The tools are good but not magic. It's always possible that they miss something (e.g. when multithreading, the bug might not manifest since the timings are all

The tools are good but not magic. It's always possible that they miss something (e.g. when multithreading, the bug might not manifest since the timings are all different when running in valgrind). But this is a red flag: "Nothing worked. It's leaking but it's not. It's leaking but the debugger says no memory was lost. It's leaking in ways that are dependent on irrelevant changes to the ordering of mundane operations. This can't be happening." This is a red flag for heap corruption - or multithreading bugs. (Stack corruption is usually a crash and a wrong stack trace). As it's not trivially reproducible, it's probably a multithreading bug. It's also easy to imagine that OP wrote a scripty-input generator in C++ to run through valgrind which gives single threaded inputs. So running under valgrind it won't be detected. So it's never fun to solve and ever since I changed languages, my multithreaded debugging skills have become a bit rusty. Hey-o! But it would be good if he amended his post to reduce the vitriol aimed at GCC. It's demonstrably false that GCC's default allocator holds a cache. It makes OP look stupid to people in the know; and makes GCC look bad to people not in the know. No one wins.