[Mlir-commits] [mlir] [mlir] Allow unroll & jam on SCF loops with results (PR #98887)

[Javier Setoain]

================
@@ -336,6 +336,33 @@ module attributes {transform.with_named_sequence} {
 
 // -----
 
+// CHECK-LABEL: @loop_unroll_and_jam_loop_with_results
+func.func @loop_unroll_and_jam_loop_with_results() -> index {
+  // CHECK:           %[[LB:.*]] = arith.constant 0
+  // CHECK:           %[[UB:.*]] = arith.constant 40
+  // CHECK:           %[[STEP:.*]] = arith.constant 8
+  %c0 = arith.constant 0 : index
+  %c40 = arith.constant 40 : index
+  %c2 = arith.constant 2 : index
+  // CHECK:           %[[RES:[a-zA-Z0-9]+]]:4 = scf.for %[[I:.*]] = %[[LB]] to %[[UB]] step %[[STEP]]
+  %sum = scf.for %i = %c0 to %c40 step %c2 iter_args(%does_not_alias_aggregated = %c0) -> (index) {
----------------
jsetoain wrote:

I've added a somewhat more complex test extracted from a common use-case. You can see the use of `iter_args` in a parallel loop. They superficially look like a loop-carried dependency, but the insertion/extraction of slices in the tensor occur in different places for each tuple of iterator values in the outermost loops; the dependency is real only in the innermost loop. The peculiar tensor "pass down" pattern in the `iter_args` should disappear after bufferization.

https://github.com/llvm/llvm-project/pull/98887