[C++ Patch] PR 59480 ("Missing error diagnostic: friend declaration specifying a default argument must be a definition")​ (Take 2)

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

Hi,

thus, as you may or may not have noticed I reverted my first try, when 
Tobias noticed that in his large codebase we were rejecting code like:

class Matrix;

Matrix rot90 (const Matrix& a, int k = 1);

class Matrix {
  friend Matrix rot90 (const Matrix&, int);
};

Matrix rot90 (const Matrix& a, int k) { return Matrix(); }

this was happening because, when duplicate_decls saw the friend 
declaration, smashed together the first two rot90 declaration and we 
ended up with a friend declaration with defaults (taken from the first 
rot90 declaration), as such rejected the third time we saw rot90. I 
think we can elegantly handle this kind of problem by exploiting the 
DECL_HIDDEN_FRIEND_P information, thus, in 
check_no_redeclaration_friend_default_args, as called by 
duplicate_decls, if the newdecl doesn't have DECL_FRIEND_P set, 
disregard an olddecl which doesn't have DECL_HIDDEN_FRIEND_P set (we 
can't do this directly because duplicate_decls is already smashing decls 
together at that point, thus we need to save the info and pass it as an 
argument). I added quite a few additional tests an also asked Tobias to 
double check on his code base. All good.

As a general arguments of why I think moving from DECL_FRIEND_P to 
DECL_HIDDEN_FRIEND_P for the olddecl is the right thing to do, if the 
olddecl is a friend but doesn't have DECL_HIDDEN_FRIEND_P set, there was 
a declaration preceding it, thus, either the friend declaration has 
default arguments and we already diagnosed that, or it doesn't , thus 
isn't interesting anymore for the diagnostic at issue.

Thanks! Paolo.

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

Attachment: CL_59480_2
Description: Text document

Attachment: patch_59480_2
Description: Text document