; Untested (check for off-by-one array indexing)
; Probably about 3x as fast as the above for large arrays
; on Intel Sandybridge-family.  Unrolling by 2 or so would help.

; prototype: unsigned long popcount(int64_t* rdi, int64_t rsi);
; requires byte count in rsi >= 8, and a multiple of 8.
ALIGN 16
global popcount
popcount:
        add    rdi, rsi           ; rdi = end of buffer
        neg    rsi                ; negative index counts up to zero
        xor    rax, rax
                                  ;do {
.cnt    mov    rdx, [rdi+rsi]      ; mov-load breaks the dependency
        popcnt rdx, rdx            ;  on rdx from last iteration.
        add    rax, rdx
        add    rsi, 8             ; index+=8
        jl    .cnt                ;}while(index<0)
        ret