[ECOS] CGY_SWAP16() seems broken to me.

[ECOS] CGY_SWAP16() seems broken to me.

[Grant Edwards]

unsigned u = CYG_SWAP16(0x1234);

What would a reasonable person (who hasn't looked at
hal_endian.h) expect u to be?

I certainly expected it to be 0x3412. 

But it's not.  It's 0x123412.  

IMO, that's just plain broken.


Here's the definition from hal_endian.h:

# define CYG_SWAP16(_x_)                                        \
    ({ cyg_uint16 _x = (_x_); ((_x << 8) | (_x >> 8)); })

Why isn't it this?

# define CYG_SWAP16(_x_)                                        \
    ({ cyg_uint16 _x = (_x_); (cyg_uint16)((_x << 8) | (_x >> 8)); })

Why would anybody expect a macro named CYG_SWAP16() to return a
"24-bit value" when passed a "16-bit value"?

-- 
Grant Edwards                   grante             Yow!  Ha ha  Ha ha Ha ha
                                  at               Ha Ha Ha Ha -- When will I
                               visi.com            EVER stop HAVING FUN?!!


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] CGY_SWAP16() seems broken to me.

[Sergei Gavrikov]

On Sat, Sep 22, 2007 at 02:56:33AM +0000, Grant Edwards wrote:
> unsigned u = CYG_SWAP16(0x1234);
> 
> What would a reasonable person (who hasn't looked at
> hal_endian.h) expect u to be?
> 
> I certainly expected it to be 0x3412. 
> 
> But it's not.  It's 0x123412.  

That is exists. I got same result

#include <cyg/hal/hal_endian.h>
#include <sys/param.h>
#include <sys/endian.h>

#include <stdio.h>

int
main (void)
{
    printf ("HAL swap16(0x1234) = 0x%x\n", CYG_SWAP16 (0x1234));
    printf ("NET swap16(0x1234) = 0x%x\n", swap16 (0x1234));
    return 0;
}

HAL swap16(0x1234) = 0x123412
NET swap16(0x1234) = 0x3412

> IMO, that's just plain broken.

... I don't know. What do eCos experts think about?

> Here's the definition from hal_endian.h:
> 
> # define CYG_SWAP16(_x_)                                        \
>     ({ cyg_uint16 _x = (_x_); ((_x << 8) | (_x >> 8)); })
> 
> Why isn't it this?
> 
> # define CYG_SWAP16(_x_)                                        \
>     ({ cyg_uint16 _x = (_x_); (cyg_uint16)((_x << 8) | (_x >> 8)); })
> 
> Why would anybody expect a macro named CYG_SWAP16() to return a
> "24-bit value" when passed a "16-bit value"?

Usually, they use yet masking in same macros. Good byteorder examples
are

net/bsd_tcpip/current/include/sys/endian.h
net/tcpip/current/include/sys/endian.h

It seems, that eCos HAL's CYG_SWAP16() either assumes what it will be to
live in "16-bit world" or it assumes that will be a developers's care. I
did find only 3 places where CYG_SWAP16() macro is used. For example,
CYG_SWAP16() macro uses itself in the eCos drivers for 16-bit Ethernet
controllers (cs8900a, ec555) "16-bit world". And I found that RedBoot's
conf_endian_fixup() (fconfig.c) uses eCos HAL's CYG_SWAP16() "as is",
because that was a "developer's care", there is there

                u16 = CYG_SWAP16(u16);

but, not :-)

		unsigned u = CYG_SWAP16(0x1234);

In any case, thank you for your report, we will be to know about.


	Sergei


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: CGY_SWAP16() seems broken to me.

[Grant Edwards]

On 2007-09-22, Sergei Gavrikov <w3sg@SoftHome.net> wrote:
> On Sat, Sep 22, 2007 at 02:56:33AM +0000, Grant Edwards wrote:
>> unsigned u = CYG_SWAP16(0x1234);
>> 
>> What would a reasonable person (who hasn't looked at
>> hal_endian.h) expect u to be?
>> 
>> I certainly expected it to be 0x3412. 
>> 
>> But it's not.  It's 0x123412.  
>
> That is exists. I got same result
>
> #include <cyg/hal/hal_endian.h>
> #include <sys/param.h>
> #include <sys/endian.h>
>
> #include <stdio.h>
>
> int
> main (void)
> {
>     printf ("HAL swap16(0x1234) = 0x%x\n", CYG_SWAP16 (0x1234));
>     printf ("NET swap16(0x1234) = 0x%x\n", swap16 (0x1234));
>     return 0;
> }
>
> HAL swap16(0x1234) = 0x123412
> NET swap16(0x1234) = 0x3412
>
>> IMO, that's just plain broken.
>
> ... I don't know. What do eCos experts think about?
>
>> Here's the definition from hal_endian.h:
>> 
>> # define CYG_SWAP16(_x_)                                        \
>>     ({ cyg_uint16 _x = (_x_); ((_x << 8) | (_x >> 8)); })
>> 
>> Why isn't it this?
>> 
>> # define CYG_SWAP16(_x_)                                        \
>>     ({ cyg_uint16 _x = (_x_); (cyg_uint16)((_x << 8) | (_x >> 8)); })
>> 
>> Why would anybody expect a macro named CYG_SWAP16() to return a
>> "24-bit value" when passed a "16-bit value"?
>
> Usually, they use yet masking in same macros. Good byteorder examples
> are
>
> net/bsd_tcpip/current/include/sys/endian.h
> net/tcpip/current/include/sys/endian.h
>
> It seems, that eCos HAL's CYG_SWAP16() either assumes what it will be to
> live in "16-bit world" or it assumes that will be a developers's care.

I think both assumptions are invalid.

> I did find only 3 places where CYG_SWAP16() macro is used. For
> example, CYG_SWAP16() macro uses itself in the eCos drivers
> for 16-bit Ethernet controllers (cs8900a, ec555) "16-bit
> world". And I found that RedBoot's conf_endian_fixup()
> (fconfig.c) uses eCos HAL's CYG_SWAP16() "as is", because that
> was a "developer's care", there is there
>
>                 u16 = CYG_SWAP16(u16);
>
> but, not :-)
>
> 		unsigned u = CYG_SWAP16(0x1234);

Right.  The existing macro works correctly if the result is
being assigned to a 16-bit lvalue, but it doesn't work if it's
used without a typecase or mask in an arithmetic expression or
if it's being assied to a 32-bit lvalue.

My fix above should be completely transparent. if the value is
being assigned to a 16 bit lvalue, the compiler will generate
exactly the same code as when using the existing "broken"
macro. If the result from CYG_SWAP16() is being masked by the
user, the fixed macro will generate exactly the same code as
the existing macro.

The only argument against my fix would be to claim that
CYG_SWAP16 is expected to return 24 bits in wider contexts, and
returning only 16 bits is going to break code or surprise a
user.

-- 
Grant Edwards                   grante             Yow!  .. I must be a
                                  at               VETERINARIAN...
                               visi.com            


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

[ECOS] Re: CGY_SWAP16() seems broken to me.

[Grant Edwards]

On 2007-09-22, Grant Edwards <grante@visi.com> wrote:

> My fix above should be completely transparent. if the value is
> being assigned to a 16 bit lvalue, the compiler will generate
> exactly the same code as when using the existing "broken"
> macro. If the result from CYG_SWAP16() is being masked by the
> user, the fixed macro will generate exactly the same code as
> the existing macro.
>
> The only argument against my fix would be to claim that
> CYG_SWAP16 is expected to return 24 bits in wider contexts, and
> returning only 16 bits is going to break code or surprise a
> user.

I've examined all uses of SWAP16 and the LE16/BE16 macros in my
source tree (which is a few months out of date), and it doesn't
look like any existing code depends on the fact that SWAP16()
produces 24 bits instead of 16.

-- 
Grant Edwards                   grante             Yow!  An INK-LING? Sure --
                                  at               TAKE one!! Did you BUY any
                               visi.com            COMMUNIST UNIFORMS??


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss

Re: [ECOS] Re: CGY_SWAP16() seems broken to me.

[Sergei Gavrikov]

On Sat, Sep 22, 2007 at 03:20:51PM +0000, Grant Edwards wrote:
> On 2007-09-22, Grant Edwards <grante@visi.com> wrote:
> 
> > My fix above should be completely transparent. if the value is
> > being assigned to a 16 bit lvalue, the compiler will generate
> > exactly the same code as when using the existing "broken"
> > macro. If the result from CYG_SWAP16() is being masked by the
> > user, the fixed macro will generate exactly the same code as
> > the existing macro.
> >
> > The only argument against my fix would be to claim that
> > CYG_SWAP16 is expected to return 24 bits in wider contexts, and
> > returning only 16 bits is going to break code or surprise a
> > user.
> 
> I've examined all uses of SWAP16 and the LE16/BE16 macros in my
> source tree (which is a few months out of date), and it doesn't
> look like any existing code depends on the fact that SWAP16()
> produces 24 bits instead of 16.

Today's score for CYG_SWAP16() is

CYG_SWAP16()      : 3 sources
CYG_CPU_TO_LE16() : 4 sources 

As you could see, ntohs(), htons() = swap16() in the little-endian world
(net/tcpip/current/include/sys/endian.h). There are tons of C lines are
using these net macros in the net stuff and at first, I just did think
about same strong rules for the HAL CYG_SWAP16() too, i.e. about type
casting (compiler care) + bit masking (CPU workaround).

I do see no criminal if you will provide a patch, which will set the
first barier - type casting. I vote for your patch. I did dislike that
0x123412.


	Sergei


-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss