String performance boost (Re: [Mono-dev] Patch to boost speedofUnicodeEncoding)

[Atsushi Eno]

Oh, forgot to mention, all my evaluations are done with Massi's
latest treeprop patch. That is, inline, copyprop, consprop and
deadce are enabled by default IIRC.

Atsushi Eno

Andreas Nahr wrote:
> Basically this also were my findings back then (you can even see some 
> functions where I did check if e.g. the length was > 10000 and then did 
> a fallback to native). (By the way: I believe in a huge number of 
> real-world cases strings will be rather short)
> Did you test using the current memcpy function? I think that it is 
> possible to do something notably faster than this method. If you are 
> using it just for copying chars I don't even think that these could 
> possibly be 1 or 3 byte aligned (Well at least I can't think of any 
> situation where they would be ;)
> But as I wrote the optimum would likely be to replace this method with 
> assembly on the most important platforms in the JIT.
> 
> Another thing that I found back then is that running the managed 
> implementations on MS .Net usually gave another significant speedup, so 
> there is still potential for there as the mono runtime gets faster 
> (Maybe better Register Allocator?) while the unmanaged versions are not 
> likely to improve much.
> 
> Also most of the implementations I wrote rely on inlining of the used 
> Char Methodcalls or they will be notably slower than native.