Speed up at_addrs with a dict.

[rainwoodman]

The old loop was O(mn). This is O(m+n).

[mgedmin]

It's hard for me to make decisions about this function since I don't use it myself. I except that for small address sets naive iteration can be faster (and should use less memory), but I don't know what the expected average size of the address set is.

Also the KeyError thing -- perhaps it makes sense, but old behavior was to suppress errors, and changing it this way would probably require a major version bump according to SemVer. Again, I don't know what makes more sense to users, as I'm not one.

[mgedmin]

This can raise KeyError.

[rainwoodman]

As there is no hard copy, the memory used by the dict should be close to 16 * number of GC objects. GC itself probably consumes as much memory as this.

I did not benchmark this in typical use cases (I only ran objgraph on small cases), though I can imagine it is difficult to find cases where O(m+n) is slower than O(mn).

For a stress test, one can probably collect new IDs with debug level of SAVEALL, and run this on the full id list.

[mgedmin]

Eh, LGTM.

Would you care to add a small changelog note in CHANGES.rst?

[klahnakoski]

I suggest

id_to_obj = {id(o): o for o in gc.get_objects()}

[mgedmin]

👍

[klahnakoski]

I suggest change these lines to

return [
    id_to_obj[i] 
    for i in address_set
    if i in id_to_obj
]

[mgedmin]

👍 except I'd put it all on a single line