Maybe it'd suffice to set maptype.valuesize to 1 for such maps.

special map fast routines sounds good. Since this should be very rare they will not usually end up in the binary and they dont complicate existing code.

I was also looking into making other special zero sized elem methods for other make methods since they are rare and we can slim down the common code paths:

• growslice
• makechan (for zero sized elems and queue)
  In both situations we have simpler overflow tests too since there are less multiplications involved.
  But i am currently working to slim down and get the current methods work with overflow correctly before splitting or duplicating them.

So specializing map functions sounds good.

So I totally forgot that we pushed the overflow pointer to the end of the bucket exactly to solve this sort of problem, and the one in #21459 . So even if key/value pointers are incremented a final time, or keys or values are zero-sized, we still can't get a beyond-the-object pointer.

Oh right. Yay! For future reference, the original bug was #9384 and fix was CL 1869.