On 08/31/2010 05:36 PM, Philip Prindeville wrote:
> On 8/31/10 1:23 AM, Andrew Haley wrote:
>> On 08/28/2010 08:41 PM, Philip Prindeville wrote:
>>> Hi.
>>>
>>> I work a bit with Asterisk, the IP-PBX open software. There's a GSM
>>> library that's included in the source tree.
>>>
>>> The problem we're encountering is that when compiled with gcc 4.1.x and
>>> 4.3 or later, it works fine.
>>>
>>> But some of our target platforms only support gcc 4.2, and when we
>>> compile against this, the optimizer in GCC exposes the fact that someone
>>> writing some ASM stubs forgot to correctly annotate a constraint, and
>>> that lack of a constraint is causing the optimizer to change the
>>> generated object more than it should.
>>>
>>> There are a lot of lines of source code, and most of my experience with
>>> asm() stubs has been rudimentary, so I'm asking for an outside pair of
>>> eyeballs that might be better skilled at identifying how the asm should
>>> be annotated than me.
>> How do you know that the asm is the problem?
>
> I don't for sure, but the last time we had similar distortion problems
> it turned out because other annotation had been left out of the asm:
>
> Index: codecs/gsm/inc/private.h
> ===================================================================
> --- codecs/gsm/inc/private.h (revision 109711)
> +++ codecs/gsm/inc/private.h (working copy)
> @@ -105,7 +105,7 @@
> __asm__ __volatile__(
>
> "addl %2,%0; jno 0f; movl $0x7fffffff,%0; adcl $0,%0; 0:"
> - : "=r" (a)
> + : "=&r" (a)
> : "0" (a), "ir" (b)
> : "cc"
Well, that's definitely an improvement, but what asm are you asking about
now? The asms in that file look fine. They are a little bit weird, in
that I would write
static __inline__ short GSM_SUB(short a, short b)
{
__asm__ __volatile__(
"subw %1,%0; jno 0f; movw $0x7fff,%0; adcw $0,%0; 0:"
: "+r" (a)
: "ir" (b)
: "cc"
);
return(a);
}
but I don't think they're wrong.
Anyway, you have pure C versions of those functions. Does it work
when you use them?
Andrew.
