qunaibit has quit [Quit: Connection closed for inactivity]
<ronan>
mattip: I don't think it's possible to get rid of dtype objects, especially considering NEP 42
<ronan>
Anyway, my main issue ATM is the stack-allocated PyArrayObjects. I believe they're fundamentally incompatible with the HPy model
<fangerer>
Do you mean _incompatible_ in the sense of we cannot allocate them on the stack in HPy? I assume this is just an optimization. But if they are stack-allocated, we could just allocate them on the heap and do the appropriate `HPy_Close` when returning from that function.
<ronan>
well, allocating on the heap would require some refactoring. Probably doable but a bit painful.
<mattip>
there is some long-standing claim that allocating on the stack is faster at runtime
<mattip>
but that slightly conflicts with the desire to move to subinterpreters and allow PyFinalize to reset everything
<mattip>
it would be nice to be able to put some benchmarking number to the claim about stack vs. heap
<ronan>
mattip: another argument against stack allocation is that with NEP 42, these malformed arrays will get passed to random user-written code