[External] : Re: premain: Possible solutions to use runtime G1 region grain size in AOT code (JDK-8335440)

[Andrew Dinn]

On 15/08/2024 23:30, Vladimir Kozlov wrote:

>> Why you still use base+offsets for constants address? Is it for Aarch64?
>>
>> What I want is something like this:
>>
>> static address grain_shift_address() { return 
>> &_aot_runtime_constants::_grain_shift; }

Hmm, I thought I had made the changes to use a direct address but they 
were not in the aarch64 version I pushed. I have pushed a new version 
that fixes the macroassembler and g1 stubs to load the field address as 
a constant. I removed the xxx_offset() methods from AOTRuntimeConstants 
and replaced them with xxx_address() methods to make sure it can only be 
done that way.

>> And you will not need `#if INCLUDE_CDS` if you put AOTRuntimeConstants 
>> into SCCache.cpp.

Yes, SCCache.hpp/cpp is a much better place for this class than 
stubRoutines.hpp/cpp. The need to generate AOT code variants that employ 
constants from the current runtime is not specific to stubGenerator 
stubs. However, it always relates to use of the AOT cache.

>> I don't see the need to separate create_field_address_list() code from 
>> field_addresses_list().

Sorry, this was a hangover from an initial implementation. The getter 
now returns the array from a static field initialized in SCCache.cpp.

>> Add assert to field_addresses_list() to check that values are 
>> initialized.
> 
> Actually you don't need assert since we need only addresses of fields.

Left as is.

> I don't see use of aot_runtime_constants_at() anymore.

It is redundant and has been removed.

> Also why you need aot_runtime_constants()? Why not have 
> _aot_runtime_constants as static field of AOTRuntimeConstants class 
> which you can initialize in (again) SCCache.cpp?
Agreed. The static singleton is now only used internally to the class. 
So, we no longer need aot_runtime_constants() any more.

Latest code:

 
https://github.com/adinn/leyden/compare/premain...adinn:leyden:JDK-8335440-load-via-codecache?expand=1

This passes tier1 tests on aarch64 and successfully improves performance 
for the javac benchmark.

regards,


Andrew Dinn
-----------