I decided to stop complaining about the speed of reflection and actually do something to help. In CGLIB CVS there is a new set of classes designed to replace common uses of java.lang.reflect. The API is intentionally similar, so you can replace code like:

Class c = String.class;
Method m = c.getMethod("indexOf", new Class[]{ String.class } );
Integer i = (Integer)m.invoke("Hello", new Object[]{ "ll" } );

with:

FastClass c = FastClass.create(String.class);
FastMethod m = c.getMethod("indexOf", new Class[]{ String.class } );
Integer i = (Integer)m.invoke("Hello", new Object[]{ "ll" } );

For even more speed there is a special invoke-by-index feature:

FastClass c = FastClass.create(String.class);
int m = c.getIndex("indexOf", new Class[]{ String.class } );
Integer i = (Integer)fc.invoke(m, "Hello", new Object[]{ "ll" } );

Reflection speed has increased quite a bit with recent JVMs, but even on 1.4.2 this code is significantly faster.

Update: By request, here are some real numbers. Two million iterations each on my Gentoo workstation. The "Overhead" column is ((Time - Direct) / Direct).