Re: [PATCH] Fix constexpr evaluation of comparisons involving pointer-to-members

[Patrick Palka <patrick at parcs dot ath dot cx>]

On Thu, Feb 4, 2016 at 12:24 PM, Patrick Palka <patrick@parcs.ath.cx> wrote:
> On Thu, Feb 4, 2016 at 11:57 AM, Jason Merrill <jason@redhat.com> wrote:
>> On 02/04/2016 10:32 AM, Patrick Palka wrote:
>>>
>>> On Thu, Feb 4, 2016 at 9:24 AM, Jason Merrill <jason@redhat.com> wrote:
>>>>
>>>> On 02/03/2016 12:51 PM, Patrick Palka wrote:
>>>>>
>>>>>
>>>>> +           && (integer_minus_onep (lhs)
>>>>> +           || integer_minus_onep (rhs)))
>>>>
>>>>
>>>>
>>>> Let's use null_member_pointer_value_p here.
>>>
>>>
>>> Done.
>>>
>>>>
>>>> Please add pointers to member functions to the testcase.
>>>
>>>
>>> This does not currently work properly because comparisons between
>>> distinct pointers to (non-virtual) member functions eventually perform
>>> a comparison between two distinct FUNCTION_DECLs, and fold_binary_loc
>>> currently does not fold away such comparisons. So any static_asserts
>>> that perform comparisons between distinct function pointers or between
>>> distinct pointers to (non-virtual) member functions fail with
>>> "non-constant condition for static assertion".  (Comparisons between
>>> pointers to distinct virtual member functions _do_ work because in
>>> that case we are ultimately just comparing integer offsets, not
>>> FUNCTION_DECLs.  And comparisons between equivalent pointers trivially
>>> work.)
>>>
>>> I think the problem may lie in symtab_node::equal_address_to, which
>>> apparently returns -1, instead of 0, for two obviously distinct
>>> FUNCTION_DECLs.
>>
>>
>> Right, because they might be defined as aliases elsewhere.  But it looks
>> like it should work if you define f and g.
>
> That makes sense.  But even when I define f and g, the comparisons
> don't get folded away:
>
> #define SA(x) static_assert ((x), #x)
>
> void f () { }
> void g () { }
>
> struct X { int a, b; void f (); void g (); };
>
> void X::f () { }
> void X::g () { }
>
> void
> foo ()
> {
>   SA (&f != &g);
>   SA (&X::f != &X::g);
> }
>
> bool bar = &f != &g;
>
> Although the static_asserts fail, the initialization of the variable
> bar _does_ get folded to 1 during gimplification.  When evaluating the
> static_asserts, symtab_node::equal_address_to again returns -1 but
> during gimplification the same call to symtab_node::equal_address_to
> returns 0.  So probably the symbols are not fully defined yet in the
> symbol table at the point when we are evaluating the static_asserts.

Indeed, symtab_node::equal_address_to returns 0 only if both symbols
have their ->analyzed flag set.  But cgraph_node::analyze() (which
sets the analyzed flag) is called during finalization of the
compilation unit -- obviously not in time for the static_asserts to
get folded.