Zsh Mailing List Archive
Messages sorted by: Reverse Date, Date, Thread, Author

Re: Idea for optimization (use case: iterate string with index parameter)

On 5 Jan 2018 at 23:23:57, Bart Schaefer (schaefer@xxxxxxxxxxxxxxxx) wrote:
> On Fri, Jan 5, 2018 at 5:38 AM, Sebastian Gniazdowski
> wrote:
> > iterating string with index parameter is quite slow, because unicode characters are  
> skipped and counted using mbrtowc().
> I can't remember the last time I needed to do that kind of iteration.

Maybe indeed it's not that common. It's one of the basic things one can do with strings but in practice, hmm. I would accumulate that optimization though, as the overall optimization starts to give effects while it's largely composed of disappointing optimizations.

> typeset -a iter=(${(s//)string})
> for ((i=1; i <= $#iter; i++)); do something with $iter[i]; done
> string=${(j//)iter} # if needed
> That is more memory-intensive, of course, but it also assists with
> cases of unordered access into the array of characters.

It might give some effects, I was doing "for letter in $iter" path blindly and missed the obvious $iter[i] way, and without index, "for letter ..." couldn't replace existing code.

> > In general, the array would hold #N (5-10 or so) last string-index requests. If new request  
> would target the same string, but index greater by 1, getarg() would call mbrtowc() once  
> (via MB_METACHARLEN macro) reusing the previous in-string pointer.
> Why only when greater by 1? If greater, scan to and record the next
> needed position. Same number of mbrtowc() conversions, overall.

Yes this should be generalized this way, I didn't want to complicate example.

I recalled yesterday that for ASCII there's a short path that returns 1 and doesn't call mbrtowc() to compute size of character. In discussion on irc this yielded a conclusion that the cache should probably be 1-element only, because it would be an overkill for simple $string[2], etc. indexing. This way the code should be very simple. The params.c part in question is: 


I'm little afraid that getarg() might be called in some generalized situations, but heck it shouldn't be called for a="$b", so the cache might well survive in many typical loops. And maybe a 2-element cache will not add much code and not slow down simple indexing.

Sebastian Gniazdowski
psprint /at/ zdharma.org

Messages sorted by: Reverse Date, Date, Thread, Author