[Mlir-commits] [mlir] [mlir][SCF] `scf.parallel`: Make reductions part of the terminator (PR #75314)

[Ivan Butygin]

Hardcode84 wrote:

> This might make it difficult for frontends to directly lower to scf.parallel, since frontend reductions like in OpenMP do not require the reduction to be the last operation.

That's a good point, but I think this is solvable in frontends, and having a clean semantics on middle level is important.

E.g. in our python compiler we support code like this:
```
sum = 0
for i in prange(size): # parallel loop
  ... some code
  sum += a[i]
  ... some more code

print(sum)
```

We are gradually uplifting `cf -> scf.while -> scf.for -> scf.parallel` and during this `scf.for -> scf.parallel` uplifting we are transforming `sum +=` to `scf.reduce` and move it to the end of the block.
And if user is doing something really weird with `sum`, we will just leave it at `scf.for` level.


https://github.com/llvm/llvm-project/pull/75314