r3995 (http://code.google.com/p/v8/source/detail?r=3995) introduce performance
regression for the case when arrat size is small (think new Array(4)).

It turns out that in those cases rep stos is slower than plain loop (apprently
due to ecx increment, but I didn't check this hypothesis.)  The next thing
to try could be direct jump into right place of long sequence of stos'es.

Review URL: http://codereview.chromium.org/2082006

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4664 ce2b1a6d-e550-0410-aec6-3dcde31c8c00