Enable 2 pending frames on ARM devices.

cc/scheduler/scheduler_state_machine.cc

Index: cc/scheduler/scheduler_state_machine.cc
diff --git a/cc/scheduler/scheduler_state_machine.cc b/cc/scheduler/scheduler_state_machine.cc
index c512987bd23f501abdaf1fceab783945f8dc1ae4..00d60147e4cfa3f0973930e9f5bc293f9ee0fad5 100644
--- a/cc/scheduler/scheduler_state_machine.cc
+++ b/cc/scheduler/scheduler_state_machine.cc
@@ -423,14 +423,6 @@ bool SchedulerStateMachine::ShouldSendBeginMainFrame() const {
   if (!HasInitializedOutputSurface())
     return false;
 
-  // SwapAck throttle the BeginMainFrames unless we just swapped.
-  // TODO(brianderson): Remove this restriction to improve throughput.

[brianderson, 2015/04/03 22:44:01]

Beware that this aspect of the patch might introduce flakiness in browser tests
and could result in regressing other use cases where it might increase the CPU
fighting. We should definitely keep an eye on the perf dashboard for this.

Do you get any benefit from just the other half of this patch? If so it might be
worth splitting them up so you don't have to revert both parts.

[alexst (slow to review), 2015/04/04 00:20:02]

On intel, just the other part was enough, on ARM, this was also required.

Maybe we should take a more cautious approach and #ifdef it out only for freon
on arm? This would let us monitor it on a smaller scale and not  cause side
effects elsewhere.

That being said, I think you are right, I'll split these 2 parts to keep impact
limited.

-  bool just_swapped_in_deadline =
-      begin_impl_frame_state_ == BEGIN_IMPL_FRAME_STATE_INSIDE_DEADLINE &&
-      perform_swap_funnel_;
-  if (pending_swaps_ >= max_pending_swaps_ && !just_swapped_in_deadline)
-    return false;
-
   if (skip_begin_main_frame_to_reduce_latency_)
     return false;