1326344 - [Perf][Gslide] 48.37%(6683 ms) slower than Chrome when creating one slide for 5 pages slide on GSlide

Status: RESOLVED INACTIVE

(Firefox :: General, defect)

Description

[Shako Ho]

# Test Case
STR
1. Open the browser
2. Open the specified Gslide document (https://goo.gl/R4XkAc)
3. Wait for loading finished
4. Press "End" Key to the end page
5. Press "Enter" to create a new slide

# Hardware
OS: Windows 7
CPU: i7-3770 3.4GMhz
Memory: 16GB Ram
Hard Drive: 1TB SATA HDD
Graphics: GK107 [GeForce GT 640]/ GF108 [GeForce GT 440/630]

# Browsers
Firefox version: 49.01
Chrome version: 55.0.2883.87

# Result
Browser | Run time (median value) 
Firefox | 13817 ms
Chrome  | 7133 ms

Comment 1

[Shako Ho]

Sorry, correct the Firefox version:  53.0a1

Comment 2

[Walter Chen[:ypwalter][:wachen]]

Attachment: test_firefox_gslide_create_slide_5_1483090664_timing.json

Comment 3

[Tom S [:evilpie]]

I looked at this in my CacheIR Analyzer, but with results that are skewed towards baseline.

SetProp:
- We seem to be missing some AddSlot ICs, I hope bug 1091978 will fix those.
- cssText fails, we need a Proxy/DOMProxy IC here
- Function.prototype is also relatively high, I think in theory we could special case this and always allow the assignment, because the resolving of Function.prototype should not be observable.

GetProp:
- We are doing incredibly well here. I only see about 800 single instance of Function.prototype, so attaching an IC would not help there. There is one place where we fail Function.prototype 19 times, so we either ran into the Ion disable limit or more likely this really only happened 19 times

GetElem:
- ~2300 Objects with an index of type string "14336023" or "8192020", so probably sparse array there. Or just doing a normal native shape check?
- ~2300 Object with an index of type |null|, no joke, probably some kind of bug? It would probably be easy to optimize some of the constant like null|undefined|true|false
- 800? Array with index -1 .... lol
- 200? Array with |undefined| ...
- various other cases which look like valid indices

Comment 4

[Tom S [:evilpie]]

Good news: Bug 1091978 indeed fixes thos AddSlot failures!
Now the biggest failures for SetProp are CSS2Properties.cssText, SVG{Image, ..}Element.J (?) and Function.prototype.

Comment 5

[:Harald Kirschner :digitarald]

[qf:-] as work happens in other bugs