;
;-----> Division loop
;
d3:     cmp     esi,14
        jb      d3a
        push    edi
        mov     edx,a[ebp][esi]         ; uj   unshifet in dx (hiWord)
        mov     eax,a[ebp][esi-4]       ; uj+1 unshifted
        mov     ebx,a[ebp][esi-8]       ; uj+2 unshifted
        mov     edi,a[ebp][esi-12]      ; uj+3 unshifted
        push    ecx
        mov     cx,d[ebp]
@@:     cmp     cx,0
        je      @F
        rcl     edi,1
        rcl     ebx,1
        rcl     eax,1
        rcl     edx,1
        dec     cx
        jmp     @B
@@:     pop     ecx
        pop     edi
        jmp     qhat
;
d3a:    mov     edx,a[ebp][esi]         ; uj   unshifted
        mov     eax,a[ebp][esi-4]       ; uj+1 unshifted
        mov     ebx,a[ebp][esi-8]       ; uj+2 unshifted
        push    ecx
        mov     cx,d[ebp]
@@:     cmp     cx,0
        je      @F
        rcl     ebx,1
        rcl     eax,1
        rcl     edx,1
        dec     cx
        jmp     @B
@@:     pop     ecx
;
;-----> Calculate and test qhat
;
qhat:   mov uj1[ebp],eax                ;  Store shifted uj+1
        mov uj2[ebp],ebx                ;  Store shifted uj+2
        mov ebx,v1[ebp]                 ;  v1 in bx
        cmp ebx,edx
        je bm1                          ; If v1 = uj then q = b - 1
;
        div ebx                         ; eax <- qhat
;                                       ; edx <- rhat: = uj*b + uj+1 -qhat*v1 = (uj*b + uj+1) mod v1
        mov q1[ebp],eax                 ; store qhat intermediately
        cmp eax,0
        jne @F
        jmp dml                         ; If q=0 process next digit
;
bm1:    mov eax,0ffffffffh              ; eax <- qhat
        mov q1[ebp],eax                 ; Store qhat intermediately
        mov edx,uj1[ebp]                ; edx <- uj+1
        add edx,v1[ebp]                 ; edx <- uj+1 + v1 =: rhat
        jc d4                           ; rhat >= b => v2 * qhat < rhat * b
;
@@:     mov ebx,edx                     ; ebx <- rhat
        mul dword ptr v2[ebp]           ; eax <- low(qhat * v2), edx <- high(qhat * v2)
        cmp edx,ebx                     ; rhat >= high(qhat * v2) ?
        jb  d4                          ; CF == 1?  Then finished
        ja  @F                          ; rhat < high(qhat * v2) => reduce qhat
        cmp eax,uj2[ebp]                ; uj+2 >= low(qhat * v2) ?
        jbe d4                          ; CF == 1 OR ZF == 1? Then finished
;
@@:     dec dword ptr q1[ebp]           ; Correcting qhat by decrementing
        add ebx,v1[ebp]                 ; ebx <- rhat + v1
        jc d4                           ; rhat >= b => v2 * qhat < rhat * b
        sub eax,v2[ebp]                 ; (qhat * v2) - v2
        sbb edx,0
;
        cmp edx,ebx                     ; rhat >= high(qhat * v2) ?
        jb  d4                          ; CF == 1?  Then finished
        ja  @B                          ; rhat < high(qhat * v2) => reduce qhat
        cmp eax,uj2[ebp]                ; uj+2 >= low(qhat * v2) ?
        ja  @B                          ; CF==0 + ZF==0? => repeat decrement of qhat
;
d4:     push    ecx                     ; Store main counter
        push    esi                     ; Store j + l(b) 
        push    edi                     ; Store j
        mov     cx,b[ebp]               ; Load counter with l(b)
        shr     cx,1
        jnc     d4a
        inc     cx
d4a:    mov     ebx,q1[ebp]             ; q in bx
        mov     esi,2
        mov     edx,0                   ; Dummy carry
;
;-----> Multiplication and subtraction
;
dms:    push    edx
        mov     eax,ebx                 ; q in ax
        mul     dword ptr b[ebp][esi]   ; hi:dx lo:ax
        sub     a[ebp][edi],eax
        adc     edx,0                   ; Carry to next subtrahend
        pop     eax
        sub     a[ebp][edi],eax
        adc     edx,0                   ; Carry to next subtrahend
        inc     edi
        inc     edi
        inc     edi
        inc     edi
        inc     esi
        inc     esi
        inc     esi
        inc     esi
        loop    dms
        sub     a[ebp][edi],edx
        jnc     dnc                     ; No correction necessary if carry = 0
;
;-----> Correction
;
        pop     edi
        push    edi
        xor     ecx,ecx
        mov     cx,b[ebp]               ; Load counter with l(b)
        shr     cx,1
        jnc     d4b
        inc     cx
d4b:    mov     esi,2
        clc
d5:     mov     eax,b[ebp][esi]         ; b[i]
        adc     a[ebp][edi],eax         ; a[j+i] + b[i]
        inc     edi
        inc     edi
        inc     edi
        inc     edi
        inc     esi
        inc     esi
        inc     esi
        inc     esi
        loop    d5                      ; Outer loop
        jnc     d51
        inc     dword ptr a[ebp][edi]
d51:    dec     dword ptr q1[ebp]       ; q = q - 1
;
;
dnc:    pop     edi                     ; Get pointers and counters
        pop     esi
        pop     ecx
dml:    mov     eax,q1[ebp]             ; Get q
        mov     q[ebp][edi],eax         ; q[j] = q
        sub     edi,4
        sub     esi,4
        dec     ecx                     ; Inner loop
        jz      d6
        jmp     d3
;
d6:     pop     edi                     ; Get l(a)-l(b)-1 (ULONG digits)
        add     edi,2                   ; di points to low-order byte of last USHORT
        mov     eax,0
@@:     cmp     ax,q[ebp][edi]          ; q[l(a)-l(b)]=0 ?
        jne     d7                      ; If so...
        dec     edi
        dec     edi                     ; ...l(q) = l(a) - l(b) - 1
        cmp     edi,0
        jne     short @B
d7:     mov     edx,edi
        shr     edx,1                   ; l(q) (#USHORTs) in dx
        mov     q[ebp],dx               ; Store l(q)
;
;-----> Calculate length of remainder
;
        mov     eax,0
        mov     ax,b[ebp]
        mov     ecx,eax
        shl     ax,1
        mov     edi,eax
        mov     ebx,0
        inc     edi
        inc     edi
d8:     dec     edi
        dec     edi                     ; di points to low-order byte of a[l(b)]
        cmp     a[ebp][edi],bx
        loope   d8
        jz      d9                      ; If a[bp][di]!= 0...
        inc     cx                      ; ...l(r) has to be incremented
d9:     mov     a[ebp],cx               ; Store l(r)
;
;-----> Store results
;
dstore: mov     esi,ebp
        mov     ebx,ebp
        pop     ebp
        push    esi
        add     esi,q
        mov     edi,_qot                ; Destination offset of quotient
        mov     ecx,0
        mov     cx,[esi]
        shr     cx,1
        jnc     @F
        inc     cx
@@:     cld
        cmp     cx,0
        je      qzero
rep     movsd
qzero:  movsw
;
;-----> Store remainder
;
        pop     esi
drs:
        mov     edi,_rst
        add     esi,a
        mov     ecx,0
        mov     cx,[esi]
        shr     cx,1
        jnc     @F
        inc     cx
@@:     cld
        cmp     cx,0
        je      rzero
rep     movsd
rzero:  movsw
        mov     eax,0                   ; Return-Value = 0: Everything OK
;
;*******************************************************************************
;       Security: Purge stack memory

divret: mov     edi,ebx
        mov     ecx,(WORKSP-4)/4
        cld
rep     stosd                           ; Overwrite with 0 or -1
;*******************************************************************************
;
        pop     esi
        pop     edi
        pop     ebx
        mov     esp,ebp
        pop     ebp
        ret
;
;
divz:   mov     eax,0                   ; Return-Value = 0: Everything OK
        mov     edi,_rst                ; Destination Offset of remainder
        mov     esi,_qot                ; Destination offset of quotient
        mov     [edi],ax                ; Remainder = 0
        mov     [esi],ax                ; Quotient = 0
        jmp     short divret
;
divbyz: mov     eax,-1
        jmp     short divret            ; Return-Value = -1: Division by zero
;
;-----> Case q = 0 and r = a
;
dra:    mov     esi,ebp
        mov     ebx,ebp
        pop     ebp
        mov     edi,_qot
        mov     word ptr [edi],0        ; q = 0
        jmp     short drs               ; Store remainder
;
;
;>>>>>> Short Division
;
;
dshort: mov     ecx,0
        mov     cx,a[ebp]
        shr     cx,1
        jnc     dsh0
        inc     cx
dsh0:   mov     edi,ecx
        shl     edi,2
        sub     edi,2
        mov     edx,0
        mov     ebx,b[ebp+2]
dsh1:   mov     eax,a[ebp][edi]
        div     ebx
        mov     q[ebp][edi],eax
        sub     edi,4
        loop    dsh1
dsh2:   mov     ecx,0
        mov     cx,a[ebp]
        mov     esi,ecx
        shl     esi,1
@@:     mov     bx,q[ebp][esi]
        cmp     bx,0
        jne     dsh3
        dec     esi
        dec     esi
        dec     ecx
        cmp     cx,0
        jne     short @B
dsh3:   mov     q[ebp],cx
        mov     a[ebp+2],edx
        mov     word ptr a[ebp],2
        cmp     word ptr a[ebp+4],0
        jne     dst
        mov     word ptr a[ebp],1
        cmp     word ptr a[ebp+2],0
        jne     dst
        mov     word ptr a[ebp],0
dst:    jmp     dstore
;
div_l  endp
END