Before incremental inlining inlines a new call site, it sanity checks the input state of the call node: if the call is in a part of the graph that is dead, there's no need to proceed with inlining and, the parser doesn't expect a strange input state anyway. One the check consists in testing the input edges to the call node that correspond to input arguments for top. Long/double arguments to a call are in 2 halves, one of which is top so incremental inlining will skip all calls with at least a long or double argument, mistaken them for dead calls.
This bug impacts the performance of the box2d test with nashorn (10-20% performance increase once fixed and LiveNodeCountInliningCutoff bumped to 30000).