Value type hash code (original) (raw)

Stuart Marks stuart.marks at oracle.com
Wed Apr 11 18:42:53 UTC 2018


Interesting stuff. Several observations:

Yes, this implies that a value type's hashCode() and equals() can change if it contains a reference to a mutable object. Sure, someone (javac?) can try to warn about this, but then we get into the business of tagging some objects as immutable. I'm not sure I want to go there.

To a certain extent, people are used to this; they use mutable objects as HashMap keys all the time. They just "know" that they mustn't mutate the object while it's in use. Occasionally, though, somebody will pop up in a huff complaining that HashMap is broken, and it's because they mutated a key... and then they'll complain that this isn't properly documented. Overall though, I don't think this is a serious problem.

s'marks

On 4/10/18 7:49 AM, David Simms wrote:

After prototyping "hashCode()" for value types, here's a few observations and thoughts...  * "The general contract of hashCode" [1] is unchanged.  * The default implementation, if no user implementation is provided, is assumed to be completely based upon the entire contents of the value (there's nothing else to go on). o The current "Object" implementation, both its generation and object header hash caching are completely inappropriate for value types. o The VM cannot pretend to know one field is more significant than another. o Large values will benefit from user implementation to provide efficiency. o Whilst the VM may provide a default implementation for safety, a "javac" generated method would be optimal (can be optimized by the JIT, includes inlining).  * Values containing references whose contents are mutable pose a problem, their hash code is only as stable as the contents of the whole object graph. o Objects may suffer similar problems, difficult to say this is any more important for values. Except to say values are supposed to be "immutable" but references may break this quality, perhaps "javac" could warn when value fields are mutable objects (not always possible, e.g. field reference to an interface). I assume a the default implementation should look something like this (only with concrete fields, not reflection): int hc = 0; for (Field field : val.getClass().getDeclaredFields()) { if (Modifier.isStatic(field.getModifiers())) continue; // Using the generic JDK hash for all types hc = (31 * hc) + Objects.hashCode(field.get(val)); } return hc; This algorithm assumes the VM implements calls to reference field's hashCode(), and encodes primitives the same as their boxed JDK counter-parts (e.g. "Long.hashCode(long l)" does not generically hash two int size chunks, rather it xors hi and lo, Boolean is another interesting example 1231 || 1237). Unclear if this is actually important...however, this example: _final ByValue class MyInt implements Comparable { final int value; //.... } The user is free to explicitly delegate to "Integer.hashCode(int val)", but is it just more natural that the default works this way ? Alternative VM implementations might hash over value data payload including field padding. With h/w support (suitable checksum instruction) there might be some performance gain for large values, but then if you introduce object references, said h/w support would get broken. Said implementation would be dependent on field layout, and not give the same result on different platforms. Whilst the Javadoc states hashing "need not remain consistent from one execution of an application to another execution of the same application." [1], I'm wondering how many folks rely on consistent hashing, more than nobody I would fear. Lastly hashing large amounts of data per value seems an unlikely general use-case.

Cheers /David Simms [1] https://docs.oracle.com/javase/10/docs/api/java/lang/Object.html#hashCode()



More information about the valhalla-dev mailing list