Intrinsic Methods

If you copy and paste some of the JVM code into your own class you might be surprised that it doesn't run as fast as when it runs in its original library.

Hotspot does this by swapping Java code out and replacing it with hand-written code that is optimized to a particular architecture. Michael Barker shows on his blog how java.lang.Integer.bitCount() is replaced with an Intel ("Nehalem and later") instruction popcnt (this appears to be defined in jdk7/hotspot/src/cpu/x86/vm/x86_*.ad)

This surprising phenomena is due to some methods being intrinsic methods. Oracle's Performance Tuning wiki page is somewhat quiet on the subject but it does point in the direction of library_call.cpp and vmSymbols.hpp.

The classes that are mentioned as having intrinsics can be found with:



$ grep do_intrinsic vmSymbols.hpp | perl -pe s/\#.*//g | perl -pe s/\\/\\/.*//g |  awk '{ print $2 }' | sort | uniq | perl -pe s/_/./g | perl -pe s/,//g

java.lang.Boolean
java.lang.Byte
java.lang.Character
java.lang.Class
java.lang.Double
java.lang.Float
java.lang.Integer
java.lang.invoke.InvokeDynamic
java.lang.invoke.MethodHandle
java.lang.invoke.MethodHandleNatives
java.lang.Long
java.lang.Math
java.lang.Object
java.lang.reflect.Array
java.lang.reflect.Method
java.lang.ref.Reference
java.lang.Short
java.lang.String
java.lang.StringBuffer
java.lang.StringBuilder
java.lang.System
java.lang.Thread
java.lang.Throwable
java.nio.Buffer
java.util.Arrays
sun.misc.AtomicLongCSImpl
sun.misc.Unsafe
sun.reflect.Reflection

[Aside: not all of them actually lead to intrinsics. I was looking at the intrinsic for java.nio.Buffer.checkIndex(...) but saw this in library_call.cpp:

  case vmIntrinsics::_checkIndex:

    // We do not intrinsify this.  The optimizer does fine with it.

    return NULL;

]

But the java.lang.Math class does seem well represented in x86_64.ad where, for instance, the sin function is mapped to an Intel opcode rather than library calls.