Improve the performance of toASCIILowerCase

[arlolra]

/cc @cscott

[cscott]

Is this really faster? The mind boggles. Regexps are usually a lot faster than char-by-char conversions, and we'd expect all-lowercase strings to be passed through unmodified (while in your patch they are rebuilt character-by-character even if there weren't any uppercase letters to begin with).

What about alternatives like:

function toASCIILowercase(s) {
  if (!/[A-Z]/.test(s)) return s;
  // whatever
}

and/or

function toASCIILowercase(s) {
  return s.replace(/[A-Z]/g, function(c) { return String.fromCharCode(c.charCodeAt(0) + 32); });
}

?

[arlolra]

Is this really faster?

Anecdotally, it seems to be.

The way I tested was with, node bin/parse --page Minneapolis --trace time < /dev/null > /dev/null and looked at "dpLoader".

Baseline	~470
This patch	~425
First suggestion above (w/ baseline)	~420
First suggestion above (w/ this patch)	~415
Second suggestion above	~480

[cscott]

Yeah, I went with "first suggestion above (w/ baseline)" in https://github.com/cscott/domino/commit/5573579bf54d1b1112429f6021fb3f297bbbbb5d since it seemed to have the same performance as the more-evil options I tried, without impacting code readability. (That's not official, that's just on my "random performance experiements" branch. https://github.com/cscott/domino/commit/08ee77fcbff4dac58f20ad224f8763af23183a07 also seemed to help.)

[arlolra]

Sounds good to me.

[cscott]

Some performance numbers. Benchmark is dpLoader time on [[User:Acer/Simple4]], measured with:

node  bin/parse.js --trace time --page "User:Acer/Simple4" --replay

Variant	Run 1	Run 2	Run 3	Avg
both	3547	3588	3550	3561
avoid calling toASCIILowerCase	4055	3854	3862	3923
speed up toASCIILowerCase	3472	3171	3685	3442
baseline	3922	4004	4051	3992

The "avoid calling" patch seemed to help in other testing, though; perhaps dpLoader is just not exercising that code path. I need to look into that further.