GCC miscompilation with __seg_fs

[Jakub Jelinek]

On Mon, Mar 13, 2023 at 06:24:47PM +0300, Sergey Bugaev wrote:
> Hello,
> 
> while exploring the generated assembly for an entirely unrelated
> reason, I found out that GCC eliminates stores through THREAD_SELF
> when it's declared using __seg_fs. This is easily reproducible outside
> of glibc; here's a sample reproducer (I've also put it on Godbolt [0]
> for easy exploration):

Smaller testcase:
typedef struct
{
  void *tcb;
  int some_member;
} tcbhead_t;

void
assign_through_self (void)
{
  (*(tcbhead_t * __seg_fs *) __builtin_offsetof (tcbhead_t, tcb))->some_member = 42;
}

>From what I can see, GCC has been DSE removing these since https://gcc.gnu.org/r6-4645
when __seg_fs support has been introduced.
One can work-around it using -fno-delete-null-pointer-checks.
You could also hide the fact that it is based on 0 pointer from the
compiler...

While GCC has targetm.addr_space.zero_address_valid hook (which AFAIK only
x86 overrides), we use it only in very few spots right now.

Feel free to file a bug report in GCC bugzilla, but that won't improve
anything on the already released compilers.

	Jakub