Need ways to distinguish Illumos from Solaris

[rorth]

Despite their common ancestry, Solaris and Illumos have diverged in some areas since Oracle closed the OpenSolaris project more than 10 years ago. Recently, the need to distinguish the too came up in several reviews. Instead of having the information spread over multiple places, I'm collecting them here. Some of this was suggested in D119695:

We should probably follow the model used in Go build tags where solaris is defined for both Solaris and Illumos, while illumos is specific to Illumos. This way considerable code can be shared while handling differences as need be.

• We need a macro (__illumos__) to distinguish the two. This needs to work both with bundled and vanilla versions of clang and gcc, so this should be done in two places.
• Define __illumos__ based on the uname -o output e.g. in llvm/CMakeLists.txt. I have a prototype for that.
• Have clang predefine it, too. This depends on...
• Provide isOSIllumos() in clang. I'm not yet certain how best to do this, though: normally the isOS*() methods just look at the triple, which is identical between the two. Maybe something like llvm/lib/Support/Unix/Host.inc (updateTripleOSVersion) can be used, but rewriting the triple to *-*-illumos would mean massive changes to the testsuite, contradicting the model above.
• The cmake files use ${CMAKE_SYSTEM_NAME} MATCHES "SunOS" tests all over the place. Not sure yet how best to handle this.

We may discover more along the way, though.

[brad0]

I had a PR pending to fix some linking issues on Illumos. #74930

But I started thinking about what else needs fixing. This touches on most of what else needs to be done.

[brad0]

So, building OpenMP on illumos I have run into a bit of an annoying issue.

llvm-project/openmp/runtime/src/z_Linux_util.cpp

Line 440 in 7c77355

status = pthread_attr_get_np(pthread_self(), &attr);

#if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
    status = pthread_attr_get_np(pthread_self(), &attr);
    KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
#else
    status = pthread_getattr_np(pthread_self(), &attr);
    KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
#endif

Solaris 11.4 implements pthread_getattr_np(). illumos implements pthread_attr_get_np().

[rorth]

That's the fundamental problem with LLVM's "cross at all times" attitude. While this is understandable (though ugly) in llvm and clang, I see no reason why target libs like openmp shouldn't use cmake tests to determine such target properties instead of trying to hardcode them. compiler-rt already does FWIW. Platforms dovary over time and between versions, like it or no.

[brad0]

I see no reason why target libs like openmp shouldn't use cmake tests to determine such target properties instead of trying to hardcode them. compiler-rt already does FWIW.

That would be better. I see no indication in the CMake files that OpenMP does any platform checks.

Platforms dovary over time and between versions, like it or no.

That's the case with all OS's.

[rorth]

I see no reason why target libs like openmp shouldn't use cmake tests to determine such target properties instead of trying to hardcode them. compiler-rt already does FWIW.

That would be better. I see no indication in the CMake files that OpenMP does any platform checks.

Probably because the project culture prefers static checks instead.

Platforms dovary over time and between versions, like it or no.

That's the case with all OS's.

True, but either the differences between versions are smaller or can be capture e.g. in OS version checks. Solaris doesn't give you that: it has been 11.4 since 2018 now (and 11 since 2011), and is going to stay that way for its remaining live time.