Ironically part of the problem was that the optimising compilers of today make different space and memory throughput tradeoffs vs CPU execution speed that don't hold for the slower memory architecture of the N64. Namely optimisations like loop unrolling that benefit modern CPUs are a detriment to the N64.
I think most of the cool stuff he did here is around restructuring the code to remove contention on the RAMBUS as it can do operations on different banks in parallel when servicing reads/writes from different components vs causing contention if they wish to read from the same bank. Also writing directly into the bank that will be used by the graphics unit from the CPU while disabling cache features of the RAMBUS, freeing up the cache for more code pages for the CPU to use.
Moving more of the code to execute on the "maxed out" GPU and actually cutting execution time because of the reduced memory bandwidth requirements was also pretty neat.