The implementation of W_HPyObject is sub-optimal
The implementation of W_HPyObject is sub-optimal for various reasons:
to allocate an instace (by calling _create_instance) we do two mallocs: one is a GC-managed malloc for W_HPyObject itself, the other is a raw-malloc to hpy_data.
We are using both lightweight and "heavy" finalizers. Ideally, we would like to call tp_destroy from the del, but probably it doesn't work out of the box because RPython actively checks that we don't call anything "weird". However we know that tp_destroy cannot do much, so we should find a workaround to convince RPython to allow it.
More info on how to solve point (1).
Ideally, we would like to allocate the memory for hpy_data together with W_HPyObject, but we think that currently RPython doesn't have the machinery to do that, i.e. to allocate an instance of a given class + some arbitrary extra amount of memory at the end.
The second problem is about moving vs non-moving memory. Ideally we would like this solution:
- we allocate W_HPyObject+extra_memory in the nursery
- we pin the memory before returning from HPy_New
- we unpin the memory when we close the handle.
However, the current pinning strategy of the RPython GC doesn't work well because it might refuse to pin memory after a while. We probably need to fix the GC in order to allow an arbitrary amount of pinned memory.
At the moment of writing, we can see the problems in the test_allocate_obj microbenchmark. Here are the results on CPython and PyPy:
hpy ------------- CPython: TestType::test_allocate_obj 673.72 us PyPY: TestType::test_allocate_obj 1492.93 us