MS fastcall 32

[nidud]

Seems the register params in 32-bit are not fixed to param index.

foo proc a1:dword, a2:real4, a3:dword

    add ecx,a1  ; add ecx,ecx
    add edx,a3  ; add edx,edx
    ret         ; ret 4

foo endp

bar proc

;   mov     edx, 3
;   push    0x40000000
;   mov     ecx, 1

    invoke foo, 1, 2.0, 3

;   push    dword ptr [ebx]

    invoke foo, ecx, [ebx], edx
    ret

bar endp

[nidud]

Watcom 16-bit

foo proc a1:word, a2:word, a3:word, a4:word, a5:word
    ret 
;   ret 2
foo endp

bar proc

;   push 5
;   * mov cx, cx 
;   * mov bx, bx
;   * mov dx, dx
;   * mov ax, ax

    invoke foo, ax, dx, bx, cx, 5
;   * add sp, 2
    ret
bar endp
end

[Baron-von-Riedesel]

Seems the register params in 32-bit are not fixed to param index.

Ok, but this isn't new - see ms32_pcheck() in proc.c. Adding line

	if ( paranode->sym.mem_type & MT_FLOAT ) return( 0 );

should fix the "float issue".

invoke foo, ecx, [ebx], edx

What's the problem here?

[nidud]

invoke foo, ecx, [ebx], edx

What's the problem here?

It is the same as the first one, so it should be:

push    dword ptr [ebx]
call    foo

but the float is assign to edx:

push    edx
mov     edx, dword ptr [ebx]
call    foo

should fix the "float issue".

I've done a rewrite of the fast-call handling in an attempt to have a more rule-based implementation rather than all these exceptions. I'm currently making some adjustment and testing, hence the reason for these "discoveries".

Another odd thing that pops up is the float handling in 64-bit for Varargs (and Unprototyped functions according to the Win64 ABI).

• Floating-point values are primarily assign to SIMD registers. Exclusively in System V, but Win64 also adds a copy to the integer register if available (may not be the case in VECTORCALL).
• The default type is double precision, so float values needs to be converted.

So you end up with something like this:

    .data
     d real8 2.0
     f real4 6.0
...
    printf(rax, d, xmm1, real4 ptr xmm2, 5.0, f)

    * cvtss2sd xmm0, dword ptr [f]
    * movsd   qword ptr [rsp+28H], xmm0
    * mov     dword ptr [rsp+20H], 0x00000000
    * mov     dword ptr [rsp+24H], 0x40140000
    * cvtss2sd xmm3, xmm2
    * movq    r9, xmm3
    * movsd   xmm2, xmm1
    * movq    r8, xmm2
    * movsd   xmm1, qword ptr [d]
    * movq    rdx, xmm1
    * mov     rcx, rax
    * call    printf