Every release I try to dig through some old bug reports and clear them 
out of the system.  This release is no different.

I'm a bit surprised I didn't tackle bz19201 in prior releases given that 
Kazu had an almost correct patch attached to the BZ.

I suspect this could be generalized to more than just clearing bytes. 
But I doubt it's worth any real effort as this is ultimately m68k specific.

Essentially Kazu's patch rewrites a load+store sequence where the load 
feeds the store's address computation.  He was missing a check to verify 
the that the target of the load was an address register.  That in turn 
causes failures such as reported by Mikael Pettersson in c#15 when he 
tried including the patch in his builds.

I was able to trigger the same kind of failure as Mikael using Kazu's 
patch and was able to confirm that adding the ADDRESS_REG_P check avoids 
those problems.

I've got a bootstrap running, but it'll be, umm, a long time before it 
finishes (both for the control build and the build with the updated 
patch).  The control build is into stage2 while the build with the patch 
is still building libraries for stage1.  But it's well past the point 
where it failed with the incorrect initial patch.

Installed on the trunk.

jeff