False unreachable case warning for inlined code

[odersky]

Compiler version

3.4.0 RC 1

Minimized example

  inline def count(inline x: Boolean) = x match
    case true => 1
    case false => 0

  assert(count(true) == 1)
  assert(count(false) == 0)
  var x = true
  assert(count(x) == 1)

Output

-- [E030] Match case Unreachable Warning: tl-count.scala:7:14 ------------------
7 |  assert(count(true) == 1)
  |         ^^^^^^^^^^^
  |         Unreachable case
  |-----------------------------------------------------------------------------
  |Inline stack trace
  |- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  |This location contains code that was inlined from tl-count.scala:5
5 |    case false => 0
  |         ^^^^^
   -----------------------------------------------------------------------------
-- [E030] Match case Unreachable Warning: tl-count.scala:8:14 ------------------
8 |  assert(count(false) == 0)
  |         ^^^^^^^^^^^^
  |         Unreachable case
  |-----------------------------------------------------------------------------
  |Inline stack trace
  |- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  |This location contains code that was inlined from tl-count.scala:4
4 |    case true => 1
  |         ^^^^
   -----------------------------------------------------------------------------

Expectation

No warnings. Generally, unreachable case warnings should be disabled for code that was inlined.

I encountered this in the wild where it did not find the source line correctly and was completely stumped until I changed the error message to give me a little bit more info. After the change I got this, which was more intelligible. I think that would be a good idea independently, but ideally we also mention previous cases as proposed in #18519.

7 |  assert(count(true) == 1)
  |         ^^^^^^^^^^^
  |         Unreachable case:
  |         Pattern false
  |         cannot possibly match a selector of type (true : Boolean)
  |         at this point in the match sequence.
  |-----------------------------------------------------------------------------
  |Inline stack trace
  |- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  |This location contains code that was inlined from tl-count.scala:5
5 |    case false => 0
  |         ^^^^^
   -----------------------------------------------------------------------------
-- [E030] Match case Unreachable Warning: tl-count.scala:8:14 ------------------
8 |  assert(count(false) == 0)
  |         ^^^^^^^^^^^^
  |         Unreachable case:
  |         Pattern true
  |         cannot possibly match a selector of type (false : Boolean)
  |         at this point in the match sequence.
  |-----------------------------------------------------------------------------
  |Inline stack trace
  |- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  |This location contains code that was inlined from tl-count.scala:4
4 |    case true => 1
  |         ^^^^
   -----------------------------------------------------------------------------

Overall, I believe that eliminating false unreachable warnings and improving error messages for real ones should be a priority for us. It's a major source of unfriendliness and head-scratching. And the other thing really needing improvement is a more robust reporting of inline traces. I noted when they come from different compilation units they are often wrong.

[nicolasstucki]

We should not check reachability for inlined code. Ideally, we should prune those cases from the tree.

[dwijnand]

And the other thing really needing improvement is a more robust reporting of inline traces. I noted when they come from different compilation units they are often wrong.

I would be good to have a reproduction of that - I've not seen that, but could be I just wasn't looking right.