Tracking issue for HashMap::raw_entry

[sfackler]

Added in #54043.

---

As of 6ecad33 / 2019-01-09, this feature covers:

impl<K, V, S> HashMap<K, V, S>
    where K: Eq + Hash,
          S: BuildHasher
{
    pub fn raw_entry(&self) -> RawEntryBuilder<K, V, S> {…}
    pub fn raw_entry_mut(&mut self) -> RawEntryBuilderMut<K, V, S> {…}
}

pub struct RawEntryBuilder<'a, K: 'a, V: 'a, S: 'a> {…} // Methods return Option<(&'a K, &'a V)>
pub struct RawEntryBuilderMut<'a, K: 'a, V: 'a, S: 'a> {…} // Methods return RawEntryMut<'a, K, V, S>
pub enum RawEntryMut<'a, K: 'a, V: 'a, S: 'a> {
    Occupied(RawOccupiedEntryMut<'a, K, V>),
    Vacant(RawVacantEntryMut<'a, K, V, S>),
}
pub struct RawOccupiedEntryMut<'a, K: 'a, V: 'a> {…}
pub struct RawVacantEntryMut<'a, K: 'a, V: 'a, S: 'a> {…}

… as well as Debug impls for each 5 new types, and their inherent methods.

[Amanieu]

What is the motivation for having separate from_hash and search_bucket methods? It seems that the only difference is whether the hash value is checked before calling is_match. However if the table does not store full hashes (i.e. hashbrown) then there is no difference between these methods.

Could we consider merging these methods into a single one? Or is there some use case where the difference in behavior is useful?

[Gankra]

I am also extremely confused by this distinction, as my original designs didn't include them (I think?) and the documentation that was written is very unclear.

[Amanieu]

cc @fintelia

[fintelia]

The reason I added search_bucket was because I wanted to be able to delete a random element from a HashMap in O(1) time, without storing an extra copy of all the keys. Basically, instead of doing something like this:

let key = map.iter().nth(rand() % map.len()).0.clone();
map.remove(&key);

I wanted to just be able to pick a random "bucket" and then get an entry/raw entry to the first element in it if any:

loop {
    if let Occupied(o) = map.raw_entry_mut().search_bucket(rand(), || true) {
        o.remove();
        break;
    }
}

(the probabilities aren't uniform in the second version, but close enough for my purposes)

[Gankra]

I continue to not want to support the "random deletion" usecase in std's HashMap. You really, really, really, should be using a linked hashmap or otherwise ordered map for that.

[Amanieu]

I have removed this method in the hashbrown PR (#56241). Your code snippet for random deletion won't work with hashbrown anyways since it always checks the hash as part of the search process.

[gdzx]

I can avoid unnecessary clones inherent to the original entry API which is nice. But unless I'm mistaken, the current raw_entry API seems to hash the keys twice in this simple use case:

#![feature(hash_raw_entry)]

use std::collections::HashMap;

let mut map = HashMap::new();

map.raw_entry_mut()
   .from_key("poneyland")
   .or_insert("poneyland", 3);

Currently I use the following function to hash once and automatically provide an owned key if necessary (somewhat similar to what was discussed in rust-lang/rfcs#1769):

use std::borrow::Borrow;
use std::collections::hash_map::RawEntryMut;
use std::hash::{BuildHasher, Hash, Hasher};

fn get_mut_or_insert_with<'a, K, V, Q, F>(
    map: &'a mut HashMap<K, V>,
    key: &Q,
    default: F,
) -> &'a mut V
where
    K: Eq + Hash + Borrow<Q>,
    Q: Eq + Hash + ToOwned<Owned = K>,
    F: FnOnce() -> V,
{
    let mut hasher = map.hasher().build_hasher();
    key.hash(&mut hasher);
    let hash = hasher.finish();

    match map.raw_entry_mut().from_key_hashed_nocheck(hash, key) {
        RawEntryMut::Occupied(entry) => entry.into_mut(),
        RawEntryMut::Vacant(entry) => {
            entry
                .insert_hashed_nocheck(hash, key.to_owned(), default())
                .1
        }
    }
}

Given k1 and k2 with the same type K such that hash(k1) != hash(k2), is there a use-case for calling RawEntryBuilderMut::from_key_hashed_nocheck with hash(k1), &k1 and then inserting with RawVacantEntry::or_insert using k2 ?

If there isn't, why not saving the hash in RawVacantEntryMut and using it inside RawVacantEntryMut::insert ? It would even be possible to assert in debug builds that the owned key has indeed the same hash as the borrowed key used to lookup the entry.

[timvermeulen]

I'm not yet very familiar with this API, but what @gdouezangrard suggested seems like a great idea to me. What even happens currently if the two hashes don't match, is the key-value pair then inserted into the wrong bucket? It's not clear to me from (quickly) reading the source code.

[sujayakar]

I submitted rust-lang/hashbrown#54 to support using a K that doesn't implement Hash via the raw entry API. See rust-lang/hashbrown#44 for the original motivation. Now that hashbrown is merged into std, could we expose this functionality on the std::collections::hash_map types as well?

If so, I'd be happy to submit a PR!

[thomcc]

This is a really great API, it's also what keeps crates (hashlink for example) using hashbrown instead of the stdlib hash map -- since hashbrown exposes this.

What could be next steps here towards stabilization?

[sanbox-irl]

Just gonna add another ping here -- what's blocking this right now?

[Amanieu]

I see a few things that need to be resolved:

• We need proper documentation for the API functions, including examples.
• hashbrown has a few extensions to RawEntry, such as https://docs.rs/hashbrown/0.9.1/hashbrown/hash_map/enum.RawEntryMut.html#method.and_replace_entry_with. We should decide whether to port these to std as well.
• The name feels a bit wierd since it isn't clear what "raw" means. But I don't have any better ideas.

I would recommend prototyping in the hashbrown crate first, which can then be ported back in the the std HashMap.

[KamilaBorowska]

I find raw_entry and raw_entry_mut methods unnecessary - unlike entry method, they don't take any parameters, they just provide access to methods that could as well be in HashMap itself. I think I would consider getting rid of those and putting raw entry APIs directly in HashMap. .raw_entry().from_key(...) is also unnecessary, unless I'm missing something it's identical to already stabilized .get_key_value(...).

I also would like to point out that RawVacantEntryMut doesn't really do much other than providing an API that allows insertion which provides a reference to inserted key and value. This structure doesn't store anything other than a mutable reference to a hash map. This particular API can be used to create unrelated keys, like in this example.

#![feature(hash_raw_entry)]

use std::collections::HashMap;

fn main() {
    let mut map = HashMap::new();
    map.raw_entry_mut().from_key(&42).or_insert(1, 2);
    println!("{}", map[&1]);
}

This is a bit like calling insert after determining an entry is vacant. I think raw_entry_mut APIs could return Options just like raw_entry APIs.

#![feature(hash_raw_entry)]

use std::collections::hash_map::{HashMap, RawEntryMut};

fn main() {
    let mut map = HashMap::new();
    if let RawEntryMut::Vacant(_) = map.raw_entry_mut().from_key(&42) {
        map.insert(1, 2);
    }
    println!("{}", map[&1]);
}

I think raw entry API is useful, but I don't think its API should be conflated with entry API.

[tkaitchuck]

As discussed here: rust-lang/hashbrown#232
Allowing the user to specify the hashed value with the contract that it is generated in the same way that the map computes that hash has two drawbacks:

1. It locks in the implementation of the Hashmap to never changing how that code is invoked. In particular this prohibits hashmap from ever using specialization. This is leaving significant performance gains on the table for types with fixed lengths and short strings. (This makes raw_entry a non-zero-cost-abstraction because the cost is incurred even if the feature is not used.)
2. It creates an opportunity for a bugs in applications that accidently do something different. If for example an application takes advantage of this to create a memoized hash or similar, and their calculation is different in some cases the results will be unexpected and lack a clear error message.

If the feature of a user specified hash is needed, it may be useful to instead provide a method on the raw entry to hash a key. That way the hashmap can implement this however it sees fit and the application code is less error prone because there is an unambiguous way to obtain the hash value if it is not known in advance.

[workingjubilee]

Doesn't this decision mean binary bloat because std::hashmap will exist because some dependency somewhere uses it and a duplicate implementation would be pulled in through hashbrown for those that need HashTable? Is there a reason not to standardize HashTable?

EDIT: To be clear, I'm not against this, but I'd like to understand if there's an interest in standardizing HashTable

The details of HashTable are extremely dependent on its particular lookup algorithm. So it goes with any "raw entry" API. We do not want to specify that sort of thing, because otherwise what happened to std::unordered_map will happen to std::collections::HashMap.

As for binary bloat, the concern you cited is why Rust allows function merging. If the implementations of two codepaths happen to be the same, their functions will simply point to the same code.

[vlovich]

Function merging happens across crates with different versions? Is it specific to functions with the same name from the same crate name or is there some kind of search algorithm to find all functions that are the same (eg if I copy paste a function to my crate, does it get deduped as well?).

[cuviper]

LLVM's MergeFunctions pass doesn't care about names or crates or versions, just whatever is in the compilation unit -- which can be everything if you use full LTO.

[vlovich]

I think that's the part I was trying to highlight - the odds of two implementations of hash table being in the same compilation unit is low and you're only going to see this maybe get resolved if you use full LTO whereas most people at most use thinLTO. I think code bloat is a valid concern and one that's not easily dismissed by "the optimizer will handle it"

[cuviper]

Much of it will be duplicated across CUs anyway by monomorphization and/or #[inline], even when it is coming from a single implementation. There's very little compiled code to share from the libhashbrown.rlib in your sysroot.