On Tue, 7 May 2024 10:46:41 GMT, Claes Redestad <redes...@openjdk.org> wrote:
>> Chen Liang has updated the pull request incrementally with one additional
>> commit since the last revision:
>>
>> Add a mixed scenario
>
> Nice improvement to the micro. It might be preferable to use random
> generation with a hardcoded or parameterized seed to reduce run-to-run
> variation, eg. `new Random(1)`.
>
> Agree that call sites are unlikely to mix `CD` and `Class`, but we're likely
> to see the bimorphicness from passing Primitive and ReferenceCDs. This is
> typically a very small and constant call overhead.
>
> The primitives path seem to be the weak link, and maybe we could take a cue
> from `sun.util.invoke.Wrapper` and set up a similar perfect hash table
> instead of a switch. This seem somewhat profitable:
>
> Name (type) Cnt Base Error Test
> Error Unit Change
> TypeKindBench.fromClassDescs PRIMITIVE 6 196,203 ± 2,469 147,898 ± 8,786
> ns/op 1,33x (p = 0,000*)
> TypeKindBench.fromClasses PRIMITIVE 6 325,336 ± 12,203 206,084 ± 2,180
> ns/op 1,58x (p = 0,000*)
>
>
> The `fromClasses PRIMITIVE` case is still a bit behind since getting the
> descriptorString takes a few extra (non-allocating) steps.
>
> Patch:
>
> diff --git a/src/java.base/share/classes/java/lang/classfile/TypeKind.java
> b/src/java.base/share/classes/java/lang/classfile/TypeKind.java
> index 5ba566b3d06..dd0a06c63ea 100644
> --- a/src/java.base/share/classes/java/lang/classfile/TypeKind.java
> +++ b/src/java.base/share/classes/java/lang/classfile/TypeKind.java
> @@ -58,6 +58,7 @@ public enum TypeKind {
>
> private final String name;
> private final String descriptor;
> + private final char basicTypeChar;
> private final int newarrayCode;
>
> /** {@return the human-readable name corresponding to this type} */
> @@ -100,6 +101,7 @@ public TypeKind asLoadable() {
> TypeKind(String name, String descriptor, int newarrayCode) {
> this.name = name;
> this.descriptor = descriptor;
> + this.basicTypeChar = descriptor.charAt(0);
> this.newarrayCode = newarrayCode;
> }
>
> @@ -154,7 +156,29 @@ public static TypeKind fromDescriptor(CharSequence s) {
> */
> public static TypeKind from(TypeDescriptor.OfField<?> descriptor) {
> return descriptor.isPrimitive() // implicit null check
> - ? fromDescriptor(descriptor.descriptorString())
> + ? ofPrimitive(descriptor.descriptorString())
> : TypeKind.ReferenceType;
> }
> +
> + // Perfect hashing for basic types, borrowed from sun.invoke.util.Wrapper
> + private static final TypeKind[] FROM_CHAR = new TypeKind[16];
> +
> + static {
> + for (TypeKind w : values()) {
> + if...
@cl4es I have tried 3 scenarios with the fixed seed: baseline before your
comment, hash table over only primitives (your proposal), hash table over all
chars (my uploaded patch)
All of these are without the specialization of `isPrimitive()`.
I observe that hash for all chars boosts the mixed scenario; cannot explain why
exactly, but I go for hashing all chars because it should benefit the
general-purpose `fromDescriptor`, used by ClassFile API's implementations too.
What do you think?
No hash table:
Benchmark (type) Mode Cnt Score Error Units
TypeKindBench.fromClassDescs PRIMITIVE avgt 6 121.863 ± 3.968 ns/op
TypeKindBench.fromClassDescs REFERENCE avgt 6 56.158 ± 1.469 ns/op
TypeKindBench.fromClassDescs MIXED avgt 6 141.508 ± 5.412 ns/op
TypeKindBench.fromClasses PRIMITIVE avgt 6 232.201 ± 3.988 ns/op
TypeKindBench.fromClasses REFERENCE avgt 6 15.663 ± 0.359 ns/op
TypeKindBench.fromClasses MIXED avgt 6 155.033 ± 1.927 ns/op
Primitive-only hash:
Benchmark (type) Mode Cnt Score Error Units
TypeKindBench.fromClassDescs PRIMITIVE avgt 6 112.731 ± 1.985 ns/op
TypeKindBench.fromClassDescs REFERENCE avgt 6 57.339 ± 1.119 ns/op
TypeKindBench.fromClassDescs MIXED avgt 6 126.926 ± 3.432 ns/op
TypeKindBench.fromClasses PRIMITIVE avgt 6 154.779 ± 4.008 ns/op
TypeKindBench.fromClasses REFERENCE avgt 6 15.619 ± 0.343 ns/op
TypeKindBench.fromClasses MIXED avgt 6 132.254 ± 2.133 ns/op
Current hash:
Benchmark (type) Mode Cnt Score Error Units
TypeKindBench.fromClassDescs PRIMITIVE avgt 6 116.446 ± 6.085 ns/op
TypeKindBench.fromClassDescs REFERENCE avgt 6 56.399 ± 1.145 ns/op
TypeKindBench.fromClassDescs MIXED avgt 6 121.087 ± 2.410 ns/op
TypeKindBench.fromClasses PRIMITIVE avgt 6 154.842 ± 4.600 ns/op
TypeKindBench.fromClasses REFERENCE avgt 6 15.897 ± 0.454 ns/op
TypeKindBench.fromClasses MIXED avgt 6 118.444 ± 3.616 ns/op
-------------
PR Comment: https://git.openjdk.org/jdk/pull/19109#issuecomment-2098712190
