Thoughts on adding getElementClass() method to StackTraceElement?
nicholas+openjdk at nicholaswilliams.net
Sat Jun 15 18:33:39 UTC 2013
Bug 4851444  was filed just over 10 years ago and has languished without so much as a comment ever since. I hate to revive something so old, but I'd like to see what everyone thinks.
I'm working on the Apache project Log4j2. Our problem is we need to get the Class object that belongs to a particular element on the stack trace. We use three different methods, failing over from one to the next, in order to get the Class object.
1. If sun.reflect.Reflection#getCallerClass(int) exists, we repeatedly call it--increasing the depth each time--to fill up a java.util.Stack, then compare it to the exception stack trace and match as many of the elements as we can to Classes returned by getCallerClass. This is surprisingly fast, but some elements don't match and we have to fail over to 3.
2. If sun.reflect.Reflection.getCallerClass(int) does not exist, we create a SecurityManager and call getClassContext(), use it to fill up a java.util.Stack, and do the same thing we would normally do with 1. This is about two times slower than 1--not terrible.
3. If the element in a stack trace doesn't match a class found in methods 1 or 2, we load the class with the ClassLoader using the class name in the StackTraceElement. This is more than TEN TIMES SLOWER than 1--terrible.
As you can see, it greatly benefits us to have some kind of efficient mechanism to match a StackTraceElement to a Class, so that's why we resort to using a non-public API (with the understanding that we must have backup options since this class will not always be present).
Where I came in is when the change was made in 1.8.0-ea-b87 to remove the parameter from sun.reflect.Reflection#getCallerClass(). Now getCallerClass() can ONLY be used to retrieve the immediate caller, not callers at various depth, making method 1 unusable in Java 8. This led me to look for an alternative.
Looking at Throwable.java and Throwable.c, getStackTrace() ultimately calls JVM_GetStackTraceElement. Looking at jvm.c from Java 6 (I can't find JVM_GetStackTraceElement in newer versions of Java), JVM_GetStackTraceElement calls CVMgetStackTraceElement.
CVMgetStackTraceElement, in turn, then gets a copy of the native version of the Class object and uses it to populate the StackTraceElement's data. This leads me to believe that the task of adding a getElementClass() method to StackTraceElement would be as simple as:
1) Adding a new constructor to StackTraceElement to preserve backwards compatibility, with the Class being null by default.
2) Add the getStackTraceElement() method and backing private field.
3) Add a CVMID_fieldWriteRef call to set the class to the backing private field.
I'm sure there are some idiosyncrasies I'm missing here, but overall this seems like a pretty trivial change. What do people think? Can this be worked in? .NET has had this ability since .NET 1.1 with the StackFrame class  and the StackTrace , which contains StackFrames. It provides more than just the Class (Type in .NET), it also provides the Method (MethodBase in .NET). It would greatly improve the efficiency of certain tasks not only in Log4j, but also in many other projects that I have found using getCallerClass() for the same purpose.
More information about the core-libs-dev