[flang] Different fir-opt tool is produced depending on FLANG_INCLUDE_TESTS

[mgorny]

The fir-opt tool that is installed unconditionally has some code conditional to FLANG_INCLUDE_TESTS. This implies that the tool installed as part of a CMake run with tests being built will be different than the tool installed otherwise.

At the very least, this goes against the principle of least surprise, and violates the Gentoo-specific policy that testing must not affect the installed package. Also, this is quite problematic for us since it implies static linking to MLIR (as the relevant test-specific MLIR libraries aren't available in the dylib — which could be acceptable for discardable test executables, but it's not fine for installed executables).

Could we perhaps build two versions of fir-opt instead? One that would not contain test-specific code and be installed, and another one with test-specific code that's only used by the test suite? I can submit a pull request if you agree with that approach.

[llvmbot]

@llvm/issue-subscribers-flang-build

Author: Michał Górny (mgorny)

The `fir-opt` tool that is installed unconditionally has some code conditional to `FLANG_INCLUDE_TESTS`. This implies that the tool installed as part of a CMake run with tests being built will be different than the tool installed otherwise.

At the very least, this goes against the principle of least surprise, and violates the Gentoo-specific policy that testing must not affect the installed package. Also, this is quite problematic for us since it implies static linking to MLIR (as the relevant test-specific MLIR libraries aren't available in the dylib — which could be acceptable for discardable test executables, but it's not fine for installed executables).

Could we perhaps build two versions of fir-opt instead? One that would not contain test-specific code and be installed, and another one with test-specific code that's only used by the test suite? I can submit a pull request if you agree with that approach.