Reference:
LLVM Atomic Instructions and Concurrency Guide
There are only
2 places need barrier:
- processing invalid queue (RMB)
- write store buffer to cache. (WMB)
That's it, period!
Other places, like read pull request, response to read request, mark as share, do NOT participate with barrier!
If using bit fields, this is an important point to note
Though adjacent bit fields are distinct objects, they’re still counted as the same memory location.
The bit fields bf1 and bf2 share a memory location, and the std::string object
s
consists of several memory locations internally,
but otherwise each member has its own memory location.
Note how the zero-length bit field marked /*bf3*/
(
the name is commented out because zero-length bitfields must be unnamed)
separates bf4 into its own memory location, but doesn't have a memory location itself.
Four important things to take away from this
- Every variable is an object, including those that are members of other objects.
- Every object occupies at least one memory location.
- Variables of fundamental type such as int or char are exactly one memory location, whatever their size, even if they’re adjacent or part of an array.
- Adjacent bit fields are part of the same memory location.
Everything hinges on those memory locations. If two threads access separate memory locations, there’s no problem: everything works fine. On the other hand, if two threads access the same memory location, then have to be careful.
