This commit fixes a situation in which, when an error happens while
decoding the content of a multiset, a temporary array would leak.

Thanks to Joshua Rogers <jrogers@opera.com> for the report.

This commit fixes a use-after-free when a delay encoded tag overrides
the container being decoded right now.

Thanks to Joshua Rogers <jrogers@opera.com> for the report.

A stray number was added to the identifier reference table.
This caused the decoder to become confused.

Fixes dumping errors for Web.SocketIO.

The mapping with encoded values is intentionally allocated with
medium sized hashtable. Shrinking it due to use of map_delete()
is just a waste of cpu.

Attempts to improve performance of encode_value() for objects,
programs and functions a bit more.

The argument order was swapped in an error callback.

The variant dispatcher often replaces the reference id to the terminating
function, which instead gets the reference id that the variant dispatcher
would have gotten otherwise.

This fixes broken encoding where the terminating function's reference id
was taken by the next function that was encoded, which in turn caused
decode_value() to fail with "Bad function identifier offset".

The decoder state wasn't freed on exit if the decoder tables contained
unfinished programs when compiled --with-debug.

deprecated on OS X.

Also, do not remember items with only 1 reference, they are rather
unlikely to show up again later on (objects notwithstanding, and they
are handled in another case).

This saves a lot of CPU time when a large array of fairly unique
strings are encoded.

This version is about 200 times faster.

Removed special case code in two places that compensated for each other.

Variant functions weren't dumped correctly if they were
overloading inherited non-variant functions.

Fixes remaining issue with the dumping of Crypto.RSA.State.

The variant dispatcher is a CFUN that may show up in pure Pike
programs, so don't fallback to the C-program encodings when
the only CFUN is the variant dispatcher.

Potentially fixes dumping problems with Crypto.RSA.State
(where generate_key() is a variant function).

The macros getdata() and getdata3() are no more.

Try to force noinline on the decode_error function.

Old style encoding was not used since about 10 years. Removing it
saves about 12 kB in decode_value2 with -Os.

Especially in error handling, and code that does a lot of other string
operations anyway the speed gain is not wort the 100+ bytes code size
of each instance of push_constant_text

This saves about 20K of code size (main pike binary), while not really
changing the speed much.

push_constant_text that seemed to be in at least pseudo-time-critical
code was kept as they were.

On a related note, push_constant_text(":") four times in a row creates
four different string variables, for obvious reasons.

Some of the remaining push_constant_text really should have a
module/file local string variable, even though that is more bothersome
to create.

It might be nice to have this in .cmod files automatically.

The Pike compiler now uses a dedicated reentrant lock instead
of the brute-force approach of using threads_disable().

This should improve concurrency with other threads.

This happens in certain programs (e.g. __builtin.*Error and some Nettle
hash states) where low_inherit is called with name == NULL.

The test looked at limits for the wrong program (the refno relates
to the program at depth depth).

Revert it for now as it breaks working code.

This reverts commit c88883e0.