Currently, the hashmap code in runtime maintains a zero buffer that is large enough for any type that has been seen in a map. Enlarging this buffer when a type larger than any before has been seen requires the code to use atomic chicanery. It would be much better to have the runtime always return a pointer to a fixed size buffer on a miss, and having the compiler handle large (fsov; perhaps larger than 256 bytes) map misses by checking the return value against this runtime.zeroptr and explicitly zeroing the value rather than copying zeros from the buffer to it.