[K42-discussion] race condition in clustered object call on unoptimized build

[Bryan S Rosenburg]

Raymond Fingas <fingas at eecg.toronto.edu> wrote on 02/16/2006 01:46:13 PM:
> David said that you wanted to see an example of the problem. This is a
> simple example with commented assembly output provided:
> 
> uval badgcc(NullObject** no){
>    return DREF(no)->Function();
> }

Thanks, Raymond.  It doesn't get any simpler than that.

> Our GCC version is:
> 
> powerpc-linux-g++ (GCC) 3.3.3 20031029 (prerelease)

I've verified that this version of the compiler (I haven't tried any other 
versions), when presented with:

        (expression)->Function();

will evaluate (expression) twice no matter how complicated it is, as long 
as it doesn't have side effects.  I still find this behavior surprising, 
but that's what it is.

The good news is that our DREF bug is easily fixed with a strategic 
"volatile" attribute.  No inline asm required.  I'm changing the 
definition of DREF in lib/libc/cobj/ObjectRefs.H to:

#define DREF(OBJ)       (*((typeof(**OBJ) * volatile *)(OBJ)))

Would you please verify that that fixes the problems you discovered?

Thanks again for tracking this problem down.

- Bryan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://ozlabs.org/pipermail/k42-discussion/attachments/20060216/2abd8f45/attachment.htm