Thanks, Mitsuharu, for the diagnosis; I think you nailed it.

Dmitry, I spotted a couple of problems with that patch.  First, any 
compiler could do that optimization and I wouldn't be surprised if some 
version of GCC does it either now or in the future, so the fix shouldn't 
be conditionalized on __clang__.  Second, making 'directory' itself 
volatile would break the (uncommon) case where GC_MARK_STACK != 
GC_MAKE_GCPROS_NOOPS, because one can't portably assign the address of a 
volatile to a pointer that isn't volatile-qualified.  Plus, there's no 
need to have every access to that local be volatile; we need only ensure 
that the value is stored somewhere.

I installed the attached patch as emacs-24 bzr 117421, as a change that 
should not break anything and most likely fixes the bug.  Please let me 
know how it works in the Clang environment.