Fix race condition in Function::optimized_graph(). (#27012)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27012
The current logic is buggy, and will fail in the following situation:
Thread A: check optimized_graph_, it is empty.
Thread A: claim the mutex in order to initialize optimized_graph_.
Thread A: copy graph_ into optimized_graph_.
Thread A: start running optimizations on optimized_graph_.
Thread B: check optimized_graph_, it is not empty.
Thread B: start using optimized_graph_.
BUG: Thread B is using the graph while it's still being mutated by
Thread A.
Test Plan: Imported from OSS
Differential Revision: D17649149
Pulled By: ZolotukhinM
fbshipit-source-id: ab82a4e76003b88733d219ecd47a42d38be2269c
Author
Mikhail Zolotukhin
