sign(float64<N>) generates incorrect code on NEON64 using gcc 9.3.0 with -ffast-math #153
Description
auto foo24(float64x2 a) { return sign(a); }
generates:
0000000000000000 <Inspiration::foo24(Inspiration::FloatVectorBase<simdpp::arch_neonfltsp::float64<4u, void> >)>:
0: 6f00e402 movi v2.2d, #0x0
4: d10383ff sub sp, sp, #0xe0
8: 910383ff add sp, sp, #0xe0
c: 4e221c00 and v0.16b, v0.16b, v2.16b
10: 4e221c21 and v1.16b, v1.16b, v2.16b
14: d65f03c0 ret
The problem seems to trace back to the way bit_and() is being processed in the implementation of _i_sign(float64):
return bit_and(a, 0x8000000000000000);
The 0x8000000'00000000 constant gets bit-converted to the floating point constant -0.0. Because of -ffast-math -0.0 and 0.0 are equivalent, so this winds up masking with 0 instead.