792490 - IonMonkey: Getting img.width 25x slower in Uint8ClampedArray benchmark

Status: RESOLVED DUPLICATE of bug 786126

(Core :: JavaScript Engine, defect)

Description

[Alon Zakai (:azakai)]

Attachment: benchmark

Attached is a benchmark that is the code the emscripten test suite uses to reftest images, basically a loop on Uint8ClampedArrays that originate from canvas images. This is 25x slower with javascript.options.ion.content on than off (or compared to before IonMonkey landed).

I don't see the same problem in the shell, so either I did something wrong when I tried to convert to a shell benchmark, or there is some interaction with the image's imageData.

To run the benchmark, load the page and look in the web console.

Comment 1

[Alon Zakai (:azakai)]

Attachment: image for benchmark

Comment 2

[Nicolas B. Pierron [:nbp]]

This bug might a duplicate of Bug 790628 which also use Uint8ClampedArrays.

Comment 3

[Alon Zakai (:azakai)]

I forgot there is a reference image the benchmark looks for. You need to download the html and image into a dir, and run the html.

Comment 4

[David Anderson [:dvander] - inactive, e-mail if emergency]

If it doesn't happen in the shell, it could be going through the DOM in which case this will be fixed by the DOM work in bug 786126.

Comment 5

[Boris Zbarsky [:bzbarsky]]

This testcase gets img.width inside the hot loop.  So yeah, bug 786126 would completely massacre performance here...

Comment 6

[Alon Zakai (:azakai)]

Thanks bz, I checked that saving img.width into a local var avoids the problem, confirming your analysis.

Comment 7

[David Anderson [:dvander] - inactive, e-mail if emergency]

Verified fixed by the recent DOM IC work.