[flang] Behavior of gfortran test depends on optimization level: pr117797.f90

[tarunprabhu]

Steps to reproduce:

Assuming that the LLVM test suite is checked out in ./llvm-test-suite, compile the test at various optimization levels

$ flang -o a0.out -O0 ./llvm-test-suite/Fortran/gfortran/regression/pr117797.f90
$ flang -o a1.out -O1 ./llvm-test-suite/Fortran/gfortran/regression/pr117797.f90
$ flang -o a2.out -O2 ./llvm-test-suite/Fortran/gfortran/regression/pr117797.f90
$ flang -o a3.out -O3 ./llvm-test-suite/Fortran/gfortran/regression/pr117797.f90
$ flang -o alto.out -flto=auto ./llvm-test-suite/Fortran/gfortran/regression/pr117797.f90

When these are run, the executables created with -O0 and -flto=auto encounter a STOP statement in the code, but those compiled with optimizations do not.

$ ./a0.out 
Fortran STOP: code 2
$ ./a1.out
$ ./a2.out
$ ./a3.out
$ ./alto.out
Fortran STOP: code 2

Expected behavior: The stop message should never be printed.

What follows is a slightly reduced from the original but that reproduces this behavior.

module foo

  type, public :: any_matrix
    class(*), allocatable :: value(:,:)
  end type

contains

  function bar(this) result(uptr)
    class(any_matrix), target, intent(in) :: this
    class(*), pointer :: uptr(:,:)
    uptr => this%value ! Seg. fault in trans-array.cc(gfc_get_array_span) here
  end function

  function build(this) result (res)
    class(*) :: this(:,:)
    type(any_matrix) :: res
    res%value = this
  end function

  function evaluate (this) result (res)
    class(*) :: this(:,:)
    character(len = 2, kind = 1), allocatable :: res(:)
      select type (ans => this)
        type is (character(*))
          res = reshape (ans, [4])
        type is (integer)
          allocate (res (8))
          write (res, '(i2)') ans
        class default
          res = ['no','t ','OK','!!']
      end select
  end

end module

  use foo
  class(*), allocatable :: up (:, :)
  integer :: i(2,2) = reshape ([1,2,3,4], [2,2])

  up = bar (build (i))
  if (any (evaluate (up) /= [' 1',' 2',' 3',' 4'])) stop 2

  deallocate (up)
end

The cases in the SELECT statement seem to have an effect on the behavior. Removing only the character(*) case results in correct behavior at all optimization levels. Removing only the class default cases causes the error to occur only at -O0 but not at any other optimization level.

[EDIT]: Remove redundant statement and clarify what effect modifying the SELECT construct has on behavior