* Use ClassValue to cache reflection arrays, avoiding clones.
* Avoid using implicit for-loop Iterator when numeric works.
* Reduce use of intermediate collections.

* Cache declared methods as in Initializer.
* Cache TypePopulator instance since some classes are used to
  populate multiple Ruby types.

Useful with -Xjit.threshold=0 to see kernel in boot stacktrace.

Also add Stringable interface to FrozenString and StringLiteral
so they can be used interchangeably when only the java.lang.String
is needed.

* Incoming scripts should be decoded from String to byte[] using
  default internal encoding, rather than trusting JDK's
  file.encoding to be appropriate.
* Outgoing streams wrapping writers should decode strings based
  on current default internal encoding.

Fixes #2403

For #3653.

MRI uses a hidden attribute on IO objects made up of two separate
read and write streams (tied_io_for_writing), when it must wrap
those streams as if they're a single IO. This is the case for
popen, and so we put the write stream into this "tied" field.

However, we're only supposed to use the tied field when there's
already a read stream. When there's just a write stream, as in

The other part of this patch fixes the close logic for non-native
popen, so that it will check if either the main stream or the
"tied" stream is open before deciding whether to close it. The old
logic only considered the primary stream.

The first part of this patch (not using "tied" when only writing)
hides the failure fixed by the second part of this patch since we
see a single-stream normal IO and the original bug does not
happen.

Note that this probably regressed in 9k because we started using
this "tied" field most places, but the non-native popen bits were
not updated to test it. If we work on getting more tests green on
Windows we should be able to find and eliminate these
discrepancies.

Fixes #3473.

* Fix handling of "--disable foo"
* Add did_you_mean to features
* Use err, not out, for printing command line errors
* Do not display usage for "--disable" with no argument

This fixes MRI's TestRubyOptions#test_enable.

…issing

…sses

…tree

* Ensure only one regexp is ever cached for //o

This is done using an AtomicReference in the compiled method for
JVM6 and a field + atomic updater in the indy call site. In both
cases, we may end up evaluating the operands twice, and the code
that produced them may still run after caching (a bug,
#2798), but we will at least guarantee to return
exactly one regexp.

* Add non-boxed paths to construct dregexp with up to 5 elements.

* Add a ThreadContext-local Encoding[1] to use for encoding
  negotiation when preprocessing the dregexp elements.

* If, at JIT time, a once-dregexp has already been encountered and
  cached in the instr, just emit that regexp directly into the
  bytecode.

This new logic is faster than what we had before, likely because
the locking I put in place for JVM6 was preventing the JVM from
jitting (punted out with "COMPILE SKIPPED: invalid parsing" due
to a flaw in my code). This new logic is lighter-weight and JITs
fine. Given the benchmark from #3735:

9.0.5:  3.87s
9.1:    0.70s
1.7.24: 0.72s

For #3735, though it has only a minor impact on a scan bench.

* This adds a hard requirement on Java 7+. However, we crossed
  that line some time ago.
* Indy can take long to warm up. However, this case is a very
  simple chain of handles.
* This improves the performance of constant lookup, since JVM6 did
  not do any caching.