We have a case of incorrect code generation by Compiler2 using the latest
(1.3.1rc1 from March 16) hotspot sources. A method in the test program
returns the wrong result using Compiler2. Compiler1 runs the example ok.

The attached zip file includes these files

  Test.class
  Test.java
  TestAlgorithm.class
  TestAlgorithm.java
  TestObject.class
  TestObject.java
  test.log
  test.sh

test.sh shows the invocation of the program as I ran it with a fastdebug VM
to produce test.log. In test.log, I have noted with "BUG:" where I think
the problem is in the generated code. It looks like the instructions to
assign the method return value into R_I0 (to correspond to the "load_local_object #3"
bytecodes) have been lost by the optimizations in Compiler2.

Further instructions from the author of the test program are:

Attached are some Java files that should reproduce the problem.
You should perform the following tests:

java Test 0
java Test 1
java Test 2
java -server Test 0
java -server Test 1
java -server Test 2

You will notice that "java -server Test 0" will fail with throwing an
exception.

The differences between the tests are the following:
Test 0: is similar to the Advisor code.
This test works with hotspot client (Compiler1), but fails with hotspot
server (Compiler2).
Test 1: is a slightly modified version of the same code, just
one if statement is different. This test works OK in both modes.
Test 2: is the same as Test 0, the only difference is that it prints
a message if it is about to return null. This test also works OK.

** See also 'Comments' section **