inside the GHC code generator

[kyra]

Bulat Ziganshin wrote:

> i prefer to see the asm code. this may be because of better high-level
> optimization strategies (reusing fib values). the scheme about i say
> will combine advantages of both worlds
no strategies, plain exponential algorithm,

ocaml:
<excerpt>
_camlFibo__fib_57:
    sub    esp, 8
L101:
    cmp    eax, 5
    jge    L100
    mov    eax, 3
    add    esp, 8
    ret
L100:
    mov    DWORD PTR 0[esp], eax
    add    eax, -4
    call    _camlFibo__fib_57
L102:
    mov    DWORD PTR 4[esp], eax
    mov    eax, DWORD PTR 0[esp]
    add    eax, -2
    call    _camlFibo__fib_57
L103:
    mov    ebx, DWORD PTR 4[esp]
    add    eax, ebx
    dec    eax
    add    esp, 8
    ret
</excerpt>

visual C++ 7.1 (next fastest):
<excerpt>
_fib    PROC NEAR
    push    esi
    mov    esi, DWORD PTR 8[esp]
    cmp    esi, 2
    jge    SHORT $L945
    mov    eax, 1
    pop    esi
    ret    0
$L945:
    lea    eax, DWORD PTR [esi-2]
    push    edi
    push    eax
    call    _fib
    dec    esi
    push    esi
    mov    edi, eax
    call    _fib
    add    esp, 8
    add    eax, edi
    pop    edi
    pop    esi
    ret    0
_fib    ENDP
</excerpt>

also, Clean is *EXACTLY* in line with ocaml. This is interesting, 
because Clean is so much similar to Haskell.