If I use SVM, is there a situation in which feature selection is better than using the Cost-Parameter?

In Literature it is often stated that SVM need no Feature Selection i.e. Joachims 1998 http://www.cs.iastate.edu/~jtian/cs573/Papers/Joachims-ECML-98.pdf. On the other hand feature selection is still studied 5 years after it i.e. http://www.machine-learning.martinsewell.com/feature-selection/Forman2003.pdf.

So if I apply SVM why should I still use feature selection? Shouldn't the cost parameter "reduce" the feature set much better?