Documentation/ deep profiler


The old Mercury profiler, mprof, was a straightforward clone of the standard Unix profiling tool, gprof. However, while gprof is quite useful in profiling C programs, we found the usefulness of mprof in profiling Mercury programs to be severely limited. The main reason is that gprof and mprof assume that all calls to a function or predicate have roughly the same cost. This is usually close enough to the truth in C programs for the output of gprof to be useful. In Mercury, it is usually very far from truth, because Mercury programs make much greater use of polymorphism than C programs do.

For example, the most expensive predicates in the Mercury compiler are usually the predicates for searching 2-3-4 trees and for inserting into 2-3-4 trees. However, these predicates are called (indirectly) from many hundreds of places in the Mercury compiler, some of which handle bigger trees than others.

We designed the deep profiler specifically to meet the needs of programs written in programming languages such as Mercury, programs whose characteristics include

The key to our solution of these problems is to associate with each profiling measurement a very detailed context: essentially a representation of the entire chain of the ancestor functions or predicates and their call sites at the time of the measurement, but compressing sequences of (mutually) recursive ancestor calls. The main challenge was keeping the required overhead within bounds. Deep profiling slows down programs by a factor between two and three, with a factor of 2.4 being reasonably typical. Since this is only slightly higher than the overheads of profiling techniques that yield significantly less detailed and less accurate data, we believe we succeeded.


Like most profilers, the Mercury deep profiler works by instrumenting the program to be profiled, having the instrumented program record its profiling data in a file, and postprocessing the contents of this file.

You can ask for a program to be compiled with deep profiling instrumentation by using the `--deep-profiling' option to `mmc', or by including `GRADEFLAGS = --deep-profiling' in your `Mmakefile'. However, please note that deep profiling is not compatible with the old Mercury profiling grades, and it is not (yet) implemented for grades (such as hlc grades) that use the compiler's MLDS back end.

Programs compiled with deep profiling can generate large amounts of data, since deep profiling yields very detailed information. We have therefore implemented the postprocessing program as a CGI-based web service. Its input is a stream of requests, with each request specifying a data file, the part the user wants to view, and his or her preferences about the format of the display. The server generates, for each request, a web page containing the requested information, which will typically contain links that generate other requests.


The Mercury deep profiler has been available since Mercury version 10.04.