RFR: 8336492: Regression in lambda serialization [v10]

[Maurizio Cimadamore]

On Mon, 29 Jul 2024 18:22:12 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

>> This PR consolidates the code for dealing with local captures in both `Lower` and `LambdaToMethod`. It does so by adding a new tree scanner, namely `CaptureScanner`, which is then subclassed by `Lower.FreeVarCollector` and also by the new `LambdaToMethod.LambdaCaptureScanner`.
>> 
>> The main idea behind the new visitor is that we can compute the set of (most) captured locals w/o relying on complex state from `Lower`, and make the logic general enough to work on *any* tree. This can be done by keeping track of all local variable declarations occurring in a given subtree `T`. Then, if `T` contains a reference to a local variable that has not been seen, we can assume that this variable is a captured variable. This simple logic works rather well, and doesn't rely on checking e.g. that the accessed variable has the same owner of the method that owns a local class (which then caused other artifacts such as `Lower::ownerToCopyFreeVarsFrom`, now removed).
>> 
>> The bigger rewrite is the one in `LambdaToMethod`. That class featured a custom visitor called `LambdaAnalyzerPreprocessor` (now removed) which maintains a stack of frames encountered during a visit, and tries to establish which references to local variables needs to be captured by the lambda, as well as whether `this` needs to be referenced by the lambda. Thanks to the new `CaptureScanner` visitor, all this logic can be achieved in a very small subclass of `CaptureScanner`, namely `LambdaCaptureScanner`.
>> 
>> This removes the need to keep track of frames, and other ancillary data structures. `LambdaTranslationContext` is now significantly simpler, and its most important field is a `Map<VarSymbol, VarSymbol>` which maps logical lambda symbols to translated ones (in the generated lambda method). We no longer need to maintain different maps for different kinds of captured symbols.
>> 
>> The code for patching identifiers in a lambda expression also became a lot simpler: we just check if we have pending lambda translation context and, if so, we ask if the identifier symbol is contained in the translation map. Otherwise, we leave the identifier as is.
>> 
>> Fixing the issue referenced by this PR is now very simple: inside `LambdaCaptureScanner`, we just need to make sure that any identifier referring to a captured local field generated by `Lower` is re-captured, and the rest just works. To make the code more robust I've added a new variable flag to denote synthetic captured fields generated by `Lower`.
>> 
>> #### C...
>
> Maurizio Cimadamore has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fix more cases where EnclosingMethodAttribute is wrong for lambdas declared in instance field inits

> After trying different approaches, I realized that the cleanest (by far) solution is to simply capture the lambda owner symbol during `Attr`, instead of trying to reconstruct this owner using various (wrong) heuristics. When doing this, I realized that there were several issues which predated my refactoring - that is, the logic in `TypeAnnotations` relies on the `BLOCK` flag to be set on static/instance initializer owner method symbols - but sometimes this flag was missing (this is now fixed). Also, the `EnclosingMethodAttribute` sometimes reported `<clinit>` being the method owner of a local class. This is deliberately against the JVMS, which state:
> 
> > In particular, method_index must be zero if the current class was immediately enclosed in source code by an instance initializer, static initializer, instance variable initializer, or class variable initializer. (The first two concern both local classes and anonymous classes, while the last two concern anonymous classes declared on the right hand side of a field assignment.)
> 
> We even had a test which enforced the _wrong_ behavior (!!).
> 
> Last, the static/instance initializer symbol used to have a `null` type. This is now fixed, as we can always rely on the `BLOCK` flag instead (which is, again, cleaner).
> 
> As a result, `CaptureSiteInfo` is now gone.

I've ended up rewriting a biggie chunk in `Attr::lambdaEnv` as that seemed out of sync with what the rest of the compiler did (and also found other instances where `EnclosingMethod` contained a method index for lambdas **not** declared inside constructors)

-------------

PR Comment: https://git.openjdk.org/jdk/pull/20349#issuecomment-2256619954