[Mlir-commits] [mlir] [mlir][x86] Add extra patterns for AMX nano-kernel lowering (acc with zero constant) (PR #222513)

[Julian Oppermann]

================
@@ -1451,81 +1476,97 @@ struct VectorContractToAMXDotProduct
         k++;
       }
     }
+
+    // Update insertion point, if acc memref after contraction loop.
+    Operation *memrefOp = srcBuffAcc.getDefiningOp();
+    if (memrefOp && outerLoop &&
+        !(memrefOp->isBeforeInBlock(outerLoop.getOperation()))) {
+      rewriter.setInsertionPoint(memrefOp);
----------------
jopperm wrote:

Thanks, the test looks good, and reverting the insertion-point-change makes sense.

https://github.com/llvm/llvm-project/pull/222513