[Mlir-commits] [flang] [mlir] [flang][openacc][openmp] Support implicit casting on the atomic interface (PR #114390)

[llvmlistbot at llvm.org]

khaki3 wrote:

Another idea is to assume a converted reference type on the LHS of `atomic.read`. The verifier checks if `x`s of `read` and `write` are the same, so we cannot convert `x` only for `read`.

```
  real :: n
  integer :: x
  !$acc atomic capture
  n = x
  x = n * n
  !$acc end atomic
```

could become

```
    %0 = fir.alloca f32 {bindc_name = "n", uniq_name = "_QFEn"}
    %1:2 = hlfir.declare %0 {uniq_name = "_QFEn"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
    %2 = fir.convert %1#1 : (!fir.ref<f32>) -> !fir.ref<i32>
    %3 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %4:2 = hlfir.declare %3 {uniq_name = "_QFEx"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
    %5 = fir.load %1#0 : !fir.ref<f32>
    %6 = fir.load %1#0 : !fir.ref<f32>
    %7 = arith.mulf %5, %6 fastmath<contract> : f32
    %8 = fir.convert %7 : (f32) -> i32
    acc.atomic.capture {
      acc.atomic.read %2 = %4#1 : !fir.ref<i32>, f32
      acc.atomic.write %4#1 = %8 : !fir.ref<i32>, i32
    }
```

My first attempt was to extend the declaration of `read` to accept the original type of `v`. But it was a destructive change. The current change in this PR adds a temp for the `read` LHS. 

https://github.com/llvm/llvm-project/pull/114390