On Tue, 16 Jul 2024 12:36:08 GMT, Roman Kennke <rken...@openjdk.org> wrote:
>> Axel Boldt-Christmas has updated the pull request incrementally with 10
>> additional commits since the last revision:
>>
>> - Remove try_read
>> - Add explicit to single parameter constructors
>> - Remove superfluous access specifier
>> - Remove unused include
>> - Update assert message OMCache::set_monitor
>> - Fix indentation
>> - Remove outdated comment LightweightSynchronizer::exit
>> - Remove logStream include
>> - Remove strange comment
>> - Fix javaThread include
>
> src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp line 674:
>
>> 672:
>> 673: // Search for obj in cache.
>> 674: bind(loop);
>
> Same loop transformation would be possible here.
I tried the following (see diff below) and it shows about a 5-10% regression in
most the `LockUnlock.testInflated*` micros. Also tried with just `num_unrolled
= 1` saw the same regression. Maybe there was some other pattern you were
thinking of. There are probably architecture and platform differences. This can
and should probably be explored in a followup PR.
diff --git a/src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp
b/src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp
index 5dbfdbc225d..4e6621cfece 100644
--- a/src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp
+++ b/src/hotspot/cpu/x86/c2_MacroAssembler_x86.cpp
@@ -663,25 +663,28 @@ void C2_MacroAssembler::fast_lock_lightweight(Register
obj, Register box, Regist
const int num_unrolled = 2;
for (int i = 0; i < num_unrolled; i++) {
- cmpptr(obj, Address(t));
- jccb(Assembler::equal, monitor_found);
- increment(t, in_bytes(OMCache::oop_to_oop_difference()));
+ Label next;
+ cmpptr(obj, Address(t, OMCache::oop_to_oop_difference() * i));
+ jccb(Assembler::notEqual, next);
+ increment(t, in_bytes(OMCache::oop_to_oop_difference() * i));
+ jmpb(monitor_found);
+ bind(next);
}
+ increment(t, in_bytes(OMCache::oop_to_oop_difference() * (num_unrolled -
1)));
Label loop;
// Search for obj in cache.
bind(loop);
-
- // Check for match.
- cmpptr(obj, Address(t));
- jccb(Assembler::equal, monitor_found);
-
+ // Advance.
+ increment(t, in_bytes(OMCache::oop_to_oop_difference()));
// Search until null encountered, guaranteed _null_sentinel at end.
cmpptr(Address(t), 1);
jcc(Assembler::below, slow_path); // 0 check, but with ZF=0 when *t == 0
- increment(t, in_bytes(OMCache::oop_to_oop_difference()));
- jmpb(loop);
+
+ // Check for match.
+ cmpptr(obj, Address(t));
+ jccb(Assembler::notEqual, loop);
// Cache hit.
bind(monitor_found);
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/20067#discussion_r1715249312
