More general dead phi elimination.

More general dead phi elimination.

This CL adds an explicit pass to eliminate dead phis. Until now, we
only removed redundant phis that occurred after load elimination locally.

Also, constant propagation may result in dead / or redundant phis which
are removed by this phase.

R=srdjan@google.com

Committed: https://code.google.com/p/dart/source/detail?r=36094

[srdjan, 2014-05-12 21:21:44]

lgtm

https://codereview.chromium.org/285483002/diff/1/runtime/vm/flow_graph_optimi...
File runtime/vm/flow_graph_optimizer.cc (right):

https://codereview.chromium.org/285483002/diff/1/runtime/vm/flow_graph_optimi...
runtime/vm/flow_graph_optimizer.cc:7308: if (join != NULL) {
if ((join != NULL ) && (join->phis_ != NULL)) {

[Vyacheslav Egorov (Google), 2014-05-13 09:27:31]

Can we merge all dead/redundant phi elimination passes together?

I really would like to avoid having more passes that do the same. 

How about having a phi emitter class that any pass that wants to emit phis into
the graph has to use?

Then this class would eliminate dead/redundant (and duplicated phis) before
actually emitting them.

[Florian Schneider, 2014-05-13 09:37:02]

On 2014/05/13 09:27:31, Vyacheslav Egorov (Google) wrote:
> Can we merge all dead/redundant phi elimination passes together?
> 
> I really would like to avoid having more passes that do the same. 
> 
> How about having a phi emitter class that any pass that wants to emit phis
into
> the graph has to use?
> 
> Then this class would eliminate dead/redundant (and duplicated phis) before
> actually emitting them.

I'll think about it. Right now, these different passes are not exactly doing the
same task.

[Florian Schneider, 2014-05-13 09:49:33]

Committed patchset #1 manually as r36094 (presubmit successful).

[Vyacheslav Egorov (Google), 2014-05-13 09:57:20]

Strictly speaking the pass that there is there for load elimination is doing
even more generic optimization because it can find duplicate phis as well. 

I strongly think we should *stop* adding passes, and start generalizing passes
where it makes sense instead.

[srdjan, 2014-05-13 14:56:11]

On 2014/05/13 09:57:20, Vyacheslav Egorov (Google) wrote:
> Strictly speaking the pass that there is there for load elimination is doing
> even more generic optimization because it can find duplicate phis as well. 
> 
> I strongly think we should *stop* adding passes, and start generalizing passes
> where it makes sense instead.

Strongly agreed. Iterating over flow graph again and again slows down the
compilation.