callgrind

callgrind - How it Works

How does it work?

• valgrind: run-time code instrumentation

• callgrind is a “tool” using valgrind infrastructure

• Call-graph analysis, optional cache and branch-prediction analysis

Compared to good old gprof …

• Sluggishly slow (the price of run-time instrumentation)

• More accurate

• Nice graphical tool ⟶ kcachegrind

callgrind - How it is Used

Compiler and Linker Calls

$ gcc -c -o program.o program.c
$ gcc -g -c -o program.o program.c # debug info
$ gcc -o program program.o

• No compiler attention needed

• Debug information only for source annotation (⟶ kcachegrind)

Producing Output: callgrind.out.<pid>

$ valgrind --tool=callgrind ./program
...
$ ls callgrind.out.*
callgrind.out.16761

callgrind - Analysis Per Commandline

Basically records the same information as gprof

• Flat profile

• Call Graph

• All sorts of counters (can detect cache misses etc.)

Most basic analysis tool: callgrind_annotate

$ callgrind_annotate callgrind.out.16761
... unreadable but informative garbage ...

callgrind - Analysis With kcachegrind

callgrind - Useful Options

Instrumented code takes very long

• Start without instrumentation

• Switch on explicitly during runtime (by PID)

Start Without Instrumentation

$ valgrind --tool=callgrind --instr-atstart=no ./program

Switch on Instrumentation

$ callgrind_control --instr=on 16761