This returns to the 2.74.0 implementation of hashing 64-bit quantities, which is more prone to collisions on little-endian platforms, but avoids hashing everything to 0 on big-endian (#2787 (closed)).
It also avoids the update to 2.74.1 breaking incorrect code that makes assumptions about hash table traversal order, which seems like something that is fine for a development branch like 2.75.x but less fine for a stable branch that is meant to be bugfix-only.
Resolves: #2787 (closed)
/cc @3v1n0 @pwithnall @wszqkzqk
Alternatively, we could cherry-pick !2994 (merged) to the 2.74.x branch, if people want to keep the reduced collisions and are OK with changing hash table iteration order on a stable branch.