[oconnor663]

> Sometimes it turns out that your code relies on some compiler behavior which has changed.

In practice this can be "undefined behavior" like dangling pointers or data races. Maybe the new version of the compiler happens to reorder a couple of instructions (which it's perfectly within it's rights to do) which turns a "benign" race into a crash or an exploitable security issue. Is it your fault for writing these bugs? Sure. But if you're a big organization, and you know you have bugs like this, is this a reason not to upgrade your compiler? Absolutely. All the real world testing you've done on your current binaries has value, and losing some of that value needs to be weighed against the benefits of upgrading.