https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106816
--- Comment #10 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Martin Jambor from comment #9)
> (In reply to Martin Liška from comment #6)
> >
> > @Martin: Do we have a declaration cloning code for functions somewhere?
>
> See e.g. cgraph_node::create_virtual_clone in cgraphclones.cc. Unless
> you want to mess with the parameters,
>
> new_decl = copy_node (old_decl);
>
> should be enough (and it should copy over the DECL_PURE and
> TREE_READLONY bit soo, I believe).
Thanks! So apparently, we call make_dispatcher_decl which makes only a partial
copy of a function_decl:
original decl:
<function_decl 0x7ffff77ce500 f
type <function_type 0x7ffff7656d20
type <void_type 0x7ffff7644f18 void VOID
align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff7644f18
pointer_to_this <pointer_type 0x7ffff764c000>>
QI
size <integer_cst 0x7ffff76460c0 constant 8>
unit-size <integer_cst 0x7ffff76460d8 constant 1>
align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff7656d20
arg-types <tree_list 0x7ffff7639d98 value <void_type 0x7ffff7644f18
void>>
pointer_to_this <pointer_type 0x7ffff77ee498>>
volatile nothrow public static decl_5 QI
/home/marxin/Programming/testcases/pr106816.c:2:6 align:8 warn_if_not_align:0
context <translation_unit_decl 0x7ffff762d168
/home/marxin/Programming/testcases/pr106816.c>
attributes <tree_list 0x7ffff77de0c8
purpose <identifier_node 0x7ffff77caa40 target
normal local bindings <(nil)>>
value <tree_list 0x7ffff77de078
value <string_cst 0x7ffff77c29e0 type <array_type 0x7ffff77ccb28>
readonly constant static "default\000">>
chain <tree_list 0x7ffff77de0a0
purpose <identifier_node 0x7ffff7650500 noreturn
normal local bindings <(nil)>>>> initial <block 0x7ffff77d1420>
result <result_decl 0x7ffff762df78 D.2761 type <void_type 0x7ffff7644f18
void>
ignored VOID /home/marxin/Programming/testcases/pr106816.c:2:1
align:8 warn_if_not_align:0 context <function_decl 0x7ffff77ce500 f>>
full-name "void f()"
struct-function 0x7ffff77df000 chain <function_decl 0x7ffff77cbc00
__cxa_call_unexpected>>
dispatcher:
(gdb) p debug_tree(func_decl)
<function_decl 0x7ffff77efd00 _Z1fv
type <function_type 0x7ffff7656d20
type <void_type 0x7ffff7644f18 void VOID
align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff7644f18
pointer_to_this <pointer_type 0x7ffff764c000>>
QI
size <integer_cst 0x7ffff76460c0 constant 8>
unit-size <integer_cst 0x7ffff76460d8 constant 1>
align:8 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7ffff7656d20
arg-types <tree_list 0x7ffff7639d98 value <void_type 0x7ffff7644f18
void>>
pointer_to_this <pointer_type 0x7ffff77ee498>>
nothrow public external QI
/home/marxin/Programming/testcases/pr106816.c:15:1 align:8 warn_if_not_align:0>
So we modify the name, drop 'target' attribute, and so on. Thus we can't use
copy_node.
>
> I am not sure at what point the duplication happens, in order to
> duplicate also all the various IPA summaries, cgraph machinery has to
> be involved and call all the summary hooks. So
> cgraph_node::create_clone should probably be used (or a part of it, or
> perhaps even one of its users?).
It can happen during the multiple_target.cc IPA pass, so lemme combine the
H.J.'s approach
and cgraph_node::create_clone.
