Improve write performance

[BridgeAR]

See nodejs/node#19289. I did not run dedicated benchmarks here.

[mscdex]

Did you check previous node branches to see how this change affects them (positively or negatively)?

[BridgeAR]

I ran those from Node.js 6 > 10-pre and either the performance is identical or better.

[BridgeAR]

4 is identical as well.

[mscdex]

I just tested these on each node branch (including master) and my versions of these functions, without the validation code, perform exactly the same as the previous implementations (before this PR). Here's what I used to benchmark:

const n = 1e9;
const b = Buffer.alloc(4);

function writeUInt32BE(buf, value, offset) {
  buf[offset++] = (value >>> 24);
  buf[offset++] = (value >>> 16);
  buf[offset++] = (value >>> 8);
  buf[offset++] = value;
  return offset;
}
console.time('original');
for (var i = 0; i < n; ++i)
  writeUInt32BE(b, 0xffff, 0);
console.timeEnd('original');

/*
function writeUInt32BE(buf, value, offset) {
  buf[offset + 3] = value;
  value = value >>> 8;
  buf[offset + 2] = value;
  value = value >>> 8;
  buf[offset + 1] = value;
  value = value >>> 8;
  buf[offset] = value;
  return offset + 4;
}
console.time('new');
for (var i = 0; i < n; ++i)
  writeUInt32BE(b, 0xffff, 0);
console.timeEnd('new');
*/

Just comment out one and uncomment the other and compare. For me, I get ~1033ms.

[BridgeAR]

To me it looks like the code is going to be inlined and this does not benchmark what should be benchmarked. Changing 0xffff to a number that is increased with each call should prevent this and yield a clearer result.

[mscdex]

Ok I modified the loop to:

  if (i % 2 === 0)
    writeUInt32BE(b, 0xffff, 0);
  else
    writeUInt32BE(b, 0xff, 0);

and now on node master, the original implementation (~3480ms) is faster than this PR (~3520ms).

Also there is a strange performance issue starting with node v9 where it takes over twice as long as previous node branches. Node v8 and older take only ~1400-1500ms for either implementation.

EDIT: switch the loop to writeUInt32BE(b, i, 0) gets performance back to node v8 and earlier.... strange. Also with that inside the loop, the new implementation does seem to be faster.

[BridgeAR]

Ok, I ran a couple of different benchmarks again and it somewhat seems to depend on the input size what function is faster. Since there is no ideal solution, I am going to close this.