Hacker News new | ask | show | jobs
by christophilus 1648 days ago
Unless the JVM is sure the logic is not effectful, it couldn’t eliminate it.
2 comments

There's an even simpler reason why the JIT compiler can't prune it: it's possible to dynamically change the logging level at run time.
That is actually a reason that a JIT compiler can prune where an AOT compiler can't. JIT can deoptimize when assumptions change.

One of the bigger wins with the JVM is assuming that a virtual method can be called statically if there are no derived classes. That's a huge win for every leaf class in the class tree. The assumption can change every time a class is loaded of course. The optimization is called devirtualization and it can be combined with inlining to get even bigger wins.

Fair point... but it seem like, even so, doing that to monitor whether a single integer variable might change sounds like a lot of added complexity. Are there other use cases that would help to justify it? Reducing the impact of failing to follow logging best practices doesn't seem like an obviously sufficient cause.
The JVM knows what string concatenation does.
`"hello" + myFooObject.toString()` Can you eliminate that? It's impossible to say without knowing whether or not `myFooObject.toString()` is effectful. Maybe the JVM can make that assessment. Maybe not. I don't know, but it's not as trivial as "knowing what string concatenation does".
A straightforward way to do this is for Java to assume built in types like strings are side effect free, but custom types may not be. This would likely cover a good portion of log lines but not all, obviously.
But what if one of the arguments is a function call? Then it isn't easy to prove there are no side effects
If it’s too big to in-line then yes.
That’s says absolutely nothing about whether the function has a side effect.
If you can inline it into the compilation unit, then you can see if it has side effects or not.
The noop determination is made during runtime within the log method called. So, it's LOG.info(obj.toString()), wherein that method inspects the log level setting, then decides whether to ignore.

Are you saying that Java does or should "optimize" in such a way that it branches to the logic within each method called during runtime, inspects which code would be executed per possibly mutable runtime values (like log level), considers possible side effects (in-method and in parameters) then decides whether to even invoke the method at all?

Isn't the point that you don't generally stuff your log of things that are only strings but of things that can become strings? (Asking since I know of your work with truffle)
I assumed things like the names of resources that are already strings because you’re using them in the actual program?
A typical logger.info("user {} did the thing", user) can skip the actual string interpolation and user value stringification, if the log level is not > info. However, logger.info("user " + user + " did the thing") cannot avoid at least the execution of user.toString(), even after jit optimizations, unless the jit could prove that toString does not have side effects. But I don't believe the jvm jit tries to do that. Am I wrong?
If user is a string, then #toString() is a trivial method, will inline, and become a no-op.
However, if user happened be an user-defined User object it would be hard to guarantee that it could be inlined