To reply to myself, it seems that it's the cross compiler that is
busted. When I compile using the native MIPS gcc the code is correct --
the .s file created via -S option to gcc does the "addiu $sp,$sp,-8"
*before* the jump to mcount, not in the branch delay slot. The cross
compiler optimizes it into the branch delay slot.

	However it looks like the .s files that come out of gcc are laid out
without taking the branch delay slot into account, ie. that appears to
come at a later stage. This leaves the cross-compiler's .s file as
wrong.
	-- Ethan