Proposal: Large arrays

Reinier Zwitserloot reinier at zwitserloot.com
Tue Mar 24 09:53:22 PDT 2009


This change has serious impact on the JVM, and the language support  
for java is clearly a patchy hack: The 'best' solution (however  
unattainable that may be) is to just ensure that arrays can be long- 
indexed, and forego the double bracket notation. You also forgot to  
mention you need rather a lot of work in the reflection API, because  
you're introducing a new magic type here (there are the primitives,  
arrays, and objects. You're adding a new one: large arrays).


I'm opposed to any solution where a clearly superior one that is  
similar is obvious, and not clearly impossible to implement (even if  
it will be implemented in java8 or beyond). If your proposal was  
somehow compatible with such a future, that'd be okay, I guess, but it  
isn't: If this double-bracket feature is added, then the extra type  
will never go away.


If you're willing to eat up 20 of the precious opcode space, you can  
instead propose something that is much simpler for all involved:

Any attempt to use arrays with a long instead of an int will work  
fine, it'll just use a different opcodes. Internally, there are such  
things as large arrays, but you can index into them using an int, and  
you can index into a plain array with a long (with the usual result if  
your long has a non-zero upper int: AIOOBException). For most other  
things including reflection, there is no discernable difference  
between large and non-large arrays. You can't grow arrays, so I don't  
really understand why you'd want a distinction in the first place.  
That's just implementation detail leaking through.

  --Reinier Zwitserloot



On Mar 24, 2009, at 17:35, james lowden wrote:

>
> Proposal: Large arrays
>
> AUTHOR(S):
>
> J. Lowden
>
> VERSION:
>
> 1.0    Initial version.
>
>
> OVERVIEW
>
> FEATURE SUMMARY:
>
> Java arrays are accessed via 32-bit ints, resulting in a maximum  
> theoretical array size of 2147483647 elements.  While
>
> this is enough for most uses, some applications that need to handle  
> large sequential data sets in memory would benefit
>
> from the ability to work with arrays indexed using 64-bit indices.   
> As "simply" changing the current array syntax to use
>
> longs as indices rather than ints would have the potential to break  
> a large amount of existing code, I'm not proposing
>
> doing any such thing.  Instead, I'd like to see a separate but  
> parallel array feature for creating and accessing both
>
> types of arrays.  A rather boring example, which simply fills a byte  
> array with hexadecimal "FF" values, is shown below:
>
> //int-indexed array
> byte [] foo = new byte [200000];
> for (int i = 0; i < foo.length; i++)
>  foo [i] = (byte) 0xff;
>
> //long-indexed array
> byte [[]] foo2 = new byte [[20000000000L]];
> for (long i = 0; i < foo2.length; i++)
>  foo2 [i] = (byte) 0xff;
>
> Syntax for declaration, instantation, instanceof and casting would  
> use doubled square brackets to differentiate it from
>
> current array access.  Single brackets can still be used for access/ 
> mutation as the compiler should have sufficient
>
> information to decide whether it can treat a specific reference as  
> an int-indexed or long-indexed ("large") array.  The length field  
> would be a long rather than an int.  Like standard arrays, large  
> arrays would derive directly from
>
> java.lang.Object.
>
>
> MAJOR ADVANTAGE:
>
> Java gains the ability to easily work with large sequential data sets.
>
>
> MAJOR BENEFIT:
>
> Declaring extremely large arrays simply isn't possible right now; in  
> cases where such a structure is desirable, something
>
> has to be hacked together.  For applications dealing with large data  
> sets, the current upper limit on array size is constricting, and  
> will only grow more so as 64-bit systems continue to become more  
> common and RAM less expensive.
>
>
> MAJOR DISADVANTAGE:
>
> This can't be implemented well solely via a compiler patch; existing  
> VMs would likely need to be changed to support this
>
> concept natively; new VM instructions parallel to the existing array  
> instructions would likely be required.  Also, a class
>
> parallel to java.util.Arrays for performing standard operations on  
> large arrays would likely be desirable (although
>
> presumably fairly simple to implement.)
>
>
> ALTERNATIVES:
>
> Build your own class for storing long-indexes sequences, possibly as  
> an array-or-arrays.
>
>
> EXAMPLES
>
> SIMPLE EXAMPLE:
>
> // calculate the first 3 billion fibonacci numbers and store them in  
> an array
>
> long [[]] fib = new long [[3000000000L]];
> fib [0] = 0;
> fib [1] = 0;
> for (long i = 2; i < fib.length; i++)
>  fib [i] = fib [i - 1] + fib [i - 2];
>
>
> ADVANCED EXAMPLE:
>
> // this doesn't really do anything particular useful, but does serve  
> to show how casting and instanceof are handled
> byte [] foo = new byte [400000];
> byte [[]] blah = new byte [[40000000000L]];
> Object temp = Math.random () < 0.5 ? foo : blah;
> if (foo instanceof byte [[]])
>  System.out.println (((byte [[]]) foo).length);
> else
>  System.out.println (((byte []) foo).length);
>
>
> DETAILS
>
> SPECIFICATION:
>
> Syntax-wise, the following expressions become legal:
>
> SomeType [[]] foo;    // declares foo as a long-indexed "large"  
> array of SomeType,
> 			// where sometype is either a primitive or reference type
>
> foo = new SomeType [[some_long_value]];   // instantiates a large  
> array of length
> 					// some_long_value, which is either a long
> 					// or some type that can upconvert or unbox
> 					// to a long
>
> long some_long = foo.length;   // large arrays will have a "length"  
> field, which is
> 				// a long indicating the number of elements in the array
>
> foo instanceof SomeType [[]]   // returns true if foo is a large array
> 				// of SomeType, false otherwise
>
> (SomeType [[]]) foo    // casts foo to a large array of SomeType.   
> If foo isn't
> 			// a large array of SomeType or a subclass of SomeType,
> 			// throws a ClassCastException
>
>
> Large arrays are not castable or assignable to or from int-indexed  
> arrays of the same element type.
>
> int [[]] p = new int [40000]; // compiler error
> int [] p = new int [[760000]]; // compiler error
> int [] foo = (int []) (new int [[4040404040L]]); // throws  
> ClassCastException
> int [[]] foo = (int [[]]) (new int [4040]); // throws  
> ClassCastException
>
>
> Canonical class names for large arrays would be generated in a  
> similar fashion to those for standard arrays, but
>
> using the opposite square bracket.  The following table shows some  
> examples.
>
> declared type     class name
> int [[]]          ]I
> int [[]][[]]      ]]I
> Object [[]]       ]Ljava.lang.Object;
>
>
> The same set of indicator character (i.e., 'I' for int, 'L' for  
> reference types, etc.) as are currently used for arrays
>
> would be used for large arrays.
>
> A set of additional VM instructions parallel to the current array  
> instructions would be added to support this feature.
>
> The following table shows the current instruction, the proposed  
> instruction for large arrays, and any semantic
>
> differences (reference http://java.sun.com/docs/books/jvms/second_edition/html/Instructions2.doc.html) 
> :
>
> array instruction   proposed instruction   differences
> aaload              lxaaload               index value becomes a long
> aastore             lxaastore              index value becomes a long
> anewarray           lxanewarray            count value becomes a long
> arraylength         lxarraylength          return value becomes a long
> baload              lxbaload               index value becomes a long
> bastore             lxbastore              index value becomes a long
> caload              lxcaload               index value becomes a long
> castore             lxcastore              index value becomes a long
> daload              lxdaload               index value becomes a long
> dastore             lxdastore              index value becomes a long
> faload              lxfaload               index value becomes a long
> fastore             lxfastore              index value becomes a long
> iaload              lxiaload               index value becomes a long
> iastore             lxiastore              index value becomes a long
> laload              lxlaload               index value becomes a long
> lastore             lxlastore              index value becomes a long
> multianewarray      lxmultianewarray       countN values become longs
> newarray            lxnewarray             count value becomes a long
> saload              lxsaload               index value becomes a long
> sastore             lxsastore              index value becomes a long
>
>
> COMPILATION:
>
> Compilation would be parallel to the present mechanism for compiling  
> arrays, but would output the lx* VM instructions listed above for  
> large arrays instead of the current array instructions.
>
> TESTING:
>
> Any test sufficient to test support for current arrays should work  
> for this as well.
>
> LIBRARY SUPPORT:
>
> It would probably be desirable to create large array versions of the  
> various java.util.Arrays methods, whether that be done by adding  
> methods to the existing java.util.Arrays class or putting them  
> somewhere new.  At some point in the future, large java.util.List<X>  
> might be a desirable library feature, but that is beyond the scope  
> of this proposal.
>
> REFLECTIVE APIS:
>
> An additional class (LargeArray or something of that sort) would  
> need to be added to the java.lang.reflect package.  This would have  
> a similar set of methods and constructors to  
> java.lang.reflect.Array, but with long "index" and "length"  
> parameters where appropriate.
>
> OTHER CHANGES:
>
> VM needs to support the above-listed large array instructions.
>
>
> MIGRATION:
>
> Existing "hacks" to provide this kind of behavior could be replaced  
> with large arrays.
>
>
> COMPATIBILITY
>
> BREAKING CHANGES:
>
> None. This feature simple doesn't exist; the proposed declaration/ 
> instantation syntax currently results in a compiler error.
>
>
> EXISTING PROGRAMS:
>
> No changes.
>
>
> REFERENCES
>
> EXISTING BUGS:
>
> 4880587 (64-Bit indexing for arrays)
>
>
> URL FOR PROTOTYPE (optional):
>
> None.
>
>
>
>




More information about the coin-dev mailing list