| Message ID | 504A6228.5030005@oracle.com |
|---|---|
| State | New |
| Headers | show |
On 09/07/2012 05:07 PM, Paolo Carlini wrote: > the regression of template/crash83.C because of excess errors. The latter happens because we have code in cp_parser_class_head which tries to improve error recovery for cases of missing 'template <>' in explicit specializations, and, for that rather broken testcase, it confuses the code I'm tentatively adding, resulting in the additional error message below: > > crash83.C:5:27: error: an explicit specialization must be preceded by ‘template <>’ > template<typename = class A<0>: > struct B {}; // { dg-error "explicit specialization|expected" } > ^ > crash83.C:5:27: error: too many template-parameter-lists for A<0> (should be 1) So we're adding a second irrelevant diagnostic to go with the first one. The error in this testcase is a simple stray ':' causing a syntax error, it has nothing to do with explicit specialization. It would be nice to improve this, but given that we're already not giving the right error, adding a second wrong error doesn't seem like a big problem. > More importantly, I'm having troubles removing the early > cp_parser_check_template_parameters check in cp_parser_class_head: one > issue is that it also checks for too few template-parameter-lists not > just too many, and thus catches rather broken testcases like, eg, > template/error7.C or template/class2.C: > > template <class T> > struct A { > struct B; > struct A::B { }; // { dg-error "" } > }; This testcase is indeed ill-formed, but its ill-formedness has nothing to do with too few template-parameter-lists. The problem is that clause 9 says If a class-head-name contains a nested-name-specifier, the class-specifier shall refer to a class that was previously declared directly in the class or namespace to which the nested-name-specifier refers, ... and the class-specifier shall appear in a namespace enclosing the previous declaration. The definition of A::B does not appear at namespace scope, so it is ill-formed. Removing the incorrect error is an improvement, but then we need to add the correct error rather than accept the testcase. > but there are definitely more problems with just checking > template_header_count < wanted too in > maybe_process_partial_specialization and removing the early check, I'm > seeing many, many fails if I just try to extend "p4" like that... Like what? Jason
Index: pt.c =================================================================== --- pt.c (revision 191082) +++ pt.c (working copy) @@ -835,6 +835,15 @@ maybe_process_partial_specialization (tree type) && !COMPLETE_TYPE_P (type)) { check_specialization_namespace (CLASSTYPE_TI_TEMPLATE (type)); + + int wanted = num_template_headers_for_class (type); + if (template_header_count > wanted) + { + error ("too many template-parameter-lists for %E " + "(should be %d)", type, wanted); + return error_mark_node; + } + SET_CLASSTYPE_TEMPLATE_SPECIALIZATION (type); if (processing_template_decl) { @@ -882,6 +891,14 @@ maybe_process_partial_specialization (tree type) tree t; tree tmpl = CLASSTYPE_TI_TEMPLATE (type); + int wanted = num_template_headers_for_class (type); + if (template_header_count > wanted) + { + error ("too many template-parameter-lists for %E " + "(should be %d)", type, wanted); + return error_mark_node; + } + if (current_namespace != decl_namespace_context (tmpl)) {