Investigate Array.push performance

[MattiSG]

Hi there,

I was surprised to see that Array.push was so much slower than direct assignment. The JSperf link you give looks like a good proof of this.

However, the latest revision of the same bench, which has much cleaner code, seems to remove this difference, and even show push as faster in recent Chrome versions (which I would have spontaneously thought, considering the JIT optimizations that can be done on predictable function calls rather than guessing numerical access).

Are you sure that this recommendation is still to be followed?

All the best,
Matti

[re1ro]

It's a bad proof of this. Take a look at my revision

Revision 10 has Math.random, preparation code and very short for loop in test cases.

[MattiSG]

Interesting.

I understand Math.random could overtime all other operations, but what I don't get is how push could perform better in r10: shouldn't it simply be at the same level as all other ops, considering the overhead of generating a random number?

[re1ro]

There is no straight correlation between code and its performance. It
depends on optimizator. Array#push is checking for array length and after
places a new value. Even if happens natively it's still two operations
until optimizator figures out how to trick it. Direct assignment is a
direct order where to put your data in memory which is one operation. Also
push() is a function execution which has its own overhead. Again until
optimizator iterpretates it differently. So in some cases push
hypotheticaly can be faster if optimizator combines it with other
operations to get less cpu instructions.

On Wednesday, November 21, 2012, Matti Schneider wrote:

Interesting.

I understand Math.random could overtime all other operations, but what I
don't get is how push could perform better in r10: shouldn't it simply be
at the same level as all other ops, considering the overhead of generating
a random number?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/27#issuecomment-10614056.