I changed the checkfile for tests/patmat/null.check as the result did seem to be wrong to me.

It seems that there has been some history on this issue, with a commit which specifically does the opposite of this PR, removing null computations because of expensive computation costs.

Would this fix be unnecessary because it incurs a large performance cost?

Maybe I need to clarify my suggestion. I mean we always project a wildcard pattern to nullable space if the type is nullable.

For example: case a => and case _ => both will match the null value in runtime. This change should also fix some bug in the existing test suit, and make the redundant check more cleaner.

Like in tests/warn/i20122.scala:

sealed trait T_B[C, D]

case class CC_A()
case class CC_B[A, C](a: A) extends T_B[C, CC_A]
case class CC_C[C, D](a: T_B[C, D]) extends T_B[Int, CC_A]
case class CC_E(a: CC_C[Char, Byte])

val v_a: T_B[Int, CC_A] = CC_B(CC_E(CC_C(null)))
v_a match
  case CC_B(CC_E(CC_C(_))) => println(0) // was warn: unreachable
  case _                   => println(1)

If we run the code, we can see the first case is executed, so it is not unreachable.

Then for the OnlyNull warning, I propose the following scheme. Issue OnlyNull warning only if:

• The target space is nullable;
• OnlyNull warning has not been issued before;
• The pattern is a wildcard pattern;
• The pattern is not covered by the previous cases, but covered by the previous cases with null.

I tried this approach, I think the code should be more cleaner: https://github.com/dotty-staging/dotty/blob/381c197ebbc1feb2f801c9df31d06e6d4d97fa6a/compiler/src/dotty/tools/dotc/transform/patmat/Space.scala

Although this may contradict to what I said in tests/explicit-nulls/warn/i21577.scala...

def f3(s: String | Null) = s match
  case s2: String =>
  case _ => // warn NullOnly or not?

This list of tests should be corrected:

tests/warn/i20122.scala failed
tests/warn/i20121.scala failed
tests/warn/i20123.scala failed

Related to #21000

cc @dwijnand

Maybe you need a rebase?

[warn] -- [E121] Pattern Match Warning: /home/runner/work/scala3/scala3/compiler/src/dotty/tools/dotc/core/SymbolLoaders.scala:340:11 
[warn] 340 |      case _ => NoAbstractFile
[warn]     |           ^
[warn]     |Unreachable case except for null (if this is intentional, consider writing case null => instead).

Looks like there's code in the other tests that use a wildcard to catch a null. Somehow that didn't throw a warning before?