[flang-rt] Fix TypeCategory for quad-precision COMPLEX

[ceseo]

Modify the TypeCategory for quad-precision COMPLEX to CFI_type_float128_Complex so it matches the TypeCode returned by SELECT TYPE lowering.

Fixes #134565

[ceseo]

I'd like to add a testcase (check the generated LLVM IR for the correct type selection). Where's the best place to put it?

[jeanPerier]

I'd like to add a testcase (check the generated LLVM IR for the correct type selection). Where's the best place to put it?

You could add a test like this using a fir.box<cpmlx.complex<f128>> type instead of the derived type. The resulting descriptor structure should contain the type code value.

[ceseo]

You could add a test like this using a fir.box<cpmlx.complex<f128>> type instead of the derived type. The resulting descriptor structure should contain the type code value.

My question was more in the lines of where in the directory structure is more adequate. Although the fix is in flang-rt, this is technically lowering. Maybe flang/test/Lower?

[jeanPerier]

You could add a test like this using a fir.box<cpmlx.complex<f128>> type instead of the derived type. The resulting descriptor structure should contain the type code value.

My question was more in the lines of where in the directory structure is more adequate. Although the fix is in flang-rt, this is technically lowering. Maybe flang/test/Lower?

Ah sorry, so you want to test SELECT TYPE lowering?

If so, I think you should make a small FIR test using fir.select_type with complex and check that you get the code you expect.

See https://github.com/llvm/llvm-project/blob/main/flang/test/Lower/select-type-2.fir as an example (except these kinds of test should live in flang/test/Fir).

You can generate your input from Fortran with bbc -emit-fir, but try to keep the input very small, focused on the testing the resulting code.

[ceseo]

You can generate your input from Fortran with bbc -emit-fir, but try to keep the input very small, focused on the testing the resulting code.

Actually, scratch that. The resulting IR will be the same with or without the fix (it will branch differently if the type is 16 or 17). I added a runtime check as you mentioned earlier.

If you think it's OK, I'll merge this.

[github-actions]

🐧 Linux x64 Test Results

• 255 tests passed
• 1 test skipped

[jeanPerier]

LGTM