RFR: 8300517: Refactor VisibleMemberTable (method members)

[Jonathan Gibbons]

On Mon, 6 Mar 2023 19:48:48 GMT, Pavel Rappo <prappo at openjdk.org> wrote:

> Please review this explorative refactoring for VisibleMemberTable (VMT).
> 
> This is the first round of refactoring for VMT. This round is about *method members*: declared (overriding and not) and inherited.
> 
> During this work I gained some insight into internal workings of VMT, fixed what was feasible and left TODOs and FIXMEs for everything else. Leaving those comments might look untidy, but leaving them out is wasteful: they clearly mark issues that should be revisited in upcoming rounds of refactoring.
> 
> As I see it today, the main issue with VMT is that implements complex and error-prone computations from Java Language Specification (JLS) by hand. For example, VMT interprets JLS rules for relations such as _inherits_, _overrides_ and _hides_. As one would imagine, sometimes VMT does it incorrectly. It would be better to eventually re-implement VMT using `javax.lang.model` as much as possible. Unlike that of `jdk.javadoc`, the day job of `javax.lang.model` is to provide JLS services.

src/jdk.javadoc/share/classes/jdk/javadoc/internal/doclets/toolkit/util/Utils.java line 1031:

> 1029:         // TODO: this computation should be eventually delegated to VisibleMemberTable
> 1030:         Set<TypeElement> alreadySeen = null;
> 1031:         assert (alreadySeen = new HashSet<>()) != null; // create set conditionally

I think this use of `assert` (here and lower down) is a step too far.  It's one thing to use asserts to verify invariants; it's too much to use them to conditionally compute state like this.

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

PR: https://git.openjdk.org/jdk/pull/12887