On Mon, 09 Dec 2019, Paul E Murphy wrote:
On 12/9/19 10:49 AM, Zack Weinberg wrote:
Credit is actually due to Gabriel; if I understand correctly, I
_introduced_ a bug here!
Zack's patch did the right thing... However, I started working on
powerpc's IEEE long double implementation before his patch landed on
master and I didn't notice the change in default behaviour. The potential
bug never really showed up, because ldbl-128ibm-compat is not yet in the
Implies file, so the __*ieee128 functions were never really exposed.
I stand corrected. I wonder how frail some of these tests might become
as they age. Is framework available today to verify tests are linking
against the desired variants of a symbol?
The tests in the subsequent patch (2/11) try to make that better, by
trying to read a long double value with %as when in ISO C99 mode. Now I'm
wondering that it should try to read a string (and allocate it
automatically) in the -D_GNU_SOURCE -std=c89 case, i.e.: also use %as, but
expect a different output.
Since each case should produce different outputs (even if the input is the
same), I think that will help verify that the correct symbol gets selected.