mlir cpu loop

// Test 8210: MLIR CPU loop — prove scf.for + iter_args lowers/links/runs AOT.
//
// Builds on 8209 (scalar |mlir): same scalar-in/scalar-out ABI, but the MLIR
// body now contains a LOOP (scf.for with an accumulator via iter_args). This is
// the lowering the kernel transform will generate, so it de-risks the next
// slice (kernel:self/pairwise -> generated MLIR). sum-upto(10) = 0+1+..+9 = 45.

~import std/io

~pub event sum-upto { n: i64 }
| total i64

// Zig fallback — defines the answer.
~proc sum-upto|zig {
    var acc: i64 = 0;
    var i: i64 = 0;
    while (i < n) : (i += 1) acc += i;
    return .{ .total = acc };
}

// MLIR implementation — a real loop (scf.for + iter_args accumulator).
~proc sum-upto|mlir {
    func.func @koru_mlir_sum_upto(%arg0: i64) -> i64 {
        %c0 = arith.constant 0 : i64
        %lb = arith.constant 0 : index
        %step = arith.constant 1 : index
        %n = arith.index_cast %arg0 : i64 to index
        %sum = scf.for %i = %lb to %n step %step iter_args(%acc = %c0) -> (i64) {
            %iv = arith.index_cast %i : index to i64
            %next = arith.addi %acc, %iv : i64
            scf.yield %next : i64
        }
        return %sum : i64
    }
}

~sum-upto|mlir(n: 10)
| total t |> std/io:print.ln("sum: {{ t:d }}")