[Python-Dev] Python VM (original) (raw)

Jakob Sievers cadr4u at gmail.com
Mon Jul 21 17:58:20 CEST 2008

Hi, I've been reading the Python VM sources over the last few afternoons and I took some notes, which I thought I'd share (and if anyone familiar with the VM internals could have a quick look at them, I'd really appreciate it).

Cheers, -jakob

Unless otherwise noted, the source file in question is Python/ceval.c.

Control Flow

The calling sequence is: main() (in python.c) -> Py_Main() (main.c) -> PyRun_FooFlags() (pythonrun.c) -> run_bar() (pythonrun.c) -> PyEval_EvalCode() (ceval.c) -> PyEval_EvalCodeEx() (ceval.c) -> PyEval_EvalFrameEx() (ceval.c).

EvalCodeEx() does some initialization (creating a new execution frame, argument processing, and some generator-specific stuff I haven't looked at yet) before calling EvalFrameEx() which contains the main interpreter loop.

Threads

PyEval_InitThreads() initializes the GIL (interpreter_lock) and sets main_thread to the (threading package dependent) ID of the current thread. Thread switching is done using PyThreadState_Swap(), which sets _PyThreadState_Current (both defined in pystate.c) and PyThreadState_GET() (an alias for _PyThreadState_Current) (pystate.h).

Async Callbacks

Asynchronous callbacks can be registered by adding the function to be called to pendingcalls[] (see Py_AddPendingCall()). The state of this queue is communicated to the main loop via things_to_do.

State

The global state is recorded in a (per-process?) PyInterpreterState struct and a per-thread PyThreadState struct. Each execution frame's state is contained in that frame's PyFrameObject (which includes the instruction stream, the environment (globals, locals, builtins, etc.), the value stack and so forth). EvalFrameEx()'s local variables are initialized from this frame object.

Instruction Stream

The instruction stream looks as follows (c.f. assemble_emit() in compile.c): A byte stream where each instruction consists of either

  1. a single byte opcode: OP
  2. a single byte opcode plus a two-byte immediate argument: OP LO HI
  3. a special opcode followed by the real opcode followed by a four byte argument: EXTENDED_ARG OP BYTE2 BYTE1 BYTE4 BYTE3

Opcode Prediction

One nice trick used to speed up opcode dispatch is the following: Using the macros PREDICT() and PREDICTED() it is sometimes possible to jump directly to the code implementing the next instruction rather than having to go through the whole loop preamble, e.g.

case FOO: // ... PREDICT(BAR); continue;

PREDICTED(BAR); case BAR: // ...

expands to

case FOO: // ... if (*next_instr == BAR) goto PRED_BAR; continue;

PRED_BAR: next_instr++; case BAR: // ...

Main Loop

Variables and macros used in EvalFrameEx()

The value stack: PyObject **stack_pointer; The instruction stream: unsigned char *next_instr; NEXTOP(), NEXTARG(), PEEKARG(), JUMPTO(), and JUMPBY() simply fiddle with next_instr. Likewise for TOP(), SET_SECOND(), PUSH(), POP(), etc. and stack_pointer.

Current opcode plus argument: int opcode; int oparg;

Error status: enum why_code why; // no, exn, exn re-raised, return, break, continue, yield int err; // non-zero is error

The environment: PyObject *names; PyObject *consts; and PyObject **fastlocals; which is accessed via GETLOCAL() and SETLOCAL().

Finally, there are some more PyObject *'s (v, w, u, and so forth, used as temporary variables) as well as PyObject *retval;

Basic structure

EvalFrameEx() { why = WHY_NOT; err = 0;

for (;;) {    <------------------+---+
    // do periodic tasks         |   |
                                 |   |
fast_next_opcode:                |   |
    opcode = NEXTOP();           |   |
    if (HAS_ARG(opcode))         |   |
        oparg = NEXTARG();       |   |
                                 |   |
dispatch_opcode:                 |   |
    switch(opcode) {             |   |
                                 |   |
    continue; -------------------+   |
                                     |
    break; ----------------------+   |
                                 |   |
    // Also, opcode prediction   |   |
    // jumps around inside the   |   |
    // switch statement          |   |
                                 |   |
    }    <-----------------------+   |
                                     |
on_error:                            |
    // no error: continue -----------+
    // otherwise why == WHY_EXCEPTION after this

fast_block_end:
    // unwind stacks if there was an error
}

// more unwinding

fast_yield: // reset current thread's exception info exit_eval_frame: // set thread's execution frame to previous execution frame return retval; }

Periodic Tasks

By checking and decrementing _Py_Ticker, the main loop executes certain tasks once every _Py_CheckInterval iterations (in fact Py_AddPendingCall() sets _Py_Ticker to zero, ensuring that pending calls are executed right after the next instruction which doesn't jump to fast_next_opcode):

Instruction implementation

Some general notes:

Nested blocks

Nested loop and try blocks are handled as follows: Each frame maintains a block stack; when entering a nested block, a SETUP_* instruction adds a PyTryBlock to the PyFrameObject's f_blockstack and registers that block's type (instruction which created the block), handler (offset into the instruction stream) and level (value stack level before the nested block was entered). When such blocks are exited normally (e.g. last iteration of a loop), the final POP_BLOCK instruction restores the value stack to the state it was in before the block. If a block is exited abnormally (e.g. a break instruction), the code following fast_block_end unwinds the value stack and jumps to the block's handler. Certain instructions, e.g. RETURN_VALUE, cause the entire block stack to be unwound (leading to multiple unwinds of the value stack). See also the comments in compile.c (compiler_try_finally()) which include nice ASCII art diagrams (incidentally, there's a typo on line 2382: s/en/an/).

Error handling

Internal errors (bad oparg, say) generally result in why being set to WHY_EXCEPTION and breaking out of the switch (if the code implementing the instruction doesn't set why, the code after on_error will). The code following fast_block_end will jump to the correct exception handler and set the global exception related variables (exception information is stored in the current execution frame and the thread state. See set_exc_info() and reset_exc_info()).

// eof

More information about the Python-Dev mailing list