Re: [Valgrind-users] inclusive profiling

[John R. <jr...@Bi...>]

> Is there any profiling tool which could count the execution time of
> basic blocks ( like gprof and oprofile) in the inclusive style?

No.

Assume that "in the inclusive style" means "program counter is in this
basic block or in the dynamic nest of subroutines called (directly or
indirectly) by this basic block."  There are two problems.  Basic blocks
tend to be short in control-dominated code [as opposed to data-dominated,
such as traditional Fortran floating-point computation or software crypto],
and there is no published technique that accounts correctly for recursion.
Short basic blocks impose high overhead for direct measurement of time,
and decrease the precision of a given duration of sampling.  Despite 26 years
since the problem of accounting for recursion was identified in the
original paper (1982), gprof has not improved its method, which is subject
to giving results that are incorrect or seriously misleading without
specific warning.

In nearly any actual executing application, just a few basic blocks are
all that matter.  Candidates can be identified by profiling at the level
of subroutines, refined by re-writing basic blocks as subroutines, then
measured by re-profiling the modified code.

My 'tsprof' proprietary application from BitWagon Software LLC solved
the problem of accounting for recursion.  The product has been withdrawn
due to insufficient market demand.

--