[libc-commits] [libc] [libc] add snmalloc as an alternative allocator to libc (PR #122284)

[Schrodinger ZHU Yifan via libc-commits]

SchrodingerZhu wrote:

@mjp41 Thanks for the suggestion. However, after investigating, I think libc's implementation is quite wrong currently. I am leaning to keep using pthread-style cleanup because it does not need me to inject `_malloc_thread_cleanup` call into startup and thread init routines.

@lntue @nickdesaulniers 

So, right inside startup code, we do a call to `atexit`, even before `call_init_array_callbacks`. 

```c++
 //...
 // We want the fini array callbacks to be run after other atexit
  // callbacks are run. So, we register them before running the init
  // array callbacks as they can potentially register their own atexit
  // callbacks.
  atexit(&call_fini_array_callbacks);

  call_init_array_callbacks(static_cast<int>(app.args->argc),
                            reinterpret_cast<char **>(app.args->argv),
                            reinterpret_cast<char **>(env_ptr));

  int retval = main(static_cast<int>(app.args->argc),
                    reinterpret_cast<char **>(app.args->argv),
                    reinterpret_cast<char **>(env_ptr));

  exit(retval);
```

When `LIBC_COPT_PUBLIC_PACKAING=On` , this call will invoke `malloc` via our `new` operators.

```
#if defined(LIBC_TARGET_ARCH_IS_GPU)
using ExitCallbackList = FixedVector<AtExitUnit, 64>;
#elif defined(LIBC_COPT_PUBLIC_PACKAGING)
using ExitCallbackList = ReverseOrderBlockStore<AtExitUnit, 32>;
#else
using ExitCallbackList = FixedVector<AtExitUnit, CALLBACK_LIST_SIZE_FOR_TESTS>;
#endif
```

This is very wrong in the sense that allocator is not even initialized. It is likely that allocators themselves will need to install `exiting` hooks. I suggest we should use mmapped memory here.

https://github.com/llvm/llvm-project/pull/122284