RFR: 8343789: Move mutable nmethod data out of CodeCache [v11]

[Dean Long]

On Tue, 18 Feb 2025 19:23:59 GMT, Boris Ulasevich <bulasevich at openjdk.org> wrote:

>> This change relocates mutable data (such as relocations, oops, and metadata) from the nmethod. The change follows the recent PR #18984, which relocated immutable nmethod data from the CodeCache.
>> 
>> The core idea remains the same: use the CodeCache for executable code while moving additional data to the C heap. The primary motivations are improving security and enhancing code density.
>> 
>> Although performance is not the main focus, testing on AArch64 CPUs, where code density plays a significant role, has shown a 1–2% performance improvement in specific scenarios, such as the CodeCacheStress test and the Renaissance Dotty benchmark.
>> 
>> The numbers. Immutable data constitutes **~30%** on the nmehtod. Mutable data constitutes **~8%** of nmethod. Example (statistics collected on the CodeCacheStress benchmark):
>> - nmethod_count:134000, total_compilation_time: 510460ms
>> - total allocation time malloc_mutable/malloc_immutable/CodeCache_alloc: 62ms/114ms/6333ms,
>> - total allocation size (mutable/immutable/nmentod): 64MB/192MB/488MB
>> 
>> Functional testing: jtreg on arm/aarch/x86.
>> Performance testing: renaissance/dacapo/SPECjvm2008 benchmarks.
>> 
>> Alternative solution (see comments): In the future, relocations can be moved to _immutable_data.
>
> Boris Ulasevich has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains 12 commits:
> 
>  - Address review comments: cleanup, move fields to avoid padding, fix CodeBlob purge to call os::free, fix nmethod::print, update Layout description
>  - add a separate adrp_movk function to to support targets located more than 4GB away
>  - Force the use of movk in combination with adrp and ldr instructions to address scenarios
>    where os::malloc allocates buffers beyond the typical ±4GB range accessible with adrp
>  - Fixing TestFindInstMemRecursion test fail with XX:+StressReflectiveCode option:
>    _relocation_size can exceed 64Kb, in this case _metadata_offset do not fit into int16.
>    Fix: use _oops_size int16 field to calculate metadata offset
>  - removing dead code
>  - a bit of cleanup and addressing review suggestions
>  - rework movoop for not_supports_instruction_patching case: correcting in ldr_constant and relocations fixup
>  - remove _code_end_offset
>  - update jvm.hotspot.code.CodeBlob class
>  - update: mutable data for all CodeBlobs with relocations
>  - ... and 2 more: https://git.openjdk.org/jdk/compare/e1d0a9c8...6c3370be

If we put the "oop pool" nearby, then we could continue to use the shorter instruction sequence.  We could put the oops in a "DataBlob" in the codecache.  With a segment codecache, they could even go in their own segment.
I suspect that many nmethods are referencing the same oops.  We could consider storing the oops in a deduplicated structure that GC could quickly scan, similar to OopStorage but reference-counted.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/21276#issuecomment-2672710391