`inline val` not inlined

[mbovel]

Bug found by @asunluer.

Compiler version

3.0
3.7.4
3.8.0-RC1-bin-20250822-658c8bd-NIGHTLY

Minimized code

inline def f(): Long =
  1L

inline def g(): Long =
  inline val x = f()
  x

inline def h(): Long = 
  inline if g() > 0L then 1L else 0L

@main def main(): Unit =
  println(h())

Output

$ scala -S 3.7.4 inline_val_bug.scala
Compiling project (Scala 3.7.4, JVM (23))
[error] ./inline_val_bug.scala:12:11
[error] Cannot reduce `inline if` because its condition is not a constant value: {
[error]   inline val x: Long = 1L:Long
[error]   x:Long
[error] }:Long > 0L
[error]   println(h())
[error]           ^^^
Error compiling project (Scala 3.7.4, JVM (23))
Compilation failed

Expectation

Compiles, and constant-fold h() to 1L.

[mbovel]

Maybe related to #24412?

cc @hamzaremmal

[bishabosha]

its the type ascription applied by inlining from a method, - transparent inline should fix

[mbovel]

Same problem with transparent:

transparent inline def f(): Long =
  1L

transparent inline def g(): Long =
  inline val x = f()
  x

transparent inline def h(): Long = 
  inline if g() > 0L then 1L else 0L

@main def main(): Unit =
  println(h())

$ scala -S 3.7.4 inline_val_bug.scala
Compiling project (Scala 3.7.4, JVM (23))
[error] ./inline_val_bug.scala:12:11
[error] Cannot reduce `inline if` because its condition is not a constant value: {
[error]   inline val x: (1L : Long) = 1L
[error]   x:Long
[error] } > 0L
[error]   println(h())
[error]           ^^^

And intuitively, I think this should work without transparent anyway, shouldn't it?

[bishabosha]

yeah why does it have x:Long if its transparent?

[mbovel]

yeah why does it have x:Long if its transparent?

The ascription is created here, in ensureNoLocalRefs:

scala3/compiler/src/dotty/tools/dotc/typer/Typer.scala

Line 1598 in 473246e

val tree1 = ascribeType(tree, avoidingType)

It is added so that x.type does not leak into the outer scope.

Would we maybe need a mechanism to strip that kind of synthetic ascriptions during inlining?

[mbovel]

Would we maybe need a mechanism to strip that kind of synthetic ascriptions during inlining?

0a416d8 seems to do something like that, but only when the expected type is a ConstantType.

[mbovel]

Would we maybe need a mechanism to strip that kind of synthetic ascriptions during inlining?

I tried it in #24425. It currently just checks if span.isSynthetic. The condition probably needs to be narrowed.

[bishabosha]

It is added so that x.type does not leak into the outer scope.

seems like this could reuse the avoidance so it only widens to 1L