boost: patch for INTERFACE_LINK_LIBS bug

[scpeters]

• Have you followed the guidelines for contributing?
• Have you checked that there aren't other open pull requests for the same formula update/change?
• Have you built your formula locally with brew install --build-from-source <formula>, where <formula> is the name of the formula you're submitting?
• Is your test running fine brew test <formula>, where <formula> is the name of the formula you're submitting?
• Does your build pass brew audit --strict <formula> (after doing brew install <formula>)?

---

In boost 1.75.0, the cmake INTERFACE_LINK_LIBRARIESS now include non-boost dependencies like libicudata from the
keg-only formula icu4c. This introduces linking errors for software that previously linked successfully in brew against boost 1.74.0 shared libraries (see #67427 and boostorg/boost_install#47).

This adds a patch to the boost formula that has been merged upstream to limit the expansion of INTERFACE_LINK_LIBRARIES to boost's static libraries, so that the linking errors with shared libraries are resolved. I've tested it locally to confirm that it resolves my issues documented in osrf/homebrew-simulation#1235.

[carlocab]

Thanks. Can you also include a comment before the patch about when we can expect it to be removed?

[scpeters]

Thanks. Can you also include a comment before the patch about when we can expect it to be removed?

I added a comment that it can be removed with the next release in b45380c

[carlocab]

I've opened a PR to fix cucumber-cpp: #67689

Unfortunately, I think it'll only fix the failure on Big Sur and not Mojave.

[carlocab]

I can't reproduce the pcl error locally (instead I am getting an entirely different build error I can't seem to fix), so my best guess given the error is to add

ENV.prepend_path "PKG_CONFIG_PATH", Formula["eigen"].opt_share/"pkgconfig"

to the start of pcl's test block. (Yes, that's where eigen's pkgconfig directory is. I don't know why either.)

[scpeters]

I can't reproduce the pcl error locally (instead I am getting an entirely different build error I can't seem to fix), so my best guess given the error is to add

ENV.prepend_path "PKG_CONFIG_PATH", Formula["eigen"].opt_share/"pkgconfig"

to the start of pcl's test block. (Yes, that's where eigen's pkgconfig directory is. I don't know why either.)

I believe it's a bug in homebrew-test-bot. I noticed in the console log of these failing jobs that just before freeling is tested, it calls brew unlink eigen because freeling conflicts with eigen. It remains unlinked when later testing pcl, which fails. Explicitly adding eigen's opt folder to the pkgconfig path would work around this

This didn't fail in #66792, because pcl had a revision bump in that PR, so it took a different code path in test-bot

[scpeters]

This didn't fail in #66792, because pcl had a revision bump in that PR, so it took a different code path in test-bot

I'll finish the thought here. I looked in the logs of #66792 and noticed the same brew unlink eigen before freeling but also a brew link eigen before pcl was built. I believe this is because there is logic for matching brew unlink for conflicting formulae with calls to brew link for unlinked dependencies (see setup_formulae_deps_instances() in lib/tests/formulae.rb).

By contrast, I believe pcl is handled as a dependent in this PR (possibly in this code?), which doesn't check for unlinked dependencies

[scpeters]

I believe it's a bug in homebrew-test-bot

I'll file an issue for this

[scpeters]

I believe it's a bug in homebrew-test-bot

I'll file an issue for this

Homebrew/homebrew-test-bot#544

[scpeters]

I can't reproduce the pcl error locally (instead I am getting an entirely different build error I can't seem to fix), so my best guess given the error is to add

ENV.prepend_path "PKG_CONFIG_PATH", Formula["eigen"].opt_share/"pkgconfig"

to the start of pcl's test block. (Yes, that's where eigen's pkgconfig directory is. I don't know why either.)

I've rebased and added this workaround in 92701db

[carlocab]

On 11.0

make[2]: *** No rule to make target `/Library/Developer/CommandLineTools/SDKs/MacOSX11.0.sdk/usr/lib/libz.tbd', needed by `lib/libpcl_surface.1.11.1.dylib'.  Stop.

One of pcl's (possibly recursive) dependencies has the SDK path baked into its compiled artifacts. It will need to be rebuilt. I believe it's [email protected] (or dependencies).

Cf. #67474, Homebrew/brew#10127

Update: Tried rebuilding only [email protected]. Didn't cut it on its own, sadly. Will look into it some more.

[scpeters]

On 11.0

make[2]: *** No rule to make target `/Library/Developer/CommandLineTools/SDKs/MacOSX11.0.sdk/usr/lib/libz.tbd', needed by `lib/libpcl_surface.1.11.1.dylib'.  Stop.

One of pcl's (possibly recursive) dependencies has the SDK path baked into its compiled artifacts. It will need to be rebuilt. I believe it's [email protected] (or dependencies).

Cf. #67474, Homebrew/brew#10127

Update: Tried rebuilding only [email protected]. Didn't cut it on its own, sadly. Will look into it some more.

I noticed a reference to MacOSX11.0.sdk in the netcdf bottle, so I've rev-bumped that and [email protected], and we'll see if that helps

[carlocab]

Tried building pcl locally again.

The only other dependency with zlib linkage is flann, so I guess that's worth giving a revision bump.

[email protected]'s dependencies with zlib linkage:

• hdf5
• libpng
• libtiff
• netcdf
• qt

I've also looked whether we need to recurse further down the dependency tree from there, and it (thankfully) doesn't look like it.

[carlocab]

Rebuilding flann seems to help, so I've given it a revision bump and rebased this against master.

pcl is still compiling as I write this, but it's gotten well past the point where CI failed now.

[carlocab]

Ok, no idea what happened here, but the CI run shouldn't be taking this long. Will try to restart the tests.

[mitchblank]

This makes me worried about my own mariadb-connector-odbc change 30e214e

I added the reference to the .tbd file because the project's cmake files expected some physically-present file for that setting (so on Big Sur /usr/lib/libiconv.dylib won't work, nor was it happy with just -liconv) It seemed that passing in the tbd file to the linker would be a harmless way of getting the -liconv behavior since the the tbd file is just a description of how to link to the library, but I guess it's possible for it to end up with the tbd as a runtime dependency?!?

[carlocab]

CI is green on the Intel Big Sur runner now. Except that the Catalina runner seems to have gotten stuck somewhere...

[BrewTestBot]

@carlocab has triggered a merge.

[scpeters]

thanks for the help @carlocab!