[flang-commits] [flang] [flang][Lower] Treat directives with nested evaluations as constructs (PR #91614)

[Mats Petersson via flang-commits]

Leporacanthicus wrote:

I think this fix is doing something good - it helps in SOME cases, but perhaps not "good enough". I'm not opposed to this going in... But a complete fix would be better, in my (and Kiran's) opinion.

It doesn't fix my original problem, I'm still seeing the STOP when it shouldn't for this code:
```
subroutine simple(x, yn)
  implicit none
  logical, intent(in) :: yn
  integer, intent(in) :: x
  integer :: i
  real(8) :: E
  E = 0d0

  if (yn) then
     !$omp parallel do private(i) reduction(+:E)
     do i = 1, x
        E = E + i
     end do
     !$omp end parallel do
  else
     stop 1
  end if
  print *, E
end subroutine simple

program p

  integer :: xx
  xx = 50

  call simple(xx, .true.)
end program p
```
Compiler generates this MLIR (only the "interesting portion" shown here):
```
    cf.cond_br %8, ^bb1, ^bb2
  ^bb1:  // pred: ^bb0
    omp.parallel {
      %10 = fir.alloca i32 {adapt.valuebyref, pinned}
      %11 = fir.declare %10 {uniq_name = "_QFsimpleEi"} : (!fir.ref<i32>) -> !fir.ref<i32>
      %12 = fir.load %5 : !fir.ref<i32>
      omp.wsloop reduction(@add_reduction_f64 %2 -> %arg2 : !fir.ref<f64>) {
        omp.loop_nest (%arg3) : i32 = (%c1_i32) to (%12) inclusive step (%c1_i32) {
          %13 = fir.declare %arg2 {uniq_name = "_QFsimpleEe"} : (!fir.ref<f64>) -> !fir.ref<f64>
          fir.store %arg3 to %11 : !fir.ref<i32>
          %14 = fir.load %13 : !fir.ref<f64>
          %15 = fir.load %11 : !fir.ref<i32>
          %16 = fir.convert %15 : (i32) -> f64
          %17 = arith.addf %14, %16 fastmath<contract> : f64
          fir.store %17 to %13 : !fir.ref<f64>
          omp.yield
        }
        omp.terminator
      }
      omp.terminator
    }
    cf.br ^bb2
  ^bb2:  // 2 preds: ^bb0, ^bb1
    %9 = fir.call @_FortranAStopStatement(%c1_i32, %false, %false) fastmath<contract> : (i32, i1, i1) -> none
    fir.unreachable
  }
```

Not sure if it's because the parallel section has a do and/or a reduction inside it?

https://github.com/llvm/llvm-project/pull/91614