Some related notes:

Array#join is a bit faster in Chrome Canary 64, but still takes more than 3 seconds.

Possible workarounds for this *specific case* that match or even outperform Firefox on the original code:

* store character codes, join in chunks (~350ms, Canary: 200ms)
  var chrs = []; 
  for (var i=0; i<40000000; i++) chrs.push(64+(i&31));

  var strs = [];
  for (var i = 0; i<chrs.length/10e3; i++)
    strs.push(String.fromCharCode.apply(null, chrs.slice(i * 10e3, (i + 1) * 10e3)));
  var str = strs.join('');

* store character codes, use TextDecoder (~100ms)
  var chrs = []; 
  for (var i=0; i<40000000; i++) chrs.push(64+(i&31));
  var str = new TextDecoder().decode(new Uint8Array(chrs));

Wooooow ... Precisely, I am parsing a 40 MB JSON file from an ArrayBuffer. First, I parse UTF8 (convert bytes into Unicode codes - integers), then I call String.fromCharCode on each code. Then I concatenate it with .join(""). Then I can give that string to JSON.parse();

I was so desperate that I wrote my own JSON parser, that works directly with char codes (avoiding a long string). It seems like TextDecoder can save me a lot of work (I never heard about it before). But it is still not supported in Edge :(

For posterity, the demo in #0 is:

				var chrs = [];
				for(var i=0; i<40000000; i++) chrs.push(String.fromCharCode(64+(i&31)));
				
				var time = Date.now();
				var str = chrs.join("");
				console.log(Date.now()-time);

Guys, I think we can close it, since TextDecoder() solves my problem, so I don't need fast string joining anymore.

FYI the new Array.p.join Torque implementation is currently under review at https://crrev.com/c/1196693.

Peter's new join impl has landed, this is what I now see locally in d8:

$ out/release/d8 tmp.js 
802

$ cat tmp.js
const chrs = [];
const chr = String.fromCharCode(64+(i&31));
for (var i=0; i<40000000; i++) chrs.push(chr);

const time = Date.now();
const str = chrs.join("");
console.log(Date.now()-time, str.length);