> Jan,
> This patch caused Bug 64982 - [5 Regression] Many g++ failures on
> x86_64-apple-darwin14 with -m32.
> Jack
Hi,
I think best way would be to move the warning to middle-end when the thunk is
being
expanded (other alternative is just to force analysis before doing the sanity
check
in cp/method.c.
I had to dro # in the error message, I am not quite sure what it means.
Jason, what do you think?
Honza
Index: cp/method.c
===================================================================
--- cp/method.c (revision 220547)
+++ cp/method.c (working copy)
@@ -418,20 +418,6 @@ use_thunk (tree thunk_fndecl, bool emit_
if (DECL_ONE_ONLY (function))
thunk_node->add_to_same_comdat_group (funcn);
- if (!this_adjusting
- || !targetm.asm_out.can_output_mi_thunk (thunk_fndecl, fixed_offset,
- virtual_value, alias))
- {
- /* If this is a covariant thunk, or we don't have the necessary
- code for efficient thunks, generate a thunk function that
- just makes a call to the real function. Unfortunately, this
- doesn't work for varargs. */
-
- if (varargs_function_p (function))
- error ("generic thunk code fails for method %q#D which uses %<...%>",
- function);
- }
-
pop_from_top_level ();
}
Index: cgraphunit.c
===================================================================
--- cgraphunit.c (revision 220547)
+++ cgraphunit.c (working copy)
@@ -1559,6 +1559,14 @@ cgraph_node::expand_thunk (bool output_a
thunk.thunk_p = false;
analyzed = false;
}
+ else if (stdarg_p (TREE_TYPE (thunk_fndecl)))
+ {
+ error ("generic thunk code fails for method %qD which uses %<...%>",
+ thunk_fndecl);
+ TREE_ASM_WRITTEN (thunk_fndecl) = 1;
+ analyzed = true;
+ return false;
+ }
else
{
tree restype;
