Bottom Up calculation for CPU profiler is wrong

[kenzieschmoll]

Recreate the example here and verify that our bottom up and call tree representations match what the android studio profilers show: https://youtu.be/v4kCRZ_O4Lc?t=1085

[kenzieschmoll]

I recreated the example from the linked Android Studio profiler video:

Here is DevTools with same profile:

Top Down (matches Android Studio)

Bottom Up (does not match Android Studio)

The difference in Android Studio's algorithm is described at this time stamp in the video (link), when the speaker describes the calculation for B.

CC @scheglov @bkonyi

[kenzieschmoll]

The Chrome DevTools CPU profiler also has slightly different results for the same data (chrome-devtools-sample-profile.json.txt):

The Chrome DevTools bottom up table does not create a bottom up root for node D (same as the Dart DevTools CPU profiler but different than the Android Studio profiler). I find it confusing that AS creates a bottom up root for node D, when no samples actually end with D.

However, the Chrome DevTools bottom up self time/total time calculations for nodes B and A matches AS for nodes B and A. At this point, I reccomend:

1. we move forward with tweaking our bottom up self time/total time calculations to match both AS and Chrome DevTools
2. we leave the behavior we currently have to make only leaf nodes in the profile bottom up roots (e.g. in the example presented in this issue, do not make D a bottom up root, which matches our current behavior and the behavior of Chrome DevTools).

One other thing to note here, both the AS profiler and our profiler nicely merge samples in the bottom up tree, so we should leave that behavior as well.

[kenzieschmoll]

Digging even deeper into this, it appears that Observatory's cpu profile doesn't match the total and self time calculations that Android studio's profiler does for a bottom up tree.

See the explanation for how AS calculates total and self time for nodes B and D here: https://youtu.be/v4kCRZ_O4Lc?t=1027. Observatory calculates the total time for B and D as the global time spent on the stack (see example below), whereas AS calculates the total time for B and D as the total time for C when called by B and D (since this is C's bottom up tree).

The total time for a bottom up root's children should sum to the total time of the bottom up root. In this case, total times for children of RenderBox.hasSize's in the bottom up tree should sum to 0.14%. Since Observatory calculates total time as global time spent on the stack for this method, this assumption is broken.

In DevTools, we will match what Android studio does.