hydra:~/c/proba$ cat templ_overload.cc namespace n { template <typename T> void f(T a) { a.f(); } template <typename T> void g(T a) { n::f(a); } } struct s {}; namespace n { void f(s) {} } int main() { n::g(s()); return 0; } hydra:~/c/proba$ g++-cvs templ_overload.cc templ_overload.cc: In function `void n::f(T) [with T = s]': templ_overload.cc:14: instantiated from `void n::g(T) [with T = s]' templ_overload.cc:31: instantiated from here templ_overload.cc:7: error: 'struct s' has no member named 'f' zsh: exit 1 g++-cvs templ_overload.cc hydra:~/c/proba$ g++-cvs -v Reading specs from /gml/shared/gcc-cvs/lib/gcc/i686-pc-linux-gnu/3.4/specs Configured with: ../gcc/configure --prefix=/gml/shared/gcc-cvs --enable-languages=c++ Thread model: posix gcc version 3.4 20030806 (experimental) hydra:~/c/proba$ g++3 templ_overload.cc hydra:~/c/proba$ g++3 -v Reading specs from /gml/shared/gcc-3.3/lib/gcc-lib/i686-pc-linux-gnu/3.3/specs Configured with: ../gcc-3.3/configure --prefix=/gml/shared/gcc-3.3 Thread model: posix gcc version 3.3 hydra:~/c/proba$ workarounds: 1. unqualify f in the definition of g. I cannot do this, because in real life n=std, f=iter_swap, g=reverse. 2. Move the definition of the overload for f before the definition of g.
This is the same problem as PR 11808. *** This bug has been marked as a duplicate of 11808 ***
Looking at PR 11808, I think it's different, since this problem vanishes if at the point of the call all overloads are seen. What's more, namespaces are not needed at all to reproduce this bug: hydra:~/c/proba$ cat templ_overload.cc template <typename T> void f(T a) {a.f();} template <typename T> void g(T a) {::f(a);} struct s {}; void f(s) {} int main() { g(s()); return 0; } hydra:~/c/proba$ g++-cvs templ_overload.cc templ_overload.cc: In function `void f(T) [with T = s]': templ_overload.cc:3: instantiated from `void g(T) [with T = s]' templ_overload.cc:11: instantiated from here templ_overload.cc:1: error: 'struct s' has no member named 'f' zsh: exit 1 g++-cvs templ_overload.cc hydra:~/c/proba$ But let's see what happens.
I still think it's somehow related, after all the global namespace is also a namespace. But be that as it may, let's reopen this report and leave a mark in the audit trail of the other PR. W.
Re-confirmed with the neat small new testcase provided. W.
Well, yes, the global namespace is a namespace, but now everything is in the same namespace. I think I've found the ultimate testcase (the filename became non-descriptive): hydra:~/c/proba$ cat templ_overload.cc template <typename T> void g(T a) {::f(a);} hydra:~/c/proba$ g++-cvs templ_overload.cc templ_overload.cc: In function `void g(T)': templ_overload.cc:1: error: `::f' has not been declared zsh: exit 1 g++-cvs templ_overload.cc hydra:~/c/proba$ (And let me add that workaround 2, though works, forces a coding style where everything is included as late as possible, that is, #include's are not grouped at the beginning of a file but scattered throughout.)
The regression in PR 11828 was introduced or exposed by the new C++ parser, added to mainline on 2003-12-28. This was verified on i686-pc-linux-gnu with cc1plus built just before and just after that merge, using both the small test case from comment #2 and the submitter's test case.
The G++ behavior is correct. Only those declarations of n::f that were available when g is defined are available for overload resolution. See [temp.nondep] in the standard for details.
*** Bug 15144 has been marked as a duplicate of this bug. ***
[temp.nondep] is only relevant if you assume that n::f is a non-dependent name, which I am not willing to assume. I believe it is dependent under the principle outlined in [temp.dep]: In an expression of the form: postfix-expression ( expression-listopt ) where the postfix-expression is an identifier, the identifier denotes a dependent name if and only if any of the expressions in the expression-list is a type-dependent expression (_temp.dep.expr_). In this testcase, the postfix-expression is not a plain identifier, but I believe that it should be handled the same way, and that this is an oversight in the standard. EDG seem to agree with me, as 3.4 accepts this testcase and those from the various duplicates of this bug.
Created attachment 6254 [details] proposed patch Here's a quick sketch of a patch for this bug, as yet untested. I'd like someone more familiar with the parser to clean it up and apply it if we can agree that this is indeed a bug.
From [temp.dep.candidate]: For a function call that depends on a template parameter, if the function name is an "unqualified-id" but not a template-id, the candidate functions are found using the usual lookup rules (3.4.1, 3.4.2) except that: I think this is more relevent than any other part. I think the unqualified-id part tells that the function name has to be unqualified.
Note ICC 6.0 in strict mode (-Xc -ansi) rejects the testcase in PR15367 with: pr11828.cc pr11828.cc(6): error: no instance of overloaded function "f" matches the argument list argument types are: (A) ::f (t); ^ detected during instantiation of "void g(T) [with T=A]" compilation aborted for pr11828.cc (code 2)
As does icc7 when given -Xc -ansi. It accepts the testcase in default mode, though. Jason, when you say you tested it with EDG, did you specify the flags for maximal standard conformance? W.
Subject: Re: [3.4/3.5 regression] qualified dependent name looked up too early On 10 May 2004 20:16:59 -0000, "bangerth at dealii dot org" <gcc-bugzilla@gcc.gnu.org> wrote: > As does icc7 when given -Xc -ansi. It accepts the testcase in default > mode, though. Jason, when you say you tested it with EDG, did you specify > the flags for maximal standard conformance? I used -A, which I believe is the relevant flag. Jason
Hm, on my linux box and with icc, -A is a preprocessor flag... W.
*** Bug 15814 has been marked as a duplicate of this bug. ***
I now agree that this is the intent of the committee.
*** Bug 18195 has been marked as a duplicate of this bug. ***
I question the apparent conclusion reached in this discussion. The reporter's testcase refers to n::f(a) where 'a' has dependent type. The conclusion so far is that 'f' is not dependent because it is qualified. The primary justification offered is that 14.6.2p1 mentions exactly two syntactic cases of dependent names: - function call with a simple 'identifier' as the function - operator in an expression However, as discussed here and elsewhere (e.g., http://gcc.gnu.org/onlinedocs/gcc/Name-lookup.html), there is general consensus that the following forms also make 'f' dependent: - x->f(a) // when x is a type-dependent expression - C::f(a) // when C is a dependent type Therefore, my interpretation is that the two cases given in 14.6.2 are *not* exhaustive, but rather merely examples. 14.6.2 first gives some general properties of dependent names (e.g., "... semantics which may differ from one instantiation to another.") that clearly apply to more situations. And 14.6.4.2 singles out unqualified names as a special case---why would they be a special case if only unqualified names could be dependent? Now, it is true that the two additional generally-agreed-upon cases above are "doubly" dependent: not only do they have arguments of dependent type, but their receiver object or qualifier is also of dependent type. Consequently, the expressions are type-dependent (14.6.2.2 p1, p3 resp.); but the standard never says that being part of a type-dependent expression automatically makes the name dependent! On the contrary, 14.6.2.2p4 gives several counterexamples. Moreover, in the simple case "f(a)", 'f' is not a type-dependent expression, and yet it is clearly a dependent name. Extrapolating from the first few sentences of 14.6.2, it seems to me that the following are exhaustive representatives of cases where the name 'foo' is dependent: A. Cases where function overload resolution is performed, and influenced by some type that might be dependent: A1. foo(a,b,c) A2. C::foo(a,b,c) A3. x->foo(a,b,c) A4. x->C::foo(a,b,c) when a or b or c is a type-dependent expression. B. Cases where lookup involves traversing a class with dependent type: B1. C::foo B2. x->foo B3. x->C::foo when C is a dependent type, or x is a type-dependent expression. C. Cases when the address of an overloaded function is matched to a target type (13.4) that is dependent. This would include all variations of C. [&][x->][C::]foo where by brackets ([]) I mean optional component. In all cases (except case C, I *think*) 'foo' could be a template-id, and would still be dependent. (Plus the case of operators in expressions where an operand is a type-dependent expr.)
*** Bug 20307 has been marked as a duplicate of this bug. ***
I think I have answered my own question: indeed, qualified lookup only considers name from the definition context, and not the instantiation context. I found this thread at google groups: http://groups-beta.google.com/group/comp.std.c++/browse_thread/thread/22521f9b57b995d0/73d79aadca8c665c Quoting from a few of the more useful messages: http://groups-beta.google.com/group/comp.std.c++/msg/ae689cfc9c99262d Moving to templates. Point of definition does both normal and argument dependent lookup, point of instantiation does ADL only (14.6.4/1.) When you disable ADL by writing N::foo(x), the only lookup that remains is normal lookup done at point of definition. Point of instantiation lookup is implicitly disabled as a consequence of the interaction between 3.4.2/1 and 14.6.4/1. http://groups-beta.google.com/group/comp.std.c++/msg/0e252a95da0f453d The OL part happens at the point of definition (phase 1), and the ADL part happens at the POI (point of instantiation; phase 2). If you remove the ADL part, you're left with a phase 1 lookup only. http://groups-beta.google.com/group/comp.std.c++/msg/6a53b35efe39fee3 The compromise that was reached was to say that lexical lookup would occur only in the definition context; only ADL would be performed in the instantiation context. Since qualified function names do not participate in ADL, any arguments in a call to a qualified name are ignored in determining whether a qualified name is dependent or not. The upshot is that, by 14.6.4p1, second-phase lookup *only* does argument-dependent lookup, which is itself disabled by qualification. This has some nasty consequences, e.g. http://groups-beta.google.com/group/comp.std.c++/msg/8379998d36b27a99 but appears to be the genuine intent of the committee.
*** Bug 26148 has been marked as a duplicate of this bug. ***