Shouldn't we be using the gp inside the kernel and potentially
-mips2 in both the kernel and userland on the hpcmips platform?

By setting something like '-G 8' for the gp setting, kernel global
registers are placed in a region which is addressed using $gp and 
a 16 bit offset saving an instruction.

The -mips2 flag (which I'm not sure of the TX3900 supports), allows
the system to assume the processor will properly stall the pipeline
on memory accesses.  

These two changes shrink the size of my kernel by about 10%, which seems
noticeable.