[fmt] Do not support Clang 20 + cppstd=20 due string literal evaluation

[uilianries]

Summary

Changes to recipe: fmt/[*]

Motivation

closes #26169

/cc @AbrilRBS

Details

It was reported by @skhaz on issue #26169, the same bug was reported to fmt directly: fmtlib/fmt#4264

Probably fixed by fmtlib/fmt#4177, but there is no milestone indicated for that hotfix.

The error ONLY occurs when building for C++20 (or newer) and using Clang 20. Here are some scenarios used for the validation:

• fmt-10.2.0-linux-cppstd20.log
• fmt-10.2.1-emscripten-cppstd20.log
• fmt-11.0.2-emscripten-cppstd14.log
• fmt-11.0.2-emscripten-cppstd17.log
• fmt-11.0.2-emscripten-cppstd20.log
• fmt-11.0.2-linux-cppstd20.log
• fmt-11.0.2-emscripten-cppstd23.log

---

• Read the contributing guidelines
• Checked that this PR is not a duplicate: list of PRs by recipe
• Tested locally with at least one configuration using a recent version of Conan

[uilianries]

Thank you for your comments, after digging a bit more, the problem is only related to the major versions of FMT: 8, 10, and 11, but when using Clang 20 only.

As reported on FMT issue 4177, the clang 20 changed its compile-time evaluation for string literals, resulting in that reported error. The same only affects some versions of FMT.

Still, the commit fmtlib/fmt@cacc310 is not enough, as commented in the same issue, but it also requires the PR fmtlib/fmt#4247.

Those changes are in the master branch, but no milestone is pointed out; the latest fmt release is from July 2024, so we do not have that fix available in binaries.

I would not backport those patches because Clang 20 is still in progress (https://clang.llvm.org/docs/ReleaseNotes.html) and may have a different result in the future. Plus, it would need to be backported for multiple versions in CCI.

I built locally the FMT by using different versions, plus using Clang 20 and Clang 19 (all using compiler.cppstd=20):

• fmt-11.0.2-linux-clang19.log: passed, no warnings
• fmt-11.0.2-linux-clang20.log: error: call to consteval function
• fmt-11.0.2-linux-clang18.log: passed, no warnings
• fmt-10.2.1-linux-clang20.log: error: call to consteval function
• fmt-10.2.1-linux-clang19.log: passed, no warnings
• fmt-10.0.0-linux-clang20.log: error: call to consteval function
• fmt-9.1.0-linux-clang20.log: format.h:4168:27: warning: identifier '_a' (passed)
• fmt-9.0.0-linux-clang20.log: format.h:4143:27: warning: identifier '_a' (passed)
• fmt-8.1.1-linux-clang20.log: format-inl.h:2538:28: error: call to consteval function
• fmt-7.1.3-linux-clang20.log: format.h:3943:48: warning: identifier '_a' (passed)
• fmt-6.2.1-linux-clang20.log: format.h:589:74: warning: identifier '_u' (passed)
• fmt-5.3.0-linux-clang20.log: format.h:423:33: warning: identifier '_u' (passed)

[uilianries]

After new test, the error only occurs with Clang 20, but using C++20 or later (I only tested using C++23, not 26). However, using C++17 all versions build without that reported error.

I updated the restriction to only invalidate the case when using Clang 20 + C++20 (or later)

Local build logs:

• fmt-5.3.0-linux-clang20-cppstd17.log: passed
• fmt-5.3.0-linux-clang20-cppstd23.log: format.h:423:33: warning: identifier '_u' (passed)
• fmt-6.2.1-linux-clang20-cppstd17.log: format.h:589:74: warning: identifier '_u' (passed)
• fmt-6.2.1-linux-clang20-cppstd23.log: format.h:589:74: warning: identifier '_u' (passed)
• fmt-7.1.3-linux-clang20-cppstd17.log: format.h:3943:48: warning: identifier '_a' (passed)
• fmt-7.1.3-linux-clang20-cppstd23.log: format.h:3943:48: warning: identifier '_a' (passed)
• fmt-8.1.1-linux-clang20-cppstd17.log: format.h:3036:27: warning: identifier '_a' (passed)
• fmt-8.1.1-linux-clang20-cppstd23.log: format-inl.h:2534:30: error: call to consteval
• fmt-9.0.0-linux-clang20-cppstd17.log: format.h:4143:27: warning: identifier '_a' (passed)
• fmt-9.0.0-linux-clang20-cppstd23.log: format.h:4143:27: warning: identifier '_a' (passed)
• fmt-9.1.0-linux-clang20-cppstd17.log: format.h:4168:27: warning: identifier '_a' (passed)
• fmt-9.1.0-linux-clang20-cppstd23.log: format.h:4168:27: warning: identifier '_a' (passed)
• fmt-10.0.0-linux-clang20-cppstd17.log: format.h:4654:27: warning: identifier '_a' (passed)
• fmt-10.0.0-linux-clang20-cppstd23.log: format-inl.h:1395:30: error: call to consteval
• fmt-10.2.1-linux-clang20-cppstd17.log: no warnings (passed)
• fmt-10.2.1-linux-clang20-cppstd23.log: format-inl.h:1382:35: error: call to consteval function
• fmt-11.0.2-linux-clang20-cppstd17.log: no warnings (passed)
• fmt-11.0.2-linux-clang20-cppstd23.log: os.cc:176:35: error: call to consteval function

[jcar87]

Thanks @uilianries for the very thorough testing.

Erring on the side of caution here: clang-20 has not been released yet, and neither has clang-19 for that matter - so I'm not sure we should add validations for unreleased versions - on the odd chance that it is a compiler bug that might be fixed before the release. I appreciate emsdk is a problem in this regard since it uses git snapshots rather than releases.

I suspect the most recent fmt 11.x already fixes this - and the check in this PR would cover a version where this is fixed, so it would be undesirable.

Not sure there is much for us to do in the recipe - these things happen when developers are on the bledeing edge: latest emsdk, at-head llvm-clang snapshot, latest fmt, c++20. If the developer's goal is to be trailblazers and discover bugs: great. If not, my advice to them would be to be a tad more conservative a use well known versions instead.