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

[Federico Bruzzone]

================
@@ -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
----------------
FedericoBruzzone wrote:

> We don't generally run canonicalize on every test, and shouldn't need to.
> 
> 
> 
> What we should avoid is make analyses and transformations more complex for handling things that are already handled in canonicalization pattern: the whole design of the compiler is the opposite, that is move things out of analyses and transformations (when possible and appropriate) into canonicalization to simplify the analyses and transformations and avoid repeating redundant work everywhere.

It perfectly makes sense. Thank you! :D 

@purnima-nlp note that what I suggested [here](https://github.com/llvm/llvm-project/pull/219665#discussion_r3886941037), then it could be simplified I think. Sorry for not saying it right away 🙏


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