Naked return value

[deltarho[1859]]

I read somewhere that we can return a Naked double via xmm0, but I cannot find the statement now.

Code: Select all

Function foo Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
    mov edx, Dword Ptr [esp+4]
    cvtsi2sd xmm0, edx
    ret
  End Asm
End Function
 
Print foo(66)
 
Sleep

I was hoping to get 66 but get -1.#IND

[jj2007]

Indeed, you can return a "naked" double via xmm0, and your code puts the double 66 correctly in xmm0, but FB print expects its double on the FPU:

Code: Select all

Function foo Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
'     mov edx, Dword Ptr [esp+4]
'     cvtsi2sd xmm0, edx
    fild Dword Ptr [esp+4]
    ret
  End Asm
End Function

[deltarho[1859]]

Thanks Jochen.

Suppose after a few lines of code we get a scalar double result in xmm0 how do we put xmm0 into the FPU?

[marcov]

On Windows, XMM for FPU is default for 64-bit, not for 32-bit.

[deltarho[1859]]

I would like to stay on topic.

Indeed, you can return a "naked" double via xmm0

How?

[jj2007]

Suppose after a few lines of code we get a scalar double result in xmm0 how do we put xmm0 into the FPU?

Code: Select all

Function foo1 Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
    fild Dword Ptr [esp+4]	' simple...
    ret
  End Asm
End Function

Function foo2 Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
    cvtsi2sd xmm0, dword Ptr [esp+4]	' get value as integer, convert to double
    sub esp, 8								' create a qword slot
    movlps qword ptr [esp], xmm0		' load double into memory
    fld qword ptr [esp]					' transfer to the FPU
    add esp, 8								' correct the stack
    ret
  End Asm
End Function

Print "directly to the FPU:", foo1(12345)
Print "with acrobatics:", foo2(12345)
Sleep

[deltarho[1859]]

Thanks Jochen.

foo2 solves the opening post perfectly.

What would be nice is 'fld xmm0' but the FPU architecture hasn't seen any work on it for years.

Code: Select all

#macro fld_xmm0
  sub esp, 8                        ' create a qword slot
  movlps qword ptr [esp], xmm0      ' load double into memory
  fld qword ptr [esp]               ' transfer to the FPU
  add esp, 8
#endmacro
 
Function foo2 Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
    cvtsi2sd xmm0, dword Ptr [esp+4]
    fld_xmm0
    ret
  End Asm
End Function

So, given some work where xmm0 is a scalar double we simply use

Code: Select all

fld_xmm0
ret

where the Naked return is Double.

fxm may like to add to the Wiki again because all the examples have Naked returning Longs only and no mention of a Double return involves the FPU.

[SARG]

With 64bit it's shorter.

Code: Select all

Function foo2 Naked ( ByVal x As Ulong ) As double
  Asm
	cvtsi2sd xmm0, ecx
	ret
  End Asm
End Function
print foo2(57897897)
sleep

[deltarho[1859]]

Ah, I wondered what marcov was getting at.

So, 'Compiler Option: -fpu' is incomplete.

We need to do a platform check to test whether my fld_xmm0 macro is needed. Life is getting too complicated - one step forward, two steps back.

[jj2007]

With 64bit it's shorter.

Codesize-wise, it's identical:

Code: Select all

Function foo Naked cdecl ( ByVal x As Ulong ) As Double
  Asm
  int 3
  #ifdef __FB_64BIT__
	cvtsi2sd xmm0, ecx
  #else
	fild Dword Ptr [esp+4]
	ret
  #endif
  nop
  End Asm
End Function

Dim as double my123
asm int 3
my123=foo(1234567)
asm nop
asm nop
Print "here it is: ";my123
Sleep

Tested with standard GAS for 32 bits and SARG's version for 64 bits. One interesting point is that the 64-bit compiler inserts the ret for you - 32-bit Gas doesn't.

[jj2007]

On Windows, XMM for FPU is default for 64-bit, not for 32-bit.

Ah, I wondered what marcov was getting at.

What he actually meant is "On Windows, XMM for returning doubles is default for 64-bit". The OS uses xmmregs for doubles. As a coder, you can use the fpu for whatever you like, though.

[SARG]

Codesize-wise, it's identical:

I agree.
I compared only foo2 functions :-)


@deltarho[1859]
XMM for returning doubles and singles on Windows and Linux.