578456 - IonMonkey: |new String() + "asdf"| 5.8x as slow as in v8

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Mandelin [:dmandelin]]

Consider this benchmark:

    var s = new String();
    for (var i = 0; i < 1000000; ++i)
        var q = g + "asdf";

Time, with empty loop time subtracted out:

  JM:   284 ms
  JSC:  153 ms
  TM:    34 ms

TM is fast because it exposes and trace-compiles the defaultValue and valueOf calls. We should be able to achieve this effect in JM with a PIC-like technique at add sites that see string inputs.

We think this will help xparb a bit. We don't know of other things it helps with yet, although it seems worth doing someday since the tracer does it.

JM wanted SS win: 3 ms

Comment 1

[David Mandelin [:dmandelin]]

A couple things I forgot:

1. This was spun off from bug 578257 comment 1, thanks to Alan.

2. For this particular bug, what we want is more of a special case in |add| to unwrap string objects rather than a general defaultValue/toString optimization.

Comment 2

[Ryan VanderMeulen [:RyanVM]]

Maybe I'm missing something here, but I'm assuming that the loop should be appending "asdf" to s, not g? Assuming that to be the case, I get the following:
Interp: 299
TM: 315
JM: 98
JM+TI: 129
d8: 26

JM appears to be faster than it used to be, but still a long way from what TM used to be and d8 still is. Also worth noting is that TI regressed this by about 32%.

Comment 3

[Till Schneidereit [:till]]

New numbers from my rMBP@2.7Ghz:

SM: 52
d8: 9
jsc: 71

Comment 4

[Guilherme Lima]

Still valid.
Nightly - 85ms
Chrome 45 - 32ms

Comment 5

[Guilherme Lima]

It seems performance regressed on this microbenchmark.
Firefox 50 - 206 ms
Chrome 54 - 17 ms

Comment 6

[Mayank Bansal]

Nightly: https://share.firefox.dev/3SL9zE2 (3.5s wallclock)
Chrome: https://share.firefox.dev/4dvhOwk (4s)

Calling this fixed.