[flang-commits] [flang] [llvm] [mlir] [flang][OpenMP] Support for "atomic compare capture" (PR #202315)

[Tom Eccles via flang-commits]

tblah wrote:

Stepping back, I think this needs a more systematic validation pass before further review.

This PR adds support for atomic compare-capture, but several valid forms close to the added tests still either abort or fail during LLVM translation. For example, postfix `<`/`>` compare-capture and complex compare-capture are both accepted by the front end but fail later. These are not obscure transformations; they are natural variants of the feature being added.

Could you please audit the supported/unsupported matrix for:

- prefix vs postfix vs fail-only capture
- integer, real, logical, and complex operands
- `==`, `<`, and `>` compare forms where accepted
- Flang lowering and MLIR-to-LLVMIR translation coverage

For each unsupported case, please add an explicit diagnostic/NYI test rather than allowing it to reach an assertion or translation failure. Please also check for regressions in the LLVM test-suite Fortran tests and the Fujitsu test-suite, since this changes OpenMP lowering/translation behavior beyond the specific new lit cases.

Once that matrix is covered and the broader Fortran test suites are checked, it will be much easier to review the remaining implementation details.


https://github.com/llvm/llvm-project/pull/202315