OK.
On Tue, Apr 3, 2018 at 5:02 AM, Paolo Carlini <paolo.carl...@oracle.com> wrote: > Hi, > > the reason why we ICE here is very simple: we pass an error_mark_node > generated in cp_parser_template_parameter_list to rewrite_template_parm > which ICEs (via get_template_parm_index). I believe it makes sense to > robustify the latter like all the other functions involved in > build_deduction_guide. The next, error recovery quality, issue is whether or > not we still want to produce diagnostic about class template argument > deduction failing: today I noticed - somewhat surprisingly - that some > compilers still do that (however the diagnostic is more terse: eg, one > additional error and two notes, not two errors and five notes!). I had > tested the below which suppresses those additional errors by returning > error_mark_node from build_deduction_guide if one of those parsing-time > error_mark_nodes is encountered in the build_deduction_guide loop - and lays > out a bit of infrastructure. > > Tested x86_64-linux. > > Thanks, Paolo. > > //////////////// >
