19:48 <antocuni> FWIW, I started a branch to implement HPyField on pypy

20:13 <antocuni> arigato, cfbolz: what are the performance of custom GC tracers in pypy?

20:14 <cfbolz> "not too bad, probably"

20:14 <cfbolz> we use them for jit frames for example

20:15 <antocuni> for HPy, some objects will need a custom tracer (if they have HPyFields). I am wondering whether I can register a custom tracer for all HPy objects (which would do nothing in most cases) or I should create two different W_ types, one with and one without a custom tracer

20:15 <cfbolz> ah, then the latter I think. but I am not sure you can have custom tracers for instances that way

20:18 <antocuni> currently we have W_HPyObject; I could create a subclass W_HPyObjectWithGc and register the custom tracer only for that

20:18 <antocuni> and it's easy to decide which class to instantiate, because there is a flag on the type

20:18 <antocuni> at least, I *think* it should work but I'm not sure

20:19 <cfbolz> ok, just saying that you cannot have custom trace hooks on instances at this point

20:19 <cfbolz> only for lltypes

20:20 <antocuni> yes, this is what I expected

20:20 <cfbolz> also custom instances have a fixed size always, I suppose you need a variable amount of space?

20:22 <antocuni> why variable?

20:23 <cfbolz> antocuni: doesn't it depend on the class definition in hpy how many fields an instances has?

20:23 <cfbolz> or is the memory managed from C and the problem is "simply" to find the GC pointers there

20:25 <antocuni> an C type declared in HPy defines how many bytes an instance needs. The memory is allocated by lltype.malloc

20:25 <antocuni> and all HPy instances (of any HPy type) are an instance of W_HPyObject (which has an hpy_data pointing to the memory)

20:25 <cfbolz> so it is variable ;-)

20:25 <cfbolz> ok

20:26 <cfbolz> I suppose in CPython you don't need this indirection

20:26 <antocuni> yes, in CPython it's just a PyObject*

20:27 <cfbolz> to get the same effect in pypy you would need a variable-sized W_HPyObject

20:27 <cfbolz> but we don't support that atm

20:28 <antocuni> yes, together with arigato we already discussed it, and I summarized here: https://foss.heptapod.net/pypy/pypy/-/issues/3542

20:28 <antocuni> I even tried to trick him into implementing the feature, but I didn't manage :)

20:53 <cfbolz> antocuni: ah, interesting

20:55 <cfbolz> Not impossible to implement

20:59 <cfbolz> I randomly read some code today that would be a problem, left a comment

21:00 <cfbolz> antocuni: another problem would be that you would like efficient support for all this in the JIT, I assume