Valgrind

Valgrind

Purpose

Guide agents through Valgrind tools: Memcheck for memory errors, Cachegrind for cache simulation, Callgrind for call graphs, and Massif for heap profiling.

Triggers

• "My program has a memory leak / use-after-free"
• "I can't use ASan — can I use Valgrind instead?"
• "How do I profile cache behaviour without perf?"
• "How do I visualize call graphs with Callgrind?"
• "How do I profile heap allocation patterns?"
• "Valgrind reports errors in third-party code I can't fix"

Workflow

1. Memcheck — memory error detection

Compile with -g -O1 for best results. -O0 is also fine; avoid -O2+ which can produce false positives.

valgrind --tool=memcheck \
         --leak-check=full \
         --show-leak-kinds=all \
         --track-origins=yes \
         --error-exitcode=1 \
         ./prog [args]

Key flags:

2. Understanding Memcheck output

==12345== Invalid read of size 4
==12345==    at 0x4007A2: foo (main.c:15)
==12345==    by 0x400846: main (main.c:30)
==12345==  Address 0x5204040 is 0 bytes after a block of size 40 alloc'd
==12345==    at 0x4C2FB0F: malloc (in /usr/lib/valgrind/vgpreload_memcheck.so)
==12345==    by 0x40074B: main (main.c:25)

• Invalid read/write: out-of-bounds access; check array bounds
• Use of uninitialised value: read before write; use --track-origins=yes
• Invalid free / double free: mismatched malloc/free; check ownership
• Definitely lost: reachable via no pointers; clear leak
• Indirectly lost: lost through a chain; usually means one root leak
• Possibly lost: might be pointing into the middle of a block; often FP with custom allocators

3. Leak kinds

For library code: --show-leak-kinds=definite,indirect reduces noise from still-reachable.

4. Suppressions

# Generate suppression for current error
valgrind --gen-suppressions=yes ./prog 2>&1 | grep -A20 '{'

# Example suppression file (valgrind.supp)
{
   openssl_uninit
   Memcheck:Cond
   fun:SHA256_Init
   ...
}

# Use suppression file
valgrind --suppressions=valgrind.supp ./prog

5. Cachegrind — cache simulation

valgrind --tool=cachegrind ./prog

# Output: cachegrind.out.PID
# Annotate source
cg_annotate cachegrind.out.12345 --auto=yes

# Diff two runs
cg_diff cachegrind.out.before cachegrind.out.after

Key metrics:

• I1mr / ILmr: L1/LL instruction cache miss rate
• D1mr / DLmr: L1/LL data read miss rate
• D1mw / DLmw: L1/LL data write miss rate

6. Callgrind — call graph profiling

valgrind --tool=callgrind --callgrind-out-file=callgrind.out ./prog

# Analyse
callgrind_annotate callgrind.out

# Visualise in KCachegrind (GUI)
kcachegrind callgrind.out

Callgrind is slower than perf but works without root and provides exact call counts.

7. Massif — heap profiling

valgrind --tool=massif ./prog

# Visualise
ms_print massif.out.PID | less

# GUI
massif-visualizer massif.out.PID

Massif shows heap usage over time; useful for finding peak allocation sites and tracking gradual leaks.

8. Performance considerations

Valgrind Memcheck runs ~10-50x slower than native. Mitigations:

• Use a shorter representative workload
• Use --error-exitcode=1 to fail fast in CI
• Use ASan (-fsanitize=address) for faster memory checking during development
• Reserve Valgrind for cases where ASan can't be used (old toolchains, production-like environments)

For a comparison of Valgrind vs ASan, see references/valgrind-vs-asan.md.

Related skills

• Use skills/runtimes/sanitizers for faster ASan/UBSan alternatives
• Use skills/profilers/linux-perf for CPU-level profiling (faster than Cachegrind)
• Use skills/profilers/flamegraphs to visualise Callgrind output

Valgrind vs AddressSanitizer

Comparison

When to use which

Use ASan when:

• You control the build process
• You want fast iteration in development
• You need to catch stack overflows and global OOB
• CI/CD integration (speed matters)

Use Valgrind when:

• You cannot recompile (testing pre-built binaries)
• You need uninitialised value detection (use --track-origins=yes)
• You need cache profiling (Cachegrind) or call graphs (Callgrind)
• You need heap usage profiling (Massif)
• You are on an older toolchain that doesn't support sanitizers

Both together

Running ASan + Valgrind on the same binary is not recommended (both intercept malloc/free). Choose one per run.

However: compile with ASan for dev CI, run Valgrind Memcheck as a separate nightly check on a non-instrumented binary for uninit-value detection.

Combining sanitizers for maximum coverage

# ASan + UBSan + LeakSanitizer (all Clang/GCC)
clang -fsanitize=address,undefined -g -O1 -o prog main.c

# MemorySanitizer (uninit reads; Clang only; all-clang build required)
clang -fsanitize=memory -g -O1 -o prog main.c

# ThreadSanitizer (data races)
clang -fsanitize=thread -g -O1 -o prog main.c