On Wed, Nov 23, 2016 at 9:56 AM, Jakub Jelinek <ja...@redhat.com> wrote: > Is there a reason why we don't return error_mark_node no matter what > complain contains?
In some cases, outside of SFINAE context we can give an error and then continue to do something reasonable. > At least on the testcase if we don't return error_mark_node for > the uses of var before deduction of auto, then we ICE later on in some > assertion that expects sane types on the variables. On the other testcase > it avoids a cascading of further errors, emits just the single error. Then it seems that in this case we should always return error_mark_node. > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? OK. > What about all the other spots in pt.c that return error_mark_node > if mark_used failed only if complain doesn't have tf_error set? They should probably be changed as well. Jason
