diff mbox

77864 Fix noexcept conditions for map/set default constructors

Message ID 4d30deb3-8720-2aa6-d0f2-94f6fca44825@gmail.com
State New
Headers show

Commit Message

François Dumont Oct. 6, 2016, 8:17 p.m. UTC
On 05/10/2016 14:13, Jonathan Wakely wrote:
> On 05/10/16 14:10 +0200, Marc Glisse wrote:
>> On Wed, 5 Oct 2016, Jonathan Wakely wrote:
>>
>>> I added conditional noexcept to maps and sets, but forgot to account
>>> for the comparison function, which could throw when constructed.
>>
>> IMO you are fighting a losing battle. We should implement 
>> noexcept(auto) (possibly with some private __noexcept_auto__ 
>> spelling) and just use that in most places where we want a 
>> conditional noexcept.
>
> That would be nice, but beyond my ability :-)
>
> So until then we can either remove the exception specs entirely, or
> make them correct.
>
>
Another approach is to rely on existing compiler ability to compute 
conditional noexcept when defaulting implementations. This is what I 
have done in this patch.

The new default constructor on _Rb_tree_node_base is not a problem as it 
is not used to build _Rb_tree_node.

I'll try to do the same for copy constructor/assignment and move 
constructor/assignment.

     * include/bits/stl_map.h (map()): Make default.
     * include/bits/stl_multimap.h (multimap()): Likewise.
     * include/bits/stl_multiset.h (multiset()): Likewise.
     * include/bits/stl_set.h (set()): Likewise.
     * include/bits/stl_tree.h (_Rb_tree_node_base()): New.
     (_Rb_tree_impl()): Make default.
     (_Rb_tree_impl::_M_initialize()): Delete.
     (_Rb_tree()): Make default.

Tested under Linux x86_64, ok to commit ?

François

Comments

Jonathan Wakely Oct. 6, 2016, 9:34 p.m. UTC | #1
On 06/10/16 22:17 +0200, François Dumont wrote:
>Another approach is to rely on existing compiler ability to compute 
>conditional noexcept when defaulting implementations. This is what I 
>have done in this patch.
>
>The new default constructor on _Rb_tree_node_base is not a problem as 
>it is not used to build _Rb_tree_node.

Why not?

>I'll try to do the same for copy constructor/assignment and move 
>constructor/assignment.

We need to make sure we don't change whether any of those operations
are trivial (which shouldn't be a problem for copy/move, because they
are definitely very non-trivial and will stay that way!)

Does this change the default constructors from non-trivial to trivial?


>--- a/libstdc++-v3/include/bits/stl_tree.h
>+++ b/libstdc++-v3/include/bits/stl_tree.h
>@@ -108,6 +108,10 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>     _Base_ptr		_M_left;
>     _Base_ptr		_M_right;
> 
>+    _Rb_tree_node_base() _GLIBCXX_NOEXCEPT
>+      : _M_color(_S_red), _M_parent(0), _M_left(this), _M_right(this)
>+    { }
>+
>     static _Base_ptr
>     _S_minimum(_Base_ptr __x) _GLIBCXX_NOEXCEPT
>     {

Another option would be:

    struct _Head_node : _Rb_tree_node_base {
      _Head_node() {
        _M_color = _S_red;
        _M_parent = _Base_ptr();
        _M_left = _M_right = this;
      }
    };

>@@ -603,23 +607,33 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>         {
> 	  _Key_compare		_M_key_compare;
> 	  _Rb_tree_node_base 	_M_header;

          _Head_node _M_header;

That way *only* this node gets the zero-initialization, not all node
bases.

With either solution we can get rid of _M_header() in every
ctor-initializer-list.


>+#if __cplusplus < 201103L
> 	  size_type 		_M_node_count; // Keeps track of size of tree.
>+#else
>+	  size_type 		_M_node_count = 0; // Keeps track of size of tree.
>+#endif
> 
>+#if __cplusplus < 201103L
> 	  _Rb_tree_impl()
> 	  : _Node_allocator(), _M_key_compare(), _M_header(),
> 	    _M_node_count(0)
>-	  { _M_initialize(); }
>+	  { }
>+#else
>+	  _Rb_tree_impl() = default;
>+#endif
> 
> 	  _Rb_tree_impl(const _Key_compare& __comp, const _Node_allocator& __a)
>-	  : _Node_allocator(__a), _M_key_compare(__comp), _M_header(),
>-	    _M_node_count(0)
>-	  { _M_initialize(); }
>+	  : _Node_allocator(__a), _M_key_compare(__comp), _M_header()
>+#if __cplusplus < 201103L
>+	  , _M_node_count(0)
>+#endif

Doing this conditionally seems pointless, why not just set it here
unconditionally?
François Dumont Oct. 8, 2016, 8:55 p.m. UTC | #2
On 06/10/2016 23:34, Jonathan Wakely wrote:
> On 06/10/16 22:17 +0200, François Dumont wrote:
>> Another approach is to rely on existing compiler ability to compute 
>> conditional noexcept when defaulting implementations. This is what I 
>> have done in this patch.
>>
>> The new default constructor on _Rb_tree_node_base is not a problem as 
>> it is not used to build _Rb_tree_node.
>
> Why not?

_Rb_tree_node_base is used in 2 context. As member of _Rb_tree_impl in 
which case we need the new default constructor. And also as base class 
of _Rb_tree_node which is never constructed. Nodes are being allocated 
and then associated value is being constructed through the allocator, 
the node default constructor itself is never invoked.

     If you think it is cleaner to create an intermediate type that will 
take care of this initialization through its default constructor I can 
do that.

>
>> I'll try to do the same for copy constructor/assignment and move 
>> constructor/assignment.
>
> We need to make sure we don't change whether any of those operations
> are trivial (which shouldn't be a problem for copy/move, because they
> are definitely very non-trivial and will stay that way!)
>
> Does this change the default constructors from non-trivial to trivial?
It would be a major compiler bug if making a constructor default was 
making it trivial.
>
>
>> --- a/libstdc++-v3/include/bits/stl_tree.h
>> +++ b/libstdc++-v3/include/bits/stl_tree.h
>> @@ -108,6 +108,10 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>>     _Base_ptr        _M_left;
>>     _Base_ptr        _M_right;
>>
>> +    _Rb_tree_node_base() _GLIBCXX_NOEXCEPT
>> +      : _M_color(_S_red), _M_parent(0), _M_left(this), _M_right(this)
>> +    { }
>> +
>>     static _Base_ptr
>>     _S_minimum(_Base_ptr __x) _GLIBCXX_NOEXCEPT
>>     {
>
> Another option would be:
>
>    struct _Head_node : _Rb_tree_node_base {
>      _Head_node() {
>        _M_color = _S_red;
>        _M_parent = _Base_ptr();
>        _M_left = _M_right = this;
>      }
>    };
If you want something like that I would rather do:

_Rb_tree_node_base() = default;
_Rb_tree_node_base(int) _GLIBCXX_NO_EXCEPT
       : _M_color(_S_red), _M_parent(0), _M_left(this), _M_right(this)
     { }

and then:

struct _Head_node : _Rb_tree_node_base {
     _Head_node() _GLIBCXX_NO_EXCEPT
     : _Rb_tree_node_base(0)
     { }
};

but as I already said the default constructor on _Rb_tree_node_base 
works just fine.

>
>> @@ -603,23 +607,33 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>>         {
>>       _Key_compare        _M_key_compare;
>>       _Rb_tree_node_base     _M_header;
>
>          _Head_node _M_header;
>
> That way *only* this node gets the zero-initialization, not all node
> bases.
>
> With either solution we can get rid of _M_header() in every
> ctor-initializer-list.
I am already preparing another patch, I will add this simplification.
>
>
>> +#if __cplusplus < 201103L
>>       size_type         _M_node_count; // Keeps track of size of tree.
>> +#else
>> +      size_type         _M_node_count = 0; // Keeps track of size of 
>> tree.
>> +#endif
>>
>> +#if __cplusplus < 201103L
>>       _Rb_tree_impl()
>>       : _Node_allocator(), _M_key_compare(), _M_header(),
>>         _M_node_count(0)
>> -      { _M_initialize(); }
>> +      { }
>> +#else
>> +      _Rb_tree_impl() = default;
>> +#endif
>>
>>       _Rb_tree_impl(const _Key_compare& __comp, const 
>> _Node_allocator& __a)
>> -      : _Node_allocator(__a), _M_key_compare(__comp), _M_header(),
>> -        _M_node_count(0)
>> -      { _M_initialize(); }
>> +      : _Node_allocator(__a), _M_key_compare(__comp), _M_header()
>> +#if __cplusplus < 201103L
>> +      , _M_node_count(0)
>> +#endif
>
> Doing this conditionally seems pointless, why not just set it here
> unconditionally
I wasn't sure about this one, just seemed clearner to not do this 0 
initialization again as already done at class definition level. If 
compiler just moved it away then ok, I can remove the check.

François
Jonathan Wakely Oct. 9, 2016, 3:14 p.m. UTC | #3
On 08/10/16 22:55 +0200, François Dumont wrote:
>On 06/10/2016 23:34, Jonathan Wakely wrote:
>>On 06/10/16 22:17 +0200, François Dumont wrote:
>>>Another approach is to rely on existing compiler ability to 
>>>compute conditional noexcept when defaulting implementations. This 
>>>is what I have done in this patch.
>>>
>>>The new default constructor on _Rb_tree_node_base is not a problem 
>>>as it is not used to build _Rb_tree_node.
>>
>>Why not?
>
>_Rb_tree_node_base is used in 2 context. As member of _Rb_tree_impl in 
>which case we need the new default constructor. And also as base class 
>of _Rb_tree_node which is never constructed. Nodes are being allocated 
>and then associated value is being constructed through the allocator, 
>the node default constructor itself is never invoked.

In C++03 mode that is true, but it's only valid because the type is
trivially-constructible. If the type requires "non-vacuous
initialization" then it's not valid to allocate memory for it and
start using it without invoking a constructor. If you add a
non-trivial constructor then we can't do that any more.

In C++11 and later, see line 550 in <bits/stl_tree.h>

        ::new(__node) _Rb_tree_node<_Val>;

This default-constructs a tree node. Currently there is no
user-provided default constructor, so default-construction does no
initialization. Adding your constructor would mean it is used for
every node.


>    If you think it is cleaner to create an intermediate type that 
>will take care of this initialization through its default constructor 
>I can do that.
>
>>
>>>I'll try to do the same for copy constructor/assignment and move 
>>>constructor/assignment.
>>
>>We need to make sure we don't change whether any of those operations
>>are trivial (which shouldn't be a problem for copy/move, because they
>>are definitely very non-trivial and will stay that way!)
>>
>>Does this change the default constructors from non-trivial to trivial?
>It would be a major compiler bug if making a constructor default was 
>making it trivial.

I must be misunderstanding you, because this is not a bug:

#include <type_traits>

struct A {
  A() { }
};

static_assert( !std::is_trivially_default_constructible<A>::value, "" );

struct B {
  B() = default;
};

static_assert( std::is_trivially_default_constructible<B>::value, "" );
Jonathan Wakely Oct. 9, 2016, 3:38 p.m. UTC | #4
On 09/10/16 16:14 +0100, Jonathan Wakely wrote:
>On 08/10/16 22:55 +0200, François Dumont wrote:
>>On 06/10/2016 23:34, Jonathan Wakely wrote:
>>>On 06/10/16 22:17 +0200, François Dumont wrote:
>>>>Another approach is to rely on existing compiler ability to 
>>>>compute conditional noexcept when defaulting implementations. 
>>>>This is what I have done in this patch.
>>>>
>>>>The new default constructor on _Rb_tree_node_base is not a 
>>>>problem as it is not used to build _Rb_tree_node.
>>>
>>>Why not?
>>
>>_Rb_tree_node_base is used in 2 context. As member of _Rb_tree_impl 
>>in which case we need the new default constructor. And also as base 
>>class of _Rb_tree_node which is never constructed. Nodes are being 
>>allocated and then associated value is being constructed through the 
>>allocator, the node default constructor itself is never invoked.
>
>In C++03 mode that is true, but it's only valid because the type is
>trivially-constructible. If the type requires "non-vacuous
>initialization" then it's not valid to allocate memory for it and
>start using it without invoking a constructor. If you add a
>non-trivial constructor then we can't do that any more.

In fact that code is highly questionable, because the element member
makes the node require non-trivial initialization. We rely on the base
being trivially-constructible, and the element being constructed by
the allocator, and assume that it's OK to then treat that memory as a
node. In fact only its base class and member have been constructed,
not the node itself. The C++11 version (using an aligned buffer) is
correct. But I'd prefer not to make the C++98 version worse.
diff mbox

Patch

diff --git a/libstdc++-v3/include/bits/stl_map.h b/libstdc++-v3/include/bits/stl_map.h
index e5b2a1b..dea7d5b 100644
--- a/libstdc++-v3/include/bits/stl_map.h
+++ b/libstdc++-v3/include/bits/stl_map.h
@@ -167,11 +167,11 @@  _GLIBCXX_BEGIN_NAMESPACE_CONTAINER
       /**
        *  @brief  Default constructor creates no elements.
        */
-      map()
-      _GLIBCXX_NOEXCEPT_IF(
-	  is_nothrow_default_constructible<allocator_type>::value
-	  && is_nothrow_default_constructible<key_compare>::value)
-      : _M_t() { }
+#if __cplusplus < 201103L
+      map() : _M_t() { }
+#else
+      map() = default;
+#endif
 
       /**
        *  @brief  Creates a %map with no elements.
diff --git a/libstdc++-v3/include/bits/stl_multimap.h b/libstdc++-v3/include/bits/stl_multimap.h
index d240427..7e86b76 100644
--- a/libstdc++-v3/include/bits/stl_multimap.h
+++ b/libstdc++-v3/include/bits/stl_multimap.h
@@ -164,11 +164,11 @@  _GLIBCXX_BEGIN_NAMESPACE_CONTAINER
       /**
        *  @brief  Default constructor creates no elements.
        */
-      multimap()
-      _GLIBCXX_NOEXCEPT_IF(
-	  is_nothrow_default_constructible<allocator_type>::value
-	  && is_nothrow_default_constructible<key_compare>::value)
-      : _M_t() { }
+#if __cplusplus < 201103L
+      multimap() : _M_t() { }
+#else
+      multimap() = default;
+#endif
 
       /**
        *  @brief  Creates a %multimap with no elements.
diff --git a/libstdc++-v3/include/bits/stl_multiset.h b/libstdc++-v3/include/bits/stl_multiset.h
index cc068a9..7fe2fbd 100644
--- a/libstdc++-v3/include/bits/stl_multiset.h
+++ b/libstdc++-v3/include/bits/stl_multiset.h
@@ -144,11 +144,11 @@  _GLIBCXX_BEGIN_NAMESPACE_CONTAINER
       /**
        *  @brief  Default constructor creates no elements.
        */
-      multiset()
-      _GLIBCXX_NOEXCEPT_IF(
-	  is_nothrow_default_constructible<allocator_type>::value
-	  && is_nothrow_default_constructible<key_compare>::value)
-      : _M_t() { }
+#if __cplusplus < 201103L
+      multiset() : _M_t() { }
+#else
+      multiset() = default;
+#endif
 
       /**
        *  @brief  Creates a %multiset with no elements.
diff --git a/libstdc++-v3/include/bits/stl_set.h b/libstdc++-v3/include/bits/stl_set.h
index 3938351..5ed9672 100644
--- a/libstdc++-v3/include/bits/stl_set.h
+++ b/libstdc++-v3/include/bits/stl_set.h
@@ -147,11 +147,11 @@  _GLIBCXX_BEGIN_NAMESPACE_CONTAINER
       /**
        *  @brief  Default constructor creates no elements.
        */
-      set()
-      _GLIBCXX_NOEXCEPT_IF(
-	  is_nothrow_default_constructible<allocator_type>::value
-	  && is_nothrow_default_constructible<key_compare>::value)
-      : _M_t() { }
+#if __cplusplus < 201103L
+      set() : _M_t() { }
+#else
+      set() = default;
+#endif
 
       /**
        *  @brief  Creates a %set with no elements.
diff --git a/libstdc++-v3/include/bits/stl_tree.h b/libstdc++-v3/include/bits/stl_tree.h
index ee2dc70..ed575d0 100644
--- a/libstdc++-v3/include/bits/stl_tree.h
+++ b/libstdc++-v3/include/bits/stl_tree.h
@@ -108,6 +108,10 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
     _Base_ptr		_M_left;
     _Base_ptr		_M_right;
 
+    _Rb_tree_node_base() _GLIBCXX_NOEXCEPT
+      : _M_color(_S_red), _M_parent(0), _M_left(this), _M_right(this)
+    { }
+
     static _Base_ptr
     _S_minimum(_Base_ptr __x) _GLIBCXX_NOEXCEPT
     {
@@ -603,23 +607,33 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
         {
 	  _Key_compare		_M_key_compare;
 	  _Rb_tree_node_base 	_M_header;
+#if __cplusplus < 201103L
 	  size_type 		_M_node_count; // Keeps track of size of tree.
+#else
+	  size_type 		_M_node_count = 0; // Keeps track of size of tree.
+#endif
 
+#if __cplusplus < 201103L
 	  _Rb_tree_impl()
 	  : _Node_allocator(), _M_key_compare(), _M_header(),
 	    _M_node_count(0)
-	  { _M_initialize(); }
+	  { }
+#else
+	  _Rb_tree_impl() = default;
+#endif
 
 	  _Rb_tree_impl(const _Key_compare& __comp, const _Node_allocator& __a)
-	  : _Node_allocator(__a), _M_key_compare(__comp), _M_header(),
-	    _M_node_count(0)
-	  { _M_initialize(); }
+	  : _Node_allocator(__a), _M_key_compare(__comp), _M_header()
+#if __cplusplus < 201103L
+	  , _M_node_count(0)
+#endif
+	  { }
 
 #if __cplusplus >= 201103L
 	  _Rb_tree_impl(const _Key_compare& __comp, _Node_allocator&& __a)
-	  : _Node_allocator(std::move(__a)), _M_key_compare(__comp),
-	    _M_header(), _M_node_count(0)
-	  { _M_initialize(); }
+	    : _Node_allocator(std::move(__a)), _M_key_compare(__comp),
+	      _M_header()
+	  { }
 #endif
 
 	  void
@@ -630,16 +644,6 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 	    this->_M_header._M_right = &this->_M_header;
 	    this->_M_node_count = 0;
 	  }
-
-	private:
-	  void
-	  _M_initialize()
-	  {
-	    this->_M_header._M_color = _S_red;
-	    this->_M_header._M_parent = 0;
-	    this->_M_header._M_left = &this->_M_header;
-	    this->_M_header._M_right = &this->_M_header;
-	  }	    
 	};
 
       _Rb_tree_impl<_Compare> _M_impl;
@@ -831,7 +835,11 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 
     public:
       // allocation/deallocation
+#if __cplusplus < 201103L
       _Rb_tree() { }
+#else
+      _Rb_tree() = default;
+#endif
 
       _Rb_tree(const _Compare& __comp,
 	       const allocator_type& __a = allocator_type())