another point where it’s doing something that to me as a Rust expert is obviously inferior: it’s using Unicode-aware string stuff although anything non-ASCII will either be ignored (if non-alphabetic) or panic (if alphabetic). It’d certainly be better to treat the input throughout the program as a sequence of bytes rather than as UTF-8.

This type of thing reminds me of the three articles ending in https://fitzgeraldnick.com/2018/02/26/speed-without-wizardry... (which has links to the first two parts of the saga), where one guy rewrote stuff in Rust for performance, another demonstrated how it was possible to make the JavaScript version faster than the Rust by some algorithm changes and by various painful and fragile tricks requiring detailed knowledge of the runtime environment, and finally the first guy applied the applicable parts of that back to the Rust, after which it handily beat the JavaScript again while also being more consistent and dependable.

[nerdponx]

It's worth distinguishing between algorithmic optimizations, optimizations that generally take advantage of the language standard/runtime, and optimizations that are highly specific for one machine/platform/implementation. It's also worth keeping track of relative programmer effort to optimize.

I think most people are moderately-optimized benchmarks, i.e. moderate effort expended relative to baseline implementation effort.

That is, people are interested in getting the most performance out of the least amount of effort.

Obviously some people want and need to care about extreme peak optimization. But if you are writing benchmarks for a wide audience, that probably should not be your priority.

[mst]

The author was pretty explicit in the article that the rust implementation was suboptimal.

I'm sure if you submitted a better implementation he'd be happy to add it in.

[dwohnitmok]

Eh kind of... The author was talking about a current iteration of the Rust code being suboptimal, not the current one, which the author believed was well-optimized.

> The objective is to get the fastest implementation possible, and having optimised Java and Rust implementations to compare against is a motivator to keep going until there really isn’t anything else that can be tried!

> Do you think Common Lisp can run as fast as well-optimised Rust? Read on if you want to find out.

[mst]

That was based on all of the rust feedback the author actually got from rust people who talked to him as opposed to people who complained in HN comments.

Hence if you want it to get even better you should provide feedback to the author, and while I absolutely respect your right to decide you can't be bothered, I don't think it's his fault that he's using code that was well-optimized according to the rustaceans who -did- talk to him.

[dwohnitmok]

Oh I'm sure there's a better way to give the author feedback than sniping in HN comments.

I'm reacting to your comment here:

> The author was pretty explicit in the article that the rust implementation was suboptimal.

This is not the way the article portrays it.

[mst]

Perhaps you missed the part where he said you'd need to rewrite it to use a trie to get a properly optimised one?

[dwohnitmok]

That's only vis-a-vis the Main.java Java implementation, which has a different algorithm, not implementation. The author is not referring to the CL implementation (note that the author explicitly calls out Main2.java, CL, and the Rust versions as being the most similar and worthy of being benchmarked). The Main.java one is more of a curio than anything else, which the author threw in because it was the first implementation they wrote (the article acknowledges it has a different algo and is therefore not directly comparable to the other two). This is not what the author is referring to when they say "well-optimised" (note that the article explicitly excludes Main.java from a lot of its valid comparisons, e.g. "Because the Rust code implements a similar algorihm to Java’s Main2, NOT Main, we should not conclude that Java can beat Rust in speed!")

The main comparison of the article is Main2.java, main.rs, and main.lisp, as the author both calls out in the article and the attached GitHub repo (as is apparent in the author's choice of optimizations; if it was a comparison of algos, then the CL and Rust versions would be rewritten to use tries as well).

The point is the author did not explicitly call out this current Rust iteration as explicitly suboptimal RE CL. The closest the article comes to calling the Rust iteration suboptimal is

> However, the Java and Rust implementations were, as CL’s, written without much thought given to performance, after all the description of the original study which introduced the problem asked participants to focus on correctness and readability, not performance.

which is referring to the previous iteration of the code, not the current one.

(It's also evident from Cryptonic and chrismorgan's comments they are talking about implementation-level concerns, not algo-level ones such as Main2.java vs the other implementations)

[dwohnitmok]

Whoops I meant "previous iteration"

[brabel]

> it’s using Unicode-aware string stuff

Rust uses UTF-8 internally for Strings, so it's very efficient to parse a file into a String, then using slices to go through it... this is probably the best you can get as parsing ASCII input as UTF-8 is very efficient (the 0-bit is always zero in ASCII, the unicode decoder only needs to check that's the case for every byte, so it's not some kind of complicated computation it's doing to decode)...

If you use bytes for everything, you will make the whole code much harder to follow and it still won't run faster.

Check for yourself: https://github.com/renatoathaydes/prechelt-phone-number-enco...

[chrismorgan]

The code will be somewhat faster (I don’t care to predict how much) from removing the variable-width character encoding in favour of bytewise access. Yes, pure ASCII stuff has some fast paths in string access, but they’re still decidedly slower than the fixed-width encoding that is [u8]. Using strings also gives the incorrect impression that it can cope with non-ASCII.

The code will be easier to follow if you use bytes throughout, because currently it’s a mixture of bytewise and charwise operations, so that you need to think a little about whether you’re dealing with char or u8 in each place (and half of them are even mislabelled); and there are suitable alternative ASCII methods for every place that uses charwise methods (e.g. char.is_digit(10) → u8.is_ascii_digit()) so that no extra burden is added. In the end the only place slightly complicated by it is printing the solutions, but more code will have been decomplicated—hotter path code, too—so that it’s easily worth it.

I don’t know where the code you’re citing came from, it’s newer than what’s on the master branch but in its changes includes some pretty bad stuff like DIGITS, using &str for something that is always a single-character ASCII digit, accessed by already having had the digit as a u8 and turning it back into a string prematurely. Admittedly the optimiser is going to fix a fair bit of that badness, but not all.

[brabel]

There are a few pull requests that claim to make the code faster, but you can run the benchmarks and see none of them actually did. Why not try to improve the code I posted above and make the apparently small changes you want to make and check if it's faster or not.

I've tried a few myself and I am almost sure your hints will not work.

> some pretty bad stuff like DIGITS, using &str for something that is always a single-character ASCII digit, accessed by already having had the digit as a u8 and turning it back into a string prematurely.

The end result needs to be printing the strings so I don't see how you can work around that. Can you at least post your code doing that in a way that won't totally destroy the performance gains you may have obtained elsewhere?