http://gcc.gnu.org/ml/gcc-patches/2006-10/msg01519.html
//===---------------------------------------------------------------------===//
+
+viterbi speeds up *significantly* if the various "history" related copy loops
+are turned into memcpy calls at the source level. We need a "loops to memcpy"
+pass.
+
+//===---------------------------------------------------------------------===//