On Thu, Feb 9, 2017 at 4:08 AM, Paolo Carlini <paolo.carl...@oracle.com> wrote: > On 14/01/2017 15:43, Jason Merrill wrote: >> >> OK. > > As you may or may not have noticed, I had to revert the patch because it > caused the regression of > tr1/2_general_utilities/shared_ptr/cons/43820_neg.cc during error recovery: > an ICE happens in remap_decl when build_variant_type_copy is called on a > TREE_TYPE (== DECL_ORIGINAL_TYPE) which is error_mark_node. A possible > straightforward and safe fix for the new issue would be simply checking for > the special condition, as I did in patch_71737_3 below, which passes > testing. Alternately, I have been fiddling also with older approaches, and > patch_71737_4 below also passes testing: it simply renounces to preserve a > typedef which names an error_mark_node type.
_4 seems consistent with the set_underlying_type behavior, I guess let's go with that. Jason
