[master] RFR: 8303813: [Lilliput] (AArch64) Use tbz instead of tst and br in load_klass()

[Roman Kennke]

On Wed, 8 Mar 2023 21:39:50 GMT, Paul Hohensee <phh at openjdk.org> wrote:

>> In the implementation of load_klass(), we have a test-and-branch idiom:
>> 
>> 
>> tst(dst, markWord::monitor_value);
>> br(Assembler::EQ, fast);
>> 
>> 
>> We can make this better:
>> 
>> `tbz(dst, exact_log2(markWord::monitor_value), fast);
>> `
>> This is not only smaller and perhaps a little faster, it also has the advantage that it doesn't touch the condition flags. This is relevant in at least one place: in verify_oop() the condition flag is alive across the load_klass() call and one jtreg test (runtime/CheckUnhandledOops/TestVerifyOops.java) is failing because of that.
>> 
>> Testing:
>>  - [x] tier1
>
> src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp line 4266:
> 
>> 4264:   // Check if we can take the (common) fast path, if obj is unlocked.
>> 4265:   ldr(dst, Address(src, oopDesc::mark_offset_in_bytes()));
>> 4266:   tbz(dst, exact_log2(markWord::monitor_value), fast);
> 
> This is a bit fragile because monitor_value must be a power of 2 for it to work.

Yes, and it is (and hopefully, always will be - or else that code would need to be changed anyway). exact_log2() has an assert that checks exactly that.

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

PR: https://git.openjdk.org/lilliput/pull/78