/* Execute compiled code */

/* XXX TO DO:
   XXX speed up searching for keywords by using a dictionary
   XXX document it!
   */

/* Note: this file will be compiled as C ifndef WITH_LLVM, so try to keep it
   generally C. */

/* enable more aggressive intra-module optimizations, where available */
#define PY_LOCAL_AGGRESSIVE

#include "Python.h"

#include "code.h"
#include "frameobject.h"
#include "eval.h"
#include "opcode.h"
#include "structmember.h"

#include "JIT/llvm_compile.h"
#include "Util/EventTimer.h"
#include <ctype.h>

#ifdef WITH_LLVM
#include "_llvmfunctionobject.h"
#include "llvm/Function.h"
#include "llvm/Support/ManagedStatic.h"
#include "llvm/Support/raw_ostream.h"
#include "JIT/global_llvm_data.h"
#include "JIT/RuntimeFeedback.h"
#include "Util/Stats.h"

#include <set>

using llvm::errs;
#endif


/* Make a call to stop the call overhead timer before going through to
   PyObject_Call. */
static inline PyObject *
_PyObject_Call(PyObject *func, PyObject *arg, PyObject *kw)
{
	/* If we're calling a compiled C function with *args or **kwargs, then
	 * this enum should be CALL_ENTER_C.  However, most calls to C
	 * functions are simple and are fast-tracked through the CALL_FUNCTION
	 * opcode. */
	PY_LOG_TSC_EVENT(CALL_ENTER_PYOBJ_CALL);
	return PyObject_Call(func, arg, kw);
}


#ifdef Py_WITH_INSTRUMENTATION
std::string
_PyEval_GetCodeName(PyCodeObject *code)
{
	std::string result;
	llvm::raw_string_ostream wrapper(result);

	wrapper << PyString_AsString(code->co_filename)
		<< ":" << code->co_firstlineno << " "
		<< "(" << PyString_AsString(code->co_name) << ")";
	wrapper.flush();
	return result;
}

// Collect statistics about how long we block for compilation to LLVM IR and to
// machine code.
class IrCompilationTimes : public DataVectorStats<int64_t> {
public:
	IrCompilationTimes()
		: DataVectorStats<int64_t>("Time blocked for IR JIT in ns") {}
};
class McCompilationTimes : public DataVectorStats<int64_t> {
public:
	McCompilationTimes()
		: DataVectorStats<int64_t>("Time blocked for MC JIT in ns") {}
};

static llvm::ManagedStatic<IrCompilationTimes> ir_compilation_times;
static llvm::ManagedStatic<McCompilationTimes> mc_compilation_times;

class FeedbackMapCounter {
public:
	~FeedbackMapCounter() {
		errs() << "\nFeedback maps created:\n";
		errs() << "N: " << this->counter_ << "\n";
	}

	void IncCounter() {
		this->counter_++;
	}

private:
	unsigned counter_;
};

static llvm::ManagedStatic<FeedbackMapCounter> feedback_map_counter;


class HotnessTracker {
	// llvm::DenseSet or llvm::SmallPtrSet may be better, but as of this
	// writing, they don't seem to work with std::vector.
	std::set<PyCodeObject*> hot_code_;
public:
	~HotnessTracker();

	void AddHotCode(PyCodeObject *code_obj) {
		// This will prevent the code object from ever being
		// deleted.
		Py_INCREF(code_obj);
		this->hot_code_.insert(code_obj);
	}
};

static bool
compare_hotness(const PyCodeObject *first, const PyCodeObject *second)
{
	return first->co_hotness > second->co_hotness;
}

HotnessTracker::~HotnessTracker()
{
	errs() << "\nCode objects deemed hot:\n";
	errs() << "N: " << this->hot_code_.size() << "\n";
	errs() << "Function -> hotness score:\n";
	std::vector<PyCodeObject*> to_sort(this->hot_code_.begin(),
					   this->hot_code_.end());
	std::sort(to_sort.begin(), to_sort.end(), compare_hotness);
	for (std::vector<PyCodeObject*>::iterator co = to_sort.begin();
	     co != to_sort.end(); ++co) {
		errs() << _PyEval_GetCodeName(*co)
		       << " -> " << (*co)->co_hotness << "\n";
	}
}

static llvm::ManagedStatic<HotnessTracker> hot_code;


// Keep track of which functions failed fatal guards, but kept being called.
// This can help gauge the efficacy of optimizations that involve fatal guards.
class FatalBailTracker {
public:
	~FatalBailTracker() {
		errs() << "\nCode objects that failed fatal guards:\n";
		errs() << "\tfile:line (funcname) bail hotness"
		       << " -> final hotness\n";

		for (TrackerData::const_iterator it = this->code_.begin();
				it != this->code_.end(); ++it) {
			PyCodeObject *code = it->first;
			if (code->co_hotness == it->second)
				continue;
			errs() << "\t" << _PyEval_GetCodeName(code)
			       << "\t" << it->second << " -> "
			       << code->co_hotness << "\n";
		}
	}

	void RecordFatalBail(PyCodeObject *code) {
		Py_INCREF(code);
		this->code_.push_back(std::make_pair(code, code->co_hotness));
	}

private:
	// Keep a list of (code object, hotness) where hotness is the
	// value of co_hotness when RecordFatalBail() was called. This is
	// used to hide code objects whose machine code functions are
	// invalidated during shutdown because their module dict has gone away;
	// these code objects are uninteresting for our analysis.
	typedef std::pair<PyCodeObject *, long> DataPoint;
	typedef std::vector<DataPoint> TrackerData;

	TrackerData code_;
};


static llvm::ManagedStatic<FatalBailTracker> fatal_bail_tracker;

// C wrapper for FatalBailTracker::RecordFatalBail().
void
_PyEval_RecordFatalBail(PyCodeObject *code)
{
	fatal_bail_tracker->RecordFatalBail(code);
}


// Collect stats on how many watchers the globals/builtins dicts acculumate.
// This currently records how many watchers the dict had when it changed, ie,
// how many watchers it had to notify.
class WatcherCountStats : public DataVectorStats<size_t> {
public:
	WatcherCountStats() :
		DataVectorStats<size_t>("Number of watchers accumulated") {};
};

static llvm::ManagedStatic<WatcherCountStats> watcher_count_stats;

void
_PyEval_RecordWatcherCount(size_t watcher_count)
{
	watcher_count_stats->RecordDataPoint(watcher_count);
}


class BailCountStats {
public:
	BailCountStats() : total_(0), trace_on_entry_(0), line_trace_(0),
	                   backedge_trace_(0), call_profile_(0),
	                   fatal_guard_fail_(0), guard_fail_(0) {};

	~BailCountStats() {
		errs() << "\nBailed to the interpreter " << this->total_
		       << " times:\n";
		errs() << "TRACE_ON_ENTRY: " << this->trace_on_entry_ << "\n";
		errs() << "LINE_TRACE: " << this->line_trace_ << "\n";
		errs() << "BACKEDGE_TRACE:" << this->backedge_trace_ << "\n";
		errs() << "CALL_PROFILE: " << this->call_profile_ << "\n";
		errs() << "FATAL_GUARD_FAIL: " << this->fatal_guard_fail_
		       << "\n";
		errs() << "GUARD_FAIL: " << this->guard_fail_ << "\n";

		errs() << "\n" << this->bail_site_freq_.size()
		       << " bail sites:\n";
		for (BailData::iterator i = this->bail_site_freq_.begin(),
		     end = this->bail_site_freq_.end(); i != end; ++i) {
			errs() << "    " << i->getKey() << " bailed "
			       << i->getValue() << " times\n";
		}

		errs() << "\n" << this->guard_bail_site_freq_.size()
		       << " guard bail sites:\n";
		for (BailData::iterator i = this->guard_bail_site_freq_.begin(),
		     end = this->guard_bail_site_freq_.end(); i != end; ++i) {
			errs() << "    " << i->getKey() << " bailed "
			       << i->getValue() << " times\n";
		}

	}

	void RecordBail(PyFrameObject *frame, _PyFrameBailReason bail_reason) {
		++this->total_;

		std::string record;
		llvm::raw_string_ostream wrapper(record);
		wrapper << PyString_AsString(frame->f_code->co_filename) << ":";
		wrapper << frame->f_code->co_firstlineno << ":";
		wrapper << PyString_AsString(frame->f_code->co_name) << ":";
		// See the comment in PyEval_EvalFrame about how f->f_lasti is
		// initialized.
		wrapper << frame->f_lasti + 1;
		wrapper.flush();

		BailData::value_type &entry =
			this->bail_site_freq_.GetOrCreateValue(record, 0);
		entry.setValue(entry.getValue() + 1);

#define BAIL_CASE(name, field) \
	case name: \
		++this->field; \
		break;

		switch (bail_reason) {
			BAIL_CASE(_PYFRAME_TRACE_ON_ENTRY, trace_on_entry_)
			BAIL_CASE(_PYFRAME_LINE_TRACE, line_trace_)
			BAIL_CASE(_PYFRAME_BACKEDGE_TRACE, backedge_trace_)
			BAIL_CASE(_PYFRAME_CALL_PROFILE, call_profile_)
			BAIL_CASE(_PYFRAME_FATAL_GUARD_FAIL, fatal_guard_fail_)
			BAIL_CASE(_PYFRAME_GUARD_FAIL, guard_fail_)
			default:
				abort();   // Unknown bail reason.
		}
#undef BAIL_CASE

		if (bail_reason != _PYFRAME_GUARD_FAIL)
			return;

		wrapper << ":";

#define GUARD_CASE(name) \
	case name: \
		wrapper << #name; \
		break;

		switch (frame->f_guard_type) {
			GUARD_CASE(_PYGUARD_DEFAULT)
			GUARD_CASE(_PYGUARD_BINOP)
			GUARD_CASE(_PYGUARD_ATTR)
			GUARD_CASE(_PYGUARD_CFUNC)
			GUARD_CASE(_PYGUARD_BRANCH)
			GUARD_CASE(_PYGUARD_STORE_SUBSCR)
			default:
				wrapper << ((int)frame->f_guard_type);
		}
#undef GUARD_CASE

		wrapper.flush();

		BailData::value_type &g_entry =
			this->guard_bail_site_freq_.GetOrCreateValue(record, 0);
		g_entry.setValue(g_entry.getValue() + 1);
	}

private:
	typedef llvm::StringMap<unsigned> BailData;
	BailData bail_site_freq_;
	BailData guard_bail_site_freq_;

	long total_;
	long trace_on_entry_;
	long line_trace_;
	long backedge_trace_;
	long call_profile_;
	long fatal_guard_fail_;
	long guard_fail_;
};

static llvm::ManagedStatic<BailCountStats> bail_count_stats;
#endif  // Py_WITH_INSTRUMENTATION


/* Turn this on if your compiler chokes on the big switch: */
/* #define CASE_TOO_BIG 1 */

#ifdef Py_DEBUG
/* For debugging the interpreter: */
#define LLTRACE  1	/* Low-level trace feature */
#define CHECKEXC 1	/* Double-check exception checking */
#endif

typedef PyObject *(*callproc)(PyObject *, PyObject *, PyObject *);

/* Forward declarations */
static PyObject * fast_function(PyObject *, PyObject ***, int, int, int);
static PyObject * do_call(PyObject *, PyObject ***, int, int);
static PyObject * ext_do_call(PyObject *, PyObject ***, int, int, int);
static PyObject * update_keyword_args(PyObject *, int, PyObject ***,
				      PyObject *);
static PyObject * update_star_args(int, int, PyObject *, PyObject ***);
static PyObject * load_args(PyObject ***, int);

#ifdef WITH_LLVM
static inline void mark_called(PyCodeObject *co);
static inline int maybe_compile(PyCodeObject *co, PyFrameObject *f);

/* Record data for use in generating optimized machine code. */
static void record_type(PyCodeObject *, int, int, int, PyObject *);
static void record_func(PyCodeObject *, int, int, int, PyObject *);
static void record_object(PyCodeObject *, int, int, int, PyObject *);
static void inc_feedback_counter(PyCodeObject *, int, int, int, int);
#endif  /* WITH_LLVM */

int _Py_ProfilingPossible = 0;

/* Keep this in sync with llvm_fbuilder.cc */
#define CALL_FLAG_VAR 1
#define CALL_FLAG_KW 2

#ifdef LLTRACE
static int lltrace;
static int prtrace(PyObject *, char *);
#endif
static int call_trace_protected(Py_tracefunc, PyObject *,
				 PyFrameObject *, int, PyObject *);
static int maybe_call_line_trace(Py_tracefunc, PyObject *,
				  PyFrameObject *, int *, int *, int *);

static PyObject * cmp_outcome(int, PyObject *, PyObject *);
static void format_exc_check_arg(PyObject *, char *, PyObject *);
static PyObject * string_concatenate(PyObject *, PyObject *,
				    PyFrameObject *, unsigned char *);

#define NAME_ERROR_MSG \
	"name '%.200s' is not defined"
#define GLOBAL_NAME_ERROR_MSG \
	"global name '%.200s' is not defined"
#define UNBOUNDLOCAL_ERROR_MSG \
	"local variable '%.200s' referenced before assignment"
#define UNBOUNDFREE_ERROR_MSG \
	"free variable '%.200s' referenced before assignment" \
        " in enclosing scope"

/* Dynamic execution profile */
#ifdef DYNAMIC_EXECUTION_PROFILE
#ifdef DXPAIRS
static long dxpairs[257][256];
#define dxp dxpairs[256]
#else
static long dxp[256];
#endif
#endif

/* Function call profile */
#ifdef CALL_PROFILE
#define PCALL_NUM 11
static int pcall[PCALL_NUM];

#define PCALL_ALL 0
#define PCALL_FUNCTION 1
#define PCALL_FAST_FUNCTION 2
#define PCALL_FASTER_FUNCTION 3
#define PCALL_METHOD 4
#define PCALL_BOUND_METHOD 5
#define PCALL_CFUNCTION 6
#define PCALL_TYPE 7
#define PCALL_GENERATOR 8
#define PCALL_OTHER 9
#define PCALL_POP 10

/* Notes about the statistics

   PCALL_FAST stats

   FAST_FUNCTION means no argument tuple needs to be created.
   FASTER_FUNCTION means that the fast-path frame setup code is used.

   If there is a method call where the call can be optimized by changing
   the argument tuple and calling the function directly, it gets recorded
   twice.

   As a result, the relationship among the statistics appears to be
   PCALL_ALL == PCALL_FUNCTION + PCALL_METHOD - PCALL_BOUND_METHOD +
                PCALL_CFUNCTION + PCALL_TYPE + PCALL_GENERATOR + PCALL_OTHER
   PCALL_FUNCTION > PCALL_FAST_FUNCTION > PCALL_FASTER_FUNCTION
   PCALL_METHOD > PCALL_BOUND_METHOD
*/

#define PCALL(POS) pcall[POS]++

PyObject *
PyEval_GetCallStats(PyObject *self)
{
	return Py_BuildValue("iiiiiiiiiiiii",
			     pcall[0], pcall[1], pcall[2], pcall[3],
			     pcall[4], pcall[5], pcall[6], pcall[7],
			     pcall[8], pcall[9], pcall[10]);
}
#else
#define PCALL(O)

PyObject *
PyEval_GetCallStats(PyObject *self)
{
	Py_INCREF(Py_None);
	return Py_None;
}
#endif


#ifdef WITH_THREAD

#ifdef HAVE_ERRNO_H
#include <errno.h>
#endif
#include "pythread.h"

static PyThread_type_lock interpreter_lock = 0; /* This is the GIL */
long _PyEval_main_thread = 0;

int
PyEval_ThreadsInitialized(void)
{
	return interpreter_lock != 0;
}

void
PyEval_InitThreads(void)
{
	if (interpreter_lock)
		return;
	interpreter_lock = PyThread_allocate_lock();
	PyThread_acquire_lock(interpreter_lock, 1);
	_PyEval_main_thread = PyThread_get_thread_ident();
}

void
PyEval_AcquireLock(void)
{
	PyThread_acquire_lock(interpreter_lock, 1);
}

void
PyEval_ReleaseLock(void)
{
	PyThread_release_lock(interpreter_lock);
}

void
PyEval_AcquireThread(PyThreadState *tstate)
{
	if (tstate == NULL)
		Py_FatalError("PyEval_AcquireThread: NULL new thread state");
	/* Check someone has called PyEval_InitThreads() to create the lock */
	assert(interpreter_lock);
	PyThread_acquire_lock(interpreter_lock, 1);
	if (PyThreadState_Swap(tstate) != NULL)
		Py_FatalError(
			"PyEval_AcquireThread: non-NULL old thread state");
}

void
PyEval_ReleaseThread(PyThreadState *tstate)
{
	if (tstate == NULL)
		Py_FatalError("PyEval_ReleaseThread: NULL thread state");
	if (PyThreadState_Swap(NULL) != tstate)
		Py_FatalError("PyEval_ReleaseThread: wrong thread state");
	PyThread_release_lock(interpreter_lock);
}

/* This function is called from PyOS_AfterFork to ensure that newly
   created child processes don't hold locks referring to threads which
   are not running in the child process.  (This could also be done using
   pthread_atfork mechanism, at least for the pthreads implementation.) */

void
PyEval_ReInitThreads(void)
{
	PyObject *threading, *result;
	PyThreadState *tstate;

	if (!interpreter_lock)
		return;
	/*XXX Can't use PyThread_free_lock here because it does too
	  much error-checking.  Doing this cleanly would require
	  adding a new function to each thread_*.h.  Instead, just
	  create a new lock and waste a little bit of memory */
	interpreter_lock = PyThread_allocate_lock();
	PyThread_acquire_lock(interpreter_lock, 1);
	_PyEval_main_thread = PyThread_get_thread_ident();

	/* Update the threading module with the new state.
	 */
	tstate = PyThreadState_GET();
	threading = PyMapping_GetItemString(tstate->interp->modules,
					    "threading");
	if (threading == NULL) {
		/* threading not imported */
		PyErr_Clear();
		return;
	}
	result = PyObject_CallMethod(threading, "_after_fork", NULL);
	if (result == NULL)
		PyErr_WriteUnraisable(threading);
	else
		Py_DECREF(result);
	Py_DECREF(threading);
}
#endif

/* Functions save_thread and restore_thread are always defined so
   dynamically loaded modules needn't be compiled separately for use
   with and without threads: */

PyThreadState *
PyEval_SaveThread(void)
{
	PyThreadState *tstate = PyThreadState_Swap(NULL);
	if (tstate == NULL)
		Py_FatalError("PyEval_SaveThread: NULL tstate");
#ifdef WITH_THREAD
	if (interpreter_lock)
		PyThread_release_lock(interpreter_lock);
#endif
	return tstate;
}

void
PyEval_RestoreThread(PyThreadState *tstate)
{
	if (tstate == NULL)
		Py_FatalError("PyEval_RestoreThread: NULL tstate");
#ifdef WITH_THREAD
	if (interpreter_lock) {
		int err = errno;
		PyThread_acquire_lock(interpreter_lock, 1);
		errno = err;
	}
#endif
	PyThreadState_Swap(tstate);
}


/* Mechanism whereby asynchronously executing callbacks (e.g. UNIX
   signal handlers or Mac I/O completion routines) can schedule calls
   to a function to be called synchronously.
   The synchronous function is called with one void* argument.
   It should return 0 for success or -1 for failure -- failure should
   be accompanied by an exception.

   If registry succeeds, the registry function returns 0; if it fails
   (e.g. due to too many pending calls) it returns -1 (without setting
   an exception condition).

   Note that because registry may occur from within signal handlers,
   or other asynchronous events, calling malloc() is unsafe!

#ifdef WITH_THREAD
   Any thread can schedule pending calls, but only the main thread
   will execute them.
#endif

   XXX WARNING!  ASYNCHRONOUSLY EXECUTING CODE!
   There are two possible race conditions:
   (1) nested asynchronous registry calls;
   (2) registry calls made while pending calls are being processed.
   While (1) is very unlikely, (2) is a real possibility.
   The current code is safe against (2), but not against (1).
   The safety against (2) is derived from the fact that only one
   thread (the main thread) ever takes things out of the queue.

   XXX Darn!  With the advent of thread state, we should have an array
   of pending calls per thread in the thread state!  Later...
*/

#define NPENDINGCALLS 32
static struct {
	int (*func)(void *);
	void *arg;
} pendingcalls[NPENDINGCALLS];
static volatile int pendingfirst = 0;
static volatile int pendinglast = 0;
static volatile int things_to_do = 0;

int
Py_AddPendingCall(int (*func)(void *), void *arg)
{
	static volatile int busy = 0;
	int i, j;
	/* XXX Begin critical section */
	/* XXX If you want this to be safe against nested
	   XXX asynchronous calls, you'll have to work harder! */
	if (busy)
		return -1;
	busy = 1;
	i = pendinglast;
	j = (i + 1) % NPENDINGCALLS;
	if (j == pendingfirst) {
		busy = 0;
		return -1; /* Queue full */
	}
	pendingcalls[i].func = func;
	pendingcalls[i].arg = arg;
	pendinglast = j;

	_Py_Ticker = 0;
	things_to_do = 1; /* Signal main loop */
	busy = 0;
	/* XXX End critical section */
	return 0;
}

int
Py_MakePendingCalls(void)
{
	static int busy = 0;
#ifdef WITH_THREAD
	if (_PyEval_main_thread &&
	    PyThread_get_thread_ident() != _PyEval_main_thread)
		return 0;
#endif
	if (busy)
		return 0;
	busy = 1;
	things_to_do = 0;
	for (;;) {
		int i;
		int (*func)(void *);
		void *arg;
		i = pendingfirst;
		if (i == pendinglast)
			break; /* Queue empty */
		func = pendingcalls[i].func;
		arg = pendingcalls[i].arg;
		pendingfirst = (i + 1) % NPENDINGCALLS;
		if (func(arg) < 0) {
			busy = 0;
			things_to_do = 1; /* We're not done yet */
			return -1;
		}
	}
	busy = 0;
	return 0;
}


/* The interpreter's recursion limit */

#ifndef Py_DEFAULT_RECURSION_LIMIT
#define Py_DEFAULT_RECURSION_LIMIT 1000
#endif
static int recursion_limit = Py_DEFAULT_RECURSION_LIMIT;
int _Py_CheckRecursionLimit = Py_DEFAULT_RECURSION_LIMIT;

int
Py_GetRecursionLimit(void)
{
	return recursion_limit;
}

void
Py_SetRecursionLimit(int new_limit)
{
	recursion_limit = new_limit;
	_Py_CheckRecursionLimit = recursion_limit;
}

/* the macro Py_EnterRecursiveCall() only calls _Py_CheckRecursiveCall()
   if the recursion_depth reaches _Py_CheckRecursionLimit.
   If USE_STACKCHECK, the macro decrements _Py_CheckRecursionLimit
   to guarantee that _Py_CheckRecursiveCall() is regularly called.
   Without USE_STACKCHECK, there is no need for this. */
int
_Py_CheckRecursiveCall(char *where)
{
	PyThreadState *tstate = PyThreadState_GET();

#ifdef USE_STACKCHECK
	if (PyOS_CheckStack()) {
		--tstate->recursion_depth;
		PyErr_SetString(PyExc_MemoryError, "Stack overflow");
		return -1;
	}
#endif
	if (tstate->recursion_depth > recursion_limit) {
		--tstate->recursion_depth;
		PyErr_Format(PyExc_RuntimeError,
			     "maximum recursion depth exceeded%s",
			     where);
		return -1;
	}
	_Py_CheckRecursionLimit = recursion_limit;
	return 0;
}

#ifdef __cplusplus
extern "C" void
#else
extern void
#endif
_PyEval_RaiseForUnboundLocal(PyFrameObject *frame, int var_index)
{
	format_exc_check_arg(
		PyExc_UnboundLocalError,
		UNBOUNDLOCAL_ERROR_MSG,
		PyTuple_GetItem(frame->f_code->co_varnames, var_index));
}

/* Records whether tracing is on for any thread.  Counts the number of
   threads for which tstate->c_tracefunc is non-NULL, so if the value
   is 0, we know we don't have to check this thread's c_tracefunc.
   This speeds up the if statement in PyEval_EvalFrameEx() after
   fast_next_opcode*/
int _Py_TracingPossible = 0;

/* for manipulating the thread switch and periodic "stuff" - used to be
   per thread, now just a pair o' globals */
int _Py_CheckInterval = 100;
volatile int _Py_Ticker = 100;

#ifdef WITH_LLVM
int _Py_BailError = 0;
#endif

PyObject *
PyEval_EvalCode(PyCodeObject *co, PyObject *globals, PyObject *locals)
{
	return PyEval_EvalCodeEx(co,
			  globals, locals,
			  (PyObject **)NULL, 0,
			  (PyObject **)NULL, 0,
			  (PyObject **)NULL, 0,
			  NULL);
}


/* Interpreter main loop */

PyObject *
PyEval_EvalFrameEx(PyFrameObject *f, int throwflag) {
	/* This is for backward compatibility with extension modules that
           used this API; core interpreter code should call
           PyEval_EvalFrame() */
	PyObject *result;
	f->f_throwflag = throwflag;
	result = PyEval_EvalFrame(f);
	f->f_throwflag = 0;
	return result;
}

PyObject *
PyEval_EvalFrame(PyFrameObject *f)
{
#ifdef DXPAIRS
	int lastopcode = 0;
#endif
	register PyObject **stack_pointer;  /* Next free slot in value stack */
	register unsigned char *next_instr;
	register int opcode;	/* Current opcode */
	register int oparg;	/* Current opcode argument, if any */
	register enum _PyUnwindReason why; /* Reason for block stack unwind */
	register int err;	/* Error status -- nonzero if error */
	register PyObject *x;	/* Temporary objects popped off stack */
	register PyObject *v;
	register PyObject *w;
	register PyObject *u;
	register PyObject *t;
	register PyObject **fastlocals, **freevars;
	_PyFrameBailReason bail_reason;
	PyObject *retval = NULL;	/* Return value */
	PyThreadState *tstate = PyThreadState_GET();
	PyCodeObject *co;
#ifdef WITH_LLVM
	/* We only collect feedback if it will be useful. */
	int rec_feedback = (Py_JitControl == PY_JIT_WHENHOT);
#endif

	/* when tracing we set things up so that

               not (instr_lb <= current_bytecode_offset < instr_ub)

	   is true when the line being executed has changed.  The
           initial values are such as to make this false the first
           time it is tested. */
	int instr_ub = -1, instr_lb = 0, instr_prev = -1;

	unsigned char *first_instr;
	PyObject *names;
	PyObject *consts;
#if defined(Py_DEBUG) || defined(LLTRACE)
	/* Make it easier to find out where we are with a debugger */
	char *filename;
#endif

/* Computed GOTOs, or
       the-optimization-commonly-but-improperly-known-as-"threaded code"
   using gcc's labels-as-values extension
   (http://gcc.gnu.org/onlinedocs/gcc/Labels-as-Values.html).

   The traditional bytecode evaluation loop uses a "switch" statement, which
   decent compilers will optimize as a single indirect branch instruction
   combined with a lookup table of jump addresses. However, since the
   indirect jump instruction is shared by all opcodes, the CPU will have a
   hard time making the right prediction for where to jump next (actually,
   it will be always wrong except in the uncommon case of a sequence of
   several identical opcodes).

   "Threaded code" in contrast, uses an explicit jump table and an explicit
   indirect jump instruction at the end of each opcode. Since the jump
   instruction is at a different address for each opcode, the CPU will make a
   separate prediction for each of these instructions, which is equivalent to
   predicting the second opcode of each opcode pair. These predictions have
   a much better chance to turn out valid, especially in small bytecode loops.

   A mispredicted branch on a modern CPU flushes the whole pipeline and
   can cost several CPU cycles (depending on the pipeline depth),
   and potentially many more instructions (depending on the pipeline width).
   A correctly predicted branch, however, is nearly free.

   At the time of this writing, the "threaded code" version is up to 15-20%
   faster than the normal "switch" version, depending on the compiler and the
   CPU architecture.

   We disable the optimization if DYNAMIC_EXECUTION_PROFILE is defined,
   because it would render the measurements invalid.


   NOTE: care must be taken that the compiler doesn't try to "optimize" the
   indirect jumps by sharing them between all opcodes. Such optimizations
   can be disabled on gcc by using the -fno-gcse flag (or possibly
   -fno-crossjumping).
*/

#if defined(USE_COMPUTED_GOTOS) && defined(DYNAMIC_EXECUTION_PROFILE)
#undef USE_COMPUTED_GOTOS
#endif

#ifdef USE_COMPUTED_GOTOS
/* Import the static jump table */
#include "opcode_targets.h"

/* This macro is used when several opcodes defer to the same implementation
   (e.g. SETUP_LOOP, SETUP_FINALLY) */
#define TARGET_WITH_IMPL(op, impl) \
	TARGET_##op: \
		opcode = op; \
		if (HAS_ARG(op)) \
			oparg = NEXTARG(); \
	case op: \
		goto impl; \

#define TARGET(op) \
	TARGET_##op: \
		opcode = op; \
		if (HAS_ARG(op)) \
			oparg = NEXTARG(); \
	case op:


#define DISPATCH() \
	{ \
		/* Avoid multiple loads from _Py_Ticker despite `volatile` */ \
		int _tick = _Py_Ticker - 1; \
		_Py_Ticker = _tick; \
		if (_tick >= 0) { \
			FAST_DISPATCH(); \
		} \
		continue; \
	}

#ifdef LLTRACE
#define FAST_DISPATCH() \
	{ \
		if (!lltrace && !_Py_TracingPossible) { \
			f->f_lasti = INSTR_OFFSET(); \
			goto *opcode_targets[*next_instr++]; \
		} \
		goto fast_next_opcode; \
	}
#else
#define FAST_DISPATCH() \
	{ \
		if (!_Py_TracingPossible) { \
			f->f_lasti = INSTR_OFFSET(); \
			goto *opcode_targets[*next_instr++]; \
		} \
		goto fast_next_opcode; \
	}
#endif

#else
#define TARGET(op) \
	case op:
#define TARGET_WITH_IMPL(op, impl) \
	/* silence compiler warnings about `impl` unused */ \
	if (0) goto impl; \
	case op:
#define DISPATCH() continue
#define FAST_DISPATCH() goto fast_next_opcode
#endif


/* Tuple access macros */

#ifndef Py_DEBUG
#define GETITEM(v, i) PyTuple_GET_ITEM((PyTupleObject *)(v), (i))
#else
#define GETITEM(v, i) PyTuple_GetItem((v), (i))
#endif

/* Code access macros */

#define INSTR_OFFSET()	((int)(next_instr - first_instr))
#define NEXTOP()	(*next_instr++)
#define NEXTARG()	(next_instr += 2, (next_instr[-1]<<8) + next_instr[-2])
#define PEEKARG()	((next_instr[2]<<8) + next_instr[1])
#define JUMPTO(x)	(next_instr = first_instr + (x))
#define JUMPBY(x)	(next_instr += (x))

/* Feedback-gathering macros */
#ifdef WITH_LLVM
#define RECORD_TYPE(arg_index, obj) \
	if(rec_feedback){record_type(co, opcode, f->f_lasti, arg_index, obj);}
#define RECORD_OBJECT(arg_index, obj) \
	if(rec_feedback){record_object(co, opcode, f->f_lasti, arg_index, obj);}
#define RECORD_FUNC(obj) \
	if(rec_feedback){record_func(co, opcode, f->f_lasti, 0, obj);}
#define INC_COUNTER(arg_index, counter_id) \
	if (rec_feedback) { \
		inc_feedback_counter(co, opcode, f->f_lasti, arg_index, \
		                     counter_id); \
	}
#define RECORD_TRUE() \
	INC_COUNTER(0, PY_FDO_JUMP_TRUE)
#define RECORD_FALSE() \
	INC_COUNTER(0, PY_FDO_JUMP_FALSE)
#define RECORD_NONBOOLEAN() \
	INC_COUNTER(0, PY_FDO_JUMP_NON_BOOLEAN)
#define UPDATE_HOTNESS_JABS() \
	do { if (oparg <= f->f_lasti) ++co->co_hotness; } while (0)
#else
#define RECORD_TYPE(arg_index, obj)
#define RECORD_OBJECT(arg_index, obj)
#define RECORD_FUNC(obj)
#define INC_COUNTER(arg_index, counter_id)
#define RECORD_TRUE()
#define RECORD_FALSE()
#define RECORD_NONBOOLEAN()
#define UPDATE_HOTNESS_JABS()
#endif  /* WITH_LLVM */


/* OpCode prediction macros
	Some opcodes tend to come in pairs thus making it possible to
	predict the second code when the first is run.  For example,
	GET_ITER is often followed by FOR_ITER. And FOR_ITER is often
	followed by STORE_FAST or UNPACK_SEQUENCE.

	Verifying the prediction costs a single high-speed test of a register
	variable against a constant.  If the pairing was good, then the
	processor's own internal branch predication has a high likelihood of
	success, resulting in a nearly zero-overhead transition to the
	next opcode.  A successful prediction saves a trip through the eval-loop
	including its two unpredictable branches, the HAS_ARG test and the
	switch-case.  Combined with the processor's internal branch prediction,
	a successful PREDICT has the effect of making the two opcodes run as if
	they were a single new opcode with the bodies combined.

    If collecting opcode statistics, your choices are to either keep the
	predictions turned-on and interpret the results as if some opcodes
	had been combined or turn-off predictions so that the opcode frequency
	counter updates for both opcodes.

    Opcode prediction is disabled with threaded code, since the latter allows
	the CPU to record separate branch prediction information for each
	opcode.

*/

#if defined(DYNAMIC_EXECUTION_PROFILE) || defined(USE_COMPUTED_GOTOS)
#define PREDICT(op)		if (0) goto PRED_##op
#define PREDICTED(op)		PRED_##op:
#define PREDICTED_WITH_ARG(op)	PRED_##op:
#else
#define PREDICT(op)		if (*next_instr == op) goto PRED_##op
#ifdef WITH_LLVM
#define PREDICTED_COMMON(op)	f->f_lasti = INSTR_OFFSET(); opcode = op;
#else
#define PREDICTED_COMMON(op)	/* nothing */
#endif
#define PREDICTED(op)		PRED_##op: PREDICTED_COMMON(op) next_instr++
#define PREDICTED_WITH_ARG(op)	PRED_##op: PREDICTED_COMMON(op) \
				oparg = PEEKARG(); next_instr += 3
#endif


/* Stack manipulation macros */

/* The stack can grow at most MAXINT deep, as co_nlocals and
   co_stacksize are ints. */
#define STACK_LEVEL()	((int)(stack_pointer - f->f_valuestack))
#define EMPTY()		(STACK_LEVEL() == 0)
#define TOP()		(stack_pointer[-1])
#define SECOND()	(stack_pointer[-2])
#define THIRD() 	(stack_pointer[-3])
#define FOURTH()	(stack_pointer[-4])
#define SET_TOP(v)	(stack_pointer[-1] = (v))
#define SET_SECOND(v)	(stack_pointer[-2] = (v))
#define SET_THIRD(v)	(stack_pointer[-3] = (v))
#define SET_FOURTH(v)	(stack_pointer[-4] = (v))
#define BASIC_STACKADJ(n)	(stack_pointer += n)
#define BASIC_PUSH(v)	(*stack_pointer++ = (v))
#define BASIC_POP()	(*--stack_pointer)

#ifdef LLTRACE
#define PUSH(v)		{ (void)(BASIC_PUSH(v), \
                               lltrace && prtrace(TOP(), "push")); \
                               assert(STACK_LEVEL() <= co->co_stacksize); }
#define POP()		((void)(lltrace && prtrace(TOP(), "pop")), \
			 BASIC_POP())
#define STACKADJ(n)	{ (void)(BASIC_STACKADJ(n), \
                               lltrace && prtrace(TOP(), "stackadj")); \
                               assert(STACK_LEVEL() <= co->co_stacksize); }
#define EXT_POP(STACK_POINTER) ((void)(lltrace && \
				prtrace((STACK_POINTER)[-1], "ext_pop")), \
				*--(STACK_POINTER))
#define EXT_PUSH(v, STACK_POINTER) ((void)(*(STACK_POINTER)++ = (v), \
                   lltrace && prtrace((STACK_POINTER)[-1], "ext_push")))
#else
#define PUSH(v)		BASIC_PUSH(v)
#define POP()		BASIC_POP()
#define STACKADJ(n)	BASIC_STACKADJ(n)
#define EXT_POP(STACK_POINTER) (*--(STACK_POINTER))
#define EXT_PUSH(v, STACK_POINTER) (*(STACK_POINTER)++ = (v))
#endif

/* Local variable macros */

#define GETLOCAL(i)	(fastlocals[i])

/* The SETLOCAL() macro must not DECREF the local variable in-place and
   then store the new value; it must copy the old value to a temporary
   value, then store the new value, and then DECREF the temporary value.
   This is because it is possible that during the DECREF the frame is
   accessed by other code (e.g. a __del__ method or gc.collect()) and the
   variable would be pointing to already-freed memory. */
#define SETLOCAL(i, value)	do { PyObject *tmp = GETLOCAL(i); \
				     GETLOCAL(i) = value; \
                                     Py_XDECREF(tmp); } while (0)

/* Start of code */

	if (f == NULL)
		return NULL;

#ifdef WITH_LLVM
	bail_reason = (_PyFrameBailReason)f->f_bailed_from_llvm;
#else
	bail_reason = _PYFRAME_NO_BAIL;
#endif  /* WITH_LLVM */
	/* push frame */
	if (bail_reason == _PYFRAME_NO_BAIL && Py_EnterRecursiveCall(""))
		return NULL;

	co = f->f_code;
	tstate->frame = f;

#ifdef WITH_LLVM
	maybe_compile(co, f);

	if (f->f_use_jit) {
		assert(bail_reason == _PYFRAME_NO_BAIL);
		assert(co->co_native_function != NULL &&
		       "maybe_compile was supposed to ensure"
		       " that co_native_function exists");
		if (!co->co_use_jit) {
			// A frame cannot use_jit if the underlying code object
			// can't use_jit. This comes up when a generator is
			// invalidated while active.
			f->f_use_jit = 0;
		}
		else {
			assert(co->co_fatalbailcount < PY_MAX_FATALBAILCOUNT);
			retval = co->co_native_function(f);
			goto exit_eval_frame;
		}
	}

	if (bail_reason != _PYFRAME_NO_BAIL) {
#ifdef Py_WITH_INSTRUMENTATION
		bail_count_stats->RecordBail(f, bail_reason);
#endif
		if (_Py_BailError) {
			/* When we bail, we set f_lasti to the current opcode
			 * minus 1, so we add one back.  */
			int lasti = f->f_lasti + 1;
			PyErr_Format(PyExc_RuntimeError, "bailed to the "
				     "interpreter at opcode index %d", lasti);
			goto exit_eval_frame;
		}
	}

	/* Create co_runtime_feedback now that we're about to use it.  You
	 * might think this would cause a problem if the user flips
	 * Py_JitControl from "never" to "whenhot", but since the value of
	 * rec_feedback is constant for the duration of this frame's execution,
	 * we will not accidentally try to record feedback without initializing
	 * co_runtime_feedback.  */
	if (rec_feedback && co->co_runtime_feedback == NULL) {
#if Py_WITH_INSTRUMENTATION
		feedback_map_counter->IncCounter();
#endif
		co->co_runtime_feedback = PyFeedbackMap_New();
	}
#endif  /* WITH_LLVM */

	switch (bail_reason) {
		case _PYFRAME_NO_BAIL:
		case _PYFRAME_TRACE_ON_ENTRY:
			if (tstate->use_tracing) {
				if (_PyEval_TraceEnterFunction(tstate, f))
					/* Trace or profile function raised
					   an error. */
					goto exit_eval_frame;
			}
			break;

		case _PYFRAME_BACKEDGE_TRACE:
			/* If we bailed because of a backedge, set instr_prev
			   to ensure a line trace call. */
			instr_prev = INT_MAX;
			break;

		case _PYFRAME_CALL_PROFILE:
		case _PYFRAME_LINE_TRACE:
		case _PYFRAME_FATAL_GUARD_FAIL:
		case _PYFRAME_GUARD_FAIL:
			/* These are handled by the opcode dispatch loop. */
			break;

		default:
			PyErr_Format(PyExc_SystemError, "unknown bail reason");
			goto exit_eval_frame;
	}


	names = co->co_names;
	consts = co->co_consts;
	fastlocals = f->f_localsplus;
	freevars = f->f_localsplus + co->co_nlocals;
	first_instr = (unsigned char*) PyString_AS_STRING(co->co_code);
	/* An explanation is in order for the next line.

	   f->f_lasti now refers to the index of the last instruction
	   executed.  You might think this was obvious from the name, but
	   this wasn't always true before 2.3!  PyFrame_New now sets
	   f->f_lasti to -1 (i.e. the index *before* the first instruction)
	   and YIELD_VALUE doesn't fiddle with f_lasti any more.  So this
	   does work.  Promise.

	   When the PREDICT() macros are enabled, some opcode pairs follow in
           direct succession without updating f->f_lasti.  A successful
           prediction effectively links the two codes together as if they
           were a single new opcode; accordingly,f->f_lasti will point to
           the first code in the pair (for instance, GET_ITER followed by
           FOR_ITER is effectively a single opcode and f->f_lasti will point
           at to the beginning of the combined pair.)
	*/
	next_instr = first_instr + f->f_lasti + 1;
	stack_pointer = f->f_stacktop;
	assert(stack_pointer != NULL);
	f->f_stacktop = NULL;	/* remains NULL unless yield suspends frame */

#ifdef LLTRACE
	lltrace = PyDict_GetItemString(f->f_globals, "__lltrace__") != NULL;
#endif
#if defined(Py_DEBUG) || defined(LLTRACE)
	filename = PyString_AsString(co->co_filename);
#endif

	why = UNWIND_NOUNWIND;
	w = NULL;

	/* Note that this goes after the LLVM handling code so we don't log
	 * this event when calling LLVM functions. Do this before the throwflag
	 * check below to avoid mismatched enter/exit events in the log. */
	PY_LOG_TSC_EVENT(CALL_ENTER_EVAL);

	if (f->f_throwflag) { /* support for generator.throw() */
		why = UNWIND_EXCEPTION;
		goto on_error;
	}

	for (;;) {
		assert(stack_pointer >= f->f_valuestack); /* else underflow */
		assert(STACK_LEVEL() <= co->co_stacksize);  /* else overflow */

		/* Do periodic things.  Doing this every time through
		   the loop would add too much overhead, so we do it
		   only every Nth instruction.  We also do it if
		   ``things_to_do'' is set, i.e. when an asynchronous
		   event needs attention (e.g. a signal handler or
		   async I/O handler); see Py_AddPendingCall() and
		   Py_MakePendingCalls() above. */

		if (--_Py_Ticker < 0) {
			if (*next_instr == SETUP_FINALLY) {
				/* Make the last opcode before
				   a try: finally: block uninterruptable. */
				goto fast_next_opcode;
			}
			if (_PyEval_HandlePyTickerExpired(tstate) == -1) {
				why = UNWIND_EXCEPTION;
				goto on_error;
			}
		}

	fast_next_opcode:
		f->f_lasti = INSTR_OFFSET();

		/* line-by-line tracing support */

		if (_Py_TracingPossible &&
		    tstate->c_tracefunc != NULL && !tstate->tracing) {
			/* see maybe_call_line_trace
			   for expository comments */
			f->f_stacktop = stack_pointer;

			err = maybe_call_line_trace(tstate->c_tracefunc,
						    tstate->c_traceobj,
						    f, &instr_lb, &instr_ub,
						    &instr_prev);
			/* Reload possibly changed frame fields */
			JUMPTO(f->f_lasti);
			assert(f->f_stacktop != NULL);
			stack_pointer = f->f_stacktop;
			f->f_stacktop = NULL;
			if (err) {
				/* trace function raised an exception */
				why = UNWIND_EXCEPTION;
				goto on_error;
			}
		}

		/* Extract opcode and argument */

		opcode = NEXTOP();
		oparg = 0;   /* allows oparg to be stored in a register because
			it doesn't have to be remembered across a full loop */
		if (HAS_ARG(opcode))
			oparg = NEXTARG();
	  dispatch_opcode:
#ifdef DYNAMIC_EXECUTION_PROFILE
#ifdef DXPAIRS
		dxpairs[lastopcode][opcode]++;
		lastopcode = opcode;
#endif
		dxp[opcode]++;
#endif

#ifdef LLTRACE
		/* Instruction tracing */

		if (lltrace) {
			if (HAS_ARG(opcode)) {
				printf("%d: %d, %d\n",
				       f->f_lasti, opcode, oparg);
			}
			else {
				printf("%d: %d\n",
				       f->f_lasti, opcode);
			}
		}
#endif

		/* Main switch on opcode */

		assert(why == UNWIND_NOUNWIND);
		/* XXX(jyasskin): Add an assertion under CHECKEXC that
		   !PyErr_Occurred(). */
		switch (opcode) {

		/* BEWARE!
		   It is essential that any operation that fails sets
		   why to anything but UNWIND_NOUNWIND, and that no operation
		   that succeeds does this! */

		/* case STOP_CODE: this is an error! */

		TARGET(NOP)
			FAST_DISPATCH();

		TARGET(LOAD_FAST)
			x = GETLOCAL(oparg);
			if (x != NULL) {
				Py_INCREF(x);
				PUSH(x);
				FAST_DISPATCH();
			}
			_PyEval_RaiseForUnboundLocal(f, oparg);
			why = UNWIND_EXCEPTION;
			break;

		TARGET(LOAD_CONST)
			x = GETITEM(consts, oparg);
			Py_INCREF(x);
			PUSH(x);
			FAST_DISPATCH();

		PREDICTED_WITH_ARG(STORE_FAST);
		TARGET(STORE_FAST)
			v = POP();
			SETLOCAL(oparg, v);
			FAST_DISPATCH();

		TARGET(POP_TOP)
			v = POP();
			Py_DECREF(v);
			FAST_DISPATCH();

		TARGET(ROT_TWO)
			v = TOP();
			w = SECOND();
			SET_TOP(w);
			SET_SECOND(v);
			FAST_DISPATCH();

		TARGET(ROT_THREE)
			v = TOP();
			w = SECOND();
			x = THIRD();
			SET_TOP(w);
			SET_SECOND(x);
			SET_THIRD(v);
			FAST_DISPATCH();

		TARGET(ROT_FOUR)
			u = TOP();
			v = SECOND();
			w = THIRD();
			x = FOURTH();
			SET_TOP(v);
			SET_SECOND(w);
			SET_THIRD(x);
			SET_FOURTH(u);
			FAST_DISPATCH();

		TARGET(DUP_TOP)
			v = TOP();
			Py_INCREF(v);
			PUSH(v);
			FAST_DISPATCH();

		TARGET(DUP_TOP_TWO)
			x = TOP();
			Py_INCREF(x);
			w = SECOND();
			Py_INCREF(w);
			STACKADJ(2);
			SET_TOP(x);
			SET_SECOND(w);
			FAST_DISPATCH();

		TARGET(DUP_TOP_THREE)
			x = TOP();
			Py_INCREF(x);
			w = SECOND();
			Py_INCREF(w);
			v = THIRD();
			Py_INCREF(v);
			STACKADJ(3);
			SET_TOP(x);
			SET_SECOND(w);
			SET_THIRD(v);
			FAST_DISPATCH();

		TARGET(UNARY_POSITIVE)
			v = TOP();
			RECORD_TYPE(0, v);
			x = PyNumber_Positive(v);
			Py_DECREF(v);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(UNARY_NEGATIVE)
			v = TOP();
			RECORD_TYPE(0, v);
			x = PyNumber_Negative(v);
			Py_DECREF(v);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(UNARY_NOT)
			v = TOP();
			RECORD_TYPE(0, v);
			err = PyObject_IsTrue(v);
			Py_DECREF(v);
			if (err == 0) {
				Py_INCREF(Py_True);
				SET_TOP(Py_True);
				DISPATCH();
			}
			else if (err > 0) {
				Py_INCREF(Py_False);
				SET_TOP(Py_False);
				DISPATCH();
			}
			STACKADJ(-1);
			why = UNWIND_EXCEPTION;
			break;

		TARGET(UNARY_CONVERT)
			v = TOP();
			RECORD_TYPE(0, v);
			x = PyObject_Repr(v);
			Py_DECREF(v);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(UNARY_INVERT)
			v = TOP();
			RECORD_TYPE(0, v);
			x = PyNumber_Invert(v);
			Py_DECREF(v);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_POWER)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Power(v, w, Py_None);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_MULTIPLY)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Multiply(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_DIVIDE)
			if (!_Py_QnewFlag) {
				w = POP();
				v = TOP();
				RECORD_TYPE(0, v);
				RECORD_TYPE(1, w);
				x = PyNumber_Divide(v, w);
				Py_DECREF(v);
				Py_DECREF(w);
				SET_TOP(x);
				if (x == NULL) {
					why = UNWIND_EXCEPTION;
					break;
				}
				DISPATCH();
			}
			/* -Qnew is in effect: jump to BINARY_TRUE_DIVIDE */
			goto _binary_true_divide;

		TARGET(BINARY_TRUE_DIVIDE)
		_binary_true_divide:
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_TrueDivide(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_FLOOR_DIVIDE)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_FloorDivide(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_MODULO)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyString_CheckExact(v))
				x = PyString_Format(v, w);
			else
				x = PyNumber_Remainder(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_ADD)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyInt_CheckExact(v) && PyInt_CheckExact(w)) {
				/* INLINE: int + int */
				register long a, b, i;
				a = PyInt_AS_LONG(v);
				b = PyInt_AS_LONG(w);
				i = a + b;
				if ((i^a) < 0 && (i^b) < 0)
					goto slow_add;
				x = PyInt_FromLong(i);
			}
			else if (PyString_CheckExact(v) &&
				 PyString_CheckExact(w)) {
				x = string_concatenate(v, w, f, next_instr);
				/* string_concatenate consumed the ref to v */
				goto skip_decref_vx;
			}
			else {
			  slow_add:
				x = PyNumber_Add(v, w);
			}
			Py_DECREF(v);
		  skip_decref_vx:
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_SUBTRACT)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyInt_CheckExact(v) && PyInt_CheckExact(w)) {
				/* INLINE: int - int */
				register long a, b, i;
				a = PyInt_AS_LONG(v);
				b = PyInt_AS_LONG(w);
				i = a - b;
				if ((i^a) < 0 && (i^~b) < 0)
					goto slow_sub;
				x = PyInt_FromLong(i);
			}
			else {
			  slow_sub:
				x = PyNumber_Subtract(v, w);
			}
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_SUBSCR)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyList_CheckExact(v) && PyInt_CheckExact(w)) {
				/* INLINE: list[int] */
				Py_ssize_t i = PyInt_AsSsize_t(w);
				if (i < 0)
					i += PyList_GET_SIZE(v);
				if (i >= 0 && i < PyList_GET_SIZE(v)) {
					x = PyList_GET_ITEM(v, i);
					Py_INCREF(x);
				}
				else
					goto slow_get;
			}
			else
			  slow_get:
				x = PyObject_GetItem(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_LSHIFT)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Lshift(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_RSHIFT)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Rshift(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_AND)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_And(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_XOR)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Xor(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(BINARY_OR)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_Or(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(LIST_APPEND)
			w = POP();
			v = POP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			err = PyList_Append(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			if (err != 0) {
				why = UNWIND_EXCEPTION;
				break;
			}
			PREDICT(JUMP_ABSOLUTE);
			DISPATCH();

		TARGET(INPLACE_POWER)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlacePower(v, w, Py_None);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_MULTIPLY)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlaceMultiply(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_DIVIDE)
			if (!_Py_QnewFlag) {
				w = POP();
				v = TOP();
				RECORD_TYPE(0, v);
				RECORD_TYPE(1, w);
				x = PyNumber_InPlaceDivide(v, w);
				Py_DECREF(v);
				Py_DECREF(w);
				SET_TOP(x);
				if (x == NULL) {
					why = UNWIND_EXCEPTION;
					break;
				}
				DISPATCH();
			}
			/* -Qnew is in effect: jump to INPLACE_TRUE_DIVIDE */
			goto _inplace_true_divide;

		TARGET(INPLACE_TRUE_DIVIDE)
		_inplace_true_divide:
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlaceTrueDivide(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_FLOOR_DIVIDE)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlaceFloorDivide(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_MODULO)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlaceRemainder(v, w);
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_ADD)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyInt_CheckExact(v) && PyInt_CheckExact(w)) {
				/* INLINE: int + int */
				register long a, b, i;
				a = PyInt_AS_LONG(v);
				b = PyInt_AS_LONG(w);
				i = a + b;
				if ((i^a) < 0 && (i^b) < 0)
					goto slow_iadd;
				x = PyInt_FromLong(i);
			}
			else if (PyString_CheckExact(v) &&
				 PyString_CheckExact(w)) {
				x = string_concatenate(v, w, f, next_instr);
				/* string_concatenate consumed the ref to v */
				goto skip_decref_v;
			}
			else {
			  slow_iadd:
				x = PyNumber_InPlaceAdd(v, w);
			}
			Py_DECREF(v);
		  skip_decref_v:
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_SUBTRACT)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			if (PyInt_CheckExact(v) && PyInt_CheckExact(w)) {
				/* INLINE: int - int */
				register long a, b, i;
				a = PyInt_AS_LONG(v);
				b = PyInt_AS_LONG(w);
				i = a - b;
				if ((i^a) < 0 && (i^~b) < 0)
					goto slow_isub;
				x = PyInt_FromLong(i);
			}
			else {
			  slow_isub:
				x = PyNumber_InPlaceSubtract(v, w);
			}
			Py_DECREF(v);
			Py_DECREF(w);
			SET_TOP(x);
			if (x == NULL) {
				why = UNWIND_EXCEPTION;
				break;
			}
			DISPATCH();

		TARGET(INPLACE_LSHIFT)
			w = POP();
			v = TOP();
			RECORD_TYPE(0, v);
			RECORD_TYPE(1, w);
			x = PyNumber_InPlaceL…