callgrind

Callgrind is a tracing profiler which records the function call history of a target program and collects the number of executed instructions. It is part of the valgrind tool suite.

Profiling data is collected by instrumentation rather than sampling of the target program.

Callgrind does not capture the actual time spent in a function but computes the inclusive & exclusive cost of a function based on the instructions fetched (Ir = Instruction read). This provides reproducibility and high-precision and is a major difference to sampling profilers like perf or vtune. Therefore effects like slow IO are not reflected, which should be kept in mind when analyzing callgrind results.

By default the profiler data is dumped when the target process is terminating, but callgrind_control allows for interactive control of callgrind.

# Run a program under callgrind.
valgrind --tool=callgrind -- <prog> [<args>]

# Interactive control of callgrind.
callgrind_control [opts]
  opts:
    -b ............. show current backtrace
    -e ............. show current event counters
    -s ............. show current stats
    --dump[=file] .. dump current collection
    -i=on|off ...... turn instrumentation on|off

Results can be analyzed by using one of the following tools

  • callgrind_annotate (cli)
    # Show only specific trace events (default is all).
    callgrind_annotate --show=Ir,Dr,Dw [callgrind_out_file]
    
  • kcachegrind (ui)

The following is a collection of frequently used callgrind options.

valgrind --tool=callgrind [opts] -- <prog>
  opts:
    --callgrind-out-file=<file> .... output file, rather than callgrind.out.<pid>
    --dump-instr=<yes|no> .......... annotation on instrucion level,
                                     allows for asm annotations

    --instr-atstart=<yes|no> ....... control if instrumentation is enabled from 
                                     beginning of the program

    --separate-threads=<yes|no> .... create separate output files per thread,
                                     appends -<thread_id> to the output file

    --cache-sim=<yes|no> ........... control if cache simulation is enabled

Trace events

By default callgrind collects following events:

  • Ir: Instruction read

Callgrind also provides a functional cache simulation with their own model, which is enabled by passing --cache-sim=yes. This simulates a 2-level cache hierarchy with separate L1 instruction and data caches (L1i/ L1d) and a unified last level (LL) cache. When enabled, this collects the following additional events:

  • I1mr: L1 cache miss on instruction read
  • ILmr: LL cache miss on instruction read
  • Dr: Data reads access
  • D1mr: L1 cache miss on data read
  • DLmr: LL cache miss on data read
  • Dw: Data write access
  • D1mw: L1 cache miss on data write
  • DLmw: LL cache miss on data write

Profile specific part of the target

Programmatically enable/disable instrumentation using the macros defined in the callgrind header.

#include <valgrind/callgrind.h>

int main() {
    // init ..

    CALLGRIND_START_INSTRUMENTATION;
    compute();
    CALLGRIND_STOP_INSTRUMENTATION;

    // shutdown ..
}

In this case, callgrind should be launched with --instr-atstart=no.

Alternatively instrumentation can be controlled with callgrind_control -i on/off.

The files cg_example.cc and Makefile provide a full example.