On Fri, Jan 10, 2025 at 11:22 AM Magnus Ihse Bursie < magnus.ihse.bur...@oracle.com> wrote:
> On 2025-01-08 22:33, David Lloyd wrote: > > Thus I'm inclined to believe that this restriction does not serve any > practical benefit, and I hope it can be reconsidered to be respecified as > a compile-time-only check. > > I agree. I was bitten by this as a Java developer in a hobby open source > project, and even if at some point I understood why things where the way > they are, I do not remember the reasoning, and I still find it confusing > that it did not work as I expected it to, nor what steps I were supposed to > take instead to resolve the problem at hand. > > I understand that this is not a very common problem or a high-priority > issue, and I would have accepted that it was down-prioritized to the point > that it will take a long time to resolve, but the current "works as > intended" approach is still a bit hard for me to swallow. > I think there's a bit of self-selection: this kind of problem mostly appears in more complex systems; the work to move complex systems to JPMS is already nonzero; complex systems do not move because of this kind of problem; therefore, the problem is considered a non-issue because nobody is complaining about it. As an experimental workaround I was trimming optional dependencies from the descriptor at run time and doing a late `addReads` instead if the dependency is present. This part works fine, however I then also have to strip out `uses` and sometimes `provides` for this to work if they correspond to packages in the optional dependency (which is a check that also incurs a run time cost), and adding those back to the module late requires hacky workarounds: for `addUses` I have to generate a class into the target module so that it can call `Module.addUses` on my behalf, and for `addProvides`, I have to access private JDK internals thus requiring `--add-exports` of a core JDK package, which is less than ideal. I did propose a patch to allow controllers to add these but it was soundly rejected and the ensuing discussion yielded little of value. So hopefully we can get around it another way: by allowing `uses` and `provides` of services whose packages are not present at run time. -- - DML • he/him
