kolibrios-gitea/kernel/trunk/blkdev/hd_drv.inc

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;                                                              ;;
;; Copyright (C) KolibriOS team 2004-2008. All rights reserved. ;;
;; Distributed under terms of the GNU General Public License    ;;
;;                                                              ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

$Revision$


; Low-level driver for HDD access
; DMA support by Mario79
; Access through BIOS by diamond

align 4
hd_read:
;-----------------------------------------------------------
; input  : eax = block to read
;          ebx = destination
;-----------------------------------------------------------
        and     [hd_error], 0
    push  ecx esi edi           ; scan cache

;    mov   ecx,cache_max         ; entries in cache
;    mov   esi,HD_CACHE+8
    call  calculate_cache
    add   esi,8

    mov   edi,1

  hdreadcache:

    cmp   dword [esi+4],0       ; empty
    je    nohdcache

    cmp   [esi],eax             ; correct sector
    je    yeshdcache

  nohdcache:

    add   esi,8
    inc   edi
    dec   ecx
    jnz   hdreadcache

    call  find_empty_slot       ; ret in edi
    cmp   [hd_error],0
    jne   return_01
; Read through BIOS?
        cmp     [hdpos], 0x80
        jae     .bios
; DMA read is permitted if [allow_dma_access]=1 or 2
        cmp     [allow_dma_access], 2
        ja      .nodma
        cmp     [dma_hdd], 1
        jnz     .nodma
        call    hd_read_dma
        jmp     @f
.nodma:
        call    hd_read_pio
        jmp     @f
.bios:
        call    bd_read
@@:
        cmp     [hd_error], 0
        jne     return_01
;    lea   esi,[edi*8+HD_CACHE]
;    push  eax
    call  calculate_cache_1
    lea   esi,[edi*8+esi]
;    pop   eax

    mov   [esi],eax             ; sector number
    mov   dword [esi+4],1       ; hd read - mark as same as in hd

  yeshdcache:

    mov   esi,edi
    shl   esi,9
;    add   esi,HD_CACHE+65536
    push  eax
    call  calculate_cache_2
    add   esi,eax
    pop   eax

    mov   edi,ebx
    mov   ecx,512/4
    cld
    rep   movsd                 ; move data
 return_01:
    pop   edi esi ecx
    ret

align 4
hd_read_pio:
    push  eax edx

    call  wait_for_hd_idle
    cmp   [hd_error],0
    jne   hd_read_error

    cli
    xor   eax,eax
    mov   edx,[hdbase]
    inc   edx
    out   dx,al   ; ATAFeatures ॣ<EFBFBD><EFBFBD><EFBFBD><EFBFBD> "<EFBFBD>ᮡ<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>⥩"
    inc   edx
    inc   eax
    out   dx,al   ; ATASectorCount <EFBFBD><EFBFBD><EFBFBD><EFBFBD>稪 ᥪ<EFBFBD><EFBFBD>஢
    inc   edx
    mov   eax,[esp+4]
    out   dx,al   ; ATASectorNumber ॣ<EFBFBD><EFBFBD><EFBFBD><EFBFBD> <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> ᥪ<EFBFBD><EFBFBD><EFBFBD><EFBFBD>
    shr   eax,8
    inc   edx
    out   dx,al   ; ATACylinder <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> 樫<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> (<EFBFBD><EFBFBD><EFBFBD><EFBFBD>訩 <EFBFBD><EFBFBD><EFBFBD><EFBFBD>)
    shr   eax,8
    inc   edx
    out   dx,al   ; <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> 樫<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> (<EFBFBD><EFBFBD><EFBFBD><EFBFBD>訩 <EFBFBD><EFBFBD><EFBFBD><EFBFBD>)
    shr   eax,8
    inc   edx
    and   al,1+2+4+8
    add   al,byte [hdid]
    add   al,128+64+32
    out   dx,al   ; <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>/<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD> <EFBFBD><EFBFBD>᪠
    inc   edx
    mov   al,20h
    out   dx,al   ; ATACommand ॣ<EFBFBD><EFBFBD><EFBFBD><EFBFBD> <EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
    sti

    call  wait_for_sector_buffer

    cmp   [hd_error],0
    jne   hd_read_error

    cli
    push  edi
    shl   edi,9
;    add   edi,HD_CACHE+65536
    push  eax
    call  calculate_cache_2
    add   edi,eax
    pop   eax

    mov   ecx,256
    mov   edx,[hdbase]
    cld
    rep   insw
    pop   edi
    sti

    pop   edx eax
    ret

disable_ide_int:
;        mov edx,[hdbase]
;        add edx,0x206
;        mov al,2
;        out dx,al
        cli
        ret

enable_ide_int:
;        mov edx,[hdbase]
;        add edx,0x206
;        mov al,0
;        out dx,al
        sti
        ret

align 4
hd_write:
;-----------------------------------------------------------
; input  : eax = block
;          ebx = pointer to memory
;-----------------------------------------------------------
    push  ecx esi edi

    ; check if the cache already has the sector and overwrite it

;    mov   ecx,cache_max
;    mov   esi,HD_CACHE+8
    call  calculate_cache
    add   esi,8

    mov   edi,1

  hdwritecache:

    cmp   dword [esi+4],0       ; if cache slot is empty
    je    not_in_cache_write

    cmp   [esi],eax             ; if the slot has the sector
    je    yes_in_cache_write

  not_in_cache_write:

    add   esi,8
    inc   edi
    dec   ecx
    jnz   hdwritecache

    ; sector not found in cache
    ; write the block to a new location

    call  find_empty_slot       ; ret in edi
    cmp   [hd_error],0
    jne   hd_write_access_denied

;    lea   esi,[edi*8+HD_CACHE]
;    push  eax
    call  calculate_cache_1
    lea   esi,[edi*8+esi]
;    pop   eax

    mov   [esi],eax             ; sector number

  yes_in_cache_write:

    mov   dword [esi+4],2       ; write - differs from hd

    shl   edi,9
;    add   edi,HD_CACHE+65536
    push  eax
    call  calculate_cache_2
    add   edi,eax
    pop   eax

    mov   esi,ebx
    mov   ecx,512/4
    cld
    rep   movsd                 ; move data
 hd_write_access_denied:
    pop   edi esi ecx
    ret

align 4
cache_write_pio:
;    call  disable_ide_int

    call  wait_for_hd_idle
    cmp   [hd_error],0
    jne   hd_write_error

    cli
    xor   eax,eax
    mov   edx,[hdbase]
    inc   edx
    out   dx,al
    inc   edx
    inc   eax
    out   dx,al
    inc   edx
    mov   eax,[esi]             ; eax = sector to write
    out   dx,al
    shr   eax,8
    inc   edx
    out   dx,al
    shr   eax,8
    inc   edx
    out   dx,al
    shr   eax,8
    inc   edx
    and   al,1+2+4+8
    add   al,byte [hdid]
    add   al,128+64+32
    out   dx,al
    inc   edx
    mov   al,30h
    out   dx,al
    sti

    call  wait_for_sector_buffer

    cmp   [hd_error],0
    jne   hd_write_error

    push  ecx esi

    cli
    mov   esi,edi
    shl   esi,9
;    add   esi,HD_CACHE+65536    ; esi = from memory position
    push  eax
    call  calculate_cache_2
    add   esi,eax
    pop   eax

    mov   ecx,256
    mov   edx,[hdbase]
    cld
    rep   outsw
    sti

;    call  enable_ide_int
    pop   esi ecx

    ret

save_hd_wait_timeout:

    push  eax
    mov   eax,[timer_ticks]
    add   eax,300               ; 3 sec timeout
    mov   [hd_wait_timeout],eax
    pop   eax
    ret

align 4
check_hd_wait_timeout:

    push  eax
    mov   eax,[hd_wait_timeout]
    cmp   [timer_ticks], eax
    jg    hd_timeout_error
    pop   eax
    mov   [hd_error],0
    ret

;iglobal
;  hd_timeout_str   db 'K : FS - HD timeout',0
;  hd_read_str      db 'K : FS - HD read error',0
;  hd_write_str     db 'K : FS - HD write error',0
;  hd_lba_str       db 'K : FS - HD LBA error',0
;endg

hd_timeout_error:

;    call  clear_hd_cache
;    call  clear_application_table_status
;    mov   esi,hd_timeout_str
;    call  sys_msg_board_str
    DEBUGF 1,"K : FS - HD timeout\n"

    mov   [hd_error],1
    pop   eax
    ret

hd_read_error:

;    call  clear_hd_cache
;    call  clear_application_table_status
;    mov   esi,hd_read_str
;    call  sys_msg_board_str
    DEBUGF 1,"K : FS - HD read error\n"
    pop   edx eax
    ret

hd_write_error:

;    call  clear_hd_cache
;    call  clear_application_table_status
;     mov   esi,hd_write_str
;     call  sys_msg_board_str
    DEBUGF 1,"K : FS - HD write error\n"
     ret

hd_write_error_dma:
;        call    clear_hd_cache
;        call    clear_application_table_status
;        mov     esi, hd_write_str
;        call    sys_msg_board_str
        DEBUGF 1,"K : FS - HD read error\n"
        pop     esi
        ret

hd_lba_error:
;    call  clear_hd_cache
;    call  clear_application_table_status
;    mov   esi,hd_lba_str
;    call  sys_msg_board_str
     DEBUGF 1,"K : FS - HD LBA error\n"
     jmp   LBA_read_ret


align 4
wait_for_hd_idle:

    push  eax edx

    call  save_hd_wait_timeout

    mov   edx,[hdbase]
    add   edx,0x7

  wfhil1:

    call  check_hd_wait_timeout
    cmp   [hd_error],0
    jne   @f

    in    al,dx
    test  al,128
    jnz   wfhil1

 @@:

    pop   edx eax
    ret


align 4
wait_for_sector_buffer:

    push  eax edx

    mov   edx,[hdbase]
    add   edx,0x7

    call  save_hd_wait_timeout

  hdwait_sbuf:                  ; wait for sector buffer to be ready

    call  check_hd_wait_timeout
    cmp   [hd_error],0
    jne   @f

    in    al,dx
    test  al,8
    jz    hdwait_sbuf

    mov   [hd_error],0

    cmp   [hd_setup],1          ; do not mark error for setup request
    je    buf_wait_ok

    test  al,1                  ; previous command ended up with an error
    jz    buf_wait_ok
 @@:
    mov   [hd_error],1

  buf_wait_ok:

    pop   edx eax
    ret

; \begin{Mario79}
align 4
wait_for_sector_dma_ide0:
        push    eax
        push    edx
        call    save_hd_wait_timeout
.wait:
        call    change_task
        cmp     [irq14_func], hdd_irq14
        jnz     .done
        call    check_hd_wait_timeout
        cmp     [hd_error], 0
        jz      .wait
        mov     [irq14_func], hdd_irq_null
        mov     dx, [IDEContrRegsBaseAddr]
        mov     al, 0
        out     dx, al
.done:
        pop     edx
        pop     eax
        ret

align 4
wait_for_sector_dma_ide1:
        push    eax
        push    edx
        call    save_hd_wait_timeout
.wait:
        call    change_task
        cmp     [irq15_func], hdd_irq15
        jnz     .done
        call    check_hd_wait_timeout
        cmp     [hd_error], 0
        jz      .wait
        mov     [irq15_func], hdd_irq_null
        mov     dx, [IDEContrRegsBaseAddr]
        add     dx, 8
        mov     al, 0
        out     dx, al
.done:
        pop     edx
        pop     eax
        ret

iglobal
align 4
; note that IDE descriptor table must be 4-byte aligned and do not cross 4K boundary
IDE_descriptor_table:
        dd      IDE_DMA
        dw      0x2000
        dw      0x8000

dma_cur_sector  dd      not 40h
dma_hdpos       dd      0
irq14_func      dd      hdd_irq_null
irq15_func      dd      hdd_irq_null
endg

uglobal
; all uglobals are zeroed at boot
dma_process     dd      0
dma_slot_ptr    dd      0
cache_chain_pos dd      0
cache_chain_ptr dd      0
cache_chain_size        db      0
cache_chain_started     db      0
dma_task_switched       db      0
dma_hdd         db      0
allow_dma_access db      0
endg

align 4
hdd_irq14:
        pushfd
        cli
        pushad
        mov     [irq14_func], hdd_irq_null
        mov     dx, [IDEContrRegsBaseAddr]
        mov     al, 0
        out     dx, al
;        call    update_counters
;        mov     ebx, [dma_process]
;        cmp     [CURRENT_TASK], ebx
;        jz      .noswitch
;        mov     [dma_task_switched], 1
;        mov     edi, [dma_slot_ptr]
;        mov     eax, [CURRENT_TASK]
;        mov     [dma_process], eax
;        mov     eax, [TASK_BASE]
;        mov     [dma_slot_ptr], eax
;        mov     [CURRENT_TASK], ebx
;        mov     [TASK_BASE], edi
;        mov     byte [DONT_SWITCH], 1
;        call    do_change_task
.noswitch:
        popad
        popfd
align 4
hdd_irq_null:
        ret

align 4
hdd_irq15:
        pushfd
        cli
        pushad
        mov     [irq15_func], hdd_irq_null
        mov     dx, [IDEContrRegsBaseAddr]
        add     dx, 8
        mov     al, 0
        out     dx, al
;        call    update_counters
;        mov     ebx, [dma_process]
;        cmp     [CURRENT_TASK], ebx
;        jz      .noswitch
;        mov     [dma_task_switched], 1
;        mov     edi, [dma_slot_ptr]
;        mov     eax, [CURRENT_TASK]
;        mov     [dma_process], eax
;        mov     eax, [TASK_BASE]
;        mov     [dma_slot_ptr], eax
;        mov     [CURRENT_TASK], ebx
;        mov     [TASK_BASE], edi
;        mov     byte [DONT_SWITCH], 1
;        call    do_change_task
.noswitch:
        popad
        popfd
        ret

align 4
hd_read_dma:
        push    eax
        push    edx
        mov     edx,[dma_hdpos]
        cmp     edx,[hdpos]
        jne     .notread
        mov     edx, [dma_cur_sector]
        cmp     eax, edx
        jb      .notread
        add     edx, 15
        cmp     [esp+4], edx
        ja      .notread
        mov     eax, [esp+4]
        sub     eax, [dma_cur_sector]
        shl     eax, 9
        add     eax, (OS_BASE+IDE_DMA)
        push    ecx esi edi
        mov     esi, eax
        shl     edi, 9
;        add     edi, HD_CACHE+0x10000
        push  eax
        call  calculate_cache_2
        add   edi,eax
        pop   eax

        mov     ecx, 512/4
        cld
        rep     movsd
        pop     edi esi ecx
        pop     edx
        pop     eax
        ret
.notread:
        mov     eax, IDE_descriptor_table
        mov     dword [eax],  IDE_DMA
        mov     word [eax+4], 0x2000
        sub     eax, OS_BASE
        mov     dx, [IDEContrRegsBaseAddr]
        cmp     [hdbase], 0x1F0
        jz      @f
        add     edx, 8
@@:
        push    edx
        add     edx, 4
        out     dx, eax
        pop     edx
        mov     al, 0
        out     dx, al
        add     edx, 2
        mov     al, 6
        out     dx, al
        call    wait_for_hd_idle
        cmp     [hd_error], 0
        jnz     hd_read_error
        call    disable_ide_int
        xor     eax, eax
        mov     edx, [hdbase]
        inc     edx
        out     dx, al
        inc     edx
        mov     eax, 10h
        out     dx, al
        inc     edx
        mov     eax, [esp+4]
        out     dx, al
        shr     eax, 8
        inc     edx
        out     dx, al
        shr     eax, 8
        inc     edx
        out     dx, al
        shr     eax, 8
        inc     edx
        and     al, 0xF
        add     al, byte [hdid]
        add     al, 11100000b
        out     dx, al
        inc     edx
        mov     al, 0xC8
        out     dx, al
        mov     dx, [IDEContrRegsBaseAddr]
        cmp     [hdbase], 0x1F0
        jz      @f
        add     dx, 8
@@:
        mov     al, 9
        out     dx, al
        mov     eax, [CURRENT_TASK]
        mov     [dma_process], eax
        mov     eax, [TASK_BASE]
        mov     [dma_slot_ptr], eax
        cmp     [hdbase], 0x1F0
        jnz     .ide1
        mov     [irq14_func], hdd_irq14
        jmp     @f
.ide1:
        mov     [irq15_func], hdd_irq15
@@:
        call    enable_ide_int
        cmp     [hdbase], 0x1F0
        jnz     .wait_ide1
        call    wait_for_sector_dma_ide0
        jmp     @f
.wait_ide1:
        call    wait_for_sector_dma_ide1
@@:
        cmp     [hd_error], 0
        jnz     hd_read_error
        mov     eax,[hdpos]
        mov     [dma_hdpos],eax
        pop     edx
        pop     eax
        mov     [dma_cur_sector], eax
        jmp     hd_read_dma

align 4
write_cache_sector:
        mov     [cache_chain_size],1
        mov     [cache_chain_pos],edi
write_cache_chain:
        cmp     [hdpos], 0x80
        jae     bd_write_cache_chain
        push    esi
        mov     eax, IDE_descriptor_table
        mov     edx,eax
        pusha
        mov     esi,[cache_chain_pos]
        shl     esi, 9
        call    calculate_cache_2
        add     esi,eax
        mov     edi, (OS_BASE+IDE_DMA)
        mov     dword [edx], IDE_DMA
        movzx   ecx, [cache_chain_size]
        shl     ecx, 9
        mov     word [edx+4], cx
        shr     ecx,2
        cld
        rep movsd
        popa
        sub     eax, OS_BASE
        mov     dx, [IDEContrRegsBaseAddr]
        cmp     [hdbase], 0x1F0
        jz      @f
        add     edx, 8
@@:
        push    edx
        add     edx, 4
        out     dx, eax
        pop     edx
        mov     al, 0
        out     dx, al
        add     edx, 2
        mov     al, 6
        out     dx, al
        call    wait_for_hd_idle
        cmp     [hd_error], 0
        jnz     hd_write_error_dma
        call    disable_ide_int
        xor     eax, eax
        mov     edx, [hdbase]
        inc     edx
        out     dx, al
        inc     edx
        mov     al, [cache_chain_size]
        out     dx, al
        inc     edx
        mov     esi, [cache_chain_ptr]
        mov     eax, [esi]
        out     dx, al
        shr     eax, 8
        inc     edx
        out     dx, al
        shr     eax, 8
        inc     edx
        out     dx, al
        shr     eax, 8
        inc     edx
        and     al, 0xF
        add     al, byte [hdid]
        add     al, 11100000b
        out     dx, al
        inc     edx
        mov     al, 0xCA
        out     dx, al
        mov     dx, [IDEContrRegsBaseAddr]
        cmp     [hdbase], 0x1F0
        jz      @f
        add     dx, 8
@@:
        mov     al, 1
        out     dx, al
        mov     eax, [CURRENT_TASK]
        mov     [dma_process], eax
        mov     eax, [TASK_BASE]
        mov     [dma_slot_ptr], eax
        cmp     [hdbase], 0x1F0
        jnz     .ide1
        mov     [irq14_func], hdd_irq14
        jmp     @f
.ide1:
        mov     [irq15_func], hdd_irq15
@@:
        call    enable_ide_int
        mov     [dma_cur_sector], not 0x40
        cmp     [hdbase], 0x1F0
        jnz     .wait_ide1
        call    wait_for_sector_dma_ide0
        jmp     @f
.wait_ide1:
        call    wait_for_sector_dma_ide1
@@:
        cmp     [hd_error], 0
        jnz     hd_write_error_dma
        pop     esi
        ret

uglobal
IDEContrRegsBaseAddr    dw      ?
endg
; \end{Mario79}

; \begin{diamond}
uglobal
bios_hdpos      dd      0       ; 0 is invalid value for [hdpos]
bios_cur_sector dd      ?
bios_read_len   dd      ?
endg
bd_read:
        push    eax
        push    edx
        mov     edx, [bios_hdpos]
        cmp     edx, [hdpos]
        jne     .notread
        mov     edx, [bios_cur_sector]
        cmp     eax, edx
        jb      .notread
        add     edx, [bios_read_len]
        dec     edx
        cmp     eax, edx
        ja      .notread
        sub     eax, [bios_cur_sector]
        shl     eax, 9
        add     eax, (OS_BASE+0x9A000)
        push    ecx esi edi
        mov     esi, eax
        shl     edi, 9
;        add     edi, HD_CACHE+0x10000
        push  eax
        call  calculate_cache_2
        add   edi,eax
        pop   eax

        mov     ecx, 512/4
        cld
        rep     movsd
        pop     edi esi ecx
        pop     edx
        pop     eax
        ret
.notread:
        push    ecx
        mov     dl, 42h
        mov     ecx, 16
        call    int13_call
        pop     ecx
        test    eax, eax
        jnz     .v86err
        test    edx, edx
        jz      .readerr
        mov     [bios_read_len], edx
        mov     edx, [hdpos]
        mov     [bios_hdpos], edx
        pop     edx
        pop     eax
        mov     [bios_cur_sector], eax
        jmp     bd_read
.readerr:
.v86err:
        mov     [hd_error], 1
        jmp     hd_read_error

bd_write_cache_chain:
        pusha
        mov     esi, [cache_chain_pos]
        shl     esi, 9
        call    calculate_cache_2
        add     esi, eax
        mov     edi, OS_BASE + 0x9A000
        movzx   ecx, [cache_chain_size]
        push    ecx
        shl     ecx, 9-2
        rep     movsd
        pop     ecx
        mov     dl, 43h
        mov     eax, [cache_chain_ptr]
        mov     eax, [eax]
        call    int13_call
        test    eax, eax
        jnz     .v86err
        cmp     edx, ecx
        jnz     .writeerr
        popa
        ret
.v86err:
.writeerr:
        popa
        mov     [hd_error], 1
        jmp     hd_write_error

uglobal
int13_regs_in   rb      v86_regs.size
int13_regs_out  rb      v86_regs.size
endg

int13_call:
; Because this code uses fixed addresses,
; it can not be run simultaniously by many threads.
; In current implementation it is protected by common mutex 'hd1_status'
        mov     word [OS_BASE + 510h], 10h             ; packet length
        mov     word [OS_BASE + 512h], cx              ; number of sectors
        mov     dword [OS_BASE + 514h], 9A000000h      ; buffer 9A00:0000
        mov     dword [OS_BASE + 518h], eax
        and     dword [OS_BASE + 51Ch], 0
        push    ebx ecx esi edi
        mov     ebx, int13_regs_in
        mov     edi, ebx
        mov     ecx, v86_regs.size/4
        xor     eax, eax
        rep     stosd
        mov     byte [ebx+v86_regs.eax+1], dl
        mov     eax, [hdpos]
        lea     eax, [BiosDisksData+(eax-80h)*4]
        mov     dl, [eax]
        mov     byte [ebx+v86_regs.edx], dl
        movzx   edx, byte [eax+1]
;        mov     dl, 5
        test    edx, edx
        jnz     .hasirq
        dec     edx
        jmp     @f
.hasirq:
        pushad
        stdcall enable_irq, edx
        popad
@@:
        mov     word [ebx+v86_regs.esi], 510h
        mov     word [ebx+v86_regs.ss], 9000h
        mov     word [ebx+v86_regs.esp], 0A000h
        mov     word [ebx+v86_regs.eip], 500h
        mov     [ebx+v86_regs.eflags], 20200h
        mov     esi, [sys_v86_machine]
        mov     ecx, 0x502
        call    v86_start
        and     [bios_hdpos], 0
        pop     edi esi ecx ebx
        movzx   edx, byte [OS_BASE + 512h]
        test    byte [int13_regs_out+v86_regs.eflags], 1
        jnz     @f
        mov     edx, ecx
@@:
        ret
; \end{diamond}