Improving DataView performance in V8 | Hacker News Reader

Improving DataView performance in V8

v8project.blogspot.com

52 pointsingve12 comments

I'm one of the Google Maps developers that worked on the typed-array version of the loading code - hooray for DataView finally being a viable alternative. :)

This is great. Let's hope they plan to also update TypedArray support in general.

NativeArrays are still like 2x faster than Float32Arrays for matrix math. Probably has to be that way ATM as the math has to be promoted to double and back?

Still, v8 seems to be about 2x slower than SpiderMonkey in this area

http://greggman.github.io/webgl-matrix-benchmarks/matrix_ben...

It seems from the code snippets that DataViews are strongly typed. So what happens when you load data with the wrong type? Does JavaScript have "undefined behavior"?

It's just a blob of bytes. If you write a 32-bit float and then read an int from the same location, you get whatever int those 32 bits form.

That generally tells me that this tool is meant for developers smarter than me to make the appropriate abstractions for me to use. Not for me to consider this a core part of my web app development toolkit. :)

Very much in keeping with the [Extensible Web Manifesto][1].

[1]: https://extensiblewebmanifesto.org/

> this tool is meant for developers smarter than me

Developers solving different problems than you. When you run into a problem that needs this tool don't hesitate to use it. Dealing with data structures like this isn't really harder than the usual JS stuff, just different and a learning curve.

And to be clear, unlike in C it's well-defined what those 32-bits are, because it's defined to be an IEEE 754 binary float32.

The data is converted to the target type. If you pass a string to setFloat64, the string will be converted to float64 first.

Great article. I didn't know DataView was much slower than TypedArrays. I always prefer DataView over TypedArrays when writing protocols and file formats, because TypedArrays follows system endianness so you cannot set endianness manually. (It is true that big-endian CPU is very rare but I cannot stop worrying about endianness.)

> we also removed support for indices or offsets that are too large (outside of Smi range) inside the TurboFan-optimized code

What exactly is Smi range? I have a feeling it is something obvious that I'm overlooking.

In V8, small machine integers are signed 31bit integers on 32bit architectures and signed 32bit integers on 64bit architectures.

Reply on news.ycombinator.com