================ @@ -0,0 +1,31 @@ +! This test checks lowering of the OpenMP tile directive +! It is done 3 times corresponding to every possible fraction of the last +! iteration before passing beyond UB. + +! RUN: %flang %flags %openmp_flags -fopenmp-version=51 -DUB=16 %s -o %t-ub16.exe ---------------- Meinersbur wrote:
The ability has been added in #150722. It is an established pattern for Clang. The OpenMP (runtime) tests are OpenMP code, for which Clang is used to generate calls to the runtime. Therefore this indeed an end2end test which are usually avoided in LLVM (See e.g. discussion at https://discourse.llvm.org/t/rfc-end-to-end-testing/53306). The alternative would have been to have test unittests that call the OpenMP runtime function as Clang would generate them (like we actually do for Flang-RT). However, these are end2end from the beginning when Intel donated the runtime to the LLVM project and no effort has been made to change it for the sake of it. Note that other runtime projects, such as [libc++](https://github.com/llvm/llvm-project/blob/main/libcxx/test/selftest/), have end2end tests as well. The fact that `!$omp tile` does not emit any runtime call IMHO is an implementation detail. It could well involved the runtime like `!$omp do`. We already have clang tests for tiling, unrolling, etc. (there was a discusscusson about it soewehere in Phabricator), so I think it is consequential to allow them for Flang as well. Some of Clang versions of those, and maybe future Flang tests, also test implementation-specific behavior, such as [how often arguments inside clauses are evaluated](https://github.com/llvm/llvm-project/blob/main/openmp/runtime/test/transform/tile/intfor.c). Those are tied to specific version of the compiler and would be a pain point to keep synchronized in the llvm-test-suite. In the past I myself argued having Flang's end-2-end tests to be in the llvm-test-suite instead of in the monorepository. Reason are that Clang does so as well, and you can only run tests for your native architecture; you would not be able to test ARM compilation when your machine is x86. For a runtime, you can only test it if you run the target arch's code, which might be why end2end tests are accepted for LLVM_ENABLE_RUNTIMES projects. I would not have added them here if OpenMP didn't already rely on end-to-end testing. I can create an RFC about end2end tests for OpenMP/runtimes. > Does this test need a requires line ensuring that flang is built? `.f90` and `.F90` file are only enabled when Flang is available: https://github.com/llvm/llvm-project/blob/c80d49590871c2060bbea758e09382bc5f9561c6/openmp/runtime/test/lit.cfg#L45-L46 https://github.com/llvm/llvm-project/pull/160298 _______________________________________________ llvm-branch-commits mailing list llvm-branch-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
