[flang][OpenMP] Custom mapper not found and crashing compilation when used in different module than which it was defined in

[jfuchs-kmt]

When using a custom mapper outside of the module it was defined, it is not found and crashes the compilation. Below is a short reproducer for the issue:

MODULE field_mod
    IMPLICIT NONE

    TYPE :: field_t
        REAL, ALLOCATABLE :: buf(:)
    END TYPE field_t
    !$omp declare mapper(custom: field_t :: t) map(t%buf)
END MODULE field_mod

PROGRAM reproducer
    USE field_mod

    IMPLICIT NONE

    REAL :: validation(5)
    TYPE(field_t) :: f_t
    ALLOCATE(f_t%buf(5))
    f_t%buf = 42.0

    !$omp target map(mapper(custom), tofrom: f_t) map(from: validation)
        validation(:) = f_t%buf(:)
    !$omp end target

    PRINT*, "Expected: 42. 42. 42. 42. 42."
    PRINT*, "Actual:", validation

    DEALLOCATE(f_t%buf)
END PROGRAM reproducer

Compiling this code (with cuda in my case) crashes...

flang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-version=52 main.F90 -O3

...with the error message:

error: loc("f_t"("/home/jfuchs/reproducer/main.F90":18:22)): invalid mapper id
error: verification of lowering to FIR failed

We can cross check the correct behavior by moving the !$omp declare mapper... line below into the PROGRAM directly. Then compilation succeeds and the values are as expected after the target region.

I am using flang version 22.0.0git (git@github.com:llvm/llvm-project.git a8057ff12956b7dd6dec44f9610c446c828c4af8).

[llvmbot]

@llvm/issue-subscribers-flang-ir

Author: None (jfuchs-kmt)

When using a custom mapper outside of the module it was defined, it is not found and crashes the compilation. Below is a short reproducer for the issue: ```fortran MODULE field_mod IMPLICIT NONE

TYPE :: field_t
    REAL, ALLOCATABLE :: buf(:)
END TYPE field_t
!$omp declare mapper(custom: field_t :: t) map(t%buf)

END MODULE field_mod

PROGRAM reproducer
USE field_mod

IMPLICIT NONE

REAL :: validation(5)
TYPE(field_t) :: f_t
ALLOCATE(f_t%buf(5))
f_t%buf = 42.0

!$omp target map(mapper(custom), tofrom: f_t) map(from: validation)
    validation(:) = f_t%buf(:)
!$omp end target

PRINT*, "Expected: 42. 42. 42. 42. 42."
PRINT*, "Actual:", validation

DEALLOCATE(f_t%buf)

END PROGRAM reproducer

Compiling this code (with cuda in my case) crashes...

flang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-version=52 main.F90 -O3

...with the error message:

error: loc("f_t"("/home/jfuchs/reproducer/main.F90":18:22)): invalid mapper id
error: verification of lowering to FIR failed

We can cross check the correct behavior by moving the `!$omp declare mapper...` line below into the PROGRAM directly. Then compilation succeeds and the values are as expected after the target region.

I am using `flang version 22.0.0git (git@<!-- -->github.com:llvm/llvm-project.git a8057ff12956b7dd6dec44f9610c446c828c4af8)`.
</details>

[kiranchandramohan]

FYI @TIFitis @agozillon. Have we added declare mapper to the module file? CC: @kparzysz

[TIFitis]

I'll look into this and report back.

[kparzysz]

I think we only put reductions in the module files at the moment.

[kiranchandramohan]

I think we only put reductions in the module files at the moment.

We do threadprivate. We probably need to create a list of all declarative directives in a meta ticket and get it all done from a module file perspective.

[agozillon]

I think we only put reductions in the module files at the moment.

We do threadprivate. We probably need to create a list of all declarative directives in a meta ticket and get it all done from a module file perspective.

Would be quite useful, I imagine something like declare target also might need to be exported to the module files to act coherently in all cases.