Zig has a very nice built-in memory leak detector in its General Purpose Allocator (GPA), but sometimes you have to break out Valgrind to get to the bottom of things.
When I first did this, I ran into a couple of problems, which I’m jotting down here along with solutions.
Thanks to mikdusan, ifreund and lemonboy for helping me figure out various issues
First of all, to get Valgrind going with Zig programs, you’ll have to switch to the C allocator. In my case, this was as easy as flipping
pub var allocator = &gpa.allocator;
to
pub var allocator = std.heap.c_allocator;
You also need to link with libc, so add this to your build.zig file:
exe.linkLibC();
Running Valgrind
I used Valgrind 3.16.1 for this; old versions of Valgrind have limited/buggy dwarf4 support which manifests itself as missing symbols in traces.
valgrind --leak-check=full --track-origins=yes \
--show-leak-kinds=all --num-callers=15 \
--log-file=leak.txt ./my-app
The --num-callers=15 is important (you can pick any suitable number of course), as otherwise you’ll get traces that are too short. For instance, I ran into a leak where the Reader interface was used, and the trace didn’t include any of my application code, only stdlib stuff.
Illegal instruction ??
I ran Valgrind on a Linux machine with avx512, and got “Illegal instruction” on an @intToFloat cast. This was solved by adding -mcpu=native-avx512f to the build command, with mcpu=baseline being the safest option if you still run into issues.
This is due to a bug in Valgrind: https://bugs.kde.org/show_bug.cgi?id=383010
Some more context here: https://github.com/ziglang/zig/issues/8615