https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64330
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jason at gcc dot gnu.org,
| |marxin at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
It is related, but in this case it looks like IPA-ICF bug to me, unless the C++
standard allows what G++ does since r216305.
struct S
{
static const unsigned short a;
static const unsigned short b;
static const unsigned short c;
};
#ifdef A
const unsigned short S::a = 2;
const unsigned short S::b = 1;
const unsigned short S::c = 2;
#endif
#ifdef B
int
main ()
{
if (&S::a == &S::c)
__builtin_abort ();
}
#endif
With g++ -O2 -o test test.C -DA -DB it succeeds, but with
g++ -O2 -c test.C -o testa.o -DA; g++ -O2 -o test -DB test.C testa.o
it fails. Works with -fno-ipa-icf. Without looking at the IPA-ICF patch, just
looking at what it generates, it sounds like it is checking TREE_ADDRESSABLE on
the variables and if they aren't addressable and are const, merges them. That
is fine if you are in LTO and can see all TUs that could ever access those
vars, but if you can't inspect all the TUs that could take the address of it,
as in the second case, but it could very well be even LTO compiled binary that
links against a shared library that references those symbols, I think you can't
merge the variables (unless -fmerge-all-constants).
