[C++ Patch] PR 70621 ("[6/7/8 Regression] ICE on invalid code at -O1 and above on x86_64-linux-gnu in record_reference, at cgraphbuild.c:64")

[Paolo Carlini <paolo dot carlini at oracle dot com>]

Hi,

in this error recovery regression, we ICE only when optimizing, while 
building the cgraph. Avoiding the reported problem seems easy, just 
check the return value of duplicate_decls in start_decl and immediately 
return back error_mark_node. However, while working on the issue, I 
noticed something slightly more interesting, IMO: we have, after the 
relevant duplicate_decls call:

-              if (decl_spec_seq_has_spec_p (declspecs, ds_constexpr)
-                  && !DECL_DECLARED_CONSTEXPR_P (field))
-                error ("%qD declared %<constexpr%> outside its class", 
field);

which I propose to remove. In fact - something I didn't really know - 
for well formed user code, duplicate_decls, near the end, does some 
memcpys which mean that its second argument (would be 'field' in the 
start_decl section we are looking at)  is adjusted to have a 
DECL_DECLARED_CONSTEXPR_P consistent with its first argument. That's of 
course because we want to accept snippets like:

struct A
{
  static const int x;
};

constexpr int A::x = 0;

In turn that means the error above is issued only when something went 
wrong in duplicate_decls in the first place. Thus the above error seems 
at least verbose. However, here in start_decl we are only handling VAR_P 
(decl), thus I don't think the message above even makes sense and could 
be misleading, given snippets like the above. Therefore, I propose to 
remove the diagnostic entirely, which overall also simplifies a patch 
dealing with c++/70621. The below passes testing as-is on x86_64-linux.

Thanks, Paolo.

////////////////////////

Attachment: CL_70621
Description: Text document

Attachment: patch_70621
Description: Text document