Since Julia does not have atomics for arrays, supporting boxed Julia objects in a hash table requires indirection (i.e., a mutable struct) somewhere at the moment because no atomic operation can be defined on Vector{Any}. Previously, LinearProbingDict was using a mutable struct only for "storage" and the actual linearization was done on a dense array using 128 bit CAS (based on Maier et al. (2019)). However, it required complex key/value storage management to actually store objects in the hash table. It turned out that just using Vector{DictSlot{K,V}} (where DictSlot{K,V} is a mutable struct) is more straightforward and beneficial in terms of performance. The important optimization to make it practical is to avoid copying DictSlot{K,V} during migration.

A nice side-effect of this approach is that it gets rid of the "torn" 64-bit load which is not likely to be supported in Julia. It is now possible to separate key and value to distinct Vectors if atomics on Vector{Any} is supported at some point. This is more likely to be supported than updating key/value location directly in Vector{Pair{Any,Any}}.