Merge pull request #787 from XJDKC/keep-order

Fix the graph operation when tensor is written by multiple independent ops