Re: warning: read-write constraint -- 2.6.15-git8 onwards

["Jan Beulich" <JBeulich@novell.com>]

Looking at the official 3.4.x, 4.0.x, and the prerelease 4.1 sources I can't see such a warning, nor am I getting it
with 3.3.3 when building on SLES9 (which makes me assume unmodified 3.3.3 doesn't have this either). I thus wonder
whether the compiler used there has an extra, ill patch (after all there's generally nothing wrong for a read-write
constraint to only allow memory on x86). Jan

>>> Andi Kleen <ak@suse.de> 19.01.06 04:03:58 >>>
On Wednesday 18 January 2006 19:25, Andy Whitcroft wrote:
> It seems that the following commit causes a bunch of warnings out of
> most of the files in the kernel tree (see below for examples).  Backing
> this out seems to cure them?




> Compiled with:
> 
>     gcc version 3.3.4 (Debian 1:3.3.4-6sarge1)
> 
> -apw
> 
>   CC      arch/x86_64/kernel/process.o
> include/asm/bitops.h: In function `default_idle':
> include/asm/bitops.h:65: warning: read-write constraint does not allow a
> register
> include/asm/bitops.h:65: warning: read-write constraint does not allow a
> register
> include/asm/bitops.h:30: warning: read-write constraint does not allow a
> register
> include/asm/bitops.h:30: warning: read-write constraint does not allow a
> register

I only tested with gcc 4.0 and 4.1 prereleases where it worked just fine.

 __asm__ __volatile__( LOCK_PREFIX
                "btrl %1,%0"
                :"+m" (ADDR)
                :"dIr" (nr));

I tried to covert them from +m to explicit input/output
arguments ("=m" (ADDR) : "0" (ADDR)) to fix your compiler 
but with that I get the the same warning as you with 3.4 with 4.0.
Don't know how else it could be written. 

Ok one could just pass the address and do a full memory clobber, but 
that would be a overly large sledgehammer for the problem.

Jan or Andreas - do you have any suggestions how to fix this or should
we revert back to the old (technically wrong) state which was

    __asm__ __volatile__( LOCK_PREFIX
                "btrl %1,%0"
                :"=m" (ADDR)
                :"dIr" (nr));


Thanks,
-Andi