Re: [PATCH] c++: equivalence of non-dependent calls [PR107461]

[Jason Merrill <jason@redhat.com>]

On 2/4/23 20:41, Jason Merrill wrote:
> On 2/4/23 20:08, Patrick Palka wrote:
>> On Sat, 4 Feb 2023, Jason Merrill wrote:
>>
>>> On 2/4/23 15:31, Patrick Palka wrote:
>>>> After r13-5684-g59e0376f607805 the (pruned) callee of a non-dependent
>>>> CALL_EXPR is a bare FUNCTION_DECL rather than ADDR_EXPR of 
>>>> FUNCTION_DECL.
>>>> This innocent change revealed that cp_tree_equal doesn't first check
>>>> dependentness of a CALL_EXPR before treating the callee as a dependent
>>>> name, which manifests as us incorrectly accepting the first two
>>>> testcases below and rejecting the third:
>>>>
>>>>    * In the first testcase, cp_tree_equal incorrectly returns true for
>>>>      the two non-dependent CALL_EXPRs f(0) and f(0) (whose CALL_EXPR_FN
>>>>      are different FUNCTION_DECLs) and so we treat #2 as a 
>>>> redeclaration
>>>>      of #1.
>>>>
>>>>    * Same issue in the second testcase, for f<int*>() and f<char>().
>>>>
>>>>    * In the third testcase, cp_tree_equal incorrectly returns true for
>>>>      f<int>() and f<void(*)(int)>() which causes us to conflate the two
>>>>      dependent specializations A<decltype(f<int>()(U()))> and
>>>>      A<decltype(f<void(*)(int)>()(U()))>, leading to a bogus error.
>>>>
>>>> This patch fixes this by making called_fns_equal treat two callees as
>>>> dependent names only if the CALL_EXPRs in question are dependent.
>>>>
>>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK 
>>>> for
>>>> trunk/12?  Patch generated with -w to ignore noisy whitespace changes.
>>>>
>>>>     PR c++/107461
>>>>
>>>> gcc/cp/ChangeLog:
>>>>
>>>>     * pt.cc (iterative_hash_template_arg) <case CALL_EXPR>: Treat
>>>>     the callee as a dependent name only if the CALL_EXPR is
>>>>     dependent.
>>>>     * tree.cc (called_fns_equal): Take two CALL_EXPRs instead of
>>>>     CALL_EXPR_FNs thereof.  As above.
>>>>     (cp_tree_equal) <case CALL_EXPR>: Adjust call to called_fns_equal.
>>>>
>>>> gcc/testsuite/ChangeLog:
>>>>
>>>>     * g++.dg/cpp0x/overload5.C: New test.
>>>>     * g++.dg/cpp0x/overload5a.C: New test.
>>>>     * g++.dg/cpp0x/overload6.C: New test.
>>>> ---
>>>>    gcc/cp/pt.cc                            |  1 +
>>>>    gcc/cp/tree.cc                          | 33 
>>>> ++++++++++++++-----------
>>>>    gcc/testsuite/g++.dg/cpp0x/overload5.C  | 12 +++++++++
>>>>    gcc/testsuite/g++.dg/cpp0x/overload5a.C | 10 ++++++++
>>>>    gcc/testsuite/g++.dg/cpp0x/overload6.C  | 16 ++++++++++++
>>>>    5 files changed, 58 insertions(+), 14 deletions(-)
>>>>    create mode 100644 gcc/testsuite/g++.dg/cpp0x/overload5.C
>>>>    create mode 100644 gcc/testsuite/g++.dg/cpp0x/overload5a.C
>>>>    create mode 100644 gcc/testsuite/g++.dg/cpp0x/overload6.C
>>>>
>>>> diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
>>>> index 255332dc0c1..c9360240cd2 100644
>>>> --- a/gcc/cp/pt.cc
>>>> +++ b/gcc/cp/pt.cc
>>>> @@ -1841,6 +1841,7 @@ iterative_hash_template_arg (tree arg, 
>>>> hashval_t val)
>>>>        case CALL_EXPR:
>>>>          {
>>>>        tree fn = CALL_EXPR_FN (arg);
>>>> +    if (TREE_TYPE (arg) == NULL_TREE)
>>>
>>> How about changing dependent_name to take the CALL_EXPR rather than the
>>> CALL_EXPR_FN?  That would mean some changes to write_expression to 
>>> move the
>>> dependent_name handling into the CALL_EXPR handling, but that doesn't 
>>> seem
>>> like a bad thing.  Other callers seem like a trivial change.
>>
>> Indeed changing dependent_name seems best, but I'm worried about such a
>> refactoring to write_expression causing unintended mangling changes at
>> this stage.  Because it seems the CALL_EXPR case of write_expression
>> isn't the user of the dependent_name branch of write_expression, at
>> least according to the following patch which causes us to ICE on
>> mangle{37,57,58,76}.C:
> 
> Yeah, I tried the same thing.  Maybe for GCC 13 better to add a new 
> function rather than change the current one.

mangle76 seems like a bug where we're producing (and testing for) the 
wrong mangling -- mangling (*this). that doesn't exist in the source. 
clang gets it right.

mangle5{7,8} has the right mangling, we're just using dependent_name to 
mangle function names that aren't dependent names (because they're 
template arguments in both cases, and qualified in the latter).

>> diff --git a/gcc/cp/mangle.cc b/gcc/cp/mangle.cc
>> index f2cda3be2cf..700857f8f3c 100644
>> --- a/gcc/cp/mangle.cc
>> +++ b/gcc/cp/mangle.cc
>> @@ -3450,6 +3450,7 @@ write_expression (tree expr)
>>       }
>>     else if (dependent_name (expr))
>>       {
>> +      gcc_unreachable ();
>>         tree name = dependent_name (expr);
>>         if (IDENTIFIER_ANY_OP_P (name))
>>       {
>> @@ -3554,7 +3555,19 @@ write_expression (tree expr)
>>           && type_dependent_expression_p_push (expr))
>>             fn = OVL_NAME (fn);
>> -        write_expression (fn);
>> +        if (tree name = dependent_name (fn))
>> +          {
>> +        if (IDENTIFIER_ANY_OP_P (name))
>> +          {
>> +            if (abi_version_at_least (16))
>> +              write_string ("on");
>> +            if (abi_warn_or_compat_version_crosses (16))
>> +              G.need_abi_warning = 1;
>> +          }
>> +        write_unqualified_id (name);
>> +          }
>> +        else
>> +          write_expression (fn);
>>         }
>>         for (i = 0; i < call_expr_nargs (expr); ++i)
>>
>> And since the CALL_EXPR case of write_expression looks through an
>> ADDR_EXPR callee before recursing, IIUC the refactoring would need to
>> make dependent_name look through an ADDR_EXPR callee as well, which
>> seems like a desirable/correct change but I'm worried that might have
>> unintended consequences as well.
>>
>>>
>>>>          if (tree name = dependent_name (fn))
>>>>            {
>>>>              if (TREE_CODE (fn) == TEMPLATE_ID_EXPR)
>>>> diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
>>>> index c1da868732b..3a57e71b76e 100644
>>>> --- a/gcc/cp/tree.cc
>>>> +++ b/gcc/cp/tree.cc
>>>> @@ -3870,16 +3870,21 @@ decl_internal_context_p (const_tree decl)
>>>>      return !TREE_PUBLIC (decl);
>>>>    }
>>>>    -/* Subroutine of cp_tree_equal: t1 and t2 are the CALL_EXPR_FNs 
>>>> of two
>>>> -   CALL_EXPRS.  Return whether they are equivalent.  */
>>>> +/* Subroutine of cp_tree_equal: t1 and t2 are two CALL_EXPRs.
>>>> +   Return whether their CALL_EXPR_FNs are equivalent.  */
>>>>      static bool
>>>>    called_fns_equal (tree t1, tree t2)
>>>> +{
>>>> +  tree fn1 = CALL_EXPR_FN (t1);
>>>> +  tree fn2 = CALL_EXPR_FN (t2);
>>>> +  if (TREE_TYPE (t1) == NULL_TREE
>>>> +      && TREE_TYPE (t2) == NULL_TREE)
>>>>        {
>>>>          /* Core 1321: dependent names are equivalent even if the 
>>>> overload
>>>> sets
>>>>         are different.  But do compare explicit template arguments.  */
>>>> -  tree name1 = dependent_name (t1);
>>>> -  tree name2 = dependent_name (t2);
>>>> +      tree name1 = dependent_name (fn1);
>>>> +      tree name2 = dependent_name (fn2);
>>>>          if (name1 || name2)
>>>>        {
>>>>          tree targs1 = NULL_TREE, targs2 = NULL_TREE;
>>>> @@ -3891,19 +3896,19 @@ called_fns_equal (tree t1, tree t2)
>>>>             of whether the function was named with a qualified- or
>>>> unqualified-id.
>>>>             Until that's fixed, check that we aren't looking at 
>>>> overload sets
>>>> from
>>>>             different scopes.  */
>>>> -      if (is_overloaded_fn (t1) && is_overloaded_fn (t2)
>>>> -      && (DECL_CONTEXT (get_first_fn (t1))
>>>> -          != DECL_CONTEXT (get_first_fn (t2))))
>>>> +      if (is_overloaded_fn (fn1) && is_overloaded_fn (fn2)
>>>> +          && (DECL_CONTEXT (get_first_fn (fn1))
>>>> +          != DECL_CONTEXT (get_first_fn (fn2))))
>>>>            return false;
>>>>    -      if (TREE_CODE (t1) == TEMPLATE_ID_EXPR)
>>>> -    targs1 = TREE_OPERAND (t1, 1);
>>>> -      if (TREE_CODE (t2) == TEMPLATE_ID_EXPR)
>>>> -    targs2 = TREE_OPERAND (t2, 1);
>>>> +      if (TREE_CODE (fn1) == TEMPLATE_ID_EXPR)
>>>> +        targs1 = TREE_OPERAND (fn1, 1);
>>>> +      if (TREE_CODE (fn2) == TEMPLATE_ID_EXPR)
>>>> +        targs2 = TREE_OPERAND (fn2, 1);
>>>>          return cp_tree_equal (targs1, targs2);
>>>>        }
>>>> -  else
>>>> -    return cp_tree_equal (t1, t2);
>>>> +    }
>>>> +  return cp_tree_equal (fn1, fn2);
>>>>    }
>>>>      bool comparing_override_contracts;
>>>> @@ -4037,7 +4042,7 @@ cp_tree_equal (tree t1, tree t2)
>>>>        if (KOENIG_LOOKUP_P (t1) != KOENIG_LOOKUP_P (t2))
>>>>          return false;
>>>>    -    if (!called_fns_equal (CALL_EXPR_FN (t1), CALL_EXPR_FN (t2)))
>>>> +    if (!called_fns_equal (t1, t2))
>>>>          return false;
>>>>          call_expr_arg_iterator iter1, iter2;
>>>> diff --git a/gcc/testsuite/g++.dg/cpp0x/overload5.C
>>>> b/gcc/testsuite/g++.dg/cpp0x/overload5.C
>>>> new file mode 100644
>>>> index 00000000000..e05b1594f51
>>>> --- /dev/null
>>>> +++ b/gcc/testsuite/g++.dg/cpp0x/overload5.C
>>>> @@ -0,0 +1,12 @@
>>>> +// PR c++/107461
>>>> +// { dg-do compile { target c++11 } }
>>>> +
>>>> +int f(...);
>>>> +template<class T> decltype(T() + f(0)) g(); // #1
>>>> +
>>>> +char f(int);
>>>> +template<class T> decltype(T() + f(0)) g(); // #2, distinct from #1
>>>> +
>>>> +int main() {
>>>> +  g<int>(); // { dg-error "ambiguous" }
>>>> +}
>>>> diff --git a/gcc/testsuite/g++.dg/cpp0x/overload5a.C
>>>> b/gcc/testsuite/g++.dg/cpp0x/overload5a.C
>>>> new file mode 100644
>>>> index 00000000000..037114f199c
>>>> --- /dev/null
>>>> +++ b/gcc/testsuite/g++.dg/cpp0x/overload5a.C
>>>> @@ -0,0 +1,10 @@
>>>> +// PR c++/107461
>>>> +// { dg-do compile { target c++11 } }
>>>> +
>>>> +template<class T> T f();
>>>> +template<class T> decltype(T() + f<int*>()) g(); // #1
>>>> +template<class T> decltype(T() + f<char>()) g(); // #2, distinct 
>>>> from #1
>>>> +
>>>> +int main() {
>>>> +  g<int>(); // { dg-error "ambiguous" }
>>>> +}
>>>> diff --git a/gcc/testsuite/g++.dg/cpp0x/overload6.C
>>>> b/gcc/testsuite/g++.dg/cpp0x/overload6.C
>>>> new file mode 100644
>>>> index 00000000000..1fbee0501de
>>>> --- /dev/null
>>>> +++ b/gcc/testsuite/g++.dg/cpp0x/overload6.C
>>>> @@ -0,0 +1,16 @@
>>>> +// PR c++/107461
>>>> +// { dg-do compile { target c++11 } }
>>>> +
>>>> +template<class T> T f();
>>>> +
>>>> +template<class> struct A { };
>>>> +
>>>> +template<class T> struct B {
>>>> +  template<class U, class = A<decltype(f<T>()(U()))>>
>>>> +  static void g(U);
>>>> +};
>>>> +
>>>> +int main() {
>>>> +  B<int> b;
>>>> +  B<void(*)(int)>::g(0); // { dg-bogus "no match" }
>>>> +}
>>>
>>>
>>
>