On 29/06/17 19:35, Alexander Monakov wrote: > I think a practical approach is to give the user a degree of control by > introducing a tri-state compiler option controlling how double-word atomics > are to be emitted: [snip] I will properly reply to the above in a bit but I thought it might help give a users view of things. Obviously, only one user with his own particular view, but it can help to have concrete examples. In lock-free data structures, the main use of DWCAS is to have a counter along with a pointer, to solve ABA. The counter is a version number for the pointer, and is incremented whenever the pointer is modified. The counter is the first element in the counter-pointer array. It is always read first, and it is a single registers worth of data, so it is read correctly. Of course, the pointer can then change after we've read the version but before we're read the pointer - but it just doesn't matter because we *have* the version already (for who-ever wrote to the pointer will also have updated the version counter first). This means when we come to write back using that pointer, our DWCAS will fail (as it should do) because the counter-pointer pair have been modified by someone else since we read them (so we have to iterate again and try once more to do whatever it was we were trying to do - we will try to read the counter-pointer pair again). On IA64, there is no DWCAS. There is a wonderful little single-word compare, double-word write. It's all that I need - for all I need to do is compare the version counter. Actually, I wonder what GCC does for CAS/DWCAS on IA64? their instruction doesn't fit the CAS/DWCAS paradym! So for me, the fact the double-word reads are not read together is absolutely irrelevant. It doesn't affect me at all.
