I've updated the original documents on the atomics page
http://gcc.gnu.org/wiki/Atomic/GCCMM to add some clarity in places
(especially in the 'synchronization models' section) and updated some
bitrot.
As mentioned last week, I've added a section on codegen which Aldy has
begun implementing: http://gcc.gnu.org/wiki/Atomic/GCCMM/CodeGen
I've also just finished an initial cut of a new page which discusses the
optimization effects the different models have. It took a while to
whittle it down to something digestible, but I think the interaction is
fairly readable now. It is not a complete analysis of each optimization
yet, but it does cover how the different models will impact the
optimizers and touches on implementation. I'll leave the real gory
details until the codegen changes are complete and there is atomic code
to play with.
The one-line version is that most optimizations will probably just need
tweaking for what to do around built-ins rather than always assuming
they are a full barrier.
WIP : http://gcc.gnu.org/wiki/Atomic/GCCMM/Optimizations
I will now pick up where Aldy left off and finish the codegen parts for
the mem_exchange instruction. (I might switch that to the load and store
operations). Then I'll create some testcases for it, and modify the
dead store optimizer as a litmus test. This will give us an
implementation sample to see that everything plays together and behaves
as expected. If that seems to work out, we'll proceed with the rest of
the atomic operations.
If you are interested in the memory model read that newest document and
let me know what isn't clear. If you know a lot about this stuff, look
at it and let me know what needs fixing. If you don't care, go read
your next email :-)
There will also be lots of time to comment when the patch is ready for
the 'cxx-mem-model' branch
Andrew
