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

[Federico Bruzzone]

================
@@ -145,6 +145,18 @@ static bool valsAreEquivalent(Value val1, Value val2,
     return false;
   if (!isMemoryEffectFree(val1DefOp) || !isMemoryEffectFree(val2DefOp))
     return false;
+
+  // Handle commuted integer additions.
+  if (auto addOp1 = dyn_cast<arith::AddIOp>(val1DefOp)) {
+    if (auto addOp2 = dyn_cast<arith::AddIOp>(val2DefOp);
+        addOp2 &&
+        val1DefOp->getAttrDictionary() == val2DefOp->getAttrDictionary() &&
+        val1.getType() == val2.getType() &&
+        valsAreEquivalent(addOp1.getLhs(), addOp2.getRhs(), loopsIVsMap) &&
+        valsAreEquivalent(addOp1.getRhs(), addOp2.getLhs(), loopsIVsMap))
+      return true;
+  }
+
   return OperationEquivalence::isEquivalentTo(
----------------
FedericoBruzzone wrote:

`OperationEquivalence::isEquivalentTo` has an additional last parameter: `checkCommutativeEquivalent` (which is a `function_ref`). See 
https://github.com/llvm/llvm-project/blob/a2600d408cfbc328131bf2317426e2de1e50eed1/mlir/lib/IR/OperationSupport.cpp#L837-L842


Can we use this parameter to cover other commutative operations as well?

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