782579 - strict mode causes a 66% slow down in simplex-noise.js execution

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Jonas Wagner]

User Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_7_4) AppleWebKit/537.4 (KHTML, like Gecko) Chrome/22.0.1229.0 Safari/537.4

Steps to reproduce:

Open http://jsperf.com/simplex-noise-strict-vs-lax
Run the benchmark.



Actual results:

In firefox, and only firefox, there is a big discrepancy in the execution speed of the code between "strict mode" and "normal mode".


Expected results:

Performance should at least be in the same area.

Comment 1

[David Teller [:Yoric] - still alive but not very active]

Same here with FF15, MacOS X.

Comment 2

[Till Schneidereit [:till]]

This is probably at least in part caused by bug 690446. There are some other sources of slowdown for strict code, but not emitting GNAME opcodes is the biggest one.

Note that, while the benchmark's setup function contains var declarations for all variables used in the benchmark code itself, those declarations are local to the setup function, so they won't affect the benchmark code. This leads to the latter using global properties.

Comment 3

[Jonas Wagner]

I created a new revision of the benchmark to address the var issue: http://jsperf.com/simplex-noise-strict-vs-lax/2

Comment 4

[Till Schneidereit [:till]]

A quick look with the JIT Inspector doesn't show any really meaningful differences. That probably doesn't mean anything, though, as with the inspector active, the times become pretty much identical.

CCing bhackett as he might know other reasons for this slowdown.

Comment 5

[Guilherme Lima]

I ran the 2nd version and I got 159 x 163. Seems to be fixed.

Comment 6

[Jonas Wagner]

Tried it in the current nightly 23.0a1 (2013-05-08) and the difference seems to be gone.

Comment 7

[Jan de Mooij [:jandem]]

Thanks guys, looks like bug 690446 was the only problem here and this bug is fixed now.