Consider cost of FP compare when performing branch optimizations, flipCode...

This issue supersedes 1147545.

Not all branches are equal especially when creating floating point compares with Java semantics for NaN cases. For example, == is cheaper than != for the DOUBLE_IFCMP operator. When performing optimizations (in particular the use of flipCode in BURS and elsewhere) we should modify the costs to reflect this lack of consistency in cost.