Thanks, then.
An intermediate step could be half-carnaval, i.e. do inline storage for struct/tuple and not for arrays. That should cause no additional nullpointer-checks and solves the most pressing API issues (but still sometimes forces people to use suboptimal illogical data-layout).
Any reason that yuyichao’s patch is not merged with #ifdefs depending on llvm version? I mean, there is a lot of llvm-version dependent conditionally compiled code flying around anyway.
(and sorry if I sound exasperated; you all are doing really cool work, can’t say this enough)