> For 3.2 it is probably OK to simply revert the patch
> after all.  I can do so if maintainer approves that.

Thanks. Please mention the reference (PR optimization/10233) in the ChangeLog 
entry. I've attached a testcase gcc.dg/i386-mmx-3.c .

> I just want to be sure that for 3.3/3.4 it is dealt correct that I am
> not quite certain about at this point :((

Perhaps instrument the code to see if invalid subregs reach the logic that 
would have triggered before your patch.

-- 
Eric Botcazou