Improve logging of GC ref map mismatches

[trylek]

As part of investigation of the GC refmap mismatch I'm fixing with

#80218

I have improved CoreCLR runtime behavior in the presence of GC refmap mismatches by adding debug-only logic to dump both the GC ref map emitted by Crossgen2 and the one calculated by the native runtime. The dump format matches the way R2RDump outputs GC ref maps next to method import cells in the R2R file.

Thanks

Tomas

/cc @dotnet/crossgen-contrib

C:\git\runtime>C:\triage\wpf\bin\Debug\net7.0-windows\win-arm64\publish\wpf.exe
GC ref map mismatch detected for method: System.Windows.Media.PathGeometry::GetPathBoundsAsRB
Runtime GC ref map: I(70) R(78) I(80)
Crossgen2 GC ref map: I(68 70) R(78) I(80)

Assert failure(PID 12596 [0x00003134], Thread: 9568 [0x2560]): false

CORECLR! CheckGCRefMapEqual + 0x1DC (0x00007fffe66fa72c) CORECLR! ExternalMethodFixupWorker + 0xA98 (0x00007fffe681b3c8)
CORECLR! DelayLoad_MethodCall + 0x58 (0x00007fffe6581a18) PRESENTATIONCORE! <no symbol> + 0x0 (0x00007fffe401ecac)
PRESENTATIONCORE! + 0x0 (0x00007fffe401eaa4) PRESENTATIONFRAMEWORK! <no symbol> + 0x0 (0x00007fffdbd34db4)
PRESENTATIONFRAMEWORK! + 0x0 (0x00007fffdbd33d6c) PRESENTATIONFRAMEWORK! <no symbol> + 0x0 (0x00007fffdbc9a85c)
PRESENTATIONCORE! + 0x0 (0x00007fffe3ecde14) PRESENTATIONFRAMEWORK! <no symbol> + 0x0 (0x00007fffdc13ce5c)
File: C:\git\runtime\src\coreclr\vm\frames.cpp Line: 2177
Image: C:\triage\wpf\bin\Debug\net7.0-windows\win-arm64\publish\wpf.exe

[trylek]

(apologies for updating the bug description, I originally copied an earlier version of the CoreCLR runtime dump where I had reversed the tags for Crossgen2 vs. Runtime)

[janvorli]

The Unix builds are failing because of the wprintf usage. We have recently removed support for wprintf from PAL, see #77852.

[trylek]

@janvorli - thanks for the heads-up, so does that mean that I should be using printf instead?

[janvorli]

so does that mean that I should be using printf instead?

Yes

[janvorli]

And also not to use %S in printf, as on Unix it would expect 32 bit char.

[janvorli]

LGTM, thank you!

[AntonLapounov]

Great improvement. I have a question regarding dumping: we skip GCREFMAP_SKIP tokens when we dump; however, in theory it is possible that the two GCRefMaps differ exactly in one of GCREFMAP_SKIP tokens, in which case the difference may not be easily seen from the output.

Nit: this may need a trailing space.

printf("POP(0x%x)", stackPop);

[trylek]

Thanks Anton for your feedback. I think that GCREFMAP_SKIP can only become an issue at the end of the GC refmap dump, otherwise it would distort the offset and trigger a mismatch for the subsequent non-skip tokens. From this viewpoint I guess we might want to log trailing GCREFMAP_SKIP entries loosely akin to how Roslyn now reports stray whitespace at line ends. I think you're also right about the POP, I "copied" that bug from R2RDump which behaves the same I believe. I'll put up a PR to fix these two issues, thanks for pointing them out!