[Bug fortran/103258] [12 Regression] ICE in gfc_typenode_for_spec, at fortran/trans-types.c:1114 since r12-4979-gee11be7f2d788e60

["sandra at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103258

--- Comment #3 from sandra at gcc dot gnu.org ---
This looks like an existing bug in error checking that was exposed by my patch
to do...  more error checking.  :-S

The problem is that gfc_set_default_type in symbol.c is setting
sym->attr.untyped even when errors are suppressed or buffered and discarded
during parsing.  Since it's marked as having already been diagnosed, we aren't
giving an error in subsequent calls where errors are *not* being suppressed or
discarded, and thus compilation continues as if nothing were wrong, until it
falls over.

Previously we were resolving only the second operand "m" in the "(n+m)"
expression in a situation where errors are buffered in the parsing stage,
leaving "n" without the attr.untyped flag set so that it was being diagnosed
later during the actual resolve phase of processing.  Now it is setting that
flag on both "n" and "m" so neither ever gives a real error.

I've got a hacky patch to avoid setting sym->attr.untyped when a real error
might not be given, but I'm not sure if it's the right solution.  Maybe a
better solution would be to emit the diagnostic somewhere other than
gfc_set_default_type, at a known point in the processing where it'll only
happen once?