libcpp.map.map 'get' is 1e6 slower than libcpp.map.map '__getitem__'

[ctb]

With this code:

from libcpp.map cimport map

cdef class int_int_map:
     cdef public map[long, long] _values

     def __getitem__(self, k):
         return self._values[k]

     def __setitem__(self, k, v):
         self._values[k] = v

     def get(self, k):
         return self._values.get(k)

I find that for 200,000 dictionary entries, the 'get' function is about 1m times slower than getitem -- please see https://gist.github.com/ctb/ecea9809b8e1b6503abda1316b0de5b1 for a full working example. I'm sure I'm doing something wrong but I'm not sure what...

[scoder]

Without trying it, I guess that cython -a would show you that calling .get() on a C++ map copies the map into a Python dict, then calls .get() on it, then throws away the dict. C++ doesn't have that method.

Works as designed.

[ctb]

ok, thanks. I guess I'm surprised that 'get' works at all, then; shouldn't it be "not defined" on the Python side if it's not defined on the C++ side, or alternatively do something smarter & faster? I'd be happy to put in a PR if the behavior should (or could) be changed.

[robertwb]

Actually, this is quite dangerous. Makes sense for primitives like complex, but imagine

cdef vector[int] v = ...
v.pop()

I think we'd be better off with errors about missing methods here.