Comment 160 for bug 1949394

To prove it's no regression, I've complied all releases from 3.10 to 3.16 to confirm it didn't work on any of them.

If I find motivation, I might try from 3.0 onward, but I'm not sure, if it'd be fruitful in terms of finding any real solution.

(Yes, I'm aware of Git's bisect functionality, which is quite useful if you know the last working version after which something broke. This might not be useful in this particular case, where I'm leaning towards this not being a regression.)