Build error with LLVM flang-new and NAG nagfor due to standard violation

[rouson]

Background

I'd like to test parallel builds of neural-fortran with LLVM flang-new 22.0.0git (the current main branch of LLVM). If my search of neural-fortran is correct, then flang-new now supports every multi-image feature that neural-fortran uses, including this_image(), num_images(), co_broadcast, and co_sum. Please let me know if I missed any features.

An apparent standard violation blocks building the current main branch of neural-fortran with the LLVM flang-new compiler and with the NAG nagfor compilers; whereas the GCC gfortran and Intel ifx compilers allow the relevant code. The feature involved was allowed by the Fortran 2003 standard but disallowed in subsequent standards.

Issue

The output of fpm test --compiler flang-new --flag -O3 includes the following:

error: Semantic errors in ././src/nf/nf_optimizers.f90
./././src/nf/nf_optimizers.f90:321:48: error: Result of pure function may not be both polymorphic and ALLOCATABLE
      class(optimizer_base_type), allocatable :: res
                                                 ^^^
<ERROR> Compilation failed for object " src_nf_nf_optimizers.f90.o "

Similarly, the output of fpm test --compiler nagfor --flag -O3 includes the following:

Error: ././src/nf/nf_optimizers.f90, line 341: Result variable of pure function GET_OPTIMIZER_BY_NAME is polymorphic allocatable
Errors in declarations, no further processing for GET_OPTIMIZER_BY_NAME

Compiler Bug Reports and the Standard

I reported this as GCC Bug 78640 in 2016, but it hasn't been fixed yet. The Fortran 2023 standard clause 4.5.3 states, "Fortran 2003 permitted the function result of a pure function to be a polymorphic allocatable variable, to have a polymorphic allocatable ultimate component, or to be finalizable by an impure final subroutine. These are not permitted by this document."

[rouson]

On macOS, brew install flang will install flang 21, which can parallelize do concurrent but doesn’t support multi-image features. Multi-image feature support requires building the main branch of LLVM from source. I just posted the following instructions for what works for me to build LLVM from source on macOS: parallel-testing-with-flang.md.

[milancurcic]

Indeed, I recall removing a few of these for flang-20 a year ago but have since forgotten about the rule, so some new ones came in.