Lessons Learned

Scala Design Failure: Implicit Numeric Conversions

Published on 2017-05-12.

TL;DR: The desire to make unrelated types act as if they were in a sub-typing relationship, which neither exists nor should exist, combined with syntax sugar that makes static dispatch look like dynamic dispatch creates a perfect storm of unintended, harmful consequences.

Implicit numeric conversions1 are a special compiler feature that adds “convenience” conversions between number types, for instance:

val num: Double = 123 // num = 123.0

It is a feature that

  • is described as a mistake by the designers of Java23
  • silently destroys data, loses numeric precision and changes semantics of overflow and division-by-zero
  • is inconsistently applied, and will become more inconsistent as new numeric types are introduced in future versions of Java
  • breaks all methods defined on numbers
  • cannot be defended against
  • cannot be deactivated

The Good: Intentions

Inferring the type of List(1, 2.3) to the useless common supertype of List[AnyVal] (as Int and Double do not have any interface in common) was deemed to be too confusing for beginners coming from Java.

Instead, compiler magic in the shape of “implicit numeric conversions” was added to convert “smaller” number types to “larger” ones, thus inferring List[Double] in the example above.

While the mechanism looks quite innocent, it rears its ugly hand in many unintuitive and unexpected circumstances, as these kinds of numeric conversions not only convert smaller integer types to larger integer types and smaller floating point number types to larger floating point number types, but also integers to floating point numbers:

val long:   Long   = 123        // Int -> Long
val double: Double = 123.45f    // Float -> Double
val wat: Float     = 123456789L // Long -> Float
//           result: 123456792.0f

The Bad: Type Inference

Scala’s type inference makes the behavior a lot more confusing compared to languages with mandatory type annotations.

Creating a list with two numbers triggers the conversion, concatenating two lists with one number each does not:

List(1, 2.3)            // List[Double]
List(1) ++ List(2.3)    // List[AnyVal]

Numbers of type Int are implicitly converted to Long, but not to BigInt:

List(1, 2L)        // List[Long]
List(1, BigInt(2)) // List[Any]

Although conversion only happens when there isn’t another unrelated type involved:

List(1.2f, 3.4d)        // List[Double]
List(1.2f, 3.4d, "abc") // List[Any]

Assigning a list of integers to a list of doubles works if done in a single line, but fails when done in two:

val nums1: List[Double] = List(1, 2, 3) // compiles

val nums2a = List(1, 2, 3)
val nums2b: List[Double] = nums5a       // fails to compile

On top of that, implicit numeric conversions also interact with type parameters. Consider this change in type inference due to a binary compatible “widening” of types in the method signature of Stream’s #:: method between Scala 2.12.1 and 2.12.2:

def fibonacci: Stream[Double] =
  0 #:: 1 #:: (fibonacci zip fibonacci.tail).map {t => t._1 + t._2}
// Compiles         in Scala 2.12.1 with def #::(hd: A): Stream[A].
// Fails to compile in Scala 2.12.2 with def #::[B >: A](hd: B): Stream[B]:
// error: type mismatch; found Stream[AnyVal], required Stream[Double]

Experienced developers understand the reasons that cause these differences, but it turns out to be quite baffling for users new to Scala.

The Ugly: Extension Methods

Another feature of Scala, extension methods, makes implicit numeric conversions much worse.

Java’s primitive types come without any methods, only operations like +, -, %, <<. Scala’s idea of minimizing the distinction between unboxed and boxed numbers means that numeric types receive “convenience” extension methods like round, floor, toBinaryString or toDegrees.

This gives rise to another set of puzzlers like the following:

123456789.round == 123456792

The reason for this behavior is that the extension methods are not consistently defined on all number types. As the compiler fails to find methods on some type (like round on Int), implicit numeric conversions are kicking in, silently converting and mangling numbers to another type that has them.

In response, another band-aid was applied. round was added to every number implicitly convertible to Float to avoid triggering the implicit conversion.

But even if all the missing methods on numbers were filled in, these efforts are easily defeated, as extension methods are statically dispatched. (Extension methods only look like instance methods, but act like static methods.) Thus the issue is just pushed down another layer:

123456789.round == 123456789 // fixed
def round(value: Float) = value.round
round(123456789) == 123456792 // unfixable

Also, this issue does not require conversions to floating point numbers. Conversions between integer types suffer from the same problem:

val bits: Byte = -1
bits.toBinaryString == "11111111111111111111111111111111" // a byte with 32 bits?

This is a design failure that cannot be fixed without abandoning implicit numeric conversions altogether.

The Worst: Having a Way Out, but not Choosing It

The far-reaching damage of implicit numeric conversions far outweighs the purported benefit of reducing beginner confusion and increasing convenience, considering how inconsistent the conversions are applied in the first place from the point of view of the target audience, new users.

Union types would have been one way out of this mess. They naturally address the concern of beginner-unfriendly type inference by removing implicit numeric conversions and simply letting type inference do its job:

List(1, 2.3) // should be List[Int|Double]

From an operational point of view, inferring List[Int|Double] is hardly more useful than List[AnyVal] before. That’s not the point, though: List[Int|Double] pinpoints the issue (mixing different number types) in a way new users can understand, whereas the old List[AnyVal] does not.

Union types enable the compiler to implement a more direct “what you see is what you get” approach instead of silently sprinkling magic over users’ code.

Disappointingly, Dotty, the next version of Scala which adds union types, barely addresses any of these issues4 and worsens the situation in some cases.

The common approach of explicitly specifying the expected supertype stopped working in Dotty:

List[Any](1, 2.3) // List[Any] = List(1.0, 2.3)

Even explicitly specifying union types does not prevent these conversions:

List[Int|Double](1, 2.3) // List[Int|Double] = List(1.0, 2.3)

Adding an unrelated type prevents the conversion though:

List(1, 2.3, "abc") // List[Any] = List(1, 2.3, abc)

Scala users that have settled on adding the imperfect Ywarn-numeric-widen to their list of compiler flags will be hit the hardest, as the warning is not implemented in Dotty and even if it were, they are now left with fewer options to address the warning in Dotty.

Bonus Quirk

Regardless of whether this problem is fixed, the implementation of round on Float and Double is still wrong and broken for unrelated reasons.

Scala repeats another mistake from Java that was originating from C. Interestingly, while the .NET team copied a lot of design decisions from Java, they considered the issue to be so egregious that they fixed it before their first release of .NET.

  1. Implicit numeric conversion is used as a term to describe the general concept in this document. In practice, various approaches have been tried to implement the concept: a) literal implicit defs in the source code which exist only for “educational purposes” and are not actually used by the compiler anymore, b) the non-implicit methods that the compiler uses instead, c) the notion of weak conformance and d) the notion of “numeric harmonization” 

  2. It would be totally delightful to go through [Java] Puzzlers, another book that I wrote with Neal Gafter, which contains all the traps and pitfalls in the language and just excise them – one by one. Simply remove them.
    There are things that were just mistakes, so for example … [misspeaks] … int to float, is a primitive widening conversion and happens silently, but is lossy if you go from int to float and back to int. You often won’t get the same int that you started with.
    Because, you know, floats, some of the bits are used for the exponent rather then the mantissa, so you loose precision. When you go to float and back to int you’ll find that you didn’t have the int you started with.
    So, you know, it was a mistake, it should corrected, it would break existing programs. So I do like the idea of essentially writing a new language which is very similar to Java which sort of fixes all these bad things. And if someone’s to call it ‘Java’, that would be great, too. Just so long as traditional Java source code can still be compiled and run against the latest VMs. […]
    Joshua Bloch, Devoxx 2008

  3. OSCON Java 2011: Josh Bloch, “Java: The Good, Bad, and Ugly Parts” 

  4. Numbers are not implicitly converted to allow extension methods calls defined on larger numbers anymore, because Dotty invented another slightly different language concept, “numeric harmonization”, which only works on a small predefined set of language constructs like if, match, try and in arguments to repeated parameters.