Remove O(N^2) algorithm from repeated field extraction.

[seancribbs]

In messages with many values for a repeated field, lists:reverse/1 is called many times, resulting in large amounts of garbage and an unreasonable amount of time spent in decoding. This moves the list reversal until the end of the decoding, only once per repeated field.

This issue was found when a customer did a 2I query returning ~150K results. Ideally this should make it into the "0.8.1" tag and Riak 1.4, although it primarily affects the client-side.

[seancribbs]

Additional info:

Customer ticket: https://basho.zendesk.com/agent/#/tickets/4704
Profiling data: https://gist.github.com/seancribbs/6656e8adbf4a7828ddac

[beerriot]

Confirmed that this drops a 61k-result 2i query on a 3-laptop-node cluster from 14 seconds down to half a second. There also seems to be no measurable slowdown to decoding of messages that do not have large repeated fields.

Note to other potential reviewers: just recompiling pokemon_pb is not enough - you must also recompile the beams generated from the .proto files, as the pokemon_pb code is effectively inlined in them.

Also confirmed that these tests still pass with this patch on both server and client side:

• client_java_verify
• client_python_verify
• mapred_*

Looks good to me. +1

[seancribbs]

Thanks @beerriot!