Closes https://github.com/ppy/osu/issues/29738.
This "regressed" in https://github.com/ppy/osu/pull/29639, but if I
didn't know better, I'd go as far as saying that this looks like a .NET
bug, because the fact that PR broke it looks not sane.
The TL;DR on this is that before the pull in question, the offending
`.Contains()` check that this commit modifies was called on a
`List<BeatmapSetInfo>`. The pull changed the collection type to
`BeatmapSetInfo[]`. That said, the call is a LINQ call, so all good,
right?
Not really. First off, the default overload resolution order means that
the previous code would call `List<T>.Contains()`, and not
`Enumerable.Contains<T>()`. Then again, why would that matter? In both
cases `T` is still `BeatmapSetInfo`, right? Well... about that...
It is difficult to tell for sure what precisely is happening here,
because of what looks like runtime magic. The end *symptom* is that the
old code ended up calling `Array<BeatmapSetInfo>.IndexOf()`, and the new
code ends up calling... `Array<object>.IndexOf()`.
So while yes, `BeatmapSetInfo` implements `IEquatable` and
the expectation would be that `Equals<BeatmapSetInfo>()` should be
getting called, the type elision to `object` means that we're back to
reference equality semantics, because that's what
`EqualityComparer.Default<object>` is.
A five-minute github search across dotnet/runtime yields this:
c4792a228e/src/coreclr/vm/array.cpp (L984-L990)
Now again, if I didn't know better, I'd see that "OPTIMIZATION:"
comment, see what transpired in this scenario, and call that
optimisation invalid, because it changes semantics. But I *probably*
know that the dotnet team knows better and am probably just going to
take it for what it is, because blame on that code looks to be years
old and it's probably not a new behaviour. (I haven't tested empirically
if it is.)
Instead the fix is just to tell the `.Contains()` method to use the
correct comparer.
The explicit detach call was removed from `updateBeatmapSet`, causing
this to occur. We could optionally add it back (it will be a noop in all
cases though).
These used to work because there was a huge blocking load operation,
which is now more asynchronous.
Note that the change made in `SongSelect` is not required, but defensive
(feels it should have been doing this the whole time).
The only exception to the rule here was "when screen isn't active apply
without debounce" but I'm not sure we want this. It would cause a
stutter on returning to song select and I'm not even sure this is a
common scenario.
I'd rather remove it and see if someone finds an actual case where this
is an issue.
This reverts a portion of https://github.com/ppy/osu/pull/9539.
The rearrangement in `SongSelect` is required to get the initial filter
into `BeatmapCarousel` (and avoid the `FilterChanged` event firing,
causing a delayed/scheduled filter application).
