For use of PyObject_GetItem

django_template3 ends up calling this a fair amount via unicode_translate
(ie it checks to see if certain entries are in the translation table).

The exceptions thrown by len itself can now be either style,
though any exceptions thrown by any called functions (ex __len__)
will still get thrown as C++ exceptions and converted if needed.

Helps in a common case of "try calling len but don't worry if no
len was defined".

list recursive printing

optimize some misc runtime functions

callable(), str(), repr(), PySequence_GetItem(),
and PyObject_HasAttrString()

Mostly by bringing in the CPython versions.

They are tricky since these are types, which means they invoke
the relatively-complicated constructor logic.  ie str() doesn't
just call __str__ on the argument: if the result is a subclass
of str, it calls result.__init__().  Similarly for unicode, except
unicode is even trickier since it takes some more arguments, one
of which is "encoding" which will have non-type-based dynamic
behavior.

I didn't realize that at first and optimized unicode() by exposing an
inner version of it that takes its arguments in registers, which we
can take advantage of using our jit-arg-rearrangement capability.
This means we have to do parts of PyArg_ParseTuple ourselves, so I
added a PyArg_ParseSingle that runs a single object through the
arg-conversion code.  PyArg_ParseSingle could be further optimized if
we want to.  Or rather, if we have functions of the form
PyArg_ParseSingle_s (which corresponds to the "s" format code) we
could skip some more of the overhead.

I had to disable most of that once I realized the encoding issue, but
I left it in since hopefully we will be able to use it again once
we have some "do some guards after mutations if we know how to resume
after a failed guard" rewriter support.

get cpython/test_tuple.py to pass

Get test_iter.py to work

Min max

assign fixed slots (vregs) to the symbols.

Use cpythons lock implementation

This switches the thread lock implementation to use a semaphore instead of a mutex.
I hope this gets rid of the threading_local.py error on travis-ci.

This removes a bottleneck of the interpreter/bjit:
most var accesses introduced a DenseMap lookup, with this change we use a fixed offset per var.
The bjit stores the pointer to the vregs array inside r14 for fast accesses.

Not having the ASTInterpreter GC allocated improves performance.
I had to add a small asm function in order to produce a special stack frame where we can easily retrieve the ASTInterpreter*,
to replace s_interpreterMaps job. This also make sure that this function really does not get inlined.
The s_interpreterMap was hard to understand and produced several times problems (duplicate entries,...)

This patch contains a hack which limits the number of variables inside a function to 512.
Because we have to make sure the are all on the stack and can't dynamically add more space.
An upcoming patch will remove this limitation and replace it with a stack alloca of the size of the actual number of variables the function uses.

This switches the thread lock implementation to use a semaphore instead of a mutex.
I hope this gets rid of the threading_local.py error on travis-ci.

Allow passing NULL for empty kwargs

This is a convention that cpython uses that is a good
performance improvement -- unlike tuples, dicts are mutable,
so we currently have to create a new dict for every call
into a kwargs-taking function even if there are no keywords
to pass.

Have to modify the receiving ends to allow NULL kwargs.  This
includes the ast interpreter (which I think will automatically get
the bjit), the llvm irgenerator, and as many runtime functions as
I can find.

Pow

Gcc 4.9 gmp fix

Improve __getattr__ handling

Reduce number of optimization passes we run

bjit: inline trivial helper functions

Release leftover chunk in the TraceStack.

The previous set was based on trying to get the absolute
best results on microbenchmarks; go back to the small set.

Fix gc visiting code for running generators