generatesSources() fails, but doesn't attempt to compare like to like

[cgruber]

generatesSources() scans all sources for all expected sources, but in the case of a failure, only reports boolean failure. In the case where code is improperly generated, but some code is generated purportedly against a given type/filename, there is no display of "expected" vs. "actual" for comparison, and the user is left to guess at the result.

I wonder if the scanning could be done at the level of filename first, seeing if a file was generated as expected at all - since JavaSourceFile does already contain a notion of filename. THEN the equality scanner could be used against the result, and if there is failure, show "expected this, but got that". The current model is better than string comparison, but still too brittle to be truly useful.

[gk5885]

So, I like the idea, but there are some issues. JavaFileObject has no notion of file path. You have the full URI (and you may remember the issues with getting the path from that…) and the file contents. That's pretty much it. I don't think that any sort of 2-tier matching is in the works, but a nice list of "These files looked kind of similar" might be really helpful.

[cgruber]

Oh... right - it's the class name, not the path that goes into it. Hmm. That sucks. I guess I don't care so much about the mechanism, as much as the outcome, and yeah, the similarity list might work well.

[JakeWharton]

This just bit me hard for about 45m. Had to debug through a few times to realize the file path had no relevance and a typo in generated source was causing the failure.

You have the full URI (and you may remember the issues with getting the path from that…)

Is there an explanation here that's relatively brief?

[gk5885]

The issue with the path is that jar URLs don't have paths in the sense that you'd hope. It would just require special-casing throughout to hand them properly.

[JakeWharton]

Would package + class name suffice as something to at least do file-based matching?

At the very least, the error message here is misleading and could use a wording change.

[gk5885]

Oh, yes. Wow. That's a pretty bad message. We should definitely change that.

[dieend]

Is it possible to actually test again the generated file with another test?