RFR: 8371164: ArrayList.addAll() optimizations [v2]
jengebr
duke at openjdk.org
Thu Nov 6 20:12:41 UTC 2025
On Tue, 4 Nov 2025 18:35:23 GMT, jengebr <duke at openjdk.org> wrote:
>> # JVM Collections Optimizations: Eliminating toArray() Performance Bottlenecks
>>
>> ## Summary
>>
>> This PR addresses performance bottlenecks in ArrayList.addAll() and Collections.SingletonSet.toArray() methods by implementing direct optimizations that bypass inefficient intermediate allocations and abstract implementations. The optimizations target high-frequency operations identified through profiling analysis, delivering 37% performance improvements for ArrayList operations and 17-43% performance improvements for SingletonSet operations under real-world conditions where multiple collection types are used.
>>
>> ## Problem Context
>>
>> ### ArrayList.addAll() Inefficiency
>> ArrayList.addAll() currently calls `c.toArray()` on the source collection to avoid iterator-based copying, but this creates unnecessary intermediate array allocation when the source is also an ArrayList. The method performs:
>>
>> 1. Call `c.toArray()` - creates intermediate array
>> 2. Call `System.arraycopy()` to copy from intermediate array to destination
>> 3. Discard intermediate array
>>
>> When both source and destination are ArrayList instances, this can be optimized to direct array copying.
>>
>> ### Collections.SingletonSet toArray() Missing Implementation
>> Collections.SingletonSet inherits the default `AbstractCollection.toArray()` implementation, which:
>>
>> 1. Creates an Object[] of the expected size
>> 2. Iterates through the collection (1 element)
>> 3. Ensures "expected" size is the actual size
>> 4. Returns the array
>>
>> For a single-element collection, this overhead is disproportionate to the actual work needed. Additionally, this implementation is vulnerable to call site poisoning, showing 74-118% performance degradation under megamorphic conditions.
>>
>> ## Optimized Methods
>>
>> ### ArrayList
>> - **`addAll(Collection<? extends E> c)`**: Added fast path for ArrayList-to-ArrayList copying using direct `System.arraycopy()` from source's internal `elementData` array, eliminating intermediate `toArray()` allocation
>>
>> ### Collections.SingletonSet
>> - **`toArray()`**: Direct implementation returning `new Object[] {element}`
>> - **`toArray(T[] a)`**: Direct implementation with proper array sizing and null termination per Collection contract
>>
>> ## Performance Impact
>>
>> | Class | Method | Size | Baseline | Optimized | Improvement |
>> |-------|--------|------|----------|-----------|-------------|
>> | ArrayList | addAll | 0 | 10.149 ns/op, 40 B/op | 3.929 ns/op, 24 B/op | **...
>
> jengebr has updated the pull request incrementally with one additional commit since the last revision:
>
> Adding direct unit tests, minor revisions to optimizations
Further investigation into the duplication/de-duplication of the fastpath shows that I *cannot reproduce* the negative results - they are clearly logged, I simply can't generate new runs with the same data. No idea why.
I've updated the code with the simplified branching, all comments are addressed.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/28116#issuecomment-3499205653
More information about the core-libs-dev
mailing list