My experience has been that it's rather heuristic based. It's a clear win when you can immediately see far enough in advance to know that it'll also decrease the amount of generated code. You can spot trivial cases where this is true at the point of inlining. However, if you stopped there, you'd leave a ton of optimizations on the table. Further optimization (e.g. DCE) will often drastically reduce code size from the inlining, but it's hard to predict in relationship to a specific inlining decision.
So, statistics and heuristics.