[Mlir-commits] [mlir] [MLIR][SCF] Handle commuted indices in parallel loop fusion (PR #219665)

[purnima shrivastava]

================
@@ -1761,3 +1761,33 @@ func.func @do_not_fuse_distinct_dynamic_bounds(%A: memref<16xf32>,
 // CHECK-LABEL: func @do_not_fuse_distinct_dynamic_bounds
 // CHECK:        scf.parallel
 // CHECK:        scf.parallel
+
+// -----
+
+func.func @fuse_commuted_indices(%arg0: memref<32xf32>,
+                                 %arg1: memref<32xf32>) {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %c16 = arith.constant 16 : index
+  %cst = arith.constant 3.000000e+00 : f32
+
+  scf.parallel (%i) = (%c0) to (%c16) step (%c1) {
+    %index = arith.addi %i, %c1 : index
+    memref.store %cst, %arg0[%index] : memref<32xf32>
+    scf.reduce
+  }
+
+  scf.parallel (%i) = (%c0) to (%c16) step (%c1) {
+    %index = arith.addi %c1, %i : index
----------------
purnima-nlp wrote:

@joker-eph 
Thanks, that makes sense. I agree that the constant-on-the-left case should be handled by canonicalization rather than adding complexity to the fusion analysis.

I’ll check whether the issue still reproduces after canonicalization using a dynamic, non-constant offset (`%iv + %offset` versus `%offset + %iv`). If it does, I’ll update the test to use that stronger case; otherwise, I’ll reconsider the approach.


https://github.com/llvm/llvm-project/pull/219665