[async] Support SFG-level DSE for scalar SNodes

[k-ye]

Sorry for this large change. It shouldn't be as complicated as the size implies.. The approach is basically to find out those scalar SNodes whose values are not read by the successor tasks. We then pass in these SNodes to the CFG pass, and instruct the live variable analysis not to add them into the final node's live_gen.

There are a few special handlings for scalar SNodes here:

1. Upon seeing a GlobalStoreStmt when constructing the task meta, we add the SNode to the input (reader) states, iff. it is not a scalar.

   taichi/taichi/program/async_utils.cpp

   Lines 127 to 130 in 3aeff7c

   	if (auto global_store = stmt->cast<GlobalStoreStmt>()) {
   	if (auto ptr = global_store->ptr->cast<GlobalPtrStmt>()) {
   	for (auto &snode : ptr->snodes.data) {
   	meta.input_states.emplace(snode, AsyncState::Type::value);

2. In SFG::optimize_dead_store(), if the state forms a flow edge, we also check if it's a scalar and if the successor task actually reads it.

   taichi/taichi/program/state_flow_graph.cpp

   Lines 800 to 801 in 3aeff7c

   	if (task->has_state_flow(s, other)) {
   	used = true;

3. atomic demotion for global values are only enabled for serial tasks. I also made them handle scalars as well.

   taichi/taichi/ir/control_flow_graph.cpp

   Line 403 in 3aeff7c

   } else if (!is_parallel_executed) {

Finally, the optimization result is cached, much like what @xumingkuan has done in #1889.

Related issue = #742

[Click here for the format server]

---

[yuanming-hu]

Thank you! I would leave the control_flow_graph part to @xumingkuan but everything else LGTM!

[yuanming-hu]

Suggested change

	for _ in range(5):
	for _ in range(2):

Does this make the test stricter?

[k-ye]

Actually the DSE works pretty deterministically :) I just wasn't sure if it was a good idea to be very specific about how many DSE has been done..

[codecov]

Codecov Report

Merging #1907 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1907   +/-   ##
=======================================
  Coverage   43.83%   43.83%           
=======================================
  Files          45       45           
  Lines        6194     6194           
  Branches     1100     1100           
=======================================
  Hits         2715     2715           
  Misses       3310     3310           
  Partials      169      169           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 21ad346...670c95c. Read the comment docs.

[xumingkuan]

The control_flow_graph part LGTM!