[JIT] Don't re run CSE on every block (#41479)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41479
Previously we were re-running CSE every time we recursed into a new block, which in turn created a new Alias Db for the whole graph. This was O(# Nodes * # Blocks).
For graphs which don't have any autodiff opportunities, such as Densenet, create_autodiff_subgraphs is now linear in number of nodes. For Densenet this pass was measured at ~.1 seconds.
This pass is still non-linear for models which actually do create autodiff subgraphs, because in the
```
bool any_changed = true;
while (any_changed) {
AliasDb aliasDb(graph_);
any_changed = false;
for (auto it = workblock.end()->reverseIterator();
it != workblock.begin()->reverseIterator();) {
bool changed;
std::tie(it, changed) = scanNode(*it, aliasDb);
any_changed |= changed;
}
}
```
loop we recreate the AliasDb (which is O(N)) every time we merge something and scan node returns. I will make that linear in next PR in the stack.
Test Plan: Imported from OSS
Reviewed By: Krovatkin
Differential Revision: D22600606
Pulled By: eellison
fbshipit-source-id: b08abfde2df474f168104c5b477352362e0b7b16
