Optimize ord implementation and signed zero canonicalization

[orlp]

These micro-optimizations significantly reduce the number of instructions comparisons take, and often makes them branchless as well. Similarly we use a trick to canonicalize signed zero to positive zero in a single instruction without branches for faster hashing.

[orlp]

For example, a <= b went from this:

example::old_leq:
        vucomiss        xmm1, xmm0
        jae     .LBB0_1
        mov     al, 1
        vucomiss        xmm0, xmm1
        jae     .LBB0_5
        mov     al, -1
        vucomiss        xmm0, xmm0
        jp      .LBB0_4
.LBB0_5:
        inc     al
        cmp     al, 2
        setb    al
        ret
.LBB0_1:
        xor     eax, eax
        vucomiss        xmm0, xmm1
        sbb     eax, eax
        inc     al
        cmp     al, 2
        setb    al
        ret
.LBB0_4:
        vucomiss        xmm1, xmm1
        setnp   al
        inc     al
        cmp     al, 2
        setb    al
        ret

to this:

example::new_leq:
        vcmpleps        xmm0, xmm0, xmm1
        vxorps  xmm2, xmm2, xmm2
        vcmpunordps     xmm1, xmm1, xmm2
        vorps   xmm0, xmm1, xmm0
        vmovd   eax, xmm0
        and     al, 1
        ret

[mbrubeck]

Thank you!

[orlp]

@mbrubeck To also give some concrete numbers, on my Apple M1 machine sorting a shuffled Vec of 1 million OrderedFloat<f64>s went from 110ms to 84ms, an 1.3x speedup. I'd expect the difference on x86-64 to be even greater.