* burak serdar <bser...@computer.org> [231114 19:09]: > I do not agree that this is because how the compiler works. A value > receiver is equivalent to pass-by-value argument, that is: > > rcp.version() > > is equivalent to: > > RPC.version(rpc) > > thus, creating the copy of the rpc variable. So, the compiler may > choose to avoid the race by not copying it, or by inlining the version > function, but according to the spec, it is passed by value, i.e., it > is copied.
I was going to say that the compiler must not optimize away the copy because it changes behavior, but thinking it through, it may optimize away the copy if it can determine that doing so will not change the behavior of a program that does not violate the language spec (or memory model). On a more esoteric note, while this is definitely a race and there are no "benign" data races, I am unaware of any current, commodity computer architectures on which the race in Dave Cheney's article will create any output other than what was expected. This is based on the assumption that current mass-marketed computers have the property that a read-during-concurrent-write will never end up with non-deterministic values anywhere but the result of the read. I.e. if memory location A is being copied to memory location B, while concurrently value Y is being written to A, then when the two concurrent operations are finished, A will have value Y, B will have a non-deterministic value, and no other memory locations will be affected. If there are CPU/architectures that do not conform to this, that are currently being produced, and to which Go has been ported, I would be interested to hear about them (even if they are not mass-marketed computers). What makes a data race _never_ benign is portability. Compiler writers must obey both the language spec and the CPU/architecture spec. A data race (according to the language spec) that is thought to be benign adds a third spec that the compiler writers _do not_ need to obey: the specific behavior of the hardware architecture when code written in that language does not obey the language spec (or memory model). The race may be benign on specific combinations of compiler version and hardware architecture version, but both hardware designers and compiler writers are free to build new versions that do not conform to the spec that the programmer implicitly created when writing the "benign" data race. ...Marvin -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/ZVTvQzxn2KQglaq8%40basil.wdw.
