Use serialisation for save_analysis

[nrc]

Currently we use an ad hoc csv format for dumping analysis data (see rustc::middle::save::recorder). It would be better to use libserialize (or serde) for the dumping step so we can easily provide data in different formats.

By default, I think save_analysis should output JSON, since that seems to be the current hotness for this sort of thing.

I expect this to be not too hard, but non-trivial, and possibly a little bit tedious.

[sinistersnare]

cc me

[gamazeps]

Willing to work on this too :)

[gamazeps]

@sinistersnare Wanna split the work ? Or did you cc just in order to see the progress ?

[sinistersnare]

Sure. this could be interesting.

[rolandshoemaker]

Is this still being worked on or abandoned?

[nrc]

@sinistersnare @gamazeps ? (See previous comment)

[sinistersnare]

Oh goodness, I forgot about this :)

I can start work on this next week, but if you want to take this @rolandshoemaker, than feel free. Sorry!

[Potpourri]

I would like to work on this.

[sinistersnare]

Please do, I have no more plans of contributing.

[Potpourri]

@nick29581 Do I need to keep csv support?

[nrc]

@Potpourri tl;dr: no.

What would be ideal would be to have a generic serialisation API which could serialise to JSON, CSV or whatever else. I'm not sure if that is possible or how easy it is (@erickt might be able to advise). I'd be pretty happy with just JSON, as long as it wasn't too hard-wired in. (You could also earn mega-love from me by converting DXR to understand JSON rather than CSV as a follow up, but I'm also happy to do that).

[steveklabnik]

Triage: looks like we're still using the CSV format.

[aochagavia]

In #31838, I introduced the Dump trait as a first step to closing this issue (see the internal docs). Outputting JSON should be as simple as creating a new implementor of Dump that dumps everything as JSON instead of CSV.

For inspiration, see csv_dumper.rs.

[nrc]

I have a PR done now which adds JSON emission. It uses serialisation for that, but it doesn't really fix this issue. I would like to see save-analysis refactored so that rather than using the Dumper trait, we use a lowering step from internal data to external data (this is done for JSON, but not CSV) rather than using span utils and so forth. Then we shouldn't need a dumper at all, just take the external data structs and serialise them suing the serialiser of your choice.

We should also use these 'external' data structures for the API, rather than exposing the internal ones like we do at the moment.

[tamird]

Looks like a bunch of this was done in 9a47160. @nrc is there anything left?

[nrc]

Oh yeah, this is all done now, thanks!