Testing hotspot's Atomic::cmpxchg(jbyte*) ?

[David Holmes]

On 19/08/2016 11:29 PM, Aleksey Shipilev wrote:
> On 08/19/2016 03:03 PM, David Holmes wrote:
>> On 19/08/2016 7:37 PM, Aleksey Shipilev wrote:
>>> On 08/19/2016 12:26 PM, David Holmes wrote:
>>>> On 19/08/2016 5:40 PM, Aleksey Shipilev wrote:
>>>>> On 08/19/2016 10:13 AM, David Holmes wrote:
>>>>>>> Now it might be cleaner to ditch Java version from Unsafe, and make
>>>>>>> native entry points like Unsafe_CompareAnd{Exchange,Swap}{Byte,Short}
>>>>>>> which would call relevant Atomic::cmpxchg-s.
>>>>>>
>>>>>> I tried commenting out the Java-ish version and defining one that
>>>>>> would
>>>>>> call Atomic::cmpxchg from unsafe.cpp in the VM. However something is
>>>>>> complaining about the intrinisics - I removed the
>>>>>> HotspotIntrinsicCandidate annotation as I don't want any intrinisics,
>>>>>> but I get a build error:
>>>>>>
>>>>>> Compiler intrinsic is defined for method
>>>>>> [jdk.internal.misc.Unsafe.compareAndSwapByte(Ljava/lang/Object;JBB)Z],
>>>>>> but the method is not annotated with @HotSpotIntrinsicCandidate.
>>>>>> Exiting.
>>>>>>
>>>>>> No idea where this is coming from or how I can disable it ??
>>>>>
>>>>> @HotSpotIntrinsicCandidate marks Java methods that are declared in
>>>>> vmSymbols.hpp, and the VM code checks it. So, removing @HSIC without
>>>>> modifying vmSymbols.hpp would blow up like that.
>>>>>
>>>>> Anyway, I think you have to leave @HSIC methods untouched, because C2
>>>>> would intrinsify them directly. To cater for the unsafe.cpp slow path,
>>>>> do a separate method:
>>>>>
>>>>>     @HotSpotIntrinsicCandidate
>>>>>     public final byte compareAndExchangeByteVolatile(Object o, long
>>>>> offset,
>>>>>                                                      byte expected,
>>>>>                                                      byte x) {
>>>>>         compareAndExchangeByte0(o, long, expected, x);
>>>>>     }
>>>>>
>>>>>     public final native byte compareAndExchangeByte0(...);
>>>>>
>>>>> ...and then do an entry point for it in unsafe.cpp.
>>>>
>>>> Okay that was the easy part.
>>>>
>>>> Now where do I find out the conditional compilation syntax for the
>>>> X-VarHandle.java.template file?
>>>
>>> The template itself is driven by Spp preprocessor, the same one used
>>> thorough JDK. It is called during build via
>>> ./jdk/make/gensrc/GensrcVarHandles.gmk.
>>
>> Yes but I don't know the syntax.
>
> If you find Spp.java in the forest, there is a comment section about the
> syntax:
>    ./jdk/make/src/classes/build/tools/spp/Spp.java
>
>
>>> But why do you need to change VH templates? VH implementations call
>>> Unsafe, see:
>>>
>>>         @ForceInline
>>>         static $type$ compareAndExchange(FieldInstanceReadWrite handle,
>>> Object holder, $type$ expected, $type$ value) {
>>>             return
>>> UNSAFE.compareAndExchange$Type$Volatile(Objects.requireNonNull(handle.receiverType.cast(holder)),
>>>
>>>                                                handle.fieldOffset,
>>>
>>> {#if[Object]?handle.fieldType.cast(expected):expected},
>>>
>>> {#if[Object]?handle.fieldType.cast(value):value});
>>>         }
>>>
>>> ...so once you do the Unsafe part above, everything should be linked
>>> together. The jtreg tests exercise VH byte ops, so you can run them.
>>
>> I was only making a change to the Byte version so the template needs to
>> be specialized for Byte to call CompareAndExchangeByte0.
>
> I get that.
>
> But I still think X-VarHandle.template change is unnecessary. The code
> for byte CAS generated from X-VarHandle.template calls
> Unsafe.compareAnd{Swap,Exchange}*. Why can't you leave VH template
> alone, and call CompareAndExchangeByte0 from
> Unsafe.compareAndExchangeByteVolatile, like I suggested before?

Ah I misunderstood. But even if I change this Java version, the 
intrinsic will do what it currently does, so this would only work -Xint 
mode. Anyway I'll figure something out on Monday.

Thanks,
David

>      @HotSpotIntrinsicCandidate
>      public final byte compareAndExchangeByteVolatile(Object o,
>                                                       long  offset,
>                                                       byte expected,
>                                                      byte x) {
>          compareAndExchangeByte0(o, long, expected, x);
>      }
>
>      public final native byte compareAndExchangeByte0(...);
>
> VH tests have a least one testing mode which does not intrinsify Unsafe
> methods, i.e. -Xint. compareAndExchangeByte0 path would be visited there.
>
> -Aleksey
>
>