It looks like I've just mistakenly overridden my own post. Sorry
A comparison with a working (G4 32bit) configuration could shed some light at this issue. Maybe you'll find some free time for doing such a comparison...
Under "comparison with G4", I mean "instruction-by-instruction" comparison starting with 0x203ce8. That's required to find the instruction that doesn't work as expected. You'll probably need to trace about 20-30 instructions, first on G4, then on G5:
trace an instruction
write down what it changes (we're interested mostly in register values)
an so on...
It's a superb tutorial on the PPC Assembly...