TO DO

• We should extend the benchmarking module to allow

  timethis(main, { MEMOIZED => [ suba, subb ] })

  What would this do? It would time main three times, once with suba and subb unmemoized, twice with them memoized.

  Why would you want to do this? By the third set of runs, the memo tables would be fully populated, so all calls by main to suba and subb would return immediately. You would be able to see how much of main's running time was due to time spent computing in suba and subb. If that was just a little time, you would know that optimizing or improving suba and subb would not have a large effect on the performance of main. But if there was a big difference, you would know that suba or subb was a good candidate for optimization if you needed to make main go faster.

• Maybe a tied-hash interface to the memo-table, which a hook to automatically populate an entry if no value is there yet?

• Perhaps memoize should return a reference to the original function as well as one to the memoized version? But the programmer could always construct such a reference themselves, so perhaps it's not necessary. We save such a reference anyway, so a new package method could return it on demand even if it wasn't provided by memoize. We could even bless the new function reference so that it could have accessor methods for getting to the original function, the options, the memo table, etc.

• The TODISK feature is not ready yet. It will have to be rather complicated, providing options for which disk method to use (GDBM? DB_File? Flat file? Storable? User-supplied?) and which stringizing method to use (FreezeThaw? Marshal? User-supplied?)

• Maybe an option for automatic expiration of cache values? (`After one day,' `After five uses,' etc.)

• Put in a better example than fibo. Show an example of a nonrecursive function that simply takes a long time to run. getpwuid for example? But this exposes the bug that you can't say memoize('getpwuid'), so perhaps it's not a very good example.

  Well, I did add the ColorToRGB example, but it's still not so good. These examples need a lot of work. factorial might be a better example than fibo.

• Add more regression tests for normalizers.

• Maybe resolve normalizer function to code-ref at memoize time instead of at function call time for efficiency? I think there was some reason not to do this, but I can't remember what it was.

• Add more array value tests to the test suite.

• Fix that `Subroutine u rededined ... line 484' message.

• Get rid of any remaining *{$ref}{CODE} or similar magic hashes.

• There should be an option to dump out the memoized values or to otherwise traverse them.

• There was probably some other stuff that I forgot.

• Here's the preliminary interface spec for the TODISK option: `memoize' takes options named SCALAR_CONTEXT and LIST_CONTEXT. Legal values are

  MEMORY
  TIE
  FAULT
  MERGE

  or a reference to a list whose first element is one of these. The default for both is MEMORY, which means that Perl's builtin hashes are used, the way they are now. FAULT means that the function should never be called in scalar/list context, and that Memoize should croak if it is. TIE means that the hash will be tied; it's usaully written as

  [TIE, packagename, argument-list]

  which specifies the package name and arguments for the tie. Memoize will load the package if appropriate. Thus

  [TIE, Storable, filename, ...]
  [TIE, DB_File, filename, flags, mode, ...]
  [TIE, MLDBM, DB_File, ... ]

  MERGE means that return values in the specified context will be stored in the same structure that is used for the other context. For example, suppose you have a function which always returns a scalar, and doesn't care whether it was called in scalar or list context. You don't want to store its list-context reutrn separately from its scalar-context return, because they're going to be the same anyway, and if you stored them separately, you'd waste a call and a cache slot. So you say LIST_CONTEXT => MERGE, and then list context is considered the same as scalar context.

  You can also use MERGE with a normalizer to get the list-context and scalar-context returns stored in the same database without conflicting with each other.

  If you specify MERGE for both, it's either an error or else you get them stored in one in-memory hash, or something.

• Include an example that caches DNS lookups.

• Make tie for Storable (Memoize::Storable)

  =item*

  Make tie for DBI (Memoize::DBI)

• Tie for SDBM doesn't work. Can't subclass SDBM? Why not?

• I think there's a bug. See `###BUG'.

• Docs / code inconsistent about SCALAR_CONTEXT vs SCALAR_CACHE. Make up your mind.

  Decision: SCALAR_CACHE.

• Storable probably can't be done, because it doesn't allow updating. Maybe a different interface that supports readonly caches fronted by a writable in-memory cache? A generic tied hash maybe?

  FETCH {
    if (it's in the memory hash) {
      return it
    } elsif (it's in the readonly disk hash) {
      return it
    } else { 
      not-there
    }
  }
  
  STORE {
    put it into the in-memory hash
  }

  Maybe `save' and `restore' methods?

• Maybe add in TODISK after all, with TODISK => 'filename' equivalent to

  SCALAR_CACHE => [TIE, Memoize::SDBM_File, $filename, O_RDWR|O_CREAT, 0666],
  LIST_CACHE => MERGE