Precision loss due to numeric conversions

[dwijnand]

Welcome to Scala.next (pre-alpha, git-hash: unknown)  (Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_121).
Type in expressions to have them evaluated.
Type :help for more information.
scala> val x = 123456789L
val x: Long = 123456789
scala> val y = 0F
val y: Float = 0.0f
scala> x - (if (true) x else y)
val res0: Float = -3.0f

[odersky]

Isn't that as spec'ed? Current Scala gives the same result, one thing we specifically do not want to do is change the runtime behavior of dotty as compared to Scala.

[dwijnand]

I agree we can't change runtime behaviour without a suitable migration path.

Talking with @aphyr we were thinking that perhaps one solution could be to:

1. by default insert runtime checks to fail early rather than return the wrong result
2. unless you can determine at compile time that the conversion is safe, so for example a 0 literal would convert without any cost, and
3. explicit conversions (e.g .toFloat) could be used in performance-critical code, avoiding the runtime check

WDYT?

[smarter]

Another possibility would be to never convert Long to Double and Int to Float, and instead use union types. I suspect this won't affect too much code since Long and Float are rarely used compared to Int and Double (where you can convert without losing precision).

[dwijnand]

Happy to defer to #2451.