Meinersbur wrote: The problems you are describing are not unique to AST dump tests, but to basically every single autogenerated tests. For instance https://github.com/llvm/llvm-project/blob/main/clang/test/OpenMP/atomic_compare_codegen.cpp : Instead of just checking the atomic instructions, why does it check the entire function? You cannot even modify Clang's CodeGen to add additional unrelated instructions because most[^1] are `CHECK-NEXT` requirements. The specificity of such tests is really bad. That the variable names are regex'ed only makes the tests even more undecipherable. If something fails it is impossible to relate the mismatched CHECK line to the `#pragma omp` line. So during review nobody looks whether those tests are even correct.
In the past I tried to improve about 3 times over this style of testing, but it turned out the community really, really likes it. I have given up. I treat tests as "no-change" tests, and instead rely on end2end tests for correctness. Instead of just removing test on a single area you care about, I would suggest to open an RFC in Discourse about how tests should look like. https://github.com/llvm/llvm-project/pull/200077 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
