RFR: 8315771: [JVMCI] Resolution of bootstrap methods with int[] static arguments [v11]

[Doug Simon]

On Tue, 19 Sep 2023 11:37:09 GMT, Sacha Coppey <duke at openjdk.org> wrote:

>> Currently, `jdk.vm.ci.meta.ConstantPool.lookupBootstrapMethodInvocation` does not support static arguments of type `int[]`.
>> 
>> Supporting those static arguments allows to correctly lookup the `BootstrapMethodInvocation` of some `InvokeDynamic` and `DynamicConstant`.
>> 
>> To lookup the constant at the index in the static arguments index list, a new class is introduced, allowing to lazily resolve the constant or obtain the constant pool index of the arguments:
>> 
>> 
>>     static class CachedBSMArgs extends AbstractList<JavaConstant> {
>>         private final JavaConstant[] cache;
>>         private final HotSpotConstantPool cp;
>>         private final int bssIndex;
>> 
>>         CachedBSMArgs(HotSpotConstantPool cp, int bssIndex, int size) {
>>             this.cp = cp;
>>             this.bssIndex = bssIndex;
>>             this.cache = new JavaConstant[size];
>>         }
>> 
>>         @Override
>>         public JavaConstant get(int index) {
>>             JavaConstant res = cache[index];
>>             if (res == null) {
>>                 int argCpi = compilerToVM().bootstrapArgumentIndexAt(cp, bssIndex, index);
>>                 res = compilerToVM().lookupConstantInPool(cp, argCpi, false);
>>                 if (res == null) {
>>                     res = JavaConstant.forInt(argCpi);
>>                 }
>>                 cache[index] = res;
>>             }
>>             return res;
>>         }
>> 
>>         @Override
>>         public int size() {
>>             return cache.length;
>>         }
>>     }
>
> Sacha Coppey has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add an extra run configuration for TestDynamicConstant instead of replacing it

While it's tempting to introduce a new type to distinguish between a primitive constant and an index in the constant pool, it would involve changing API that is already used in Graal.
What's more, we're (ab)using the fact that `java.lang.invoke.ConstantGroup.get(int index)` always returns a boxed value. Boxed Java values in JVMCI are always represented by a `JavaConstant` for which `c.getJavaKind() == JavaKind.Object`. So we can use `PrimitiveConstant` here as a side channel to represent a constant pool index.

@Zeavee can you please update the javadoc of `BootstrapMethodInvocation.getStaticArguments()` one more time to make this crystal clear:


        /**
         * Gets the static arguments with which the bootstrap method will be invoked.
         *
         * The {@linkplain JavaConstant#getJavaKind kind} of each argument will be
         * {@link JavaKind#Object} or {@link JavaKind#Int}. The latter represents an
         * unresolved {@code CONSTANT_Dynamic_info} entry. To resolve this entry, the
         * corresponding bootstrap method has to be called first:
         *
         * <pre>
         * List<JavaConstant> args = bmi.getStaticArguments();
         * List<JavaConstant> resolvedArgs = new ArrayList<>(args.size());
         * for (JavaConstant c : args) {
         *     JavaConstant r = c;
         *     if (c.getJavaKind() == JavaKind.Int) {
         *         // If needed, access corresponding BootstrapMethodInvocation using
         *         // cp.lookupBootstrapMethodInvocation(pc.asInt(), -1)
         *         r = cp.lookupConstant(c.asInt(), true);
         *     } else {
         *         assert c.getJavaKind() == JavaKind.Object;
         *     }
         *     resolvedArgs.append(r);
         * }
         * </pre>


No need to re-run testing after this change as it's only a doc change.

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

PR Comment: https://git.openjdk.org/jdk/pull/15588#issuecomment-1727370299