Bump minimum toolchain version

[thieta]

Summary

RFC: https://discourse.llvm.org/t/rfc-increasing-the-gcc-and-clang-requirements-to-support-c-17-in-llvm

Following the policy here: https://llvm.org/docs/DeveloperPolicy.html#toolchain

This forum post here will be updated with the timeline and status: https://discourse.llvm.org/t/important-new-toolchain-requirements-to-build-llvm-will-most-likely-be-landing-within-a-week-prepare-your-buildbots/61447

[xbolva00]

GCC X.1 / 2 releases tend to be quite buggy. Maybe 7.5 is better option?

[cor3ntin]

LGTM

[tschuett]

Do you want to say something about buildbots, C++ features, and that you are not committing until foo.

[efriedma]

Did anyone actually try building LLVM with the "minimum required" compilers? I recall having problems with an old build config using clang 5 not too long ago...

[mehdi_amini]

I have a bot building with clang-5, and one building with gcc 5.x, but I’ll happily upgrade them to whatever is the new minimum

[thieta]

In D122976#3424609, @xbolva00 wrote:

GCC X.1 / 2 releases tend to be quite buggy. Maybe 7.5 is better option?

While that sounds nice - the versions in this change is the ones we have discussed in the RFC and the higher we bump versions the more people we are going to lock out - so I think we'll keep 7.1 for now and if we encounter specific breakages / bugs in that version we might update the document and suggest people use the latest 7.x version instead.

In D122976#3424764, @tschuett wrote:

Do you want to say something about buildbots, C++ features, and that you are not committing until foo.

Here is the rough plan I am thinking right now and I will post in the forums soon:

• Announcement in the forums that this is happening today or tomorrow.
• Let's allow a week for buildbot owners to upgrade before we merge this commit. Maybe merge it end of this week or maybe beginning of next.
• The soft error will live in main until we branch for LLVM 15.x
• When that has been branched - we will convert it to a hard error in main and up the CXX_VERSION to 17.
• LLVM 16.x will require the new toolchain and you will be allowed to use C++17 features.

[thieta]

Added note about toolchain requirements to release notes

[jhenderson]

LGTM, but:

Announcement in the forums that this is happening today or tomorrow.

I'd give it maybe a few more days than that (perhaps the end of the week or start of next), to ensure opportunity for those who have a bit of time off to chime in, in case they didn't see the original RFC/realise it had reached a conclusion.

[thieta]

In D122976#3425649, @jhenderson wrote:

LGTM, but:

Announcement in the forums that this is happening today or tomorrow.

I'd give it maybe a few more days than that (perhaps the end of the week or start of next), to ensure opportunity for those who have a bit of time off to chime in, in case they didn't see the original RFC/realise it had reached a conclusion.

Yeah I have no idea why I wrote today or tomorrow here when I was thinking end of this week or start of next one - as I wrote in the post in the forum as well. Must have been before my second cup of coffee. Good spot though!

[jyknight]

llvm/cmake/modules/CheckCompilerVersion.cmake

22

Replace this with set(LIBSTDCXX_SOFT_ERROR 7) and use per comment below.

100–101

As this change increments the minimum GCC version to 7, we should switch to using _GLIBCXX_RELEASE to check the libstdc++ version. This is better than using __GLIBCXX__ dates, because the dates are updated for every patch release on every branch, so it cannot actually distinguish between branches properly.

The hack of checking iostream_category can go away, as well (it was just a workaround for the unreliability of GLIBCXX).

So, the whole check below should be able to be simplified to:

#if defined(__GLIBCXX__)
#if !defined(_GLIBCXX_RELEASE) || _GLIBCXX_RELEASE < ${LIBSTDCXX_SOFT_ERROR}
#error Unsupported libstdc++ version
#endif
#endif
int main() { return 0; }

[thieta]

Simpilify the libstdc++ check - as suggested by jyknight

[thieta]

@jyknight Addressed your comments - I did test it on my system to fudge the required versions a bit. But if someone have access to a GCC < 7.1 system it might be good to test it there. Otherwise I'll try to figure out how to test it in a docker container maybe.

[thieta]

Gentle Ping - I am planning to land this tomorrow, unless I get feedback to the opposite.

[h-vetinari]

llvm/cmake/modules/CheckCompilerVersion.cmake

110

Was this additional indent intentional? Shouldn't then the whole content of check_cxx_sources_compiles be indented?

[thieta]

Fixed un-intentional indent.

[thieta]

llvm/cmake/modules/CheckCompilerVersion.cmake

110

that was not intentional no - probably my editor doing something funny. Thanks for spotting that.

[jyknight]

LG, just some minor comments.

llvm/cmake/modules/CheckCompilerVersion.cmake

100–101

Delete the comment; it's now incorrect, and no longer necessary since the test is now obvious.
Delete unnecessary #include <iosfwd>.

[thieta]

llvm/cmake/modules/CheckCompilerVersion.cmake

100–101

iosfwd was needed in order for _GLIBCXX_RELEASE to be defined - maybe we could include a smaller header to get that define - but I couldn't delete it.

[thieta]

Removed unnecessary comment.