Implicit struct _GUID on Windows crashes DeclRewriter::detectInlineStruct

[mattmccutchen-cci]

By default on Windows, Clang 11 injects an implicit struct _GUID declaration into every translation unit; this is new since Clang 9. This declaration has an invalid source location, so it causes an assertion failure in DeclRewriter::detectInlineStruct in 129 of our regression tests. We should probably just skip inline struct detection if the struct source location is invalid.

While the struct _GUID injection is enabled by default only on Windows, we can enable it on other OSes for testing purposes using the -fms-extensions compiler option. Thus, 3c ms-extensions.c -- -fms-extensions reproduces the problem if ms-extensions.c contains:

int *x;

Some options for testing:

1. Just add the above as a regression test.
2. Add an -fms-extensions regression test containing a manually curated wider variety of constructs.
3. Develop a convenient way to run the entire regression test with -fms-extensions (likely easier with Refactor test RUN commands into a separate tool #355) and do that occasionally (we probably don't want to spend the time to run it on every change).
4. Just rely on real testing on Windows to catch issues with -fms-extensions and anything else that differs on Windows.

I'm not sure which one offers the best trade-off between value and work.

[kyleheadley]

I don't know any more than you've written here, so I vote 1. Maybe 2 with some structs if you have further reason to believe there will be a conflict with code. But 3 is too expensive and 4 catches problems too late.

[mattmccutchen-cci]

Option 1 sounds good. Thanks Kyle. PR coming soon.