Currently ::malloc() can be called by many places, including

ResourceObj::operator new()
-> AllocateHeap()
  -> os::malloc(size_t, MEMFLAGS flags, const NativeCallStack&)

Inside os::malloc(), many checks are done related to NMT (and after JDK-8253495, checking DumpSharedSpaces).

During the review of JDK-8253495, it was suggested that we should try to reduce the overheads in os::malloc(). Maybe we should have a fast path when both NMT and DumpSharedSpaces, we should directly call into os::malloc() and avoid all the unnecessary checks.

I.e., add something like this at various places??

char* AllocateHeap(size_t size,
                   MEMFLAGS flags,
                   const NativeCallStack& stack,
                   AllocFailType alloc_failmode /* = AllocFailStrategy::EXIT_OOM*/) {

+ if (_os_malloc_fast_path) {
+ return os::fast_malloc(size);
+ }

  char* p = (char*) os::malloc(size, flags, stack);
  if (p == NULL && alloc_failmode == AllocFailStrategy::EXIT_OOM) {
    vm_exit_out_of_memory(size, OOM_MALLOC_ERROR, "AllocateHeap");
  }
  return p;
}

(Note: it's unclear how much this would save.)