We've recently encountered a problem with the c2 compiler in hotspot 1.7/1.6 when
the method "public native long getAddress(long address)" of class "sun.misc.Unsafe"
gets inlined as an intrinsic call on 32-bit platforms.
The API documentation of "sun.misc.Unsafe.getAddress" states that "it fetches a native
pointer from a given memory address" and that "if the native pointer is less than 64
bits wide, it is extended as an unsigned number to a Java long".
Unfortunately, the intrinsic version of this method in hotspot 1.7/1.6 extends the
native pointer as a signed number to a Java long, thus resulting in a discrepancy
between interpreted execution and compiled execution.
Let's consider the following example:
  import sun.misc.Unsafe;
  public class UnsafeGetAddressTest {
      public static void main(String[] args) {
          Unsafe unsafe = Unsafe.getUnsafe();
          long address = unsafe.allocateMemory(unsafe.addressSize());
          unsafe.putAddress(address, 0x0000000080000000L);
          // from sun.misc.Unsafe.getAddress' documentation:
          // "If the native pointer is less than 64 bits wide, it is
          // extended as an unsigned number to a Java long."
          result = unsafe.getAddress(address);
          System.out.printf("1: was 0x%x, expected 0x%x\n", result,
                  0x0000000080000000L);
          for (int i = 0; i < 1000000; i++) {
              result = unsafe.getAddress(address);
          }
          System.out.printf("2: was 0x%x, expected 0x%x\n", result,
                  0x0000000080000000L);
      }
      static volatile long result;
  }
and one execution of the example in interpreted-mode:
   $ /tmp/jdk1.7.0/bin/java -Xint -showversion -Xbootclasspath/a:UnsafeGetAddressTest.jar -XX:+PrintCompilation UnsafeGetAddressTest
  java version "1.7.0-ea"
  Java(TM) SE Runtime Environment (build 1.7.0-ea-b24)
  Java HotSpot(TM) Server VM (build 12.0-b01, interpreted mode)
   1: was 0x80000000, expected 0x80000000
  2: was 0x80000000, expected 0x80000000
 and one execution in mixed-mode:
   $ /tmp/jdk1.7.0/bin/java -Xmixed -server -showversion -Xbootclasspath/a:UnsafeGetAddressTest.jar -XX:+PrintCompilation UnsafeGetAddressTest
  java version "1.7.0-ea"
  Java(TM) SE Runtime Environment (build 1.7.0-ea-b24)
  Java HotSpot(TM) Server VM (build 12.0-b01, mixed mode)
   1: was 0x80000000, expected 0x80000000
  ---   n   sun.misc.Unsafe::getAddress
    1%      UnsafeGetAddressTest::main @ 65 (113 bytes)
  2: was 0xffffffff80000000, expected 0x80000000
 The mixed-mode execution shows that the native pointer 0x80000000 gets extended as
a signed number to the Java long 0xffffffff80000000L after the method has been
on-stack-replaced. The interpreted execution always returns 0x80000000L.
The problem arises from the deletion of opto's CastP2LNode class and the new usage
of CastP2XNode and ConvX2L in "hotspot/src/share/vm/opto/library_call.cpp" in method
"LibraryCallKit::inline_unsafe_access" which is not equivalent to the old implementation.