Eliot Miranda a écrit :
> 
>     Also, lastChunkSize is not set to a correct size.
>     It should be:
>            (lastChunkSize := size - base + 1 \\ chunkSize) = 0 ifTrue:
>                    [lastChunkSize := chunkSize].
>     which can be more efficiently written:
>            lastChunkSize := size - base \\ chunkSize + 1.
> 
>     Well, I guess every one use with base = 1...
> 
> 
> No that's not right.  base determines the first index, so that e.g. 
>     SparseLargeArray
> new: (2 raisedToInteger: 32)
> chunkSize: 32 * 1024
> arrayClass: ByteArray
> base: 0
> defaultValue: 0
> 
> provides a bye-addressible array of 2^32 elements with indices 0 through 
> (2 raisedTo: 32) - 1.
> 

That's interesting because that's exactly how I understood 'base' first.
But that's not how it did work before your patches.
SparseLargeTable inherits from ArrayedCollection for which the whole 
protocol require indices between 1 and self size...

And that is effectively how SparseLargeTable at: and at:put: did handle 
the indices... (mySparseTable at: 0) did not work.

A logical explanation was that base was here for saving bytes, but that 
was only a deduction, i did not discuss that with Yoshiki.

That's also why the (self basicAt: anIndex) were protected in #noChekAt:
the first elements (before base) and also the last ones (see how 
#zapDefaultOnlyEntries use #privateSize: ) were potentially out of self 
"basicBounds".

Now, the modifications you made are appealing, but beware not using one 
of the 1-based inherited message on self. That's rather not bug-proof.

Nicolas