define pope pop
use32

; [80386]

mov   eax,$01234567
mov   ecx,eax
mov   edx,eax
mov   esi,eax

; bswap esi ; Prohibited 

nop
nop

; 22 Bytes, 7 instructions, 22+y cycles, uses stack,
; __NO__ register trashed (except stack), EAX
push  eax        ; 2
shl   eax,8      ; 3
mov   al,[esp+1] ; 4 ?
shl   eax,16     ; 3
mov   ah,[esp+2] ; 4 ?
mov   al,[esp+3] ; 4 ?
add   esp,4      ; 2 | POPE nothing 

nop
nop
mov   edi,eax
nop
nop

; 14 Bytes, 8 instructions, 24+4x cycles, uses stack,
; 1 register trashed, EAX
push  eax    ; 2
pope  ax     ; 4+x
pope  bx     ; 4+x
xchg  al,ah  ; 3
xchg  bl,bh  ; 3
push  ax     ; 2+x
push  bx     ; 2+x
pope  eax    ; 4

nop
nop

; 15 Bytes, 6 instructions, 16 cycles, 1 register trashed, ECX
xor   ebx,ebx    ; 2 | MOVNTQ EBX,0
xchg  cl,ch      ; 3
shrd  ebx,ecx,16 ; 3 | ECX unmodified !!!
shr   ecx,16     ; 3
xchg  cl,ch      ; 3
or    ecx,ebx    ; 2 | Add high 16 bits into result

nop
nop

; 14 Bytes, 5 instructions, 14+x cycles, 1 register trashed,
; result in different register: BSWAP EBX<-EDX
xchg  dl,dh      ; 3
shrd  ebx,edx,16 ; 3 | EDX unmodified !!!
shr   edx,16     ; 3
xchg  dl,dh      ; 3
mov   bx,dx      ; 2+x | Add low 16 bits into result

nop
nop

xor   edx,edx
ud2
    

macro bswap reg {
    if reg eq eax
   xchg    ah,al
       ror     eax,16
      xchg    ah,al
    else if reg eq ebx
     xchg    bh,bl
       ror     ebx,16
      xchg    bh,bl
    else if reg eq ecx
     xchg    ch,cl
       ror     ecx,16
      xchg    ch,cl
    else if reg eq edx
     xchg    dh,dl
       ror     edx,16
      xchg    dh,dl
    else
   xchg    reg,eax
     xchg    ah,al
       ror     eax,16
      xchg    ah,al
       xchg    reg,eax
    end if
}