Without the GroupV2Params caches, processing multiple large group
updates is extremely wasteful, decrypting the same profile keys and
UUIDs over and over again. The size of these caches has also been
*increased* to match the group size limit, or else there's no benefit
for large groups.
Other LRUCache instances are still off by default in the NSE by virtue
of the "NSE max" defaulting to 0.
When Data is put into an NSCache it gets converted to an NSData, which
is a CFData underneath. CFData's hash function only looks at a prefix
of the data, and the secret params were taking up a good chunk of that
prefix, leading to hash collisions. By putting more variable data
first, we get better hashes *and* exit memcmp faster in the case of a
collision.
If there are several group changes being returned at once,
deserializing all of them up front can lead to excessive memory use,
especially for very large groups. This commit switches to
deserializing each change as it is used, which doubles the amount of
CPU work (because we loop through the changes twice) in exchange for
only ever having one Change object graph in memory at a time. It does
also mean we can get a protobuf deserialization error a bit later in
the process if the data is ever corrupted, but that shouldn't cause
additional issues at these particular call sites.
Previously we would only do this if we could handle *all* unknown
fields, but these days we sometimes have more than one unknown field
and learn about one at a time.
Matthew previously stripped this down in our fork of AFNetworking; at
this point there's no benefit to it living in a separate repo and
separate target. With this we can remove the AFNetworking target
completely.
OWSHTTPSecurityPolicy was originally a subclass of AFSecurityPolicy
specifically for use with Signal-run services, but it already contains
the part of AFSecurityPolicy that we use elsewhere.
