The current behavior of static_assert when inside a template definition is to detect and attempt to evaluate the assert during the parser stage, before any instantiations. (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52809) "A static_assert declaration that does not depend on template parameters will be detected and reported while parsing the template, before any instantiation. This is OK because such a template would have no valid instantiation, which immediately renders the program ill-formed." It is my opinion that this is not true. It is possible that the template might have valid specializations that occur later on in the compilation unit, making it impossible to determine whether or not it is possible to construct an instantiation of the template when parsing. The relevant part of the standard says this " Refer to 14.6p8 in the spec If no valid specialization can be generated for a template definition, and that template is not instantiated, the template definition is ill-formed, no diagnostic required. " This seems at first to imply that the current behavior is legal. However, the current behavior evaluates the static_assert and emits a message EVEN if valid specializations CAN still be generated for a template definition. Consider this code: template<bool a> struct s { static_assert(0, "uhoh"); }; template<> struct s<false> { }; s< false > q; This should compile, because s<false> is a valid specialization of s, meaning that s<a> does not get instantiated. However, on gcc 4.5.1 and gcc 4.7, we get prog.cpp:3:5: error: static assertion failed: "uhoh" Strangely, because of this behavior, we can 'fix' it by making the constant expression dependent on a, even if we know it will always be false template<bool a> struct s { static_assert(a!=a, "uhoh" ); }; template<> struct s<false> { }; Compiles. The current behavior of GCC in this case MAY be conformant with spec, but it seems to go against the intent of the spec even if it doesn't go against the letter. It also makes it difficult to use static_assert in certain metaprogramming applications(like the example above, which could be used to evaluate if a metafunction returned true). Also, as bug 52809 demonstrates, it is confusing to users.
(In reply to comment #0) > It is my opinion that this is not true. It is true. > It is possible that the template might > have valid specializations that occur later on in the compilation unit, making > it impossible to determine whether or not it is possible to construct an > instantiation of the template when parsing. No it isn't. You can generate valid instantiations from a *different* template definition, but the wording is in terms of template definitions, not templates. > The relevant part of the standard says this " Refer to 14.6p8 in the spec > > If no valid specialization can be generated for a template definition, and that > template is not instantiated, the template definition is ill-formed, no > diagnostic required. > " It specifically says "for a definition". The primary definition of a template is a definition, an explicit specialization provides a different definition. If a valid instantiation can be generated from the second definition that doesn't change the fact that if no valid instantiation can be generated for the primary definition then that primary definition is ill-formed. > This seems at first to imply that the current behavior is legal. However, the > current behavior evaluates the static_assert and emits a message EVEN if valid > specializations CAN still be generated for a template definition. > > Consider this code: > > template<bool a> > struct s { > static_assert(0, "uhoh"); > }; > > template<> > struct s<false> { > > }; > > s< false > q; > > This should compile, because s<false> is a valid specialization of s, meaning > that s<a> does not get instantiated. No, the primary template is a template definition but no valid specialization can be generated from it. The explicit specialization is a second definition, its existence doesn't change the fact the primary definition cannot ever be used (also the explicit specialization hasn't even been seen when the primary definition is first parsed.) > However, on gcc 4.5.1 and gcc 4.7, we get > > prog.cpp:3:5: error: static assertion failed: "uhoh" And also on clang++ and EDG. > Strangely, because of this behavior, we can 'fix' it by making the constant > expression dependent on a, even if we know it will always be false > > template<bool a> > struct s { > static_assert(a!=a, "uhoh" ); > }; That is a bit strange at first sight, but actually that doesn't fix anything, that definition is still ill-formed because no valid instantiation can be generated from it. It would be conforming for a compiler to reject that program, but it's not required to (it's "ill-formed, no diagnostic required.") The solution is to not write it like that. A more sensible definition would be: template<bool a> struct s { static_assert(!a, "uhoh"); }; Or even just template<bool a> struct s; (Though the static assertion does give the option of a nicer message.) > template<> > struct s<false> { > > }; > > Compiles. > > The current behavior of GCC in this case MAY be conformant with spec, but it It definitely is. Read 14.6p8 again. It doesn't matter if the template is never instantiated, it's ill-formed. > seems to go against the intent of the spec even if it doesn't go against the > letter. I don't see how it goes against the intent -- the standard clearly says a template definition from which no valid specialisation can be generated is ill-formed. Your primary template is such a definition. So it's ill-formed. > It also makes it difficult to use static_assert in certain > metaprogramming applications(like the example above, which could be used to > evaluate if a metafunction returned true). So make it dependent, or leave the primary template undefined. The alternative is to change three major compilers to accommodate a highly questionable style of metaprogramming that the standard says is ill-formed. Don't do it. > Also, as bug 52809 demonstrates, it > is confusing to users. It doesn't demonstrate that at all, it just asserts it.
Closing.
The trunk branch has been updated by Jason Merrill <jason@gcc.gnu.org>: https://gcc.gnu.org/g:9944ca17c0766623bce260684edc614def7ea761 commit r13-6133-g9944ca17c0766623bce260684edc614def7ea761 Author: Jason Merrill <jason@redhat.com> Date: Fri Feb 10 16:16:45 2023 -0800 c++: static_assert (false) in template [DR2518] For a long time, people have expected to be able to write static_assert (false) in a template and only have it diagnosed if the template is instantiated, but we (and other implementations) gave an error about the uninstantiated template because the standard says that if no valid instantiation of the template is possible, the program is ill-formed, no diagnostic required, and we try to diagnose IFNDR things when feasible. At the meeting last week we were looking at CWG2518, which wanted to specify that an implementation must not accept a program containing a failing #error or static_assert. We also looked at P2593, which proposed allowing static_assert in an uninstantiated template. We ended up combining these two in order to avoid requiring implementations to reject programs with static_assert (false) in uninstantiated templates. The committee accepted this as a DR, so I'm making the change to all standard modes. This behavior was also conformant previously, since no diagnostic was required in this case. We continue to diagnose non-constant or otherwise ill-formed conditions, so no changes to existing tests were needed. DR 2518 PR c++/52809 PR c++/53638 PR c++/87389 PR c++/89741 PR c++/92099 PR c++/104041 PR c++/104691 gcc/cp/ChangeLog: * semantics.cc (finish_static_assert): Don't diagnose in template context. gcc/testsuite/ChangeLog: * g++.dg/DRs/dr2518.C: New test.