For the vast majority of cases I've found it possible (and very much worthwhile) to just think a bit harder and avoid race conditions by construction (for instance, reformulating an algorithm so that writes only occur to different cache lines in a single pass, etc.). That's obviously not always feasible, but in my experience people give up far too quickly.

For the remaining few cases - yes, manually inserting the minimal necessary amount of synchronization (on x86-64 that's often a single instruction). On GPUs you obviously need memory barriers fairly frequently, but that's much less of an issue than on CPUs. Anyway, I don't have any faith in "smart" compilers or language features (which have large costs elsewhere) that try to free me from the simple necessity to carefully think through what my algorithm actually does before I implement it. If you care about getting close to optimal performance, you'll have to do the thinking anyway.