Regression in type parameter inference in 2.13

[smarter]

When a non-Any upper bound is given to a type parameter, 2.13 instantiates the parameter to the upper bound even when a more precise type is known in some situations (unlike 2.12 and Dotty). Compare:

Welcome to Scala 2.12.4 (OpenJDK 64-Bit Server VM, Java 1.8.0_162).
Type in expressions for evaluation. Or try :help.

scala> def id0[T](x: T): T = x
id0: [T](x: T)T

scala> id0(if (1 == 1) 1 else 2)
res0: Int = 1

scala> def id[T <: AnyVal](x: T): T = x
id: [T <: AnyVal](x: T)T

scala> id(if (1 == 1) 1 else 2)
res1: Int = 1

And:

Welcome to Scala 2.13.0-M3 (OpenJDK 64-Bit Server VM, Java 1.8.0_162).
Type in expressions for evaluation. Or try :help.

scala> def id0[T](x: T): T = x
id0: [T](x: T)T

scala> id0(if (1 == 1) 1 else 2)
res0: Int = 1

scala> def id[T <: AnyVal](x: T): T = x
id: [T <: AnyVal](x: T)T

scala> id(if (1 == 1) 1 else 2)
res1: AnyVal = 1

[adriaanm]

Regressed with scala/scala#5310 (-Yliteral-types):

[info] Running scala.tools.nsc.MainGenericRunner -usejavacp
Welcome to Scala 2.13.0-local-2cdd67e (Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_172).
Type in expressions for evaluation. Or try :help.

scala> def id[T <: AnyVal](x: T): T = x 
id: [T <: AnyVal](x: T)T

scala>  id(if (1 == 1) 1 else 2)
res0: Int = 1

> scala -Yliteral-types
[info] Running scala.tools.nsc.MainGenericRunner -usejavacp -Yliteral-types
Welcome to Scala 2.13.0-local-2cdd67e (Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_172).
Type in expressions for evaluation. Or try :help.

scala> def id[T <: AnyVal](x: T): T = x 
id: [T <: AnyVal](x: T)T

scala>  id(if (1 == 1) 1 else 2)
res0: AnyVal = 1

[adriaanm]

Specifically, it's because of bounds propagation in TypeVar's deriveConstraint. I thought I had a PR that disabled that already... EDIT: scala/scala#6789

[adriaanm]

Note that this is sensitive to the expected type because typedIf assigns the expected type when it's fully defined. This is not in the spec -- instead, it's said to be the weak lub, as you'd imagine. (Same for match -- using the expected type is also not spec'ed)