Reading
mig21’s weblog, in which I often find really interesting stuff,
I found Ian Lance Taylor’s article “
Single Threaded Memory Model“, which kind of bothers me a bit. It reports a recent discussion on the
gcc and
LKML mailing lists about how C compilers, gcc in this case, optimize for single threaded code, sometimes leading to counter-intuitive results which won’t work properly in multi-threaded software (leading, for example, to race conditions).
The interesting part is that the C language standard, which in general describes a single threaded model, apparently says nothing about when values have to be written to memory. Thus, this kind of optimization seem to be perfectly valid.
Linus Torvalds complains about gcc developers taking more care about “what the spec says” than about real problems, but the fact is that if those kind of optimizations are valid according to the specs, there’s no guarantee at all that it won’t happen with whatever C compiler you might be using.
To prevent this kind of behaviour in standard C, according to Ian, the variable needs to be marked as volatile, or it needs to use an explicit memory barrier. The moral in this case is that multithreading might not be as easy as it seems and should be handled with caution.