https://issues.dlang.org/show_bug.cgi?id=13532
--- Comment #3 from Dmitry Olshansky <[email protected]> --- (In reply to Vladimir Panteleev from comment #0) > The first surprise for me was that declaring a regex object (either Regex or > > StaticRegex) with "enum" was so much slower. It makes sense now that I > think > about it: creating a struct literal inside a loop will be more > expensive than > referencing one already residing somewhere in memory. > Perhaps it might be > worth mentioning in the documentation to avoid using > enum with compiled regexes. It's a common anti-pattern, it's the same issue with array literals, it's the same issue with anything that takes some time to compute or allocates. regex function call does both. It's worth adding a note though, fell free t create a pull. I'm not sure I'll get to it soon. (In reply to Vladimir Panteleev from comment #2) > Well, it's slower for this particular case, not necessarily in general. > CCing Dmitry. That's right. Problem is simple backtracking engine of CTFE version which is an unfortunate historical point as I'd pick the other engine of the two if I could go back in time. (In reply to hsteoh from comment #1) > ctRegex is slower than regular regex?! Whoa. That just sounds completely > wrong. What's the cause of this slowdown? I thought the whole point of > ctRegex is to outperform runtime regex by making use of compile-time > optimization. Whatever happened to that?? If this is the case, we might as > well throw ctRegex away. I'm fully aware of this. Unfortunately adding yet another engine (C-T "robust" engine) is increasingly a maintenace disaster. Consider also that working on compile-time generated regex is a nightmare of ~5-10 minutes to run all tests and constant out of memory conditions. Duplicating the amount of work done at CTFE is something DMD CAN'T handle at the moment. Another problem is regex accumulated a lot of technical debt, and needs a serious amount of refactoring before pling up more stuff. Then with modular design (the one I roughly outlined in my talk) we can put more and more components into it. Sadly all of this goes very slooowly. --
