Valgrind

Notes sur Valgrind, issuees d'une formation de Julian Seward, et de mon expérience.

Le site de Valgrind est valgrind.org.

Sommaire :

• Memcheck
• Callgrind
• Cachegrind
• Other other profiling tools
• Massif
• Helgrind

Memcheck

Memcheck (+ Valkyrie UI): memory error detector:

• - addressing errors
• - invalid uses of malloc and free
• - memory leaks
• - use of undefined values

Recommended compiling options:

• -g (to see line numbers)
• -O (avoid -O2 and above)
• -fno-inline-functions (maybe)

Run Memcheck:

valgrind --tool=memcheck --xml=yes --leak-check=full --show-reachable=yes --xml-file=results.xml program [with its arguments]

Options:

• --num-callers=[x]: to see more frames in errors stack traces
• --track-origins=yes: to find the sources of undefinedness (slower)
• --suppressions=[filename] (--gen-suppressions=yes/all to start)
• for complex programs: --trace-children=yes, --child-silent-after-fork=yes, --log-file=logfile%p.txt, --trace-children-skip=patt1,patt2

Make requests from code:

#include <memcheck/memcheck.h>

Callgrind

Callgrind (+ KCachegrind UI): low level time profiler and function-call profiler, simulating CPU caches.

Recommended compiling options:

• -g (to see line numbers)
• maximum optimisation flags
• -march=native
• -fno-inline-functions: disable inlining to get clearer results

Run Callgrind:

valgrind --tool=callgrind --simulate-cache=yes --branch-sim=yes --cacheuse=yes \
--callgrind-out-file=callgrind.out program [with its arguments]
callgrind_annotate callgrind.out [source files to annotate] | grep -v callgrind_annotate

Manual selective profiling:

• - launch with --instr-atstart=no
• - start profiling: callgrind_control -i on [pid]
• - stop profiling: callgrind_control -i off [pid]
• - dump profile data: callgrind_control -d [pid]
• - reset profile counters: callgrind_control -z [pid]

Automatic selective profiling:

valgrind --tool=callgrind --simulate-cache=yes \
--toggle-collect="*MyClass::MyMethod*" \
--dump-after="MyClass::MyMethod(arguments)" \
program [with its arguments]

To get the exact signature of a function:

nm -l -C libdll.so | grep MyMethod

One file per thread: --separate-threads=yes.

Cachegrind

Cachegrind: simpler than Callgrind. Does branch predictions.

Run Cachegrind:

valgrind --tool=cachegrind --branch-sim=yes --cache-sim=yes program [with its arguments]
cg_annotate cachegrind.out.[pid] [source files to annotate]  # or --auto=yes

Other other profiling tools

• - oprofile
• - strace

Massif

Massif: heap memory space profiler. Takes periodic snapshots of heap.

Run Massif:

valgrind --tool=massif program [with its arguments]

Display Massif data:

ms_print massif.out.valgrind.[pid]

ms_print options:

• --x=[x]: graph number of columns
• --y=[y]: graph number of rows

Tool similar to Massif: DHAT.

Helgrind

Helgrind: thread error detector.

• - Detects race conditions, and potential deadlocks in multithreaded code.
• - Finds shared data accessed without adequate locking.

Similar tool: DRD.

La dernière mise à jour de cette page date d'avril 2019.

Le contenu de ce site est, en tant qu'œuvre originale de l'esprit, protégé par le droit d'auteur.
Pour tout commentaire, vous pouvez m'écrire à xavier.lamorlette@gmail.com.