[Flang][OpenMP] Incorrect result of an addition of a firstprivate array in a workshare construct

Summary

When an array is firstprivate in an OpenMP workshare construct, an addition using it in the construct returns an incorrect result.

Reproducer

$ cat test.f90

program test
  integer(4),parameter::n=7*7
  integer(4),dimension(n)::z,a
  z=10
  a=0
  call sub1(a,z,n)
  print *,'pass'
end

subroutine sub1(a,z,n)
  integer(4)::n
  integer(4),dimension(n)::z,a
!$omp parallel workshare firstprivate(z)
  a=z+1
!$omp end parallel workshare
  if (any(a/=z+1) ) then
    print *,"fail",a
    error stop
  end if
end subroutine

$ flang --version
flang version 23.0.0git (https://github.com/llvm/llvm-project.git 58208a0cc165c23d2fad4ed46950b342156875ac)
Target: aarch64-unknown-linux-gnu
Thread model: posix
InstalledDir: /path/to/llvm/build/bin
Build config: +assertions
$ flang -fopenmp -O0 test.f90 && env OMP_NUM_THREADS=4 ./a.out
 fail 11 11 11 11 11 11 11 11 11 11 11 11 11 65536 -1476406295 65536 6 1
 -1468463655 65536 -1468511103 65536 0 1 -1468464959 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1
Fortran ERROR STOP: code 1

IEEE arithmetic exceptions signaled: INEXACT

Elements of array a from index 14 onwards show incorrect values.

Non-problematic Conditions

The issue does not occur under the following conditions:

• Disabling OpenMP (e.g., -fno-openmp, or OMP_NUM_THREADS=1).
• Enabling optimization (i.e., specifying -O1 or higher).
• Changing the firstprivate variable to a shared variable.
• Removing the addition in the construct:

    !$omp parallel workshare firstprivate(z)
  -   a=z+1
  +   a=z
    !$omp end parallel workshare
  

Other Compilers

• GFortran

  $ gfortran --version
  GNU Fortran (GCC) 15.2.0
  Copyright (C) 2025 Free Software Foundation, Inc.
  This is free software; see the source for copying conditions.  There is NO
  warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  
  $ gfortran -fopenmp -O0 test.f90 && env OMP_NUM_THREADS=4 ./a.out
   pass
  

• Intel Fortran

  $ ifx --version
  ifx (IFX) 2024.2.0 20240602
  Copyright (C) 1985-2024 Intel Corporation. All rights reserved.
  
  $ ifx -fopenmp -O0 test.f90 && env OMP_NUM_THREADS=4 ./a.out
   pass
  

Start with the test.f90 reproducer and compile it using flang -fopenmp -O0, then run it with OMP_NUM_THREADS=4 to reproduce the incorrect array values. The work is done when this firstprivate array addition produces the expected result and the program prints "pass" under the reported conditions.