[llvm-branch-commits] [flang] [mlir] [Flang][mlir][OpenMP] Support affinity clause codegen in Flang (PR #182222)

[Tom Eccles via llvm-branch-commits]

https://github.com/tblah commented:

Thanks for the updates. I think there is a remaining issue for non-default lower bounds. For this code example

```
  subroutine task_affinity_nondefault_lb()
    implicit none
    integer, parameter :: n = 8
    integer :: a(0:n)
    integer :: i

    !$omp parallel
    !$omp single
    !$omp task affinity(iterator(i = 0:n) : a(i))
      a(0) = 1
    !$omp end task
    !$omp end single
    !$omp end parallel
  end subroutine
```

We get
```
        %c0_i32 = arith.constant 0 : i32
        %c8_i32 = arith.constant 8 : i32
        %8 = fir.convert %c0_i32 : (i32) -> index
        %9 = fir.convert %c8_i32 : (i32) -> index
        %c1 = arith.constant 1 : index
        %10 = omp.iterator(%arg0: index) = (%8 to %9 step %c1) {
          %11 = fir.convert %arg0 : (index) -> i32
          %12 = fir.convert %11 : (i32) -> i64
          %13 = fir.convert %12 : (i64) -> index
          %14 = fir.shape %c9 : (index) -> !fir.shape<1>
          %15 = fir.array_coor %3#0(%14) %13 : (!fir.box<!fir.array<9xi32>>, !fir.shape<1>, index) -> !fir.ref<i32>
          %16 = fir.box_elesize %3#0 : (!fir.box<!fir.array<9xi32>>) -> index
          %17 = fir.convert %16 : (index) -> i64
          %18 = fir.convert %15 : (!fir.ref<i32>) -> !fir.ref<i8>
          %19 = omp.affinity_entry %18, %17 : (!fir.ref<i8>, i64) -> !omp.affinity_entry_ty<!fir.ref<i8>, i64>
          omp.yield(%19 : !omp.affinity_entry_ty<!fir.ref<i8>, i64>)
        } -> !omp.iterated<!omp.affinity_entry_ty<!fir.ref<i8>, i64>>
```

That array_coor is getting a fir.shape not a fir.shapeshift so it won't know that it is dealing with non-default lower bounds so the 0 index is going to produce invalid code.

As a more minor thing, fir.box_elesize is overkill here because the size of the element can easily be determined at compile time. But perhaps that is better left for a later patch adding a canonicalization from a fir.box_elesize to a constant where possible.

https://github.com/llvm/llvm-project/pull/182222