Ave provides a default partitioning of behaviors at the spatial domain (i.e., partition the processors that are involved in different behaviors at the p-chain level), which has already been described. The user retains the ability to partition behaviors across any level, with respect to the intrinsic (or derived) attributes of the abstract events of that level.
A new partition imposes an additional structure on the tree. The user can investigate the behavior through this new structural transformation: A partition is thus equivalent to the imposition of multiple views on the program behavior [12]. We provide a uniform interface for such structural transformations. The user can view the match tree through the structure imposed by the new (or default) partitions.
Example: Binary Image Compression (Continued) The user can
ask Ave to display the match tree through the default partition
by selecting the View button in the menu and specifying that
the match tree should be viewed through the default partition to
get a display shown in Figure 
. Each new partition
has a name and a list of variables that are used in the predicate expression
used to create the partition. The default partition is
named default, and has a variable called pattern. Needless
to say, both are keywords in Ariadne.
Examples showing the usefulness of computing new partitions based on attribute values are presented in []. The usefulness of imposing such structural transformation on the match tree will become clear when we discuss the integration of event- and state-based approaches. Partitioning with respect to attributes is also present in performance debuggers, where a frequency domain filter can divide a set of processors into a number of categories, and show the category statistics to illuminate some global behavior [8][][1]. However in the performance visualization, the attributes along which partitioning needs to be done do not vary across application domains, and can thus be provided as in-built functions. In our domain of correctness debugging, the user has to specify the attribute of interest, and may need to compute it explicitly. Our spreadsheet approach thus subsumes the functionality that is currently provided by the performance visualization systems [8][][1], and can be used as a unified debugging platform.
