Collection attributes, as defined by Boyland, can be used as a mechanism for concisely specifying cross-referencelike properties such as callee sets, subclass sets, and sets of variable uses. We have implemented collection attributes in our declarative meta programming system JastAdd, and used them for a variety of applications including devirtualization analysis, metrics, and flow analysis. We propose a series of evaluation algorithms for collection attributes, and compare their performance and applicability. The key design criteria for our algorithms are 1) that they work well with demand evaluation, i.e., defined properties are computed only if they are actually needed for a particular source code analysis problem and a particular source program, and 2) that they work in the presence of circular (fixed-point) definitions that are common for many source code analysis problems, e.g., flow analysis. We show that the best algorithms work well on large practical problems, includin...