Safe Initialization

Dotty implements experimental safe initialization check, which can be enabled by the compiler option -Ycheck-init.

A Quick Glance

To get a feel of how it works, we first show several examples below.

Parent-Child Interaction

Given the following code snippet:

abstract class AbstractFile {
   def name: String
   val extension: String = name.substring(4)
}

class RemoteFile(url: String) extends AbstractFile {
   val localFile: String = url.hashCode + ".tmp"       // error: usge of `localFile` before it's initialized
   def name: String = localFile
}

The checker will report:

-- Warning: tests/init/neg/AbstractFile.scala:7:4 ------------------------------
7 |	val localFile: String = url.hashCode + ".tmp"  // error
  |	    ^
  |    Access non-initialized field value localFile. Calling trace:
  |     -> val extension: String = name.substring(4)	[ AbstractFile.scala:3 ]
  |      -> def name: String = localFile            	[ AbstractFile.scala:8 ]

Inner-Outer Interaction

Given the code below:

object Trees {
  class ValDef { counter += 1 }
  class EmptyValDef extends ValDef
  val theEmptyValDef = new EmptyValDef
  private var counter = 0  // error
}

The checker will report:

-- Warning: tests/init/neg/trees.scala:5:14 ------------------------------------
5 |  private var counter = 0  // error
  |              ^
  |             Access non-initialized field variable counter. Calling trace:
  |              -> val theEmptyValDef = new EmptyValDef    [ trees.scala:4 ]
  |               -> class EmptyValDef extends ValDef       [ trees.scala:3 ]
  |                -> class ValDef { counter += 1 }	     [ trees.scala:2 ]

Functions

Given the code below:

abstract class Parent {
  val f: () => String = () => this.message
  def message: String
}
class Child extends Parent {
  val a = f()
  val b = "hello"           // error
  def message: String = b
}

The checker reports:

-- Warning: tests/init/neg/features-high-order.scala:7:6 -----------------------
7 |  val b = "hello"           // error
  |      ^
  |Access non-initialized field value b. Calling trace:
  | -> val a = f()                              	[ features-high-order.scala:6 ]
  |   -> val f: () => String = () => this.message	[ features-high-order.scala:2 ]
  |    -> def message: String = b	                [ features-high-order.scala:8 ]

Design Goals

We establish the following design goals:

By reasonable usage, we include the following use cases (but not restricted to them):

Principles and Rules

To achieve the goals, we uphold three fundamental principles: stackability, monotonicity and scopability.

Stackability means that objects are initialized in stack order: if the object b is created during the initialization of object a, then all fields of b should become initialized before or at the same time as a. Scala enforces this property in syntax by demanding that all fields are initialized at the end of the primary constructor, except for the language feature below:

var x: T = _

Control effects such as exceptions may break this property, as the following example shows:

class MyException(val b: B) extends Exception("")
class A {
  val b = try { new B } catch { case myEx: MyException => myEx.b }
  println(b.a)
}
class B {
  throw new MyException(this)
  val a: Int = 1
}

In the code above, the control effect teleport the uninitialized value wrapped in an exception. In the implementation, we avoid the problem by ensuring that the values that are thrown must be transitively initialized.

Monotonicity means that the initialization status of an object should not go backward: initialized fields continue to be initialized, a field points to an initialized object may not later point to an object under initialization. As an example, the following code will be rejected:

trait Reporter { def report(msg: String): Unit }
class FileReporter(ctx: Context) extends Reporter {
    ctx.typer.reporter = this                // ctx now reaches an uninitialized object
    val file: File = new File("report.txt")
    def report(msg: String) = file.write(msg)
}

In the code above, suppose ctx points to a transitively initialized object. Now the assignment at line 3 makes this, which is not fully initialized, reachable from ctx. This makes field usage dangerous, as it may indirectly reach uninitialized fields.

Monotonicity is based on a well-known technique called heap monotonic typestate to ensure soundness in the presence of aliasing [1]. Otherwise, either soundness will be compromised or we have to disallow the usage of already initialized fields.

Scopability means that an expression may only access existing objects via formal parameters and this. More precisely, given any environment ρ (which are the value bindings for method parameters and this) and heap σ for evaluating an expression e, if the resulting value reaches an object o pre-existent in σ, then o is reachable from ρ in σ. Control effects like coroutines, delimited control, resumable exceptions may break the property, as they can transport a value upper in the stack (not in scope) to be reachable from the current scope. Static fields can also serve as a teleport thus breaks this property. In the implementation, we need to enforce that teleported values are transitively initialized.

With the established principles and design goals, following rules are imposed:

  1. In an assignment o.x = e, the expression e may only point to transitively initialized objects.

    This is how monotonicity is enforced in the system. Note that in an initialization val f: T = e, the expression e may point to an object under initialization. This requires a distinction between mutation and initialization in order to enforce different rules. Scala has different syntax for them, it thus is not an issue.

  2. References to objects under initialization may not be passed as arguments to method calls or constructors.

    Escape of this in the constructor is commonly regarded as an anti-pattern, and it's rarely used in practice. This rule is simple for the programmer to reason about initialization and it simplifies implementation. The theory supports safe escape of this with the help of annotations, we delay the extension until there is a strong need.

  3. Local definitions may only refer to transitively initialized objects.

    It means that in a local definition val x: T = e, the expression e may only evaluate to transitively initialized objects. The same goes for local lazy variables and methods. This rule is again motivated for simplicity in reasoning about initialization: programmers may safely assume that all local definitions only point to transitively initialized objects.

Modularity

For modularity, we forbid subtle initialization interaction beyond project boundaries. For example, the following code passes the check when the two classes are defined in the same project:

class Base {
  private val map: mutable.Map[Int, String] = mutable.Map.empty
  def enter(k: Int, v: String) = map(k) = v
}
class Child extends Base {
  enter(1, "one")
  enter(2, "two")
}

However, when the class Base and Child are defined in two different projects, the check will emit a warning for the calls to enter in the class Child. This restricts subtle initialization within project boundaries, and avoids accidental violation of contracts across library versions.

We impose the following rules to enforce modularity:

  1. A class or trait that may be extended in another project should not call virtual methods on this in its template/mixin evaluation, directly or indirectly.

  2. The method call o.m(args) is forbidden if o is not transitively initialized and the target of m is defined in an external project.

  3. The expression new p.C(args) is forbidden, if p is not transitively initialized and C is defined in an external project.

Theoretically, we may analyze across project boundaries based on tasty. However, from our experience with Dotty community projects, most subtle initialization patterns are restricted in the same project. As the rules only report warnings instead of errors, we think it is good to first impose more strict rules, The feedback from the community is welcome.

Theory

The theory is based on type-and-effect systems [2]. We introduce two concepts, effects and potentials:

π = C.this | Warm(C, π) | π.f | π.m | π.super[D] | Cold | Fun(Π, Φ) | Outer(C, π)
ϕ = π↑ | π.f! | π.m!

Potentials (π) represent values that are possibly under initialization.

Effects are triggered from potentials:

To ensure that the checking always terminate and for better performance, we restrict the length of potentials to be finite (by default 2). If the potential is too long, the checker stops tracking it by checking that the potential is actually transitively initialized.

For an expression e, it may be summarized by the pair (Π, Φ), which means evaluation of e may produce the effects Φ and return the potentials Π. Each field and method is associated with such a pair. We call such a pair summary. The expansion of proxy potentials and effects, such as π.f, π.m and π.m!, will take advantage of the summaries. Depending on the potential π for this, the summaries need to be rebased (asSeenFrom) before usage.

The checking treats the templates of concrete classes as entry points. It maintains the set of initialized fields as initialization progresses, and check that only initialized fields are accessed during the initialization and there is no leaking of values under initialization. Virtual method calls on this is not a problem, as they can always be resolved statically.

More details can be found in a forthcoming paper.

Back Doors

Occasionally you may want to suppress warnings reported by the checker. You can either write e: @unchecked to tell the checker to skip checking for the expression e, or you may use the old trick: mark some fields as lazy.

Caveats

The system cannot handle static fields, nor does it provide safety guarantee when extending Java or Scala 2 classes. Calling methods of Java or Scala 2 is always safe.

References