https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82004
--- Comment #25 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Martin Liška from comment #24) > I can confirm that the Jakub's commit is fixing that. Revering the patch I > see a miscomparison. That said, I would close it as resolved. Andrey? That is weird, because we've figured that in the actual SPEC test it is used in a loop with a variable exponent. If it no longer reproduces, isn't that some other change? For a real fix I was hoping somebody would provide new, faster and precise exp10{,f,l} implementations to glibc and then we could just use exp10 rather than exp.