Merely traverses the reifiee and records local symbols along with their metalevels.
Merely traverses the reifiee and records local symbols along with their metalevels.
Makes sense of cross-stage bindings.
Makes sense of cross-stage bindings.
Analysis of cross-stage bindings becomes convenient if we introduce the notion of metalevels. Metalevel of a tree is a number that gets incremented every time you reify something and gets decremented when you splice something. Metalevel of a symbol is equal to the metalevel of its definition.
Example 1. Consider the following snippet:
reify { val x = 2 // metalevel of symbol x is 1, because it's declared inside reify val y = reify{x} // metalevel of symbol y is 1, because it's declared inside reify // metalevel of Ident(x) is 2, because it's inside two reifies y.eval // metalevel of Ident(y) is 0, because it's inside a designator of a splice }
Cross-stage bindings are introduced when symbol.metalevel != curr_metalevel. Both bindings introduced in Example 1 are cross-stage.
Depending on what side of the inequality is greater, the following situations might occur:
1) symbol.metalevel < curr_metalevel. In this case reifier will generate a free variable that captures both the name of the symbol (to be compiled successfully) and its value (to be run successfully). For example, x in Example 1 will be reified as follows: Ident(newFreeVar("x", IntClass.tpe, x))
2) symbol.metalevel > curr_metalevel. This leads to a metalevel breach that violates intuitive perception of splicing. As defined in macro spec, splicing takes a tree and inserts it into another tree - as simple as that. However, how exactly do we do that in the case of y.eval? In this very scenario we can use dataflow analysis and inline it, but what if y were a var, and what if it were calculated randomly at runtime?
This question has a genuinely simple answer. Sure, we cannot resolve such splices statically (i.e. during macro expansion of reify),
but now we have runtime toolboxes, so noone stops us from picking up that reified tree and evaluating it at runtime
(in fact, this is something that
Expr.eval and
Expr.value do transparently).
This is akin to early vs late binding dilemma. The prior is faster, plus, the latter (implemented with reflection) might not work because of visibility issues or might be not available on all platforms. But the latter still has its uses, so I'm allowing metalevel breaches, but introducing the -Xlog-runtime-evals to log them.
As we can see, the only problem is the fact that lhs'es of eval can be code blocks that can capture variables from the outside. Code inside the lhs of an eval is not reified, while the code from the enclosing reify is.
Hence some bindings become cross-stage, which is not bad per se (in fact, some cross-stage bindings have sane semantics, as in the example above). However this affects freevars, since they are delicate inter-dimensional beings that refer to both current and next planes of existence. When splicing tears the fabric of the reality apart, some freevars have to go single-dimensional to retain their sanity.
Example 2. Consider the following snippet:
reify { val x = 2 reify{x}.eval }
Since the result of the inner reify is wrapped in an eval, it won't be reified together with the other parts of the outer reify, but will be inserted into that result verbatim.
The inner reify produces an Expr[Int] that wraps Ident(freeVar("x", IntClass.tpe, x)). However the freevar the reification points to will vanish when the compiler processes the outer reify. That's why we need to replace that freevar with a regular symbol that will point to reified x.
Example 3. Consider the following fragment:
reify { val x = 2 val y = reify{x} y.eval }
In this case the inner reify doesn't appear next to eval, so it will be reified together with x. This means that no special processing is needed here.
Example 4. Consider the following fragment:
reify { val x = 2 { val y = 2 val z = reify{reify{x + y}} z.eval }.eval }
The reasoning from Example 2 still holds here - we do need to inline the freevar that refers to x. However, we must not touch anything inside the eval'd block, because it's not getting reified.
An (unreified) path that refers to definition with given fully qualified name
An (unreified) path that refers to definition with given fully qualified name
Creator for last portion of name (either TermName or TypeName)
Reifies any supported value.
Reifies any supported value.
For internal use only, use reified instead.
Reify a case object defined in Mirror
Reify a case object defined in Mirror
Reify a reference to a symbol
Reify a reference to a symbol
Reify a tree.
Reify a tree.
For internal use only, use reified instead.
Reify a type.
Reify a type.
For internal use only, use reified instead.
Rolls back certain changes that were introduced during typechecking of the reifee.
Rolls back certain changes that were introduced during typechecking of the reifee.
These include: * Replacing type trees with TypeTree(tpe) * Transforming Modifiers.annotations into Symbol.annotations * Transforming Annotated annotations into AnnotatedType annotations * Transforming Annotated(annot, expr) into Typed(expr, TypeTree(Annotated(annot, _)) * Non-idempotencies of the typechecker: https://issues.scala-lang.org/browse/SI-5464
An (unreified) path that refers to term definition with given fully qualified name
An (unreified) path that refers to term definition with given fully qualified name
An (unreified) path that refers to type definition with given fully qualified name
An (unreified) path that refers to type definition with given fully qualified name