If you have them 8 bits in memory, it's better to shift right AFTER addition (only one shift, better precision). Maybe ggn was speaking about having them pre-shifted in memory, ie 7 bits?
For adding channels together, pre-shifted data is a requirement to have any form of performance.
Another method is multiplex. Eg running the soundchip at 50 kHz but scale the samples to 25 kHz. Then you have the full 8-bit resolution intact and the mixloop is still quick. Before you object that it won't sound any good. Look at Protracker STe "hi speed mode" or EPSS :)