Improve load forwarding:

runtime/vm/compiler.cc

Index: runtime/vm/compiler.cc
diff --git a/runtime/vm/compiler.cc b/runtime/vm/compiler.cc
index dc2e042a7305b0a24498473c529923a768aaacdd..63882a5080eb2442230c154a80d960f20f453fe5 100644
--- a/runtime/vm/compiler.cc
+++ b/runtime/vm/compiler.cc
@@ -291,7 +291,14 @@ static bool CompileParsedFunctionHelper(const ParsedFunction& parsed_function,
         propagator.Propagate();
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
       }
-      optimizer.Canonicalize();
+      if (optimizer.Canonicalize()) {
+        // To fully remove redundant boxing (e.g. BoxDouble used only in
+        // environments and UnboxDouble instructions) instruction we
+        // first need to replace all their uses and then fold them away.
+        // For now we just repeat Canonicalize twice to do that.
+        // TODO(vegorov): implement a separate representation folding pass.
+        optimizer.Canonicalize();
+      }
       DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
       // Perform register allocation on the SSA graph.