<% function binOpS1S2(op) { %>
        cmp    r8d, 0
        <% const s1s2_solveNeg = global.tmpLabel() %>
        js     <%=s1s2_solveNeg%>

        cmp    r9d, 0
        js     <%=s1s2_solveNeg%>
        xor    rdx, rdx         ; both ops are positive so do the op and return
        mov    edx, r8d
        <%=op%>    edx, r9d
        mov    [rdi], rdx       ; not necessary to adjust so just save and return
        ret

<%=s1s2_solveNeg%>:
<%=     global.setTypeDest("0x80"); %>
<%=     global.toLong_b() %>
<%=     global.toLong_a() %>
<%=     binOpL1L2(op)       %>


<% } %>

<% function binOpS1L2(op) { %>
        cmp    r8d, 0
        <% const s1l2_solveNeg = global.tmpLabel() %>
        js     <%=s1l2_solveNeg%>
        movsx  rax, r8d
        <%=op%> rax, [rdx +8]
        mov    [rdi+8], rax
<%   for (let i=1; i<n64; i++) { %>
        xor    rax, rax
        <%=op%> rax, [rdx + <%= (i*8)+8 %>]
<%     if (i== n64-1) { %>
        and    rax, [lboMask]
<%     }  %>
        mov    [rdi + <%= (i*8)+8 %> ], rax
<%   } %>
        ret

<%=s1l2_solveNeg%>:
<%=     global.toLong_a() %>
<%=     global.setTypeDest("0x80"); %>
<%=     binOpL1L2(op)       %>

<% } %>

<% function binOpL1S2(op) { %>
        cmp    r9d, 0
        <% const l1s2_solveNeg = global.tmpLabel() %>
        js     <%=l1s2_solveNeg%>
        movsx  rax, r9d
        <%=op%> rax, [rsi +8]
        mov    [rdi+8], rax
<%   for (let i=1; i<n64; i++) { %>
        xor    rax, rax
        <%=op%> rax, [rsi + <%= (i*8)+8 %>];
<%     if (i== n64-1) { %>
        and    rax, [lboMask] ;
<%     }  %>
        mov    [rdi + <%= (i*8)+8 %> ], rax;
<%   } %>
        ret

<%=l1s2_solveNeg%>:
<%=     global.toLong_b() %>
<%=     global.setTypeDest("0x80"); %>
<%=     binOpL1L2(op)       %>

<% } %>

<% function binOpL1L2(op) { %>
<%   for (let i=0; i<n64; i++) { %>
        mov rax, [rsi + <%= (i*8)+8 %>]
        <%=op%> rax, [rdx + <%= (i*8)+8 %>]
<%     if (i== n64-1) { %>
        and    rax, [lboMask]
<%     }  %>
        mov    [rdi + <%= (i*8)+8 %> ], rax
<%   } %>
        ret
<% } %>


<% function binOp(op) { %>
;;;;;;;;;;;;;;;;;;;;;;
; b<%= op %>
;;;;;;;;;;;;;;;;;;;;;;
; Adds two elements of any kind
; Params:
;   rsi <= Pointer to element 1
;   rdx <= Pointer to element 2
;   rdi <= Pointer to result
; Modified Registers:
;    r8, r9, 10, r11, rax, rcx
;;;;;;;;;;;;;;;;;;;;;;
<%=name%>_b<%=op%>:
        mov    r8, [rsi]
        mov    r9, [rdx]
        bt     r8, 63          ; Check if is short first operand
        jc     <%=op%>_l1
        bt     r9, 63          ; Check if is short second operand
        jc     <%=op%>_s1l2

<%=op%>_s1s2:
<%=     binOpS1S2(op)       %>


<%=op%>_l1:
        bt     r9, 63          ; Check if is short second operand
        jc     <%=op%>_l1l2


<%=op%>_l1s2:
        bt     r8, 62          ; check if montgomery first
        jc     <%=op%>_l1ms2
<%=op%>_l1ns2:
<%=     global.setTypeDest("0x80"); %>
<%=     binOpL1S2(op)       %>

<%=op%>_l1ms2:
<%=     global.setTypeDest("0x80"); %>
        push   r9              ; r9 is used in montgomery so we need to save it
<%=     global.fromMont_a()       %>
        pop    r9
<%=     binOpL1S2(op)       %>


<%=op%>_s1l2:
        bt     r9, 62          ; check if montgomery first
        jc     <%=op%>_s1l2m
<%=op%>_s1l2n:
<%=     global.setTypeDest("0x80"); %>
<%=     binOpS1L2(op)       %>

<%=op%>_s1l2m:
<%=     global.setTypeDest("0x80"); %>
        push   r8             ; r8 is used in montgomery so we need to save it
<%=     global.fromMont_b()       %>
        pop    r8
<%=     binOpS1L2(op)       %>


<%=op%>_l1l2:
        bt     r8, 62          ; check if montgomery first
        jc     <%=op%>_l1ml2
        bt     r9, 62          ; check if montgomery first
        jc     <%=op%>_l1nl2m
<%=op%>_l1nl2n:
<%=     global.setTypeDest("0x80"); %>
<%=     binOpL1L2(op)       %>

<%=op%>_l1nl2m:
<%=     global.setTypeDest("0x80"); %>
<%=     global.fromMont_b()       %>
<%=     binOpL1L2(op)       %>

<%=op%>_l1ml2:
        bt     r9, 62          ; check if montgomery first
        jc     <%=op%>_l1ml2m
<%=op%>_l1ml2n:
<%=     global.setTypeDest("0x80"); %>
<%=     global.fromMont_a()       %>
<%=     binOpL1L2(op)       %>

<%=op%>_l1ml2m:
<%=     global.setTypeDest("0x80"); %>
<%=     global.fromMont_a()       %>
<%=     global.fromMont_b()       %>
<%=     binOpL1L2(op)       %>
<% } %>

<%= binOp("and") %>
<%= binOp("or") %>
<%= binOp("xor") %>


;;;;;;;;;;;;;;;;;;;;;;
; bnot
;;;;;;;;;;;;;;;;;;;;;;
; Adds two elements of any kind
; Params:
;   rsi <= Pointer to element 1
;   rdi <= Pointer to result
; Modified Registers:
;    r8, r9, 10, r11, rax, rcx
;;;;;;;;;;;;;;;;;;;;;;
<%=name%>_bnot:
        <%=     global.setTypeDest("0x80"); %>

        mov     r8, [rsi]
        bt      r8, 63          ; Check if is long operand
        jc      bnot_l1
bnot_s:
        <%=     global.toLong_a() %>
        jmp     bnot_l1n

bnot_l1:
        bt     r8, 62          ; check if montgomery first
        jnc    bnot_l1n

bnot_l1m:
<%=     global.fromMont_a()       %>

bnot_l1n:
<%   for (let i=0; i<n64; i++) { %>
        mov    rax, [rsi + <%= i*8 + 8 %>]
        not    rax
<%     if (i== n64-1) { %>
        and    rax, [lboMask]
<%     }  %>
        mov    [rdi + <%= i*8 + 8 %>], rax
<%   } %>
        ret