[RFC patch] avoid warning on accesses to hardwired address

[Martin Sebor]

Thanks to recent code refactoring in GCC 12, -Warray-bounds has
started to diagnose accesses to constant addresses just like many
other flow based warnings do (e.g., -Wstringop-overflow).
The warnings are meant to help detect accesses resulting from
invalid arithmetic on null pointers.  There may be a better way
to detect them but GCC doesn't have the detection yet.  This
warning change has in turn exposed Glibc's uses of this trick
in the implementation of the THREAD_SELF macro.

I have tried a few approaches to avoid the warning but none worked
or seemed satisfactory:

1) Using #pragma GCC diagnostic doesn't work in macros.
2) Using _Pragma doesn't work there either due to a GCC limitation.
3) Declaring the pointer volatile works but prevents accesses to
    it from being eliminated when the result isn't used (something
    Glibc apparently cares about).
4) Defining a simple static inline function to wrap the code and
    the #pragmas doesn't work because the header is #included in
    files where struct pthread isn't defined.
5) Introducing a global variable with the address and initializing
    it in some .c file seems too heavy-weight.
6) Falling back on the pre-GCC 6 asm would be safer but seems like
    a step back.

Finally I have come up with the attached solution that combines (4)
and (1).  If (6) is preferable I'm happy to go that route.  If there
are other alternatives I'd be glad to consider them as well.

Thanks
Martin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: glibc-thread_self.diff
Type: text/x-patch
Size: 974 bytes
Desc: not available
URL: <https://sourceware.org/pipermail/libc-alpha/attachments/20210708/111983b8/attachment.bin>