Optimize psign

Reference issue

What does this implement/fix?

Optimizes the generic sign function for floating point types.

Previously (float):

• 3 constants
• 3 comparisons
• 3 logicals
• 1 blend

Now:

• 2 constants
• 1 comparison
• 2 logicals (+1 absolute value) = ~3 logicals
• 1 blend

AVX2 disassembly for float

Eigen::internal::test_old(float __vector(8)):
        vxorps  xmm1, xmm1, xmm1
        vcmpps  ymm3, ymm0, ymm0, 0
        vbroadcastss    ymm4, DWORD PTR .LC1[rip]
        vcmpps  ymm2, ymm1, ymm0, 17
        vcmpps  ymm1, ymm0, ymm1, 17
        vandps  ymm2, ymm2, ymm4
        vbroadcastss    ymm4, DWORD PTR .LC3[rip]
        vandps  ymm1, ymm1, ymm4
        vorps   ymm1, ymm2, ymm1
        vblendvps       ymm0, ymm0, ymm1, ymm3
        ret

Eigen::internal::test_new(float __vector(8)):
        vbroadcastss    ymm1, DWORD PTR .LC5[rip]
        vbroadcastss    ymm3, DWORD PTR .LC1[rip]
        vandps  ymm2, ymm0, ymm1
        vandnps ymm1, ymm1, ymm0
        vxorps  xmm0, xmm0, xmm0
        vcmpps  ymm0, ymm0, ymm2, 17
        vorps   ymm1, ymm1, ymm3
        vblendvps       ymm0, ymm2, ymm1, ymm0
        ret

Additional information