Insert Special Code Sequence After Object allocation

[Xin Tong]

On Wed, Nov 28, 2012 at 1:35 AM, Krystal Mo <krystal.mo at oracle.com> wrote:
> On 11/28/2012 02:02 PM, Xin Tong wrote:
>>
>> On Wed, Nov 28, 2012 at 12:58 AM, Xin Tong <xerox.time.tech at gmail.com>
>> wrote:
>>>
>>> On Tue, Nov 27, 2012 at 11:23 PM, Krystal Mo <krystal.mo at oracle.com>
>>> wrote:
>>>>
>>>> Hi Xin,
>>>>
>>>> Depends on what the code sequence is and what it does.
>>>>
>>>> If the code sequence you're inserting can be represented in C2's IR
>>>> nodes,
>>>> then you can try to add those nodes in either
>>>> PhaseMacroExpand::expand_allocate_common() or
>>>> PhaseMacroExpand::initialize_object(), or somewhere else around.
>>>>
>>> It is a few assembly instructions. can i wrap the assemblies in a
>>> function that hasa prologue and epilogue to save everything. and does
>>> the IR has any way of calling a function ?
>>>
>>> Xin
>>
>> Or another thing i am willing to do is to force the JIT to always fail
>> on the inlined allocation and force it to the out-of-line ( VM
>> managed) allocation. I just do not want to turn the JIT completely off
>> as that have an big impact on the behaviours of the program.  Is this
>> doable ?
>
> Sure, if you're willing take the performance hit.
>
> It's easy: in PhaseMacroExpand::expand_allocate_common(), just make sure the
> local variables initial_slow_test and always_slow are assigned NULL and true
> respectively, and there you go. It'll call into
> OptoRuntime::new_instance_C().
>

I change the   bool always_slow = true; it runs ok. but when i change
 Node* initial_slow_test = NULL ;
//alloc->in(AllocateNode::InitialTest);
the VM crashes. Is it necessary to make initial_slow_test to NULL ? is
not initial_slow_test used to initialize allocattion from slow path ?

# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0x00007fd6d0e58572, pid=18333, tid=140560399947520
#
# JRE version: 7.0
# Java VM: OpenJDK 64-Bit Server VM (23.2-b09 mixed mode linux-amd64
compressed oops)
# Problematic frame:
# V  [libjvm.so+0x7b6572]  BoolNode::make_predicate(Node*, PhaseGVN*)+0x22
#
# Failed to write core dump. Core dumps have been disabled. To enable
core dumping, try "ulimit -c unlimited" before starting Java again
#
# An error report file with more information is saved as:
# /home/xtong/fastcache/benchmarks/SPECJBB2005/hs_err_pid18333.log
#
# If you would like to submit a bug report, please visit:
#   http://bugreport.sun.com/bugreport/crash.jsp
#
./run.sh: line 13: 18333 Aborted                 (core dumped) java
-Xms1024m -Xmx2048m spec.jbb.JBBmain -propfile SPECjbb.props


Xin



> - Kris
>
>
>> Xin
>>
>>>> If your code sequence cannot be represented in C2's IR, well, that will
>>>> take
>>>> some effort. The AllocateNode is lost after PhaseMacroExpand, so you
>>>> can't
>>>> use it as a mark to insert your code. You might want to use
>>>> prefetchAlloc in
>>>> the AD files as a place to insert code, if that works for your purpose.
>>>>
>>>> That said, though, if the code sequence could be written in Java
>>>> bytecode,
>>>> that'll be the easiest solution: don't have to touch the VM, just do
>>>> bytecode instrumentation.
>>>>
>>>> - Kris
>>>>
>>>>
>>>> On 2012/11/28 10:36, Xin Tong wrote:
>>>>>
>>>>> I would like to insert a special binary code sequence after the
>>>>> allocation of a java object in the opto JIT code. can anyone tell me
>>>>> how to do that ?
>>>>>
>>>>> Xin
>
>