Dynamic Analysis and Profiling of Multithreaded Systems - Advanced Operating Systems and Kernel Applications - page 169

Box 5.

nt main(int argc, char* argv[])
{
executable* exec = new executable(argv[1]);
exec->read_contents();
routine* r;

foreach_routine (r, exec->routines())

{

instrument(r);

while(!exec->hidden_routines()->is_empty())

{

r = exec->hidden_routines()->irst();

exec->hidden_routines()->remove(r);

instrument(r);

exec->routines()->add(r);

}

}

addr x = exec->edited_addr(exec->start_address());
exec->write_edited_executable(st_cat(argv[1], “.count”), x);
return (0);
}

void instrument(routine* r)
{
static long num = 0;
cfg* g = r->control_low_graph();
bb* b;

foreach_bb (b, g->blocks())

{

if (1 < b->succ()->size())

{

edge* e;

foreach_edge (e, b->succ())

{

// incr_count is the user-defined code snippet

e->add_code_along(incr_count(num));

num += 1;

}

}

}
r->produce_edited_routine();
r->delete_control_low_graph();
}

The Paradyn work from the University of
Wisconsin, Madison (Miller, Callaghan, Car-
gille, Hollingsworth, Irvin, & Karavanic, 1995)
was designed specifically for measuring the
performance of parallel programs. Paradyn uses

dynamic instrumentation to apply trace function-
ality according to a set of resource hierarchies,
as shown in Figure 3 (shaded nodes represent an
example focus, all spin locks in CPU#1, in any
procedure). Entities within the resource hierar-