Add support for modular build structure.

[grafikrobot]

This is part of the effort to make the Boost libraries "modular" for build and consumption. See https://lists.boost.org/Archives/boost/2024/01/255704.php and https://github.com/grafikrobot/boost-b2-modular/blob/b2-modular/README.adoc for more information.

This PR depends on the following other PRs being merged to both develop and master branches of the respective repos:

• Add support for modular build structure. boost#854

This PR will be changed to ready for review, i.e. not draft, when the above are merged. Do not merge this one until that time.

[grafikrobot]

Please review and merge this PR at your earliest convenience.

[HDembinski]

I didn't see this, because I am not getting notifications when a PR is opened unless someone mentions me with @HDembinski. I changed something in the options of this repo, putting my email address in explicitly under email notifications. Perhaps that is what was missing.

@grafikrobot Thank you for the changes to the build. There seems to be an issue with boost.random in the slow tests, a linker error. Do you have an idea how to fix this? You also remove some extra warning flags cxxflags=-Wno-unused-parameter cxxflags=-Wno-uninitialized, why?

[grafikrobot]

You also remove some extra warning flags cxxflags=-Wno-unused-parameter cxxflags=-Wno-uninitialized, why?

Those are for avoiding warnings in Math (IIRC) that cause errors because of the warnings-as-errors option.

[grafikrobot]

@HDembinski

There seems to be an issue with boost.random in the slow tests, a linker error. Do you have an idea how to fix this?

Have those slow tests ever passed? Asking as I don't see them in previous commits.

[HDembinski]

@HDembinski

There seems to be an issue with boost.random in the slow tests, a linker error. Do you have an idea how to fix this?

Have those slow tests ever passed? Asking as I don't see them in previous commits.

Yes, they used to pass. I only run them on PRs, but not on pushes, that's why they don't show up in the list of commits.

[HDembinski]

@grafikrobot Thank you for the work, I will look into this on the weekend. The problem for me also is that I am not developing on a Mac anymore, I have a Windows machine now (urgh). I think I can fix all these three issues.

[HDembinski]

So, I finally had a closer look.

* The coverage fails after doing the building. I'm not familiar with the coverage it's doing to decipher why the lcov file is missing.

It looks like ../../b2 $B2_OPTS toolset=gcc-${GCC_VERSION} cxxstd=latest coverage=on test//all doesn't compile anything. No tests are run and no coverage data is created. Any idea why? The alias test//all should run all tests.

If I remember correctly, lcov only works with a matching version of gcc. If you know a better way to measure coverage in C++, I'd love to switch to another tool. You upgraded from gcc-11 to gcc-15, I will probably need to upgrade lcov as well.

* The slow clang build detects ODR violations.

That's because of boost-serialization if you look closely, which you added as a constant dependency to this build. boost-serialization causes lots of issues, that's why I keep it as an optional dependency that is only compiled for some tests that are not as strict. I run harder tests on boost-histogram itself.

* The slow gcc test detects dangling temp reference error/warning in the main test.

That's a newly discovered bug - it seems.

[HDembinski]

@grafikrobot Progress.

The coverage test still fails, because it doesn't compile anything. test//all works in other circumstances. Weird.

The gcc slow test still fails because of that "possibly dangling pointer". As far as I understand, the pointer shouldn't be dangling because of NRVO. The issue is not detected by another other compiler and not caught by the UBSAN / ASAN tests, so perhaps a false positive. As a workaround, we could turn off that warning for gcc for that line.

[grafikrobot]

The coverage test still fails, because it doesn't compile anything. test//all works in other circumstances. Weird.

I can't repro that locally, with the same compiler. So definitely strange why in CI it would not build the tests in this case.

[grafikrobot]

@HDembinski FYI.. I'm now thinking there's a bug in B2 parsing the CLI arguments in this particular instance. As it's behaving as if the test//all, or test//quick, argument isn't there.

[HDembinski]

Bizarr. I haven't tried to reproduce locally, but the same issue should show up then if I use the same args. Will try this.

[HDembinski]

@grafikrobot Everything fixed now. I will do some finishing touches, then merge.

[grafikrobot]

@HDembinski many, many, thanks for figuring out the build problems to get this merged!

[HDembinski]

Thanks for all the help!