A question about register constraint of inline assembly

A question about register constraint of inline assembly

[Lin Chun-Zhi]

Hi there,

I got a question about register constraint in kernel source.
Here is the cut of bios32_service()

597         __asm__("lcall (%%edi); cld"
598                 : "=a" (return_code),
599                   "=b" (address),
600                   "=c" (length),
601                   "=d" (entry)
602                 : "" (service),
603                   "1" (0),
604                   "D" (&bios32_indirect));

In line 602, its register constraint is a NULL string. I can't find how
compiler arranges the registers in gcc.info.

Does anyone know the further information?

Kindly regards,
Roach
--
Open WebMail Project (http://openwebmail.org)

Re: A question about register constraint of inline assembly

[Ian Lance Taylor]

"Lin Chun-Zhi" <lcz90@cs.ccu.edu.tw> writes:

> I got a question about register constraint in kernel source.
> Here is the cut of bios32_service()
> 
> 597         __asm__("lcall (%%edi); cld"
> 598                 : "=a" (return_code),
> 599                   "=b" (address),
> 600                   "=c" (length),
> 601                   "=d" (entry)
> 602                 : "" (service),
> 603                   "1" (0),
> 604                   "D" (&bios32_indirect));
> 
> In line 602, its register constraint is a NULL string. I can't find how
> compiler arranges the registers in gcc.info.

If there are no constraints, the compiler does not constrain the
operand.  It could be anywhere.  The statement will use "service",
though; assignments to it will not be deleted.

By the way, in linux-2.6.8.1 that statement looks like this:

	__asm__("lcall *(%%edi); cld"
		: "=a" (return_code),
		  "=b" (address),
		  "=c" (length),
		  "=d" (entry)
		: "0" (service),
		  "1" (0),
		  "D" (&bios32_indirect));

Ian

Re: A question about register constraint of inline assembly

[Lin Chun-Zhi]

On 01 Jun 2005 23:23:46 -0400, Ian Lance Taylor wrote
> "Lin Chun-Zhi" <lcz90@cs.ccu.edu.tw> writes:
> 
> > I got a question about register constraint in kernel source.
> > Here is the cut of bios32_service()
> > 
> > 597         __asm__("lcall (%%edi); cld"
> > 598                 : "=a" (return_code),
> > 599                   "=b" (address),
> > 600                   "=c" (length),
> > 601                   "=d" (entry)
> > 602                 : "" (service),
> > 603                   "1" (0),
> > 604                   "D" (&bios32_indirect));
> > 
> > In line 602, its register constraint is a NULL string. I can't find how
> > compiler arranges the registers in gcc.info.
> 
> If there are no constraints, the compiler does not constrain the
> operand.  It could be anywhere.  The statement will use "service",
> though; assignments to it will not be deleted.
> 
> By the way, in linux-2.6.8.1 that statement looks like this:
> 
> 	__asm__("lcall *(%%edi); cld"
> 		: "=a" (return_code),
> 		  "=b" (address),
> 		  "=c" (length),
> 		  "=d" (entry)
> 		: "0" (service),
> 		  "1" (0),
> 		  "D" (&bios32_indirect));
> 
> Ian

Thank you, Ian. I got it.

But there is another wield thing. I can't pass the compiling stage use the 
code below.

int main(void)
{
        int __value, __port;

        asm volatile ( "movl %1,%0" : "=a" (__value), "=d" (__port)
                                   : ""   (__value), "1"  (__port)
                     );

        return 0;
}

The error message is
[root@littleroach tmp]# gcc a.c
a.c: In function `main':
a.c:10: warning: asm operand 2 probably doesn't match constraints
a.c:10: error: impossible constraint in `asm'

But the same code in the kernel works very well.
Do I miss something?

Roach

--
Open WebMail Project (http://openwebmail.org)

Re: A question about register constraint of inline assembly

[Lin Chun-Zhi]

Thanks for Ian's relpy. It's very helpful for me.

On 02 Jun 2005 09:37:29 -0400, Ian Lance Taylor wrote
> "Lin Chun-Zhi" <lcz90@cs.ccu.edu.tw> writes:
> 
> Please reply to the list, not just to me.
> 
> > But there is another wield thing. I can't pass the compiling stage use 
the 
> > code below.
> > 
> > int main(void)
> > {
> >         int __value, __port;
> > 
> >         asm volatile ( "movl %1,%0" : "=a" (__value), "=d" (__port)
> >                                    : ""   (__value), "1"  (__port)
> >                      );
> > 
> >         return 0;
> > }
> > 
> > The error message is
> > [root@littleroach tmp]# gcc a.c
> > a.c: In function `main':
> > a.c:10: warning: asm operand 2 probably doesn't match constraints
> > a.c:10: error: impossible constraint in `asm'
> > 
> > But the same code in the kernel works very well.
> > Do I miss something?
> 
> It would probably work when optimizing.  Otherwise gcc is complaining
> because it has a memory address and it doesn't know what type of
> register to put it into.
> 
> You should specify a constraint, though.  It's not a good idea to not
> specify one.  If you want to specify that the value is changed by the
> asm (which is not the case in your example) use '+' instead of '='.
> 
> Ian


--
Open WebMail Project (http://openwebmail.org)