RFR: 8338544: Dedicated Array class descriptor implementation

[Chen Liang]

On Fri, 4 Oct 2024 17:38:51 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

>> @cl4es discovered that Stack Map generation in ClassFile API uses `componentType` and `arrayType` for `aaload` `aastore` instructions, which are currently quite slow. We can split out array class descriptors from class or interfaces to support faster `arrayType` and `componentType` operations.
>> 
>> Tentative, as I currently have no way to measure the actual impact of this patch on the startup performance; however, this made the `ClassDesc` implementations much cleaner.
>
> src/java.base/share/classes/jdk/internal/constant/ArrayClassDescImpl.java line 116:
> 
>> 114:         sb.append(componentDesc);
>> 115:         return sb.toString();
>> 116:     }
> 
> Is there really that much benefit in lazily computing the descriptor? `ReferenceClassDescImpl` doesn't do this either... Maybe we can keep things simple and initialize the descriptor in the constructor?

This laziness is actually the main reason I had this implementation: In stack map generation, we need to compute the array descriptors of many types yet not using them in the end; the string computation involved a lot of allocations, especially with frequent `arrayType()` and `componentType()` calls.

> src/java.base/share/classes/jdk/internal/constant/ArrayClassDescImpl.java line 126:
> 
>> 124:         Class<?> clazz = element.resolveConstantDesc(lookup);
>> 125:         for (int i = 0; i < rank; i++)
>> 126:             clazz = clazz.arrayType();
> 
> Looking at the implementation of `arrayType`, it reflectively creates an array and then returns its class. Just makes me think we need a better we to look up an array class directly in the JDK :D

Indeed; and it's more alarming that `Class.forName("[[[[[Ljava.lang.Object;")` is slower than `Object.class.arrayType().arrayType().arrayType().arrayType().arrayType()`.

> src/java.base/share/classes/jdk/internal/constant/MethodTypeDescImpl.java line 89:
> 
>> 87: 
>> 88:     private static ClassDesc validateArgument(ClassDesc arg) {
>> 89:         if (requireNonNull(arg) == CD_void)
> 
> Is it safe to make this change? Are instances of `PrimitveClassDescImpl` canonicalized now? (not in this patch, but it looks that way in the source code).

Yes, they are canonicalized now for performance. `PrimitiveClassDescImpl` extends `DynamicConstantDesc` so creation and comparison of non-unique instances was costly.

> src/java.base/share/classes/jdk/internal/constant/ReferenceClassDescImpl.java line 58:
> 
>> 56:         if (descriptor.charAt(0) == '[') {
>> 57:             return ArrayClassDescImpl.ofValidatedDescriptor(descriptor);
>> 58:         }
> 
> I think arrays should be handled separately by the caller, or `ArrayClassDescImpl` should be made a sub-type of  `ReferenceClassDescImpl`. This factory in `ReferenceClassDescImpl` handling a sibling type seems a bit confusing.

Indeed, since the factory usages are widespread, I only did a hotfix migration. This class should now be called `ClassOrInterfaceDescImpl` (or "type klass" in hotspot terms or "declared types" in javac terms)

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

PR Review Comment: https://git.openjdk.org/jdk/pull/20665#discussion_r1788128649
PR Review Comment: https://git.openjdk.org/jdk/pull/20665#discussion_r1788125022
PR Review Comment: https://git.openjdk.org/jdk/pull/20665#discussion_r1788126100
PR Review Comment: https://git.openjdk.org/jdk/pull/20665#discussion_r1788127459