Oliver, Cypress has many DPM's, they range from $2 to $many depending on size. My method has been to reserve 1/2 for a to b and the other half for b to a. a writes a block of data to its write section and then stores the length of the block in the flag byte. When b sees the flagbyte being non-zero, it reads the data block and clears the flagbyte. The overhead for this method is one memory write time (movx, inc dptr, djnz) per BYTE and thus beats any serial scheme by at leat a factor of 10. If you already are partitioning the data memory (e.g. for memory mapped I/O, there is virtually no cost associated but for the cost of the DPM.
Ben, DMA (in the case of UART) will relieve your processor from processing every byte you are transmitting; however, it has nothing to do with memory sharing, and the comm speed will be the same.