[master] RFR: 8303769: [Lilliput] Fix interpreter asymmetric fast-locking [v2]

[Thomas Stuefe]

On Wed, 8 Mar 2023 18:57:15 GMT, Roman Kennke <rkennke at openjdk.org> wrote:

>> Currently we get the asymmetric locking check in the interpreters wrong:
>> 
>> ldr(header_reg, Address(rthread, JavaThread::lock_stack_current_offset()));
>> cmpoop(header_reg, obj_reg);
>> br(Assembler::NE, slow_case);
>> 
>> The intention is to load the top of the lock-stack, and compare it to the unlocked object, and, if not equal, branch to the slow-path to handle it. However, what it really does is, it loads the *address* of the top of lock-stack, and compares that to the unlocked object. This can never succeed, and therefore we always call the slow-path. Additionally, the address is not the address of the topmost object, it is the address of the next free slot. What we really want to load is the element at -1 oop from that address. This is not incorrect, but it's unnecessarily slow.'
>> 
>> I've already pushed that fix to the upstream fast-locking PR https://github.com/openjdk/jdk/pull/10907
>> 
>> Testing:
>>  - [x] tier1 (x86_64, x86_32, aarch64)
>>  - [ ] tier2 (x86_64, x86_32, aarch64)
>
> Roman Kennke has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Alias tmp registers, also fix RISCV (hopefully)

+1. Thanks for humoring my register name obsession.

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

Marked as reviewed by stuefe (Committer).

PR: https://git.openjdk.org/lilliput/pull/76