Improve SminessPropagator to propagate sminess across cycles of phis.

Improve SminessPropagator to propagate sminess across cycles of phis.

R=fschneider@google.com

Committed: https://code.google.com/p/dart/source/detail?r=12630

[Florian Schneider, 2012-09-20 08:54:02]

Can you look at the following example again? All phis should be smis here, but
v22 is somehow missed.

 checksmi v5
           |
           v 
v22 <- phi(v5, v62)

v35 <- phi(v22, v59)
v55 <- smi-op
v58 <- smi-op
v59 <- phi(v55, v58, v35)
v62 <- phi(v35, v22)

https://codereview.chromium.org/10950035/diff/1/runtime/vm/flow_graph_optimiz...
File runtime/vm/flow_graph_optimizer.cc (right):

https://codereview.chromium.org/10950035/diff/1/runtime/vm/flow_graph_optimiz...
runtime/vm/flow_graph_optimizer.cc:1101:
!in_worklist_->Contains(input->definition()->ssa_temp_index())) {
What about operands defined by phis that depend (possibly indirectly) on this
phi? They may not be in the worklist because initially only phis that change
because of a check instruction are added to the worklist.

[Vyacheslav Egorov (Google), 2012-09-20 11:22:43]

You are right, I need to transitively propagate optimistic marking before I
start unmarking phis.

Please take another look.

[Florian Schneider, 2012-09-20 11:48:50]

LGTM.

https://codereview.chromium.org/10950035/diff/5001/runtime/vm/flow_graph_opti...
File runtime/vm/flow_graph_optimizer.cc (right):

https://codereview.chromium.org/10950035/diff/5001/runtime/vm/flow_graph_opti...
runtime/vm/flow_graph_optimizer.cc:1113: // Place all operands that are
potentially smi phis to worklist.
I don't see how input operands can be affected by the current phi. Shouldn't it
be enough to add all uses of this phi to the worklist?

https://codereview.chromium.org/10950035/diff/5001/runtime/vm/flow_graph_opti...
runtime/vm/flow_graph_optimizer.cc:1133: AddToWorklist(phi_use);
Maybe rename this to AppendToWorkList since the algorithm relies on the fact
that new phis are added at the end of the list.