So I think I'm missing something obvious but I've ...

# arrow-meta

[dalexander]

So I think I'm missing something obvious but I've been stuck on this problem for a while, and can't find any examples solving a similar problem.
What I'm doing is generating a type adapter that takes classes of As and turns them into classes of B. This has gone pretty well, but I'm running issues with the next step I want to take, which is generating a "B provider" using the adapters. It seems like this requires knowing what all the adapters I've already created are, so I need a second phase of file generation, and this is where I've been running into issues. Are there any examples of generating files in the analysisCompleted phase, or modifying files in that phase? I haven't been able to figure out how to create a Transformation.newSources from within that function, or how to create a new KtFile successfully (the KtFile I've created and added to the files list have led to a NPE with the message 'VirtualFile is null for KtFile'). Any insight into how to create the provider in a second phase of analysis would be helpful, or a way of incrementally building a file during analysis that would only be written out when analysis was completed.

[raulraja]

Not sure if it helps but the current proof system finds providers of instances based on extension functions which is how you go there from a type to another. Do you have some sample code? What you are describing sounds close to our proof system in terms of impl

[dalexander]

That does sound somewhat similar. I can share some code in DM depending on what you're interested in seeing.

[dalexander]

I didn't mention in the initial info dump but I also just tried regenerating the provider file every time I create a new adapter, but even though there's only a single output file the compiler seems to find N instances of the file, which leads to an error message about redeclaring the "DataProvider" class.

[dalexander]

To answer my own question 😅 it appears as thought KtPsiFactory#createFile will potentially cause this NPE about the VirtualFile being null, the solution appears to be to call createPhysicalFile instead, which seems to resolve the issue.

[dalexander]

I'm still running into other complications about package paths, I'll try and share the solution when I find it so other people who need to do something similar will have some guideposts. It is notable that Transform.newSources<KtFile>(fileText.file(fileName, filePath)) in the analysisCompleted phase doesn't seem to work because internally it creates a KtFile that doesn't have a virtualPath, so trying to use that seems to cause a NPE I mentioned in the original message.

[raulraja]

that is most likely a bug, if you don’t mind creating an issue, it should be an easy fix where we can create instead a temp file

[dalexander]

I'll see about opening a ticket with the stack trace and what I learned about the underlying issue.

[dalexander]

fun Meta.reanalyze(
        shouldReanalyze: () -> Boolean,
        onReanalyze: () -> Unit
): ExtensionPhase =
        analysis(
                doAnalysis = { _, _, _, _, _, _ -> null },
                analysisCompleted = { project, module, bindingTrace, files ->
                    if (shouldReanalyze()) {
                        onReanalyze()
                        AnalysisResult.RetryWithAdditionalRoots(
                                bindingTrace.bindingContext,
                                module,
                                emptyList(),
                                listOfNotNull(),
                                addToEnvironment = true
                        )
                    } else {
                        null
                    }
                }
        )

In case anyone else out there is trying to manage a many-to-one mapping of source files/classes to output files, adding this particular extension phase to the extension phases in a meta() will cause analysis to be re-performed, then you can use quote syntax to generate new sources. Hopefully this saves someone some time.