We should leverage profiling info around acmp to speculate that one operand is a value object of known type. This would then enable the optimization added by JDK-8228361.


Original report:

acmp of value types from non-flattened arrays is 5x times slower than on flattened.

Benchmark Mode Cnt Score Error Units
o.o.b.v.a.array.Value032.branch_obj_equals025 avgt 5 4.647 ± 0.543 ns/op
o.o.b.v.a.array.Value032.branch_val_equals025 avgt 5 0.607 ± 0.030 ns/op

branch_val_equals025 - flattened array.
branch_obj_equals025 - non-flattened, runtime type of the array is "Object[]".
The reason:

  33.71% Unknown, level 0 java.lang.invoke.MethodHandle::invokeBasic, version 1, compile id 550
  32.54% C2, level 4 java.lang.runtime.ValueObjectMethods::isSubstitutable, version 4, compile id 953
   8.08% C2, level 4 org.openjdk.bench.valhalla.acmp.array.Value032::cmp_branch_obj, version 3, compile id 923
   3.67% C2, level 4 java.lang.invoke.LambdaForm$MH.0x000000007d08b400::invoke, version 2, compile id 922