Re: net net/len (0.0.0.0)

[Felipe Kellermann <stdfk () terra com br>]

On Sun, 28 May 2006 1:14pm  -0700, Guy Harris wrote:

> Felipe Kellermann wrote:
>
> > This fixes the problem to me. Does anyone know why 0xffffffff << (32 - 0)
> > is resulting in 0xffffffff in mcode?
>
> To quote the ANSI C89 standard, section 3.3.7 "Bitwise Shift Operators":
>
>     If the value of the right operand is negative or is greater than or
> equal to the width in bits of the promoted left operand, the behavior
> is undefined.

Yes, I did know that that shift resulted in a UB.  GCC (4.1.1) even warns 
about this operation when it can calculate the rside in compilation time.


> This was done, I think, because some processors' shift instructions take
> the shift count modulo the register length, so shifting left by 32 on a
> 32-bit processor leaves the result unchanged, and other processors' shift
> instructions don't, so a shift count of 32 on a 32-bit processor clears
> the result.  On 64-bit processors, there might be an additional issue of
> 64-bit shift instructions, which would presumably, when used to shift a
> 32-bit quantity, clear the result (as the lower 32 bits would be shifted
> into the upper 64 bits, and 32 bits of 0 shifted in) vs. 32-bit shift
> instructions, if the processor has them, which might take the shift count
> modulo 32.

Yes.  Though I still haven't looked at the resulting instructions, it is 
something in these lines.  I will try to write a test-case when time 
permits.


> The OS X machine you tried has a PowerPC processor, which might handle
> this differently from x86.  The compiler used could also make a
> difference.
>
> The fix is to have gen_mcode() check for a masklen of 0, and explicitly
> set m to zero in that case.  I've checked that fix into the main and x.9
> branches.

Thanks.  I will look at the resulting assembly in the different toolchains 
and processos and systems and eventually discuss this subject with the GCC 
or binutils people.

-- 
Felipe Kellermann
-
This is the tcpdump-workers list.
Visit https://lists.sandelman.ca/ to unsubscribe.