In SERVER-91116, we allowed decorations to have extended alignment requirement, by making the allocation of the underlying decoration buffer alignment aware. This caused a small regression in our benchmark of hot-path instructions. SERVER-91892 resolves nearly all of that regression, but there is still a potential small regression of 16 instructions in the 32-thread benchmark and ~30 in the 1-thread benchmark, which is still about a 0.1% regression from the baseline.
This ticket tracks investigating that remaining regression, if we want to, and seeing if there are any additional savings to be had by optimizing this code path.