Significant difference in performance of JSArrays vs TypedArrays when performing arithmetic in a loop

[Suraj Sharma]

Hello everyone,

I was running a micro benchmark(attached at the bottom) and noticed something interesting.

There is a significant difference in performance  between TypedArray and regular JSArray when performing arithmetic in a loop,

When timing the function foo() on release build of d8 on x64 machine we get following result:

JSArray: 616 ms

TypedArray: 1261 ms

I did some investigation and found that we always perform an overflow check for loop variable i when using TypedArray irrespective of the value of length.

I have made some progress in addressing this issue, but need some help in putting some pieces together.

Here is the link summarizing my investigation.

PTAL

Thanks 

Suraj

[Jakob Kummerow]

Isn't this an off-by-one error of sorts? I'm by no means an expert on Turbofan range analysis, but it seems that if TF realized that i is strictly less than the array length, then that would be enough to optimize away the overflow check on the increment?

--
--
v8-dev mailing list
v8-...@googlegroups.com
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to v8-dev+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/v8-dev/c2c96e24-ce44-4afb-9a98-ab2829c9cae2n%40googlegroups.com.

[Georg Neis]

I don't think so. If the typed array is large (length >= 2^31) then the (signed) overflow can and will happen when i = 2^31 - 1. [side note: even unsigned overflow (carry flag) would happen when i = 2^32 - 1, which the max length of 2^32 allows for.]

Suraj, I'll take a closer look at this when I find time and will comment in your document.

Thanks!

 Georg

To view this discussion on the web visit https://groups.google.com/d/msgid/v8-dev/CAKSzg3QKPe-3CwFm%2BU626FX2JTaNrT1u6ZD%3DF1izQfnsYwgyaA%40mail.gmail.com.

[Jakob Kummerow]

Oh, right. It was my thinking that was off by one, twice: on the signed bit length, and on the maximum value! Sorry for any confusion caused.

[Suraj Sharma]

Precisely! :) Thanks Georg and Jakob for the feedback.
Feel free to reach out if you need any clarification or more information :)

Thanks

Suraj