[Bug ipa/84833] [8 Regression] target_clones regression since r251047

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84833

--- Comment #5 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #2)
> Just a note, as can be seen e.g. on
> static volatile int a;
> static int my_foo1 (void) { return 1; }
> static int my_foo2 (void) { return 2; }
> typedef int (*F) (void);
> static F resolve_foo (void) { if (a) return my_foo1; else return my_foo2; }
> static int foo (void) __attribute__ ((ifunc ("resolve_foo")));
> 
> int
> main ()
> {
>   if (foo () != 2)
>     __builtin_abort ();
> }
> there is no reason why any of the target clones, resolver or the ifunc
> function should be .globl, they can all be local to the TU, IFUNC work just
> fine that way too.  So, if the function with target_clones attribute is not
> TREE_PUBLIC, neither should be any of the multiple_targets.c created symbols.

If I see correctly, we do not make these global:

nm pr84833-2.o 
0000000000000000 b a
                 U abort
000000000000000c i foo
0000000000000022 T main
0000000000000000 t my_foo1
0000000000000006 t my_foo2
000000000000000c t resolve_foo

Note that it was subject of PR81213.