Two - pass heap pro ling : a matter of life

A heap proole is a chart showing the contents of heap memory throughout a computation. Contents are depicted abstractly by showing how much space is occupied by memory cells in each of several classes. A good heap prooler can use a variety of attributes of memory cells to deene a classiication. EEective prooling usually involves a combination of attributes. The ideal prooler gives full support for combination in two ways. First, a section of the heap of interest to the programmer can be speciied by constraining the values of any combination of cell attributes. Secondly, no matter what attributes are used to specify such a section, a heap proole can be obtained for that section only, and any other attribute can be used to deene the classiication. Achieving this ideal is not simple for some combinations of attributes. A heap proole is derived by interpolation of a series of censuses of heap contents at diierent stages. The obvious way to obtain census data is to traverse the live heap at intervals throughout the computation. This is ne for static attributes (eg. What type of value does this memory cell represent?), and for dynamic attributes that can be determined for each cell by examining the heap at any given moment (eg. From which function closures can this cell be reached?). But some attributes of cells can only be determined retrospectively by post-mortem inspection as a cell is over-written or garbage-collected (eg. Is this cell ever used again?). Now we see the problem: if a prooler supports both live and post-mortem attributes, how can we implement the ideal of unrestricted combinations? That is the problem we solve in this paper. We give techniques for prooling a heap section speciied in terms of both live and post-mortem attributes. We show how to generate live-attribute prooles of a section of the heap speciied using post-mortem attributes, and vice versa. 1 Variety and combination in heap prooling We begin by reviewing the diierent kinds of memory-cell attributes on which heap proolers to-date have been based, and how these attributes have been combined. The original heap prooler of the kind we are concerned with in this paper dealt with just two attributes of heap cells RW93]. A producer proole classiies cells by