Hacker News Reader

Comment by ok_dad | Hacker News Reader

> The regex methods are unbelievably fast regardless of string length.

I immediately pinned regex as the winner, and here is why:

In Python, you can almost always count on specialized functionality in the stdlib to be faster than any Python code you write, because most of it has probably been optimized in CPython by now.

Second, I have to ask, why would someone think that regular expressions, a tool specifically designed to search strings, would be slower than any other tool? Of course it's going to be the fastest! (At least in a language like Python.)

View on HN

But it's not the fastest, it's actually incredibly slow and in general regexs are very slow. The key insight is that string processing in Python is very slow, and that regex here outperforms everything else because it's the only approach that is implemented in C.

With that insight it should follow that using another implementation in C should outperform even the regex, and indeed the following simple Python method that this article for whatever reason ignored vastly outperforms everything:

    def contains_vowel_find(s):
        for c in "aeiouAEIOU":
            if s.find(c) != -1:
                return True
        return False

That's because s.find(c) is implemented in C.

In my benchmark this approach is 10 times faster than using a regex:

https://gist.github.com/kranar/24323e81ea1c34fb56aff621f6c09...

Very nice find (pun intended). I added an update at the end with your solution and a link to this comment.

The more you learn about regexes, the more you learn that you can’t make general statements about their performance. Performance is contingent upon the engine, its algorithms, its implementation, the patterns, and the input.

I've always found regexes to do a string search faster than other methods, but there is always more to learn! Thanks for the lesson today.

Playing around, I found that the generator approach was competitive with this if you permuted the loop nest, and often even slightly faster (at the 1000 character length):

    def any_gen_perm(s):
        return any(c in s for c in "aeiouAEIOU")

I think the crux is that you want the inner loop inside the fast C-implemented primitive to be the one iterating over the longer string, and to leave the outer loop in Python to iterate over the shorter string. With both my version and yours, the Python loop only iterates and calls into the C search loop 10 times, so there's less interpreter overhead.

I suspect that permuting the loop nest in similar variations will also see a good speed up, and indeed trying just now:

    def loop_in_perm(s):
        for c in "aeiouAEIOU":
            if c in s:
                return True
        return False

seems to give the fastest result yet. (Around twice as fast as the permuted generator expression and your find implementation on my machine, with 100 and 1000 character strings.)

This is beautiful!!! Adding it to the end of the post right now.

For English samples, this should prove faster:

     def loop_in_perm(s):
        for c in "eaoiuEAOIU":
            if c in s:
                return True
        return False

Oh, that's a nice optimization!

> in general regexs are very slow

I really don't think this is true. If you assume that the string is ASCII, a well-optimized regex for a pattern of this type should be a tight loop over a few instructions that loads the next state from a small array. The small array should fit completely in cache as well. This is basically branchless. I expect that you could process 1 character per cycle on modern CPUs.

If the string is short (~100 characters or less, guessing), I expect this implementation to outperform the find() implementation by far as find() almost certainly will incur at least one more branch mispredict than the regex. For longer strings, it depends on the data, as the branchless regex implementation will scan through the whole string, so find() will be faster if there is an vowel early on in the string. Find still might be faster even if there are no vowels; the exact time in that case depends on microarchitecture.

For non-ASCII, things are a bit trickier, but one can construct a state machine that is not too much larger.

Putting benchmarks where my mouth is: https://github.com/tylerhou/benchmarks/blob/main/vowels-benc...

Writing implementations in C++, using a reasonable encoding of how an efficient regex compiler might compile the vowel regex, the regex implementation outperforms the loop-based implementations significantly in all cases except for long strings with vowels on M1 Max MBP.

Strings were generated randomly following the regex [0-9A-Za-z]. Short strings were between 5-20 chars. Long strings were 5000 chars.

The speedups are:

- Short strings & with vowels: regex is 2.6x faster

- Short strings & no vowels: regex is 5.9x faster

- Long strings & with vowels: loop is 329x faster. This is expected, as because the regex implementation is branchless, it must always search through the whole string.

- Long strings & no vowels: regex is 1.5x faster.

If you add an early return to the regex implementation, the regex with early return becomes strictly faster than the loop versions. The early return is slower than the no early return for the short strings & no vowels case because the extra branch has a cost. Full outputs in a comment at the bottom of the linked file.

Why do you loop over haystack multiple times though? If you iterate over long string once and write fixed loop with vowels checks in way that will be friendly to autovectorize optimization it might be faster and more idiomatic C or C++.

Thanks for reminding me, I meant to also benchmark the interchanged version. I updated the file above with the new benchmark results.

The original commenter called `find()` once per vowel, so that's why I benchmarked the regex against the less-idiomatic code.

The interchanged version (loop over haystack outside, over vowels inside) is (mildly) faster than all the regex versions except for short strings & no vowels.

One thing to note is that all the loop versions are not easily generalizable to non-ASCII strings, while the regex version is fairly easily.

I poked that loop in compiler explorer for few minutes and I think its allergic to autovectorization. Lets assume that properly vectorized regexp from proper language in the future wins :)

Sure, if you make a bunch of assumptions and manually implement how you think a regex will compile your code, allowing the optimizer to take your compile time implementation and make a finely tuned algorithm specifically for one use case, you can make something that outperforms what is kind of a dumb algorithm.

But if you use an actual regex the way people actually use them, using either the one provided by their standard library or one that is readily available then regex is pretty slow. Certainly, in principle a regex is neither fast or slow, it's a declarative description of a set of strings. Any claim about its performance rests on particular implementations.

For example you wrote out a benchmark that presumably is a lot faster than a naive search... but notice you didn't use the standard library <regex>, instead you manually implemented an algorithm to perform a search and then just decided that this is what a regex implementation would have done anyways (ignoring that by implementing it directly in source code, the optimizer can then fine tune it).

So now... there is a 10% chance that you genuinely didn't know that C++ has a <regex> library that you could have used, or you did in fact know that C++ has such a library and you chose not to you use it because... you also know it's very slow.

I did the benchmark myself using the standard library regex, boost regex, and PCRE2 regex, and all of them are about 5-15x slower than the simple loop:

https://gist.github.com/kranar/a3187cba00d57ba630b74b84f09b4...

> Sure, if you make a bunch of assumptions and manually implement how you think a regex will compile your code, allowing the optimizer to take your compile time implementation and make a finely tuned algorithm specifically for one use case,

But that's not what is happening. The only thing that the optimizer does is unrolls the loop that advances the state. Otherwise, the sequence of instructions is standard: two loads and some pointer arithmetic. There is no finely tuned algorithm. https://godbolt.org/z/fqdb5bssc

> But if you use an actual regex the way people actually use them, using either the one provided by their standard library or one that is readily available then regex is pretty slow.

Yes, most regex implementations are not designed to be efficient, they are designed for ergonomics (including supporting exponential-time features like lookaround). (Both boost::regex and PCRE2 support exponential-time features; std::regex is just not optimized at all.) This is well-known. https://swtch.com/~rsc/regexp/regexp1.html. That's why I said "well-optimized regex." Most regexes, in practice, are not well-optimized!

Maybe this is a semantic distinction between what you and I consider a "regex." I interpreted "regex" as how people in CS theory understand it; i.e., a finite automata that decides whether a string is in a given language. I think that's perfectly reasonable here, as the given problem is finding the fastest algorithm that decides whether a string is in the language generated by ASCII characters without the vowels. I think it's reasonable for me to use a (not hand-optimized) implementation that obeys that definition in my comparison.

>The only thing that the optimizer does is unrolls the loop that advances the state.

The idea that you'd dismiss loop unrolling as some kind of non-issue is incredibly baffling, you need only compare the benchmark built without optimizations enabled to the one with optimizations enabled, and you'll see that the unoptimized build is about half as fast!

Speaking of Python the most recent Python release has some significant performance increases in some scenarios because of some additional loop unrolling that has been exploited.

>That's why I said "well-optimized regex." Most regexes, in practice, are not well-optimized!

So then what exactly is the point of your argument? I point out that using regex's are slow, which almost anyone would reasonably assume means that using regex libraries tend to be slow.

Is the entire point of your comment that you can take a regex and then manually implement a special purpose algorithm for it that isn't slow? Was that the point of your comment, because if so you could have made that clear initially and saved both of us a lot of time.

>I think that's perfectly reasonable here, as the given problem is finding the fastest algorithm that decides whether a string is in the language generated by ASCII characters without the vowels.

No it's not reasonable at all because if you take your definition then there is nothing to use. Being a regular language is not something you can "use" or "not use", it's a property of a set of strings, you don't get to use it. The language consisting of all strings that have at least one vowel in them is a regular language, case closed, usage has nothing to do with it. In the sense that you're using it, any algorithm whatsoever that returns true or false for a string that matches some arbitrary regular language can be considered an implementation of a regex, for some fixed regular language. But of course this is a completely trivial argument which is why there's absolutely no point in discussing it.

In the non-trivial sense that everyone else uses it... regex's are libraries that people use that let them write a string representing a pattern and then the library returns an object that can be used to test whether or not a string matches that pattern, or can be used to search for a substring that matches a pattern along with a host of functionality. The claim in my original post is that these libraries, including the one used by the blog post, tend to be slow. They are, as you point out, designed for ergonomics and convenience, not for performance.

It is not at all reasonable to argue that some arbitrary implementation you divined for a particular regular language constitutes some argument that regular expressions are a high performance means of performing string matching, but at least you have clarified your position and shown it to be absolutely trivial and meaningless, and hopefully anyone who has read your post won't be misled into thinking you meant that regex libraries that are commonly used by actual engineers like Python's, or PCRE2, or boost or the standard library are fast... but I have my doubts about that.

> Second, I have to ask, why would someone think that regular expressions, a tool specifically designed to search strings, would be slower than any other tool? Of course it's going to be the fastest! (At least in a language like Python.)

I know that in Go doing string operations "manually" is almost always faster than regexps. In a quick check, about 10 times faster for this case (~120ns vs. ~1150ns for 100 chars where the last is a vowel).

Of course Python is not Go, but I wouldn't actually expect simple loops in Python to be that much slower – going from "10x faster" to "2x slower" is quite a jump.

Perhaps "yeah duh obvious" if you're familiar with Python and its performance characteristics, but many people aren't. Or at least I'm not. Based on my background, I wouldn't automatically expect it.

It's worth noting here that Go has a notoriously slow regexp implementation.

>Second, I have to ask, why would someone think that regular expressions, a tool specifically designed to search strings, would be slower than any other tool? Of course it's going to be the fastest! (At least in a language like Python.)

Flexibly, search string flexibly.

Why you think that we cannot achieve faster search in non flexible fashion?

Reply on news.ycombinator.com