At the end of the intro of the documentation, it is mentioned that the implementation relies on a C++ implementation of SwissTables. This page contains illustrations about two variants: « flat » and « node » based.
The main difference between these two variants is pointer stability. In the « node » based version, the key-value pairs, once inserted, keep their address in memory even if the hash is reorganised. In the « flat » version, some insertions/removals can make the previous key-value pairs be moved in memory.
When it comes to the Rust implementation, I am not experienced enough to be certain of any specific detail, but I tried this simple example based on yours.
use std::collections::HashMap; fn main() { let name = vec![ String::from("Charlie"), String::from("Winston"), String::from("Brian"), String::from("Jack"), ]; let age = vec![50, 5, 7, 21]; let mut people_ages: HashMap<String, i32> = name.into_iter().zip(age.into_iter()).collect(); let mut keys = Vec::new(); let mut values = Vec::new(); for (key, value) in &people_ages { keys.push(key); values.push(value); let key_addr = key as *const String as usize; let value_addr = value as *const i32 as usize; println!("{:x} {:x} {}: {}", key_addr, value_addr, key, value); } // people_ages.insert("Bob".to_owned(), 4); // mutable and immutable borrow println!("keys: {:?}", keys); println!("values: {:?}", values); } /* 55e08ff8bd40 55e08ff8bd58 Brian: 7 55e08ff8bd20 55e08ff8bd38 Charlie: 50 55e08ff8bd00 55e08ff8bd18 Winston: 5 55e08ff8bce0 55e08ff8bcf8 Jack: 21 keys: ["Brian", "Charlie", "Winston", "Jack"] values: [7, 50, 5, 21] */
The commented out line (insertion) is rejected because we cannot alter the hashmap while keeping references to its content. Thus, I guess (I'm not certain) that the implementation does not rely on the « node » based variant since we cannot take benefit of the pointer stability it provides (due to the ownership model in Rust), and probably it relies on the « flat » variant.
This means that we can expect that the key-value pairs associated with the same hash are tightly packed in memory, and iterating over them should be very similar to iterating over a vector: regular progression (with some skips however) very friendly with cache prefetch. Printing the addresses tends to confirm that guess (however the test is not complete enough), and shows a backward progression.
