hyperscan

djson), although others (chiefly Daniel Lemire and John Keiser) now do all the real work. It's still my design under the hood (although the "On Demand" stuff is significantly new and different). I built the first prototypes of simdjson because Daniel trolled me with a paper from Microsoft and I was bored.

ions to that, but fundamentally I bet Hyperscan could be made to slow down in a similar (linear) way to re2.

ls) to pattern analysis for better literal optimizations and more involved SIMD algorithms and better performance in the face of Unicode. (In addition to supporting a lot more of Unicode than GNU grep does.) Making programs run faster often leads to greater code complexity. Such a phenomenon, ironically, is easy to observe when comparing GNU tools with other coreutils implementations from BSD or busybox. Thirdly, the way in which you counted lines is weirdly arbitrary. For example, GNU grep uses a lot of code from gnulib, including but not limited to, an entire regex engine. (It used to be part of GNU grep proper, so if you had done this analysis a few years ago, your LOC count would be at least double than what you came up with. Which supports my previous point: GNU grep's special lazy DFA internal regex engine only exists for performance reasons.) Fourthly, I made a decision to make a lot of ripgrep's internals generic over both the regex engine and the printer. This was to make it possible for others to use those libraries, but also to make it easier to support plugging other regex engines into ripgrep. The official ripgrep distribution, for example, comes with both Rust's regex engine and PCRE2. But it goes beyond that. Someone even maintains a fork of ripgrep[1] that uses Hyperscan[2] to make it faster in some cases. The fact that this is even plausible to do is only because much of ripgrep's core is, in fact, generic. Making things generic tends to require more code and more complexity. I made this trade off intentionally. Fifthly, your LOC counts include a ton of tests for ripgrep, where as your GNU grep LOC count presumably does not. For example, about half the LOC in the `printer` crate are tests. And that's just at a library level. Does GNU grep even have library level tests at all? A cursory glance suggests it does not. Sixthly, in the course of developing ripgrep, I made a conscious decision to contribute back to the Rust library ecosystem. As a result, several of ripgrep's libraries are now used in Cargo and rustc itself, among many other tools. GNU grep has no such goal beyond the GNU project and their gnulib itself. Writing more tightly coupled application level code would make the code much tighter. (Which you can observe for yourself by looking at the evolution of ripgrep. It started off with a lot more tightly coupled application level code.) There are probably more reasons why your comparison falls flat, but this is what I could come up with quickly before going to bed. In particular, the choice of programming language is but one factor in overall code complexity. Design goals and other such things arguably have even more of an impact. Thus, your analysis is shallow, and the lack of self-awareness of this fact combined with your inflammatory marks and seeming certainty of your position are why you are a troll. [1] - https://sr.ht/~pierrenn/ripgrep/ [2] - https://github.com/intel/hyperscan