[ci] build with clingtest flag active if interpreter folder was touched

[ferdymercury]

This Pull request:

Changes or fixes:

Fixes #15230
Fixes #6957

Sibling PR in root-project/roottest#1232

Checklist:

• tested changes locally
• updated the docs (if necessary)

[dpiparo]

This PR is a great attempt to introduce a useful new feature. It seems some information is still needed before starting to test it thoroughly.

[dpiparo]

@vgvassilev

[vgvassilev]

These changes are too advanced for me. Maybe @mcbarton could take a look...

[github-actions]

Test Results

    18 files      18 suites   4d 5h 33m 48s ⏱️
 2 690 tests  2 665 ✅ 0 💤 25 ❌
46 567 runs  46 519 ✅ 0 💤 48 ❌

For more details on these failures, see this check.

Results for commit 3db45e0.

♻️ This comment has been updated with latest results.

[ferdymercury]

It's working now.

Should we also:

• set LLVM_BUILD_TYPE=Debug ?
• add more platforms ? Right now: both windows, one mac, one ubuntu, one linux

[hahnjo]

This looks really (too) complicated and IMHO is not what we actually want:

• check-cling doesn't seem to work in-tree on Windows.
• On Apple Silicon, we (still) cannot catch exceptions from interpreted code and therefore many (all?) NullDeref tests fail. We probably want to start with an x86 machine.
• For Linux, as I predicted in #6957 (comment), exposing the LLVM symbols from libCling interferes with other libraries, in this case TensorFlow (but I think we have seen problems with Mesa in the past as well).
• roottest-root-meta-countIncludePaths checks the number of include paths, which is larger with clingtest=ON to make the tests pass.

In general, as was discussed before, we still want to run all PR builds because this is what matters for ROOT's functionality. On top of this, we want extra builds with clingtest that only execute check-cling. This can be accomplished much easier with a separate workflow that triggers on pull_request with paths.

[ferdymercury]

* On Apple Silicon, we (still) cannot catch exceptions from interpreted code and therefore many (all?) `NullDeref` tests fail. We probably want to start with an x86 machine.

Changed.

* `roottest-root-meta-countIncludePaths` checks the number of include paths, which is larger with `clingtest=ON` to make the tests pass.

This should be fixed in my opinion in the CMakeLists using an if(clingtest) --> pass argument of extra include paths. (build_root.py should forward this flag to roottest as -Dclingtest when cmaking) and adapt the ctest limit.

This can be accomplished much easier with a separate workflow that triggers on pull_request with paths.

I hope someone can implement that variant soon ;) and I will close this PR then.

[ferdymercury]

Closing in favour of #18127