On Mon, 10 Jun 2024 at 09:52, Jonathan Wakely <jwak...@redhat.com> wrote:
>
> On Mon, 10 Jun 2024 at 05:38, François Dumont <frs.dum...@gmail.com> wrote:
> >
> > Hi
> >
> > libstdc++: [_Hashtable] Optimize destructor
> >
> > Hashtable destructor do not need to call clear() method that in addition to
> > destroying all nodes also reset all buckets to nullptr.
> >
> > libstdc++-v3/ChangeLog:
> >
> > * include/bits/hashtable.h (~_Hashtable()): Replace clear call with
> > a _M_deallocate_nodes call.
> >
> > Tested under Linux x64, ok to commit ?
>
> This will only matter at -O0 because the compiler will optimize out
> the memset and zeroing in the destructor, because those are dead
> stores.
>
> The call to memset in the destructor is incorrect anyway, that would
> be undefined behaviour for fancy pointers of non-trivial type.
It should be something like this, which the compiler can still
optimize for raw pointers:
--- a/libstdc++-v3/include/bits/hashtable.h
+++ b/libstdc++-v3/include/bits/hashtable.h
@@ -2582,8 +2582,8 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
clear() noexcept
{
this->_M_deallocate_nodes(_M_begin());
- __builtin_memset(_M_buckets, 0,
- _M_bucket_count * sizeof(__node_base_ptr));
+ for (size_t __i = 0; __i < _M_bucket_count; ++__i)
+ _M_buckets[__i] = nullptr;
_M_element_count = 0;
_M_before_begin._M_nxt = nullptr;
}
>
> OK for trunk.
