This is a really cool algorithm I found on-line, and I'm trying to
clean up the code and make it more useful.  Arg1 is the source address
of the plaintext (8 bytes) and arg2 is the destination.  delta, dis,
and n are just parts of the algorithm.  k is the 16 byte key.

There are 2 things I'd like to improve.  First, I'm not sure it uses
the best register choice in several places.  I'd like any thoughts you
might have on that.  Second, and this ties in with the first, the key
value should really be a parameter.  But since it requires 16 bytes, it
wouldn't fit in a register.  I'd like to avoid breaking it up into 4
parts.  But it seems that every register is already being used.  Maybe
locals is the solution

So, is there a better choice of register usage.  And would it make
sense just to set up 4 locals with each of the 4 dwords from the key
before the algorithm starts, and just use them as necessary?

delta   equ 0x9e3779b9
dis     equ 0xc6ef3720
n       equ 32

section .data

k       dd 123,456,789,112

section .text

; --------------------------------------
tea_encode:
    BEGIN
    mov eax,arg1
    mov ebx,[eax+4]
    mov eax,[eax]
    mov ecx,n       ; n (count)
    xor edx,edx     ; sum
..enstart:
    mov esi,ebx
    mov edi,esi     ; esi, edi = z
    shl esi,4
    shr edi,5
    xor edi,esi
    add edi,ebx
    mov esi,edx
    and esi,3
    mov esi,[k+esi*4]
    add esi,edx
    xor esi,edi
    add eax,esi
    add edx,delta
    mov esi,eax
    mov edi,esi
    shl esi,4
    shr edi,5
    xor edi,esi
    add edi,eax
    mov esi,edx
    shr esi,11
    and esi,3
    mov esi,[k+esi*4]
    add esi,edx
    xor esi,edi
    add ebx,esi
    loop .enstart
    mov edi,arg2
    mov [edi+0*4],eax
    mov [edi+1*4],ebx
    END

; --------------------------------------

Jonathon,

This does not answer your question, but just wanted to let you know
that if this is the basic TEA algorithm, there is an improved version
called XTEA.
This is more secure. Maybe you should use that.

Kind regards
Eddy