optimize setLanguage, setScript, shapeWithTrace

[chearon]

I'm seeing createCString as a dominator in a project I'm working on. Nodejs text encoding/decoding is slow: nodejs/node#39879. Browsers don't show much difference with this change.

Here's how I profiled it in harfbuzzjs:

diff --cc examples/hbjs.example.js
index b7718e9,b7718e9..75b4c3a
--- a/examples/hbjs.example.js
+++ b/examples/hbjs.example.js
@@@ -7,8 -7,8 +7,15 @@@ function example(hb, fontBlob, text) 
    font.setScale(1000, 1000); // Optional, if not given will be in font upem
  
    var buffer = hb.createBuffer();
++  console.time('c strings');
++  for (let i = 0; i < 1000; i++) {
++    buffer.setLanguage('en');
++    buffer.setScript('Latn');
++  }
++  console.timeEnd('c strings');
    buffer.addText(text || 'abc');
--  buffer.guessSegmentProperties();
++  buffer.setDirection('ltr');
++  buffer.setClusterLevel(1);
    // buffer.setDirection('ltr'); // optional as can be set by guessSegmentProperties also
    hb.shape(font, buffer); // features are not supported yet
    var result = buffer.json(font);

[chearon]

By the way, I also profiled harfbuzzjs against canvas's measureText in Chrome and Firefox yesterday. I was surprised at how well it performed. Unsurprisingly measureText is faster, but doing word wrapping that way necessarily involves measuring words individually. When you compare that to shaping whole paragraphs in harfbuzzjs, it is often only slower by a few milliseconds before the browser's shape cache kicks in (I measured ~10 paragraphs).

I've also experimented with shaping word-by-word like Firefox does, but that's too many shape() calls. Calling into WASM incurs a cost. I think if whole paragraphs were shaped until enough data is collected, then switch to shaping by word, harfbuzzjs becomes a serious choice to layout text in a canvas.

[khaledhosny]

Thanks, the use of TextEncoder here always felt like an overkill.

[khaledhosny]

It probably does not matter in practice, but you can turn the string into a tag yourself and use hb_script_from_iso15924_tag()

[chearon]

I may try that and open a followup. I also wonder if there would be a big difference if you can pass an hb_tag directly to setScript and setLanguage (so it accepts string | number).