Performance in Qt 5.10

We need to investigate benchmark regressions in qmlbench when comparing Qt 5.10 to Qt 5.9.0.

This task is a collection of data on the affected tests.

1. animations/comparison/moving-images-simple.qml
When testing locally, Qt 5.9.0 averages at 384 frames, while 5.10 averages at 352 (8% regression). Interestingly, an equivalent improvement is visible in animations/comparisons/moving-images-animators.qml and no change visible in the other two animation types (script and animations)

2. creation/qml.modelaccess/modelaccess_jsvalue.qml
Qt 5.9.0 averages at 748 frames, and Qt 5.10 at 720, so 3.7% regression. May not be statistically significant, but many of the modelaccess benchmarks show the same trend. edit: The regression in modelaccess_jsvalue.qml seems to be due to the random generator dip, as can be seen from this graph: https://testresults.qt.io/grafana/dashboard/db/qmlbench-branches?panelId=122&fullscreen&orgId=1&from=1507873293946&to=1510054806774&var-suite=creation%2Fqml.modelaccess&var-benchmark=All&var-branch=v5.9.0&var-branch=5.10&var-hardwareId=eskil_linux_tx1 At October 31st, when the replacement for the RNG landed, the gap went from 3.7% to 0.7%. It is still always 1 frame below, which is odd, but not worth investigating further.

3. quick controls 2
Most of the Qt Quick Controls 2 test show lower performance in Qt 5.10, but indication is that for at least a few this is due to correctness fixes (so the previous runs were not measuring the correct behavior). An interesting result is that the tumbler benchmark actually shows improved performance on Ubuntu desktop, but quite the opposite on TX1. This was also one where we actually identified a correctness fix that may have caused the regression. So this is a puzzle, really.

4. creation/quick.text/delegates_text_plain.qml
This is a bit unstable, but Qt 5.9.0 clocks in at about 680 frames, whereas Qt 5.10 is at 615. There does definitely seem to be a reproducible regression at least, and the same issue may be contributing to the results in other tests as well. It is also quite visible on TX1

5. There is also a dip around August 23rd in several locations which is very visible on TX1. I am having trouble reproducing this on desktop, but the changes that went in at that time are quite general and, if they affect performance, they would probably affect multiple tests. It is tracked individually by --QTBUG-63922. edit: Looking at the desktop results (where there are more results, hence a clearer picture) dev/5.10 is now outperforming 5.9.0 for the quick.item creation tests, so this seems to have been fixed by changing to a pregenerated random number generator in the tests.

In general, there does seem to be some minor things such as this which we need to keep an eye on. It is possible that multiple changes have accumulated to bring performance a little bit down, or it is possible that these results are red herrings.