On Mon, 29 Jan 2024, Jakub Jelinek wrote:

> Most of the other tests (except the tests for side-effects in arguments)
> have constant arguments and so can be folded very early during compilation,
> even the side-effects in arguments tests can be folded soon when optimizing,
> so these functions were just an attempt to make sure the macros can be
> compiled/linked when the arguments are certainly not known.
> But if you think it isn't worth checking that or if it is already tested by
> some other test, I can remove that.

I think the existing tests adequately cover non-constant arguments 
(non-constant unless you unroll the test loop and extract the arguments 
from a const array of structures then propagate from the variable storing 
them after loading from the array, that is).  What the existing tests 
don't actually cover for standard types (and this new test does cover) is 
constant arguments.

Note that as we're currently in release freeze for glibc 2.39, a new 
feature change such as this should probably only go in before 2.39 is 
released with the approval of Andreas K. Hüttel as release manager (while 
if going in after 2.39 is released, the new feature of unsigned __int128 / 
unsigned _BitInt support should get a NEWS entry for 2.40).

-- 
Joseph S. Myers
josmyers@redhat.com