Crossgen /createpdb option doesn't have enough test coverage

[MichalStrehovsky]

I would expect we would at least try to use createpdb with Pri0 tests and make sure it's not bogus.

The option didn't work when default interface methods are present.

[mangod9]

@trylek -- Assume this is still relevant for crossgen2.

[trylek]

The test coverage is still poor; perhaps we could put something together as part of the 6.0.0 quality week of 8/23 (if I remember the date correctly). An interesting question is what exactly we need to proclaim that the PDB "is not bogus" in the sense of issue description. We'd either need to run some debugger tests or use the PDB manually to translate ETW information or some such; perhaps @hoyosjs might have some ideas or suggestions what could be the sweet spot on the balance between complexity of the testing process and the level of PDB validation.

[MichalStrehovsky]

An interesting question is what exactly we need to proclaim that the PDB "is not bogus"

The original bug I ran into when I filed this issue was that we would straight up crash while generating the PDB. Looks like the bug @trylek ran into in #70407 would be detectable by checking whether the PDB was generated at all?

If so, maybe the min bar would be just to: ask crossgen to generate a PDB + check some PDB is on disk after crossgen is done.

[trylek]

Sadly in this case the PDB actually got generated, it was just bogus due to not being updated upon refcount on the DIA writer going to zero. At the first glance I would be inclined to hand-compile either a test or a framework assembly (SPC being the easiest choice), use DIA to enumerate its symbols and check it for a few hand-picked known methods. I'm certainly open to other suggestions.

[trylek]

Another option might be to leverage the existing Crossgen determinism test by modifying it to build a PDB on Windows and compare the symbols in the two versions (in addition to checking for some known entries).

[MichalStrehovsky]

Sadly in this case the PDB actually got generated, it was just bogus due to not being updated upon refcount on the DIA writer going to zero

Yeah, that's sad. Looks like DIA is the only option to figure out if the contents is sane.

[agocke]

We could probably use PdbToXml to validate that the contents of a small assembly are what we expect.