RFR: JDK-8217909: Make unused r12 register (without compressed oops) available to regalloc in C2

[dean.long at oracle.com]

Hi Nils.  Yes, I was just seeing if I could get around the 
set_stack_version problem, so I didn't do a complete
cleanup.  I was leaving that to Roman :-)

dl

On 2/14/19 12:41 AM, Nils Eliasson wrote:
> HI,
>
> The reg_classes that was introduced with PreservedFramePointer 
> (*_with_rbp, *_no_rbp) is still there in x86_x64.ad. How about 
> handling rbp in the same way as you suggest for r12 (patching the mask 
> in reg_mask_init).
>
> // Nils
>
> On 2019-02-14 07:32, dean.long at oracle.com wrote:
>> How about something like this:
>>
>> http://cr.openjdk.java.net/~dlong//8217909/webrev/
>>
>> dl
>>
>> On 2/13/19 5:05 AM, Roman Kennke wrote:
>>> Apparently, this is a limitation of CodeSnippetRegClass, in 
>>> formsopt.hpp:
>>>
>>>    void set_stack_version(bool flag) {
>>>      assert(false, "User defined register classes are not allowed to
>>> spill to the stack.");
>>>    }
>>>
>>> I believe this means I cannot use this approach.
>>>
>>> I also don't feel like extending the AD syntax, parser, etc at this 
>>> point.
>>>
>>> Do you have any other suggestions? Or should we go with my original 
>>> patch?
>>>
>>> Roman
>>>
>>>> I started a version based on Dean's approach:
>>>>
>>>> http://cr.openjdk.java.net/~rkennke/JDK-8217909/webrev.01/
>>>>
>>>> Unfortunately it barks on ptr_reg class:
>>>>
>>>> assert fails
>>>> /home/rkennke/src/openjdk/jdk-jdk/src/hotspot/share/adlc/formsopt.hpp
>>>> 246: User defined register classes are not allowed to spill to the 
>>>> stack.
>>>>
>>>> Interestingly, it doesn't do that when I only handle any_reg class.
>>>> However, I am not sure what that means or how to fix it. Any hints?
>>>>
>>>> Thanks,
>>>> Roman
>>>>
>>>>> There is a way to do it without cloning all the register class 
>>>>> variants.
>>>>> For the arm64 port we massaged the register masks in
>>>>> Compile::pd_compiler2_init():
>>>>>
>>>>> http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b756e7a2ec33/src/cpu/arm/vm/arm_64.ad#l638 
>>>>>
>>>>>
>>>>>
>>>>> It might be worth considering, as the number of variants grows.
>>>>>
>>>>> dl
>>>>>
>>>>> On 2/11/19 3:02 AM, Roman Kennke wrote:
>>>>>> When running with compressed oops, the r12 register holds the 
>>>>>> heapbase,
>>>>>> and thus is not available to register allocation in C2. However, 
>>>>>> when
>>>>>> *not* running with compressed oops, it is still not available and
>>>>>> remains unused. It should be made available to register 
>>>>>> allocation in
>>>>>> this case.
>>>>>>
>>>>>> This patch implements this by introducing with_r12 and no_r12 
>>>>>> variants
>>>>>> of basically all register classes, and add dynamic reg classes to 
>>>>>> select
>>>>>> one or the other, based on current settings. I needed to add 
>>>>>> UseZGC in
>>>>>> those flags because ZGC asserts to not get r12 in its barriers. 
>>>>>> Not sure
>>>>>> that this is necessary.
>>>>>>
>>>>>> Note that we might want to use the r12 register in Shenandoah 
>>>>>> later to
>>>>>> keep GC state. In this case, we'd need to add UseShenandoahGC or 
>>>>>> such to
>>>>>> the test. Do we want to abstract this whole check? Not sure that 
>>>>>> this is
>>>>>> possible/feasible to do in .ad though...
>>>>>>
>>>>>> Bug:
>>>>>> https://bugs.openjdk.java.net/browse/JDK-8217909
>>>>>> Webrev:
>>>>>> http://cr.openjdk.java.net/~rkennke/JDK-8217909/webrev.00/
>>>>>>
>>>>>> Testing: tier1 no regressions locally, eyeball generated code, 
>>>>>> yes it
>>>>>> does use r12 now.
>>>>>>
>>>>>> Can I please get reviews?
>>>>>>
>>>>>> Roman
>>