08:22 <cfbolz> cool, the pypy binary got 800KB smaller :-)

08:57 <arigato> gcref-based-type-erasure? great, finally :-)

08:58 <cfbolz> arigato: yep, it's pretty cool :-)

08:59 <cfbolz> before, we had an rpython type "list of strings" and one "list of instances"

08:59 <cfbolz> now it's all the same

08:59 <cfbolz> and similarly for dicts (for values in general, and for keys that are identity-based too)

09:00 <arigato> do all tests now require a non-'ref' gc?

09:00 <cfbolz> arigato: no, just the very specific ones that test the type erasure

09:00 <cfbolz> arigato: the llinterp tests are anyway fine, but if you want to generate C code with refcounting you can't do the gcref type erasure

09:01 <cfbolz> I am not really sure keeping refcounting around is actually that sensible in general

09:01 <arigato> agreed, but what occurs now? it falls back to the old more limited erasure in case of 'ref'?

09:01 <cfbolz> arigato: yes, exactly

09:02 <arigato> cool, but also meh, more kinda-pointless logic

09:02 <cfbolz> yes, maybe it's time to kill refcounting

09:05 <cfbolz> arigato: nowadays it would also make sense to make all the GcStructs have a typeptr, even those of str and list etc. it doesn't cost extra space for incminimark and it would remove a bunch of strange cases a bit everywhere. plus it would make all pointers inspectable in the same way

09:11 <cfbolz> I was saying nonsense above, of course. for dict *keys*, only instances can be type erased, and that was always the case. all other gcrefs either have specific equality and hashing, like strings and tuples, or they are unhashable, like lists. so nothing changes for keys. but for values the gcref-based type erasure still applies, and helps quite a bit to reduce the number of dict variants