Re: Cheaper RAS algorithm for CAS on VAX?

[Matt Thomas <matt%3am-software.com@localhost>]

On Nov 12, 2010, at 3:20 PM, Johnny Billquist wrote:

> On 2010-11-12 18:11, Matt Thomas wrote:
>> 
>> I was thinking of how AVR32 implements LL/SC sematics and wondering
>> if I apply the same ideas to the VAX.
>> 
>> Basically, everytime AVR32 takes an exception it clears a flag which
>> was set by the load-locked instruction.  Store-conditional sees the
>> flag was cleared and fails the store.
>> 
>> Instead of a flag, I think we can do the same thing using indirect
>> addressing.  Instead of a flag, use a known address CASADDR to store
>> the address for which the CAS is being performed.  After the CAS, has
>> finished CASADDR is always to a value that when dereferenced indirectly
>> will cause a trap.
>> 
>> the RAS would be approximately something like:
>> 
>>      .word   0
>>      movl    4(ap), CASADDR
>>      movl    @CASADDR, %r0
>>      cmpl    %r0, 8(ap)
>>      bne     1f
>>      movl    12(ap), @CASADDR
>> 1:
>>      movl    $CASMAGIC, CASADDR
>>      ret
>> 
>> Now all we have to do in the trap handler is check if the VA is
>> CASADDR and the PC is within the routine, and if so, just set PC
>> to beginning of routine and return from trap.  No messy check in
>> the exception entries or exits.
>> 
>> As long as you don't have the CAS routine interrupted, there's
>> only a tiny bit of overhead.
> 
> Without doing any deeper analysis - what about multiprocessor systems?

This can't work on MP since concurrent writes to the same location could
be performed by different CPUs without detection.  This forces the slow
way currently implemented.