On Mar 12, 2020, at 12:02, Marco Sulla <mail.python.org@marco.sulla.e4ward.com> wrote:
On Thu, 12 Mar 2020 at 19:52, Andrew Barnert <abarnert@yahoo.com.via.e4ward.com> wrote:
I suppose it could track calls out to C functions as they happen and mark every value that was live before the call, and then instead of numpy checking refs==1 is could check refs==1 && !c_flagged, and then it wouldn’t need the C frame hackery. But that seems like it would slow down everything in exchange for occasionally helping numpy and a few other C extension libs.
The author of the patch says this is already implemented in string concatenation in Python itself. Maybe we should look at the implementation, but I don't know where and what to search.
Strings, unlike numpy arrays, are immutable. (You can’t mutate them from Python; you can mutate them from the C API but it’s against the rules to do so after they’ve been shared with the interpreter.) And strings, unlike numbers (which numpy arrays act as), don’t have any C API functions that steal references. So they don’t require any kind of tracking or stack-fumbling. For a simpler parallel: the CPython compiler can merge two strings into a single object, and the CPython interpreter can intern strings at runtime, because this is guaranteed to be safe, but obviously neither can do the same thing for arbitrary objects.