I have not benchmarked any.
The allocation-based ones are fairly obvious wins. We did use allocation tracker to discover some hot points which were not using some of these (as mentioned in the talk) that we observed disappear after changing. We saw the GC running slightly less just by watching logs in these instances, but no real numbers.
As to the ones around caching method and field lookups, these are really tiny so you'd probably have to be using something like JMH to get numbers that were true and accurate (but then not relevant to Android). They're based around knowledge of how the VM works and simply choosing to do something less (once vs. many/N times).
I would like to see actual numbers, but I fear that sharing actual numbers would diminish people's opinion as to whether these are necessary. They're bound to be small! I forgot to mention in the talk, but in hot places like the draw path things like caching method/field lookups are seen all over AOSP. A frame's time is as scarce of a resource as the battery so doing something once vs. many/N times is going to help keep you on that path to 60fps.