https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123207
--- Comment #15 from Chen Chen <chz0808 at gmail dot com> --- (In reply to Chen Chen from comment #14) > On Loongarch64 I saw the same issue. Tested with gcc16-20251214 and > gcc16-20251221 snapshots on an Loongson 3A6000 (la664 microarchitecture) > machine, miscompare appears for gcc16-20251221 snapshot when built with "-O3 > -ffast-math" or "-Ofast". For gcc16-20251214 snapshot it is OK. For the above statement "-march=native" should also be combined with "-O3 -ffast-math" or "-Ofast". I also tested gcc16-20251221 snapshot on an AMD Zen4 machine. Miscompare persists with "-O3 -march=native" but disappears with "-O2 -march=native" for COPTIMIZE. The optimization level for CXXOPTIMIZE does not matter.
