[flang-commits] [clang] [flang] [flang][OpenMP] Upstream `do concurrent` loop-nest detection. (PR #127595)

[Valentin Clement バレンタイン クレメン via flang-commits]

clementval wrote:

> Thanks @skatrak and @bhandarkar-pranav for the approval.
> 
> @kiranchandramohan @clementval I think there is a pretty simple solution that enables us to mark multi-range loop nests. I think we can add an optional attribute to the `fir::DoLoopOp` to store the loop nest depth: `nest_depth(n)`. So for the following input:
> 
> ```fortran
> do concurrent (i=1:10, j=1:10)
> end do
> ```
> 
> The MLIR would look like this:
> 
> ```mlir
>     fir.do_loop %arg0 = %10 to %11 step %c1 unordered nest_depth(2) {
>       %14 = fir.convert %arg0 : (index) -> i32
>       fir.store %14 to %3#1 : !fir.ref<i32>
>       fir.do_loop %arg1 = %12 to %13 step %c1_2 unordered {
>         %15 = fir.convert %arg1 : (index) -> i32
>         fir.store %15 to %1#1 : !fir.ref<i32>
>         ....
>       }
>     }
> ```
> 
> So the new attribute would be attached only to the outermost loop op in the nest. I think this is a non-distruptive change to the op that enables us to model loop nests more easily. So this is similar to what @kiranchandramohan suggested above but I think more self-contained: one attribute to tie the whole nest together. WDYT?

I would prefer a proper operation than trying to patch the current one. We have plenty of example of loop operation with multiple ranges so let's model it the proper way. 


https://github.com/llvm/llvm-project/pull/127595