Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Replace HashMap implementation with SwissTable #56241

Closed
wants to merge 20 commits into from

Conversation

Amanieu
Copy link
Member

@Amanieu Amanieu commented Nov 26, 2018

The implementation is from the hashbrown crate.

This is mostly complete, however it is missing 2 features:

  • Adaptive early resizing. -- doesn't seem to be needed
  • try_reserve -- DONE

cc @pietroalbini @gankro

@rust-highfive
Copy link
Collaborator

r? @sfackler

(rust_highfive has picked a reviewer for you, use r? to override)

@rust-highfive rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Nov 26, 2018
@pietroalbini
Copy link
Member

@bors try

cc @rust-lang/infra, this should get a perf run as soon as the try build finishes (it's needed for the second day of impl days at rustfest rome)

@bors
Copy link
Contributor

bors commented Nov 26, 2018

⌛ Trying commit cfd3225 with merge 5aaebb98d2620392ef4b74147a89b4ec1b024455...

@rust-highfive

This comment has been minimized.

@pietroalbini
Copy link
Member

@bors try delegate+

@bors
Copy link
Contributor

bors commented Nov 26, 2018

✌️ @Amanieu can now approve this pull request

@bors
Copy link
Contributor

bors commented Nov 26, 2018

⌛ Trying commit 03db6f4 with merge 73919a1b0a20e544eac9e8b30869b34170590d8e...

@rust-highfive

This comment has been minimized.

@rust-highfive

This comment has been minimized.

@bors
Copy link
Contributor

bors commented Nov 26, 2018

💔 Test failed - status-travis

@bors bors added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Nov 26, 2018
@rust-highfive

This comment has been minimized.

@Amanieu
Copy link
Member Author

Amanieu commented Nov 27, 2018

It seems that the new HashMap is causing errors in name resolution. However I am not familiar with the code. Could someone point me at the HashMaps that are likely to cause the errors above?

cc @petrochenkov

@petrochenkov
Copy link
Contributor

petrochenkov commented Nov 27, 2018

@Amanieu
use fmt::Debug blocks resolution of #[derive(Debug)].
Apparently we can't determine that it doesn't import anything named Debug in macro namespace soon enough.

The workaround is to remove use fmt::Debug; and use fmt::Debug in code.

@petrochenkov
Copy link
Contributor

Oh, wait a second, this PR doesn't actually modify rawfp.rs and libcore in general.
That's worse.

@petrochenkov
Copy link
Contributor

petrochenkov commented Nov 27, 2018

Did the previous HashMap have any guarantees about "order of elements == order of insertion", or something like that, accidentally?

@Amanieu
Copy link
Member Author

Amanieu commented Nov 27, 2018

@petrochenkov There are no ordering guarantees in HashMap, so it is very likely that the ordering has changed when the HashMap implementation changed. Is there anything in name resolution that relies on the insertion order?

I am currently trying to narrow down the exact hash map that is causing the issue. I can confirm that switching only librustc_resolve to the new hashmap reproduces this issue (everything else still uses the old hashmap).

Note that it is also possible that there is a bug in the new HashMap.

@petrochenkov
Copy link
Contributor

petrochenkov commented Nov 27, 2018

There are no ordering guarantees in HashMap, so it is very likely that the ordering has changed

Fix: order of elements with a hash collision, not the general order.
But that probably doesn't matter because our hash maps cannot be used as multi-maps.

I am currently trying to narrow down the exact hash map that is causing the issue. I can confirm that switching only librustc_resolve to the new hashmap reproduces this issue (everything else still uses the old hashmap).

Yes, it's better to bisect all uses of FxHashMap in rustc_resolve and find out what exact map is responsible, then it will be clearer what happens on the resolve side.

@stokhos
Copy link

stokhos commented Jan 14, 2019

Ping from triage @sfackler have you had time to review this pr?

@Amanieu
Copy link
Member Author

Amanieu commented Jan 14, 2019

@alexcrichton suggested that instead of copying hashbrown into libstd, we could have libstd import hashbrown directly as an extern crate. libstd would then just have an API shim with documentation and stability attributes that wraps the actual implementation in the hashbrown crate.

@TimNN
Copy link
Contributor

TimNN commented Jan 22, 2019

Ping from triage @Amanieu / @alexcrichton: What are the plans for this PR?

@Gankra
Copy link
Contributor

Gankra commented Jan 22, 2019

(still on medical leave)

i do not think we should extern hashbrown without the review we would require to merge this code

@gnzlbg
Copy link
Contributor

gnzlbg commented Jan 23, 2019

I would prefer to:

  • wait for @gankro's review before changing libstd to use hashbrown,
  • in the mean time,
    • change this PR to add hashbrown as an an external crate since that would allow us to set up some more extensive CI for it, instead of adding that to the CI budget of libstd (e.g. do some fuzzing, etc.). This would mean that @gankro 's review would need to continue in the hashbrown repo though

    • do a crater run once this PR is green (have we done that already?). I think that for this PR is ok to do a couple of crater runs. One crater run now might uncover issues, and we should definitely do a crater run right before merging after all review comments have been addressed to make sure that none of the fixes broke anything.

@gankro get well soon and don't stress yourself about the review !

@Amanieu
Copy link
Member Author

Amanieu commented Jan 23, 2019

I'm going to leave the PR as it is until the review is finished. I will be integrating any feedback back into the hashbrown repo.

@panaman67
Copy link

Really excited for this commit. Great work!

@TimNN TimNN added S-blocked Status: Marked as blocked ❌ on something else such as an RFC or other implementation work. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Jan 29, 2019
@TimNN
Copy link
Contributor

TimNN commented Jan 29, 2019

(still on medical leave)

i do not think we should extern hashbrown without the review we would require to merge this code

Thanks for the update!

@gankro get well soon and don't stress yourself about the review !

Couldn't agree more, get better soon!


I'm marking this as blocked for the moment, so it doesn't show up during the regular PR triage.

Copy link
Contributor

@alkis alkis left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is exciting to see SwissTable making it into the Rust standard library! Thanks @Amanieu for doing all the work.

}
fn make_hash<K: Hash + ?Sized>(hash_builder: &impl BuildHasher, val: &K) -> u64 {
let mut state = hash_builder.build_hasher();
val.hash(&mut state);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One potential mitigation to the quadradic insertion bug is to always hash the pointer to the buckets as well as the key. This will completely eliminate the quadradic behavior on the merge_dos benchmark but it will slow down resizing the buckets array because it will effectively reshuffle all elements on every resize and this will have worse cache performance than the current approach.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I considered that but I don't feel comfortable mixing address bits into the hash key. I think that this could allow an attacker to defeat address randomization by inferring address bits from looking at the iteration order of a hash table.

// Gotta resize now.
// Ideally we would put this in VacantEntry::insert, but Entry is not
// generic over the BuildHasher and adding a generic parameter would be
// a breaking change.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why can't you move the reserve(1) inside the else block?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because the reserve could invalidate the result of the lookup if it rehashes the table.

Copy link
Contributor

@alkis alkis Feb 4, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is ok if it does. In most cases there is no need to rehash. If a rehash happens, the cost of an additional probe round is insignificant. Right now this reserve(1) is breaking the API guarantees. reserve(N) followed by N entry(K).insert(V) will cause a rehash even though it is specified that it will not.

Edit: also the way this is coded you can defer reserve(1) until the entry is inserted, right?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry, you were right, this reserve call should indeed be in the else block.

We can't defer the reserve call until the entry is inserted because Entry would need to hold a reference to the hasher, which is not possible since in libstd Entry is Entry<K, V> instead of Entry<K, V, S> and changing that would be a breaking change.

key: K,
elem: VacantEntryState<K, V, &'a mut RawTable<K, V>>,
table: &'a mut RawTable<(K, V)>,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To get the VacantEntry we already searched the table and we know exactly in which Bucket we will insert. I think this struct needs to hold a Bucket otherwise insert() will need to probe again.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There are different constraints when searching for an element vs searching for an insertion slot. In particular, we ignore (skip over) deleted buckets in the former case while in the latter case we can insert on top of a deleted bucket.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Apologies for the confusing comment. I think the Entry API allows two implementations here. On return:

  1. VacantEntry already points to the actual bucket and insert() is trivial.
  2. VacantEntry caches the key and hash and insert() does the actual insert/grow dance.

In either case the optimal way to implement find_or_insert semantics needed by entry() -> insert() is:

  • find probe, if found return it (insert() won't be called)
  • insert probe, if bucket is deleted or if growth_left > 0 return it (this probe costs very little because we are touching at most the same cachelines we touched in the previous probe - a deleted bucket can only make us return earlier)
  • rehash and do another insert probe to find where to insert it (the cost of an additional probe here is insignificant - we already did N probes to rehash the container and copied N key, value tuples)

src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
// This may panic.
let hash = hasher(item.as_ref());

// We can use a simpler version of insert() here since there are no
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The reason find_insert_slot() can be used is because 1) we know there is enough space for all entries and 2) we know all items are unique.

src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
@Amanieu
Copy link
Member Author

Amanieu commented Feb 21, 2019

I opened a new PR (#58623) which makes libstd depend on hashbrown from crates.io.

Copy link
Contributor

@Gankra Gankra left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am back from leave and am resuming the review. This batch represents my review of RawTable, which contains a few serious correctness concerns.

#[inline]
pub unsafe fn drop(&self) {
self.ptr.as_ptr().drop_in_place();
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Unless you need some kind of interior mutability, you should make the mutating methods take &mut self just so &SomethingContainingABucket can't accidentally do interior mutability things. Also just makes it clear to the reader that there isn't anything weird going on here.

}

// Branch prediction hint. This is currently only available on nightly but it
// consistently improves performance by 10-15%.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reminder that this is a thing (status change?)

src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
src/libstd/collections/hash/raw/mod.rs Show resolved Hide resolved
Amanieu added a commit to rust-lang/hashbrown that referenced this pull request Mar 5, 2019
@Amanieu
Copy link
Member Author

Amanieu commented Mar 5, 2019

I have addressed most of the feedback in this commit: Amanieu/hashbrown@5d1d4e9

Note that I'm not updating this PR any more, all future development towards integrating hashbrown into libstd is in these two PRs:

@Gankra
Copy link
Contributor

Gankra commented Mar 6, 2019

Since this has gotten pretty out-of-sync with the actual code we should probably drop this review, and I'll just do an "offline" review of map.rs (since I expect it to be pretty straightforward, esp. since you moved most logic into the raw_table).

Also I'm not sure how important handling size_of<(K, V)>() == 0 particularly gracefully can matter? There's no way to feed in state for hashing to be anything other than random, right? Like sure we shouldn't do UB, but otherwise I think we can spin our wheels or maybe even just panic?

@Amanieu
Copy link
Member Author

Amanieu commented Mar 6, 2019

The handling of ZSTs should be fixed now (see my changes to Bucket).

I'm going to close this PR, any further reviews should be on one of the two PRs linked above.

@Amanieu Amanieu closed this Mar 6, 2019
Amanieu added a commit to rust-lang/hashbrown that referenced this pull request Mar 6, 2019
Amanieu added a commit to rust-lang/hashbrown that referenced this pull request Mar 8, 2019
Amanieu added a commit to rust-lang/hashbrown that referenced this pull request Mar 26, 2019
Amanieu added a commit to rust-lang/hashbrown that referenced this pull request Mar 29, 2019
bors added a commit that referenced this pull request Apr 24, 2019
Replace HashMap implementation with SwissTable (as an external crate)

This is the same as #56241 except that it imports `hashbrown` as an external crate instead of copying the implementation into libstd.

This includes a few API changes (all unstable):
- `try_reserve` is added to `HashSet`.
- Some trait bounds have been changed in the `raw_entry` API.
- `search_bucket` has been removed from the `raw_entry` API (doesn't work with SwissTable).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
S-blocked Status: Marked as blocked ❌ on something else such as an RFC or other implementation work.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet