Valgrind has been complaining on x86_64 for a long time about the get_obj_key function, and since it was extremely noisy for me and obscured other (more important) errors I decided to track it down.

Basically, the previous logic was too simple to work.  We were walking off the end of d.bytes by iterating up to i < sizeof(void*)*2, where d.bytes is just 0 < sizeof(void*).  Why it wasn't complaining on x86 probably comes down to something else living on the stack there that was assigned.

To be correct we separate iteration for each repetition of the address from iteration over the components of d.