Addressed perf regression in Browsermark2.0 array blur test

test/mjsunit/array-literal-feedback.js

Index: test/mjsunit/array-literal-feedback.js
diff --git a/test/mjsunit/array-literal-feedback.js b/test/mjsunit/array-literal-feedback.js
index 93ed3bce9988bb95dd7e7b6ac6b665a1cdc96534..3e95eab1d0f1655d89407c4e4760381c62e11c25 100644
--- a/test/mjsunit/array-literal-feedback.js
+++ b/test/mjsunit/array-literal-feedback.js
@@ -93,11 +93,15 @@ if (support_smi_only_arrays) {
   }
 
   get_literal(3);
-  get_literal(3);
-  %OptimizeFunctionOnNextCall(get_literal);
+  // It's important to store a from before we crankshaft get_literal, because
+  // mementos won't be created from crankshafted code at all.
   a = get_literal(3);
+  %OptimizeFunctionOnNextCall(get_literal);
+  get_literal(3);
   assertOptimized(get_literal);
   assertTrue(%HasFastSmiElements(a));
+  // a has a memento so the transition caused by the store will affect the
+  // boilerplate.
   a[0] = 3.5;
 
   // We should have transitioned the boilerplate array to double, and