https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116891
--- Comment #6 from Walter Mascarenhas <walter.mascarenhas at gmail dot com> --- Hi Andrew, The proper optimization in this case would be to use the instruction vfnmsub132pd followed by a change of sign. It could be something like fma_ru: vfnmsub132pd xmm0, xmm2, xmm1 vmovddup xmm1, QWORD PTR .LC1[rip] vxorpd xmm0, xmm0, xmm1 ret .LC1: .long 0 .long -2147483648 On Mon, Sep 30, 2024 at 3:30 PM pinskia at gcc dot gnu.org < gcc-bugzi...@gcc.gnu.org> wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116891 > > Andrew Pinski <pinskia at gcc dot gnu.org> changed: > > What |Removed |Added > > ---------------------------------------------------------------------------- > Summary|[12/13/14/15 Regression] |[12/13/14/15 Regression] > |invalid optimization of |invalid optimization of > |-fma(-x,y,-z) when -03 and |-fma(-x,x,-z) when -03 and > |-frounding-math are used |-frounding-math are used > > --- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> --- > Simplified testcase: > ``` > double f(double ae, double ax) > { > return -__builtin_fma( -ax, ax, -ae ); > } > ``` > > That is pushing the outer negate into FMA is not valid due to > addition/multiply > being in infinite precision. > > -- > You are receiving this mail because: > You reported the bug.
