Failing FFI example about cffi-lua HOT 6 CLOSED

landed the big memory management rework, so i will close this

from cffi-lua.

i'm not sure how you got it to work to that point, as broken handling of variable-length structs (c9fa529) would prevent it from even allocating properly (it would initialize read_idx to 16 and leave the buffer empty, leading to assertion errors)

that said, after that fix, i was able to reproduce your condition and pushed some fixes, can you test that on your side?

from cffi-lua.

Could it be the Arm64 architecture that I'm using? Will test asap :)

from cffi-lua.

there is also an issue i just discovered; cdata on most platforms must be 16-aligned, but lua_newuserdata returns pointers that are 8-aligned; in practice this will usually not manifest (hence no failing tests in CI even on platforms with strict alignment requirements) but UBsan (rightly) complains about that (so i need to figure out some way to fix it, and probably add UBsan to CI)

the obvious way is to overallocate by 16 bytes and then align the pointer (while making sure to adjust any potential calls to userdata rawlen, because that needs the original pointer and may return wrong size), but that's both a waste and bug-prone, so i'll need to think about this

from cffi-lua.

and no, i don't think so

from cffi-lua.

from cffi-lua.