NoCollectionsViewHolder to Compose caused 333ms regression in cold main start up (5/10/22)

[mcomella]

From our nightly regression detection:

• 5/9: 1.464s
• 5/10: 1.797s

We regressed 333ms – we should find out the cause and address it.

┆Issue is synchronized with this Jira Task

[Mugurell]

Following the above graph these are all the merged between 8-10 of May.

[Amejia481]

We are backing out #24455 with PR #25519 temporarily, as we think it could have caused this regression. We are going to give a higher priority to #25259 to identify if this is something that is affecting real users or it's something that we only see in local builds that don't have baseline profiles .

[mcomella]

After the PR #25519, we appear to have reverted the regression:

Data points:

• Nightly 5/9: 1.464s
• Nightly 6/5: 1.480s

There is a difference of ~16ms between the two which is likely a symptom of the bimodal behavior introduced on 5/28 #25545 (i.e. since half our replicates are a higher run, it will pull the median up from its previous single mode value) but it could also be a small regression.

Note: the graph pasted here is difficult to read due to the bimodal behavior introduced on 5/28 #25545 and a temporary regression from 5/30 until 6/2 exclusive.

[Amejia481]

Just to add more info:

@mcomella and I did some research, we found the time increased looks like, it's coming from adding a new compose UI piece earlier to the startup path. At the moment, we can't 100% state that this new compose component UI is slower than the previous one, as the builds we were using to compare, don't reflect what our users really are experiencing,as the previous view toolkit has some built-in optimizations that in compose are added by the Google play store, as a result we can't compare apples to apples. As mentioned here #25253 (comment) we are going to give a higher priority to #25259 to identify if this is something that is affecting real users or it's something that we only see in local builds that don't have baseline profiles .

When we started researching we took two profiles, one for 5/9(Before the regression) and one for 5/10(After the regression) the major thing that caught our attention towards to 7b895ab was the compose theming work (ProvideFirefoxColors) was happening much earlier on the startup path and possible delaying it. Additionally, we did a side by side comparison where we were able to see the longer daily on the startup first frame on the 5/10 build.

Left: Before 5/9 and Right:after 5/10.

combined.1.mp4