callgrind is a profiling tool built on Valgrind's instrumentation engine. Unlike perf, which samples at fixed intervals, callgrind counts instructions deterministically—every instruction is counted. This gives reproducible results even for very short-running programs.

valgrind --tool=callgrind ./program
callgrind_annotate callgrind.out.12345

perf record is fast but requires many samples to be statistically accurate. callgrind is slower but deterministic—perfect for analyzing short-lived test cases or programs where sampling might miss small but important functions.

Generating profiles:

valgrind --tool=callgrind --dump-instr=yes ./program   # detailed instruction counts
callgrind_annotate callgrind.out.PID

By default, callgrind records per-function instruction counts and call graphs. --dump-instr=yes adds per-instruction counts (larger data file, more detail).

Output: callgrind_annotate displays the call graph sorted by instruction count. Each line shows a function, number of instructions executed, and number of times it was called:

Ir                 Calls  Fn name
1,234,567          5,000  main
  1,000,000        1,000  process
    500,000        1,000  compute_kernel
      250,000      1,000  matrix_multiply

“Ir” is instruction references (instructions executed). This shows main executed 1.2M instructions and called process 5000 times.

Advantages over perf: 1. Deterministic: same results every run (no sampling noise) 2. Complete: captures every instruction, never misses small functions 3. Reliable: short programs that don't generate enough samples under perf are profiled accurately

Disadvantages: 1. Slow: 10-50x overhead from instrumentation 2. Not real-time: can't profile production workloads 3. Memory-heavy: large programs produce gigabyte-sized profile files

Use cases: Test suite profiling, understanding performance of short algorithms, verifying that optimizations are effective (deterministic comparisons), profiling on systems where perf isn't available.

Comparing callgrind profiles:

valgrind --tool=callgrind --outfile=profile1.out ./program_v1
valgrind --tool=callgrind --outfile=profile2.out ./program_v2
callgrind_annotate profile1.out profile2.out

This compares two profiles side-by-side, showing which functions changed performance.