[cmake] remove fail-on-missing=OFF path for all dependencies that back opt-in features

[ferdymercury]

meaning features that have no builtin_option as backup, and that are mostly OFF by default.

With the new behavior, an error is raised instead of that feature being automatically turned OFF. The users is forced to actively install the dependency or to manually turn back that feature to OFF.

Split from #23015 as suggested by @guitargeek

[guitargeek]

Thank you very much!

We should make sure that every error path has an actionable error message.

Also, you're missing to give the same treatment to find_package(Arrow) and find_package(DCAP), which are dependencies of the opt-in features arrow and dcache.

[guitargeek]

Sorry @ferdymercury for these multiple iterations of review, but this should hopefully be the final one, at least from my side!

[guitargeek]

LGTM! Thanks for the improvements and you patience with my comments.

I think this is a great step forward:

• You're not getting the options that you explicitly set ON silently disabled anymore even without fail-on-missing, which I think was quite annoying
• I think the only possible counterargument to this PR is that it affects uses who set -Dall=ON and rely on features getting auto-disabled. To that I would say:
  • We're using SEND_ERROR for the error messages, so one gets a clear configuration error with all the accumulated errors about missing dependencies to act on in one go. This is not a major inconvenience, and we document this in the release notes.
  • The all option should not be encouraged anyway, as we're doing not the best job in syncing its logic with the actual build flags. tmva-sofie, experimental_adaptivecpp, experimental_genvectorx, or roofit_multiprocess are not enabled, for example. That's also because the expected semantics are not clear. Should all include experimental preview features?
• The CMake code is greatly simplified

Please squash the commit history before merging or squash on merge

[guitargeek]

I have approved the PR, but it would be good to get a second review or approval from other build system experts before merging: @hageboeck, @bellenot, @dpiparo, what do you think?

[hageboeck]

Hello, I think this is going in a good direction! Eventually, I would even like fail-on-missing to be removed.

For me, the only question is how we can be more helpful. In every message, I would hope to see:

• The package that CMake failed to find
• The option that was asking for it
• A hint to <PackageName>_ROOT and/or CMAKE_PREFIX_PATH

Should we craft a standard message that is issued everywhere? Here is a first draft:

message(SEND_ERROR "libuuid not found while option -Ddaos=On or -Ddaos_mock=On. "
  "Disable this option, install the dependency, or make it visible to CMake by setting <PackageName>_ROOT or adding it to CMAKE_PREFIX_PATH.")

We could even think about putting a macro at the top of SearchInstalledSoftware.cmake, but we wouldn't save a lot of lines.

[ferdymercury]

Eventually, I would even like fail-on-missing to be removed.

That's #23015 ;)

Should we craft a standard message that is issued everywhere?

Sounds good, but in my opinion we should go first with the functionality change (agreeing on how much we remove) and in a last step we can focus on message standardization once the number of code lines has decreased.

There is also #23020 that will help in reducing number of lines before standardizing.

[github-actions]

Test Results

    22 files      22 suites   3d 13h 39m 56s ⏱️
 3 856 tests  3 856 ✅ 0 💤 0 ❌
75 691 runs  75 691 ✅ 0 💤 0 ❌

Results for commit 190c09d.

[guitargeek]

Thanks @hageboeck for the additional comments and to @ferdymercury for addressing them!

Indeed, I look forward to see improved error messages, once we have converged to the set of error paths that we actually want to keep.