Re: [PATCH] Fix ODR violations in code using <ext/atomicity.h>

[Jonathan Wakely <jwakely@redhat.com>]

On 21/06/19 18:08 +0100, Jonathan Wakely wrote:
>On 21/06/19 13:01 -0400, Nathan Sidwell wrote:
>>On 6/21/19 12:01 PM, Jonathan Wakely wrote:
>>>Nathan noticed that the 'static inline' functions in <ext/atomicity.h>
>>>cause ODR violations when used from inline functions or templates (see
>>>[basic.def.odr] p12 bullet (12.2)). His modules branch now diagnoses
>>>those violations, so we need a fix.
>>>
>>>Looking at the history (r114044 by Paolo) I believe the idea was indeed
>>>to allow different definitions to be used in different TUs. I think
>>>__attribute__((__always_inline__)) is a better match for that than
>>>'static inline', and doesn't violate the ODR (at least, not if all TUs
>>>have the same values for __GTHREADS and _GLIBCXX_ATOMIC_BUILTINS).
>>>
>>>These functions still violate this rule in [dcl.inline]:
>>>
>>>C++17: "If a function with external linkage is declared inline in one
>>>translation unit, it shall be declared inline in all translation units
>>>in which it appears; no diagnostic is required."
>>>
>>>C++2a WP: "If a function or variable with external or module linkage
>>>is declared inline in one translation unit, there shall be a reachable
>>>inline declaration in all translation units in which it is declared;
>>>no diagnostic is required."
>>>
>>>But that doesn't seem to be diagnosable by today's implementations.
>>>
>>>Does this change seem reasonable?
>>
>>yes, thanks!
>
>Actually, I think I prefer this version. This produces identical code
>to the always_inline version, but without the indirection to
>additional inline functions, i.e. just inline the relevant code into
>the _dispatch functions.
>
>Tests are still running but no failures so far, as expected.

Oops, I spoke too soon:

FAIL: tr1/2_general_utilities/shared_ptr/thread/default_weaktoshared.cc (test for excess errors)
UNRESOLVED: tr1/2_general_utilities/shared_ptr/thread/default_weaktoshared.cc compilation failed to produce executable
FAIL: tr1/2_general_utilities/shared_ptr/thread/mutex_weaktoshared.cc (test for excess errors)
UNRESOLVED: tr1/2_general_utilities/shared_ptr/thread/mutex_weaktoshared.cc compilation failed to produce executable

Those tests explicitly use the __atomic_add function that the patch
removes. But those would be easy to adjust.