.486
.model flat, stdcall
 option casemap: none
 include windows.inc

.code

DllEntry PROC hINSTDLL:HINSTANCE, reason:DWORD, reserved1:DWORD
        mov eax, TRUE
        ret
DllEntry ENDP

PreSieveAtkinAsm proc stdcall sieve: PTR, limit: DWORD, x: DWORD, root: DWORD

    mov esi, sieve
    mov ebx, 0

mainloop:
    inc ebx ;increment y
    cmp ebx, root ;compare root with y
    jg mainloopend  ;y <= root

    ;n - eax
    ;y - ebx
    ;n = (4*x*x)+(y*y)

part1:
    mov eax, x ;x
    mul eax ;x*x
    mov ecx, 4
    mul ecx ;4*x*x
    mov ecx, eax ;4*x*x in ecx
    mov eax, ebx ;y
    mul eax ;y*y
    add eax, ecx ;4*x*x+y*y in eax

    cmp eax, limit
    jle part2

    push eax ;push n to stack

    mov ecx, 12
    div ecx
    pop eax ;pop n from stack
    mov eax, 1
    cmp eax, edx
    je xor1 ;n
    mov eax, 5
    cmp eax, edx
    je xor1
    jmp part2

    ;n in eax

xor1:
    not BOOL PTR [esi + 1 * eax]

part2:
    ;n - eax
    ;y - ebx
    ;n = (3*x*x)+(y*y)

    mov eax, x ;x
    mul eax ;x*x
    mov ecx, 3
    mul ecx ;4*x*x
    mov ecx, eax ;4*x*x in ecx
    mov eax, ebx ;y
    mul eax ;y*y
    add eax, ecx ;4*x*x+y*y in eax

    cmp eax, limit
    jle part3

    push eax ;push n to stack

    mov ecx, 12
    div ecx
    pop eax ;pop n from stack
    mov eax, 7
    cmp eax, edx
    je xor2 ;n
    jmp part3

    ;n in eax

xor2:
    not BOOL PTR [esi + 1 * eax]

part3:
    ;n - eax
    ;y - ebx
    ;n = (4*x*x)-(y*y)
    cmp x, ebx
    jle mainloop

    mov eax, x ;x
    mul eax ;x*x
    mov ecx, 4
    mul ecx ;3*x*x
    mov ecx, eax ;3*x*x in ecx
    mov eax, ebx ;y
    mul eax ;y*y
    sub eax,ecx ;4*x*x-y*y in eax

    cmp eax, limit
    jle mainloop

    push eax ;push n to stack

    mov ecx, 12
    div ecx
    pop eax ;pop n from stack
    mov ecx, 11
    cmp ecx, edx
    je xor3
    jmp mainloop

    ;n in eax

xor3:
    not BOOL PTR [esi + 1 * eax]

mainloopend:
    leave
    ret

PreSieveAtkinAsm endp

END DllEntry