Remove enumerator allocation in for (HashSet, HashSet) comparisons in SetEquals and IsProperSupersetOf. #78613
Conversation
SetEquals and IsProperSupersetOf.
|
Tagging subscribers to this area: @dotnet/area-system-collections Issue DetailsWhen using HashSet.SetEquals, I noticed that the fast path didn't seem to be avoiding enumerator allocation.
|
| namespace System.Collections.Generic | ||
| { | ||
| internal ref struct BitHelper | ||
| internal readonly ref struct BitHelper |
There was a problem hiding this comment.
Is there a reason not to have readonly here?
There was a problem hiding this comment.
It doesn't help or hurt anything, given its implementation and use. With a different implementation, where some or all of the bit state is stored in the instance, it wouldn't be appropriate. Logically given its surface area it doesn't make sense for it to be readonly, but it's also internal, so, doesn't really matter either way.
There was a problem hiding this comment.
How would you prefer I leave it?
There was a problem hiding this comment.
I'd prefer not readonly, as it's confusing seeing a readonly struct with mutating methods (e.g. MarkBit).
|
|
||
| // Already confirmed that the sets have the same number of distinct elements, so if | ||
| // one is a superset of the other then they must be equal. | ||
| return ContainsAllElements(otherAsSet); |
There was a problem hiding this comment.
After these changes, there is now only one caller for ContainsAllElements(IEnumerable<T>) (IsSupersetOf). Should the helper method just be inlined there?
|
@madelson could you please provide benchmark numbers for the gain we are getting from this change? |
src/libraries/System.Private.CoreLib/src/System/Collections/Generic/HashSet.cs
Show resolved
Hide resolved
@tarekgh I didn't see any benchmarks in dotnet/performance for this, so I just make a quick one which hits the path of 2 hash sets (100 ints) with the same comparer:
We see similar gains across both methods and that all allocation has been eliminated. |
When using HashSet.SetEquals, I noticed that the fast path didn't seem to be avoiding enumerator allocation.