This fixes an oversight in my previous patch here. I used biggest_mode 
in the assumption that if the reg was used in the function, it would be 
set to something other than VOIDmode, but that fails if we have a 
multiword access - only the first hard reg gets its biggest_mode 
assigned in that case.

Bootstrapped and tested on x86_64-linux, ran (just) the new arm testcase 
manually with arm-eabi. Ok?

(The testcase seems to be from glibc. Do we keep the copyright notices 
on the reduced form)?


Bernd