1910925 - Modified Codepen demo (https://codepen.io/MittenedWatchmaker/pen/mzGdLM) gets relatively slower on Nightly compared to Chrome

Status: NEW

(Core :: JavaScript Engine, task, P3)

Description

[Mayank Bansal]

Open the modified Codepen demo

Nightly: https://share.firefox.dev/3LRugu8 (19s)
Chrome: https://share.firefox.dev/4fpSeug (11s)

As you keep on increasing the size of the "init" in the js, the difference between Nightly and Chrome widens.

Comment 1

[Mayank Bansal]

Attachment: Modified Testcase 320x320.html

Comment 2

[Jan de Mooij [:jandem]]

This is all in Ion JIT code. Likely some guards we fail to optimize or something similar.

Comment 3

[Iain Ireland [:iain]]

Taking a quick look at a profile and the iongraph, this might have something to do with V8's optimization to have special representation for integer arrays. The hot loop is this code:

while(directions[x][y]!=5) {
  //console.log(x,y,directions[x][y])
  x-=(directions[x][y]==1)-(directions[x][y]==2);
  y-=(directions[x][y]==3)-(directions[x][y]==4);
  if(flux[x][y]==0){fluxval++}
  flux[x][y]+=fluxval;
}

And it looks from the profile like we spend a significant fraction of our time unboxing the contents of `directions.

I'm not sure we have an easy way to speed this up.

Comment 4

[Iain Ireland [:iain]]

I took another quick look at this to see whether GVN was failing to common up all the loads of directions[x][y], but that turns out not to be the case. We load it once at the top of the loop, and then again after modifying x in the x-= line. So GVN is working as expected here.

Comment 5

[Mayank Bansal]

Nightly: https://share.firefox.dev/4hlYHGm (17s)