RFR(M): optional flattening - interpreter/runtime part

[Frederic Parain]

The patch has been tested by forcing the JVM into interpreted mode what
ever command line argument is specified, and always disabling passing/returning
values in registers. All tests in hotspot_valhalla suite pass but ValueTypeTestBench
which fails because of a time out.

Fred


> On Sep 19, 2017, at 17:24, Frederic Parain <frederic.parain at oracle.com> wrote:
> 
> Here’s an updated webrev:
> 
> http://cr.openjdk.java.net/~fparain/optional-flattening-push/webrev.02/index.html
> 
> Changes:
>  - code has been updated according to the recent interpreter changes
>  - the issue with oop_map generation in layout_fields() has now a correct fix
>  - A test in ValueOops has been commented because it has to be revisited
>    with the introduction of optional flattening, but I’d like to discuss this topic
>    with David Simms on a different CR.
> 
> Thank you,
> 
> Fred
> 
> 
>> On Sep 14, 2017, at 11:03, Tobias Hartmann <tobias.hartmann at oracle.com> wrote:
>> 
>> Hi Fred,
>> 
>> I've just verified that the problem is solved with your latest webrev!
>> 
>> Thanks,
>> Tobias
>> 
>> On 14.09.2017 16:50, Frederic Parain wrote:
>>> Tobias,
>>> Thank you for catching that.
>>> There’s a bug in the layout computation, optional flattening code moves non-flattened
>>> values from the value types section of the layout to the nonstatic_oop section, but the size
>>> of the later was not updated accordingly.
>>> Here’s an updated webrev fixing this issue:
>>> http://cr.openjdk.java.net/~fparain/optional-flattening-interp/webrev.01/index.html
>>> Regards,
>>> Fred
>>>> On Sep 14, 2017, at 03:46, Tobias Hartmann <tobias.hartmann at oracle.com> wrote:
>>>> 
>>>> Hi Fred,
>>>> 
>>>> there seems to be a problem with the field layout:
>>>> 
>>>> compiler.valhalla.valuetypes.MyValue1: field layout
>>>> @ 16 --- instance fields start ---
>>>> @ 24 "x" I
>>>> @ 16 "y" J
>>>> @ 32 "z" S
>>>> @ 40 "o" Ljava.lang.Integer;
>>>> @ 48 "oa" [I
>>>> @ 56 "v1" Qcompiler.valhalla.valuetypes.MyValue2;
>>>> @ 64 "v2" Qcompiler.valhalla.valuetypes.MyValue2;
>>>> @ 28 "c" I
>>>> @ 56 --- instance fields end ---
>>>> @ 56 --- instance ends ---
>>>> @176 --- static fields start ---
>>>> @192 "s" I
>>>> @184 "sf" J
>>>> @176 "v3" Qcompiler.valhalla.valuetypes.MyValue2;
>>>> @200 --- static fields end ---
>>>> 
>>>> OopMapBlocks:   1  /  4
>>>>   Offset:  40  - 64 Count:   4
>>>> 
>>>> "instance fields end" points to offset 56 which is not correct because the non-flattened field v2 is at offset 64.
>>>> 
>>>> Thanks,
>>>> Tobias
>>>> 
>>>> On 13.09.2017 21:41, Frederic Parain wrote:
>>>>> Greetings,
>>>>> Please review this changeset implementing optional field flattening
>>>>> in the interpreter and the runtime. The support for C2 is provided by
>>>>> Tobias in a separate changeset.
>>>>> http://cr.openjdk.java.net/~fparain/optional-flattening-interp/webrev.00/index.html
>>>>> Note that trying to run some programs/tests with this changeset and
>>>>> not Tobias’ changeset is likely to cause crashes unless interpreted mode
>>>>> is forced and passing/returning values in registers is disabled.
>>>>> Thank you,
>>>>> Fred
>