[aarch64-port-dev ] Field too big for insn issue with Shenandoah

[Roman Kennke]

Am 10.09.2018 um 20:03 schrieb Andrew Haley:
> On 09/10/2018 04:11 PM, Roman Kennke wrote:
>> This only happens when running with Shenandoah. I suspect that the extra
>> barriers that we emit with Shenandoah make the code too big and the PC
>> too far away from the constants, but I am not sure.
> 
> Nice catch! I always knew that this was a theoretical possibility, but
> I never could find a way to trigger it.
> 
>> Does anybody have an idea why that happens and how to possibly fix it?
> 
> I'll have a look. It's going to be very awkward if loads can't reach the
> constant pool. The load offset is 19 bits, shifted by 2, so that's +/- a
> megabyte. I don't really believe that any method will generate a megabyte
> of code, so I suspect that there may be some other reason. Perhaps enabling
> debug mode just generates too much checking code.
> 
> Which test was running?
> 

It was happening with basically any program running under fastdebug + C1
+ Shenandoah. It occured because we tripled NMethodSizeLimit when
running with Shenandoah to make space for extra code for barriers. And
we 'fixed' it by leaving that flag alone:

http://hg.openjdk.java.net/shenandoah/jdk/rev/5d035e4eb822

You might want to crank up that flag to trigger the bug (if it is one)
in upstream (no Shenandoah). A proper fix might be to enforce an upper
bound for that flag that ensures offsets from code can always reach the
constants.

Thanks,
Roman