[11] 8207383: [Graal] SelfChangedCDS.java fails with "guarantee(disp == (intptr_t)(jint)disp) failed: must be 32-bit offset"

[dean.long at oracle.com]

On 7/22/18 10:28 AM, Vladimir Kozlov wrote:
> Dean,
>
> On 7/21/18 9:57 PM, dean.long at oracle.com wrote:
>> On 7/21/18 5:37 PM, Vladimir Kozlov wrote:
>>> To be clear. We don't use far call for AOT because in AOT code we 
>>> call through trampoline anyway. Right?
>>
>> Yes, we call through a "PLT".
>>
>>>
>>> My question is do we do that for all calls? Or only runtime calls?
>>>
>>
>> I believe we call through the PLT for everything except calls to code 
>> that we know is in the AOT library, like Graal-generated stubs.
>> This is a direct call:
>
> Sorry, I was not clear. I asked about code generated by Graal JIT not 
> for AOT. Your changes seem generate indirect call in all cases when 
> scratch register is specified:
>
> if (scratch != null && !GeneratePIC.getValue(crb.getOptions())) {
>

Yes, and that was also the old behavior.  The new changes are:
1) to disallow using the scratch register for AOT calls
2) allowing directJmp to make a far jump
3) using a scratch register for far calls to foreign functions 
ENABLE_STACK_RESERVED_ZONE and THROW_DELAYED_STACKOVERFLOW_ERROR. Other 
calls are left unchanged.
> What about calls from nmethod to nmethod? They should not be far 
> calls. I am not familiar with Graal and that is why I am asking.
>

For Java invokes, they can be direct or indirect.  For direct invokes on 
x64, Graal is using near calls with a null scratch register.

http://hg.openjdk.java.net/jdk/jdk11/file/d5138f8da1ba/src/jdk.internal.vm.compiler/share/classes/org.graalvm.compiler.lir.amd64/src/org/graalvm/compiler/lir/amd64/AMD64Call.java#l104

dl

> Thanks,
> Vladimir
>
>>
>>      2883:       e8 98 1a 00 00          callq  4320 
>> <Stub<ExceptionHandlerStub.exceptionHandler>>
>>
>> This calls through a PLT:
>>
>>      2c17:       e8 b4 15 00 00          callq  41d0 
>> <M1_759_java.lang.invoke.MethodHandle.asType(Ljava/lang/invoke/MethodType;)Ljava/lang/invoke/MethodHandle;_plt.entry> 
>>
>>
>> Finally, we use trampolines when an nmethod needs to make a far call 
>> to AOT code.
>>
>> dl
>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 7/21/18 11:52 AM, dean.long at oracle.com wrote:
>>>> https://bugs.openjdk.java.net/browse/JDK-8207383
>>>> http://cr.openjdk.java.net/~dlong/8207383/webrev/
>>>>
>>>> This fixes a regression caused by JDK-8181855.  The fix is to use a 
>>>> scratch register to make far calls, but force a near call for AOT.
>>>>
>>>> Changeset based on contribution by Tom Rodriguez.
>>>>
>>>> dl
>>