On Tue, 17 Mar 2026 10:07:58 GMT, Marc Chevalier <[email protected]> wrote:
> To do `acmp`, if the pointer comparison fails, we check nullity of operands,
> whether operands have the same class, and if this class is a value class. And
> if everything works, we call `isSubstitutable`.
>
> Currently, when we have profiling information, we use it to speculate that
> operands are null or of an identity class (in which case, pointer comparison
> is all we should do). This patch proposes to take advantage of the profiling
> when it hints that operands are value objects. By speculating that at least
> one operand is of a given value class, the call to `isSubstitutable` can be
> intrinsified later, and thus, spare the Java call.
>
> Speculating on one operand is enough since we will check that operands are of
> the same type.
>
>
> Some relevant microbenchmark in `valhalla.acmp.array.Value032` (in ns/op
> (lower is better)):
>
> | | branch_obj_equals000 | branch_val_equals000 |
> branch_obj_equals025 | branch_val_equals025 |
> |---------------|----------------------|----------------------|----------------------|----------------------|
> | Before | 4.361 ± 0.153 ns/op | 0.907 ± 0.015 ns/op | 3.332 ± 0.040
> ns/op | 0.845 ± 0.032 ns/op |
> | With this fix | 0.911 ± 0.019 ns/op | 0.899 ± 0.021 ns/op | 0.838 ± 0.023
> ns/op | 0.806 ± 0.012 ns/op |
>
>
>
> <details>
> <summary>Full results of <tt>micro:valhalla.acmp.array.Value032</tt></summary>
> In more details, before:
>
>
> Benchmark Mode Cnt Score Error
> Units
> Value032.branch_obj_equals000 avgt 15 4.087 ± 0.122
> ns/op
> Value032.branch_obj_equals025 avgt 15 3.117 ± 0.067
> ns/op
> Value032.branch_obj_equals050 avgt 15 2.235 ± 0.083
> ns/op
> Value032.branch_obj_equals075 avgt 15 1.292 ± 0.077
> ns/op
> Value032.branch_obj_equals100 avgt 15 0.316 ± 0.004
> ns/op
> Value032.branch_val_equals000 avgt 15 0.891 ± 0.024
> ns/op
> Value032.branch_val_equals025 avgt 15 0.839 ± 0.033
> ns/op
> Value032.branch_val_equals050 avgt 15 0.869 ± 0.087
> ns/op
> Value032.branch_val_equals075 avgt 15 0.816 ± 0.023
> ns/op
> Value032.branch_val_equals100 avgt 15 0.906 ± 0.032
> ns/op
> Value032.result_obj_equals000 avgt 15 4.115 ± 0.185
> ns/op
> Value032.result_obj_equals025 avgt 15 3.326 ± 0.346
> ns/op
> Value032.result_obj_equals050 avgt 15 2.353 ± 0.516 n...
For the first prototype, and significant refactoring.
test/hotspot/jtreg/compiler/valhalla/inlinetypes/TestMissingOptAcmp.java line
45:
> 43: */
> 44:
> 45: value class MyValue1MissingOptAcmp {
After [8368988: [lworld] Rename classes in valhalla/inlinetypes to be
distinct](https://bugs.openjdk.org/browse/JDK-8368988), the class in
`TestNewAcmp.java` was copy-pasted here. I'm just redoing the fix as it makes
my IDE complain again.
-------------
PR Comment: https://git.openjdk.org/valhalla/pull/2237#issuecomment-4073784885
PR Review Comment:
https://git.openjdk.org/valhalla/pull/2237#discussion_r2945736272
