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

[jakub at gcc dot gnu.org]

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

            Bug ID: 84833
           Summary: [8 Regression] target_clones regression since r251047
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: ipa
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
                CC: marxin at gcc dot gnu.org
  Target Milestone: ---

__attribute__((noipa)) void
baz (void (*fn) (void))
{
  asm volatile ("" : "+g" (fn) : : "memory");
  fn ();
}

__attribute__((target_clones("arch=sandybridge", "default"))) static void
bar (void)
{
}

__attribute__((target_clones("arch=sandybridge", "default"))) void
foo (void)
{
  baz (bar);
}

int
main ()
{
}

no longer links.  Even previously it has been quite broken, as
    57: 0000000000400ae0    35 FUNC    GLOBAL DEFAULT   13
bar._GLOBAL___baz.resolver
    59: 0000000000400ae0    35 IFUNC   GLOBAL DEFAULT   13 bar
were exported symbols even when the function actually were declared static, but
no it doesn't even link.

Another (though preexisting) bug is lack of optimization, if we have a call or
reference of a target_clones dispatcher in a function with the same
target_clones attribute (or maybe even just target attribute or similar), we
should optimize it at compile time into a reference to a particular target
clone.

Got into this from PR78808, the above is just an approximation of why PR78808
doesn't work on the trunk either.