[Bug c++/106902] [13/14/15/16 Regression] Program compiled with -O3 -mfma produces different result

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106902

--- Comment #43 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
For reference, the last iteration is at
https://inbox.sourceware.org/gcc-patches/7cd2a907-68fb-d726-f129-ff2aacf3b107@ispras.ru/

For x86, I believe I need to clone some existing tests, adding
-ffp-contract=fast, to keep some coverage of the current behavior (and I
dropped the ball on that, sorry!)

I still welcome advice how to coordinate this with maintainers of all the other
FMA-capable targets in GCC without burning too much time or remaining goodwill
(if any): if the above patch is applied, they can get a few testsuite
regressions.