In current proton protocol processing code make heavy use of the pn_data_t data structure for coding and decoding AMQP frames to send and receive. Unfortunately the data structure is complex and the code which uses it is not very efficient. This means that a lot of CPU is consumed maniputing these data structures unnecessarily during the critical path operations of proton. During frame processing pn_data_t is not even really necessary as it is only used as an intermediary to extract necessary frame parameters out of the frames (or as an intermediary to construct the frames from the necessary frame parameters.
This means that is entirely feasible to go directly from/to the parameters to/from the frames without using an intermediate pn_data_t at all.
The most complicated aspect of this and one which will be deferred somewhat is that of using non scalar structured data as part of AMQP frames.