Attempt to fix performance regression from #20120

[jchyb]

Attempts to fix #20120.
I'll try to prepare an explanation later today/tomorrow, for now let's see if it will pass CI (it does fix the performance in the minimization)
Also, there may be better ways of achieving the same effect (like not producing redundant TypeBounds before committing the typerState, I'll look into this later).

[jchyb]

While it currently fails certain tests, it does fix the performance. I'll try to look into why those type bounds are created for intersection types and whether we can fix this by not creating them there, preserving the tests.

[EugeneFlesselle]

This is a very interesting observation.
I encountered a similar problem when working on #19955.

I believe the part of the code attempting to do this logic is

scala3/compiler/src/dotty/tools/dotc/core/ConstraintHandling.scala

Lines 304 to 311 in 63810dc

	val bound = legalBound(param, rawBound, isUpper)
	val oldBounds @ TypeBounds(lo, hi) = constraint.nonParamBounds(param)
	val equalBounds = (if isUpper then lo else hi) eq bound
	if equalBounds && !bound.existsPart(_ eq param, StopAt.Static) then
	// The narrowed bounds are equal and not recursive,
	// so we can remove `param` from the constraint.
	constraint = constraint.replace(param, bound)
	true

[jchyb]

I found that this happens after we search for implicits and, in result, commit the new typerstate, in the version without the intersection type the constraint was consisting of a single type, so it was easily removed. For intersection types, we had that weird TypeBound with equal lower and upper bounds, so the constraint was kept and we later compiled with an unfulfilled type variable, causing the bad performance

[odersky]

What was the TypeBound looking like?

It's generally not recommended to use == for types, since it has no good meaning. Types are highly recursive. == risks either running into an infinite recursion or missing some equalities. So the only sanctioned equalities are eq (fast, incomplete) and =:= (slow, complete, but prone to cycles).

So maybe one can use a custom mixture of structural comparison of AndTypes and eq comparison for their operands.

[jchyb]

It's TypeBounds(Repro.ObjectId & Endpoints.this.Foo, Repro.ObjectId & Endpoints.this.Foo), or TypeBounds(AndType(TypeRef(ThisType(TypeRef(ThisType(TypeRef(NoPrefix,module class <empty>)),module class Repro$)),trait ObjectId),TypeRef(ThisType(TypeRef(ThisType(TypeRef(ThisType(TypeRef(NoPrefix,module class <empty>)),module class Repro$)),class Endpoints)),trait Foo)),AndType(TypeRef(ThisType(TypeRef(ThisType(TypeRef(NoPrefix,module class <empty>)),module class Repro$)),trait ObjectId),TypeRef(ThisType(TypeRef(ThisType(TypeRef(ThisType(TypeRef(NoPrefix,module class <empty>)),module class Repro$)),class Endpoints)),trait Foo)))

The eq comparison in place of == is a sufficient fix for the reproduction in #20120, but I am wondering where this constraint (TX1 >: Repro.ObjectId & Endpoints.this.Foo <: Repro.ObjectId & Endpoints.this.Foo) is created (and will try to investigate that)

[odersky]

It could be that one of the sides contained a type variable that got instantiated. But then I think @EugeneFlessele's latest fix would apply. Unless Eugene finds the root cause and fixes it otherwise, we can go with the eq comparison in instType. Maybe that turns out to be the simplest solution. Does eq also suffer from the #6505 failure?

[odersky]

It's just a single test that fails, and looking at the history we never really understood why it works now and failed previously. So it's worth digging some more here.

In principle, it would be nice if we could do the improvement in this PR. Or maybe thge weaker form I suggested above:

So maybe one can use a custom mixture of structural comparison of AndTypes and eq comparison for their operands.

[jchyb]

Fixed in #20399