If a method's type contains a non-primitive value class then it has two
signatures: one before erasure and one after erasure. Before this
commit, we checked if this was the case using `isAnyValSubtype`, but
this is too crude since primitive value classes are also subtypes of
`AnyVal` but do not change signature after erasure.
This commit replaces `isAnyValSubtype` by `isDerivedValueClass` which
excludes primitive value classes.
In practice, for an empty class, this reduces the size of the output of
`DefaultShowAPI` from 65 lines to 25 lines.
Before:
https://gist.github.com/smarter/cf1d6fe58efda88d6ee6#file-old-api
After:
https://gist.github.com/smarter/cf1d6fe58efda88d6ee6#file-new-api
Before this commit, we did not do the invalidation for methods with
multiple parameter list, the comment above `hasValueClassAsReturnType`
said:
Note: We only inspect the "outermost type" (i.e. no recursion) because
we don't need to inspect after erasure a function that would, for
instance, return a function that returns a subtype of AnyVal.
But this is wrong: a method with signature:
def foo(a: A)(b: B): C
is erased to:
def foo(a: A, b: B): C
and not, as the comment in the code suggest, to:
def foo(a: A): B => C
so we do need to inspect the final result type of methods, because they
can be value classes that will be erased to their underlying value.
Since pickled annotated types and symbols only mention
static annotations, whereas compilation from source
sees all annotations, we must explicitly filter annotations
in the API representation using the same criteria as the pickler,
so that we generate the same API when compiling from source
as when we're loading classfiles.
Specialize two implementations for each value of the `inherit` boolean argument.
Also use a more direct way of distinguishing declared and inherited members.
backwards compat for source-dependencies/inherited-dependencies
For refinement types, the Structure was already restricted
to declarations (and not inherited members), but all base types
were still included for a refinement's parents, which would
create unwieldy, and even erroneous (cyclic) types by expanding
all constituents of an intersection type to add all base types.
Since the logic already disregarded inherited members, it seems
logical to only include direct parents, and not all ancestor types.
```
class Dep {
def bla(c: Boolean) = if (c) new Value else "bla"
}
class Value extends java.lang.Comparable[Value] { def compareTo(that: Value): Int = 1 }
```
Motivated because we want to make it more robust & configurable.
Original motivation was to diagnose a cyclic type representation,
likely due to an f-bounded existential type, as illustrated by the following:
```
class Dep {
// The API representation for `bla`'s result type contains a cycle
// (an existential's type variable's bound is the existential type itself)
// This results in a stack overflow while showing the API diff.
// Note that the actual result type in the compiler is not cyclic
// (the f-bounded existential for Comparable is truncated)
def bla(c: Boolean) = if (c) new Value else "bla"
}
class Value extends java.lang.Comparable[Value] { def compareTo(that: Value): Int = 1 }
```
Limit nesting (`-Dsbt.inc.apidiff.depth=N`, where N defaults to `2`),
and number of declarations shown for a class/structural type
(via `sbt.inc.apidiff.decls`, which defaults to `0` -- no limit).
Limiting nesting is crucial in keeping the size of api diffs of large programs
within a reasonable amount of RAM...
For example, compiling the Scala library, the API diff with nesting at `4`
exhausts 4G of RAM...
The only aspect of the self variable that's relevant for
incremental compilation is its explicitly declared type,
and only when it's different from the type of the class that declares it.
Technically, any self type that's a super type of the class could be ignored,
as it cannot affect external use (instantiation/subclassing) of the class.
Call `initialize` in case symbol's `info` hadn't been completed
during normal compilation.
Also, normalize to the class symbol immediately.
Add a TODO regarding only looking at class symbols,
and thus ignoring the term symbol for objects,
as the corresponding class symbol has all the relevant info.
Mention that private members are being extracted and included in the api
structures but ignored in many other parts of incremental compiler. I've
made a mistake of assuming that private members are ignored at api
extraction time. This manifested itself as bug #2324.
The signatures of methods that have value classes as arguments or return
type change during the erasure phase. Because we only registered
signatures before the erasure, we missed some API changes when a class
was changed to a value class (or a value class changed to a class).
This commit fixes this problem by recording the signatures of method
before and after erasure.
Fixessbt/sbt#1171
Fix the problem with unstable names synthesized for existential
types (declared with underscore syntax) by renaming type variables
to a scheme that is guaranteed to be stable no matter where given
the existential type appears.
The sheme we use are De Bruijn-like indices that capture both position
of type variable declarion within single existential type and nesting
level of nested existential type. This way we properly support nested
existential types by avoiding name clashes.
In general, we can perform renamings like that because type variables
declared in existential types are scoped to those types so the renaming
operation is local.
There's a specs2 unit test covering instability of existential types.
The test is included in compiler-interface project and the build
definition has been modified to enable building and executing tests
in compiler-interface project. Some dependencies has been modified:
* compiler-interface project depends on api project for testing
(test makes us of SameAPI)
* dependency on junit has been introduced because it's needed
for `@RunWith` annotation which declares that specs2 unit
test should be ran with JUnitRunner
SameAPI has been modified to expose a method that allows us to
compare two definitions.
This commit also adds `ScalaCompilerForUnitTesting` class that allows
to compile a piece of Scala code and inspect information recorded
callbacks defined in `AnalysisCallback` interface. That class uses
existing ConsoleLogger for logging. I considered doing the same for
ConsoleReporter. There's LoggingReporter defined which would fit our
usecase but it's defined in compile subproject that compiler-interface
doesn't depend on so we roll our own.
ScalaCompilerForUnit testing uses TestCallback from compiler-interface
subproject for recording information passed to callbacks. In order
to be able to access TestCallback from compiler-interface
subproject I had to tweak dependencies between interface and
compiler-interface so test classes from the former are visible in the
latter. I also modified the TestCallback itself to accumulate apis in
a HashMap instead of a buffer of tuples for easier lookup.
An integration test has been added which tests scenario
mentioned in #823.
This commit fixes#823.
This way we have a little bit more clear separation
between compiler phase logic and the core logic responsible for
processing each compilation unit and extracting an api for it.
As added benefit, we have a little bit less of mutable state
(e.g. sourceFile doesn't need to be a var anymore).
The API extraction logic contains some internal caches that are
required for correctness. It wasn't very clear if they have to
be maintained during entire phase run or just during single compilation
unit processing. It looks like they have to be maintained during
single compilation unit processing and refactored code both
documents that contracts and implements it in the API phase.
Martin Duhem
Guillaume Martres
Adriaan Moors
Krzysztof Romanowski
Grzegorz Kossakowski
Pierre DAL-PRA
Brian McKenna
Eugene Yokota