diff mbox series

Avoid allocator operator== in _Safe_container

Message ID 6edc763b-e463-a676-0232-7b2193fb403b@gmail.com
State New
Headers show
Series Avoid allocator operator== in _Safe_container | expand

Commit Message

François Dumont Aug. 9, 2021, 10:23 a.m. UTC
Some newly introduced tests in 
23_containers/unordered_map/cons/default.cc revealed that we are forcing 
the allocator type to have a operator==.

It can be avoided by checking allocator_traits::is_always_equal.

     libstdc++: [_GLIBCXX_DEBUG] Avoid allocator operator== when always 
equal

     Use std::allocator_traits::is_always_equal to find out if we need 
to compare
     allocator instances on safe container allocator aware move constructor.

     libstdc++-v3/ChangeLog:

             * include/debug/safe_container.h
             (_Safe_container(_Safe_container&&, const _Alloc&, 
std::true_type)): New.
             (_Safe_container(_Safe_container&&, const _Alloc&, 
std::false_type)): New.
             (_Safe_container(_Safe_container&&, const _Alloc&)): Use 
latters.

Tested under Linux x86 Debug mode.

Ok to commit ?

François

Comments

Jonathan Wakely Aug. 9, 2021, 10:33 a.m. UTC | #1
On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
<libstdc++@gcc.gnu.org> wrote:
>
> Some newly introduced tests in
> 23_containers/unordered_map/cons/default.cc revealed that we are forcing
> the allocator type to have a operator==.

All allocators are required to have operator== so that should not be a
problem. What is the error?
Jonathan Wakely Aug. 9, 2021, 10:34 a.m. UTC | #2
On Mon, 9 Aug 2021 at 11:33, Jonathan Wakely wrote:
>
> On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
> <libstdc++@gcc.gnu.org> wrote:
> >
> > Some newly introduced tests in
> > 23_containers/unordered_map/cons/default.cc revealed that we are forcing
> > the allocator type to have a operator==.
>
> All allocators are required to have operator== so that should not be a
> problem. What is the error?

OK, I see it. I just forgot to define operator== and operator!= for
the custom allocator in that new test, and that should be added.
Jonathan Wakely Aug. 9, 2021, 10:45 a.m. UTC | #3
On Mon, 9 Aug 2021 at 11:34, Jonathan Wakely wrote:
>
> On Mon, 9 Aug 2021 at 11:33, Jonathan Wakely wrote:
> >
> > On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
> > <libstdc++@gcc.gnu.org> wrote:
> > >
> > > Some newly introduced tests in
> > > 23_containers/unordered_map/cons/default.cc revealed that we are forcing
> > > the allocator type to have a operator==.
> >
> > All allocators are required to have operator== so that should not be a
> > problem. What is the error?
>
> OK, I see it. I just forgot to define operator== and operator!= for
> the custom allocator in that new test, and that should be added.

Fixed like this instead. Tested x86_64-linux with -D_GLIBCXX_DEBUG.
Pushed to trunk.
commit 2eff2a3cb521c58212885a3dca638764285b5691
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Mon Aug 9 11:36:07 2021

    libstdc++: Make allocator equality comparable in tests
    
    libstdc++-v3/ChangeLog:
    
            * testsuite/23_containers/unordered_map/cons/default.cc: Add
            equality comparison operators to allocator.
            * testsuite/23_containers/unordered_set/cons/default.cc:
            Likewise.

diff --git a/libstdc++-v3/testsuite/23_containers/unordered_map/cons/default.cc b/libstdc++-v3/testsuite/23_containers/unordered_map/cons/default.cc
index d64d078a7da..7a785e980b1 100644
--- a/libstdc++-v3/testsuite/23_containers/unordered_map/cons/default.cc
+++ b/libstdc++-v3/testsuite/23_containers/unordered_map/cons/default.cc
@@ -18,6 +18,9 @@ template<typename T>
 
     void deallocate(T *p, std::size_t n)
     { std::allocator<T>().deallocate(p, n); }
+
+    bool operator==(const NoDefaultConsAlloc&) const { return true; }
+    bool operator!=(const NoDefaultConsAlloc&) const { return false; }
   };
 
 using Map = std::unordered_map<int, int, std::hash<int>, std::equal_to<int>,
diff --git a/libstdc++-v3/testsuite/23_containers/unordered_set/cons/default.cc b/libstdc++-v3/testsuite/23_containers/unordered_set/cons/default.cc
index 41281d3d774..fb87c96ce9d 100644
--- a/libstdc++-v3/testsuite/23_containers/unordered_set/cons/default.cc
+++ b/libstdc++-v3/testsuite/23_containers/unordered_set/cons/default.cc
@@ -18,6 +18,9 @@ template<typename T>
 
     void deallocate(T *p, std::size_t n)
     { std::allocator<T>().deallocate(p, n); }
+
+    bool operator==(const NoDefaultConsAlloc&) const { return true; }
+    bool operator!=(const NoDefaultConsAlloc&) const { return false; }
   };
 
 using Set = std::unordered_set<int, std::hash<int>, std::equal_to<int>,
François Dumont Aug. 9, 2021, 11:52 a.m. UTC | #4
On 09/08/21 12:45 pm, Jonathan Wakely wrote:
> On Mon, 9 Aug 2021 at 11:34, Jonathan Wakely wrote:
>> On Mon, 9 Aug 2021 at 11:33, Jonathan Wakely wrote:
>>> On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
>>> <libstdc++@gcc.gnu.org> wrote:
>>>> Some newly introduced tests in
>>>> 23_containers/unordered_map/cons/default.cc revealed that we are forcing
>>>> the allocator type to have a operator==.
>>> All allocators are required to have operator== so that should not be a
>>> problem. What is the error?
>> OK, I see it. I just forgot to define operator== and operator!= for
>> the custom allocator in that new test, and that should be added.
> Fixed like this instead. Tested x86_64-linux with -D_GLIBCXX_DEBUG.
> Pushed to trunk.

Ok, I thought my change was better because we have many allocator types 
in tests without operator ==/!= (see 
23_containers/*/cons/noexcept_default_construct.cc). But of course the 
tests are not making any use of it for the moment, no big deal.

So this patch is just an optimization, may I still commit it ? Unless 
you like the fact that Debug mode is checking that those operators are 
provided when allocator-aware move constructor is being used.
François Dumont Aug. 9, 2021, 12:08 p.m. UTC | #5
On 09/08/21 1:52 pm, François Dumont wrote:
> On 09/08/21 12:45 pm, Jonathan Wakely wrote:
>> On Mon, 9 Aug 2021 at 11:34, Jonathan Wakely wrote:
>>> On Mon, 9 Aug 2021 at 11:33, Jonathan Wakely wrote:
>>>> On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
>>>> <libstdc++@gcc.gnu.org> wrote:
>>>>> Some newly introduced tests in
>>>>> 23_containers/unordered_map/cons/default.cc revealed that we are 
>>>>> forcing
>>>>> the allocator type to have a operator==.
>>>> All allocators are required to have operator== so that should not be a
>>>> problem. What is the error?
>>> OK, I see it. I just forgot to define operator== and operator!= for
>>> the custom allocator in that new test, and that should be added.
>> Fixed like this instead. Tested x86_64-linux with -D_GLIBCXX_DEBUG.
>> Pushed to trunk.
>
> Ok, I thought my change was better because we have many allocator 
> types in tests without operator ==/!= (see 
> 23_containers/*/cons/noexcept_default_construct.cc). But of course the 
> tests are not making any use of it for the moment, no big deal.
>
> So this patch is just an optimization, may I still commit it ? Unless 
> you like the fact that Debug mode is checking that those operators are 
> provided when allocator-aware move constructor is being used.
>
>
A simpler version, using _Safe_container move constructor when allocator 
instances are always equals.
Jonathan Wakely Aug. 9, 2021, 2:24 p.m. UTC | #6
On Mon, 9 Aug 2021 at 13:08, François Dumont <frs.dumont@gmail.com> wrote:
>
> On 09/08/21 1:52 pm, François Dumont wrote:
> > On 09/08/21 12:45 pm, Jonathan Wakely wrote:
> >> On Mon, 9 Aug 2021 at 11:34, Jonathan Wakely wrote:
> >>> On Mon, 9 Aug 2021 at 11:33, Jonathan Wakely wrote:
> >>>> On Mon, 9 Aug 2021 at 11:26, François Dumont via Libstdc++
> >>>> <libstdc++@gcc.gnu.org> wrote:
> >>>>> Some newly introduced tests in
> >>>>> 23_containers/unordered_map/cons/default.cc revealed that we are
> >>>>> forcing
> >>>>> the allocator type to have a operator==.
> >>>> All allocators are required to have operator== so that should not be a
> >>>> problem. What is the error?
> >>> OK, I see it. I just forgot to define operator== and operator!= for
> >>> the custom allocator in that new test, and that should be added.
> >> Fixed like this instead. Tested x86_64-linux with -D_GLIBCXX_DEBUG.
> >> Pushed to trunk.
> >
> > Ok, I thought my change was better because we have many allocator
> > types in tests without operator ==/!= (see
> > 23_containers/*/cons/noexcept_default_construct.cc). But of course the
> > tests are not making any use of it for the moment, no big deal.

Those tests are all OK, the allocators inherit from std::allocator so
can use its equality operators.

But if I did mess up any more test allocators then we should fix those ones too.

> > So this patch is just an optimization, may I still commit it ? Unless
> > you like the fact that Debug mode is checking that those operators are
> > provided when allocator-aware move constructor is being used.
> >
> >
> A simpler version, using _Safe_container move constructor when allocator
> instances are always equals.

Yes, this optimization is good. OK for trunk, thanks!
diff mbox series

Patch

diff --git a/libstdc++-v3/include/debug/safe_container.h b/libstdc++-v3/include/debug/safe_container.h
index d9636c29e9b..b06c33dac7e 100644
--- a/libstdc++-v3/include/debug/safe_container.h
+++ b/libstdc++-v3/include/debug/safe_container.h
@@ -57,7 +57,12 @@  namespace __gnu_debug
       _Safe_container(const _Safe_container&) = default;
       _Safe_container(_Safe_container&&) = default;
 
-      _Safe_container(_Safe_container&& __x, const _Alloc& __a)
+    private:
+      _Safe_container(_Safe_container&& __x, const _Alloc& __a, std::true_type)
+      : _Safe_container()
+      { _Base::_M_swap(__x); }
+
+      _Safe_container(_Safe_container&& __x, const _Alloc& __a, std::false_type)
       : _Safe_container()
       {
 	if (__x._M_cont().get_allocator() == __a)
@@ -65,6 +70,12 @@  namespace __gnu_debug
 	else
 	  __x._M_invalidate_all();
       }
+
+    protected:
+      _Safe_container(_Safe_container&& __x, const _Alloc& __a)
+      : _Safe_container(std::move(__x), __a,
+		      typename std::allocator_traits<_Alloc>::is_always_equal{})
+      { }
 #endif
 
     public: