On 09/28/2018 01:03 PM, Rob wrote: > If it helps, this is likely related to a bugfix that went into freebsd's libm a > while back after they switched to clang. I can't remember the details, but I > believe the bug was to do with clang incorrectly constant-folding floating > point expressions in the wrong rounding mode, or perhaps how it interacted with > fesetround() and the fenv access pragmas. > > The mirror here [1] shows the fix. Yeah, I see. It's about setting the flags in the FPU so that the division by zero is delayed to runtime rather than evaluated as a constant by the compiler. - return -two25/zero; /* log(+-0)=-inf */ + return -two25/vzero; /* log(+-0)=-inf */ -- Andrew Haley Java Platform Lead Engineer Red Hat UK Ltd. <https://www.redhat.com> EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671
