I know the mask bit itself won't fit in the context pointer on 32-bit, but you could store the mask bit index in the upper 16 bits. For example, STATX_ATIME is 0x20, so you'd store 5 in the upper 16 bits, and recover the mask bit as 1 << i in the getter. That way you'd only need one getter per type, rather than one per member, which is a lot less code.

(You probably already know, but just in case, you can get the bit index with __builtin_ctz, and both GCC and Clang will evaluate __builtin_ctz at compile time when the arguments are constant.)

I prefer to avoid bad surprises on 32-bit systems when Linux will get more bits in mask. If all bits of mask are used, there will be no remaining place for an offset. It sounds more future proof to do as i did, one function per member and put the mask and "offset" there.

Perhaps I wasn't clear. I'm not proposing to store the mask in the context pointer; I'm proposing to store the index of the mask bit, which is a number from 0 to 31. Even if the mask were somehow extended to 64 bits in the future, the index would then be from 0 to 63, which would still easily fit.

CPython doesn't really care about code size, so I'm fine with the way you did it.

Oh ok. The problem is that I'm not sure that all compilers supported by Python on Linux can compute the bit position at build time.

I don't think that code size is an issue here.