cpython/InternalDocs/interpreter.md

The bytecode interpreter

This document describes the workings and implementation of the bytecode interpreter, the part of python that executes compiled Python code. Its entry point is in Python/ceval.c.

At a high level, the interpreter consists of a loop that iterates over the bytecode instructions, executing each of them via a switch statement that has a case implementing each opcode. This switch statement is generated from the instruction definitions in Python/bytecodes.c which are written in a DSL developed for this purpose.

Recall that the Python Compiler produces a CodeObject, which contains the bytecode instructions along with static data that is required to execute them, such as the consts list, variable names, exception table, and so on.

When the interpreter's PyEval_EvalCode() function is called to execute a CodeObject, it constructs a Frame and calls _PyEval_EvalFrame() to execute the code object in this frame. The frame hold the dynamic state of the CodeObject's execution, including the instruction pointer, the globals and builtins. It also has a reference to the CodeObject itself.

In addition to the frame, _PyEval_EvalFrame() also receives a Thread State object, tstate, which includes things like the exception state and the recursion depth. The thread state also provides access to the per-interpreter state (tstate->interp), which has a pointer to the per-runtime (that is, truly global) state (tstate->interp->runtime).

Finally, _PyEval_EvalFrame() receives an integer argument throwflag which, when nonzero, indicates that the interpreter should just raise the current exception (this is used in the implementation of gen.throw.

By default, _PyEval_EvalFrame() simply calls [_PyEval_EvalFrameDefault()] to execute the frame. However, as per PEP 523 this is configurable by setting interp->eval_frame. In the following, we describe the default function, _PyEval_EvalFrameDefault().

Instruction decoding

The first task of the interpreter is to decode the bytecode instructions. Bytecode is stored as an array of 16-bit code units (_Py_CODEUNIT). Each code unit contains an 8-bit opcode and an 8-bit argument (oparg), both unsigned. In order to make the bytecode format independent of the machine byte order when stored on disk, opcode is always the first byte and oparg is always the second byte. Macros are used to extract the opcode and oparg from a code unit (_Py_OPCODE(word) and _Py_OPARG(word)). Some instructions (for example, NOP or POP_TOP) have no argument -- in this case we ignore oparg.

A simplified version of the interpreter's main loop looks like this:

    _Py_CODEUNIT *first_instr = code->co_code_adaptive;
    _Py_CODEUNIT *next_instr = first_instr;
    while (1) {
        _Py_CODEUNIT word = *next_instr++;
        unsigned char opcode = _Py_OPCODE(word);
        unsigned int oparg = _Py_OPARG(word);
        switch (opcode) {
        // ... A case for each opcode ...
        }
    }

This loop iterates over the instructions, decoding each into its opcode and oparg, and then executes the switch case that implements this opcode.

The instruction format supports 256 different opcodes, which is sufficient. However, it also limits oparg to 8-bit values, which is too restrictive. To overcome this, the EXTENDED_ARG opcode allows us to prefix any instruction with one or more additional data bytes, which combine into a larger oparg. For example, this sequence of code units:

EXTENDED_ARG  1
EXTENDED_ARG  0
LOAD_CONST    2

would set opcode to LOAD_CONST and oparg to 65538 (that is, 0x1_00_02). The compiler should limit itself to at most three EXTENDED_ARG prefixes, to allow the resulting oparg to fit in 32 bits, but the interpreter does not check this.

In the following, a code unit is always two bytes, while an instruction is a sequence of code units consisting of zero to three EXTENDED_ARG opcodes followed by a primary opcode.

The following loop, to be inserted just above the switch statement, will make the above snippet decode a complete instruction:

    while (opcode == EXTENDED_ARG) {
        word = *next_instr++;
        opcode = _Py_OPCODE(word);
        oparg = (oparg << 8) | _Py_OPARG(word);
    }

For various reasons we'll get to later (mostly efficiency, given that EXTENDED_ARG is rare) the actual code is different.

## Jumps

Note that when the switch statement is reached, next_instr (the "instruction offset") already points to the next instruction. Thus, jump instructions can be implemented by manipulating next_instr:

• A jump forward (JUMP_FORWARD) sets next_instr += oparg.
• A jump backward sets next_instr -= oparg.

Inline cache entries

Some (specialized or specializable) instructions have an associated "inline cache". The inline cache consists of one or more two-byte entries included in the bytecode array as additional words following the opcode/oparg pair. The size of the inline cache for a particular instruction is fixed by its opcode. Moreover, the inline cache size for all instructions in a family of specialized/specializable instructions (for example, LOAD_ATTR, LOAD_ATTR_SLOT, LOAD_ATTR_MODULE) must all be the same. Cache entries are reserved by the compiler and initialized with zeros. Although they are represented by code units, cache entries do not conform to the opcode / oparg format.

If an instruction has an inline cache, the layout of its cache is described in the instruction's definition in Python/bytecodes.c. The structs defined in pycore_code.h allow us to access the cache by casting next_instr to a pointer to the relevant struct. The size of such a struct must be independent of the machine architecture, word size and alignment requirements. For a 32-bit field, the struct should use _Py_CODEUNIT field[2].

The instruction implementation is responsible for advancing next_instr past the inline cache. For example, if an instruction's inline cache is four bytes (that is, two code units) in size, the code for the instruction must contain next_instr += 2;. This is equivalent to a relative forward jump by that many code units. (In the interpreter definition DSL, this is coded as JUMPBY(n), where n is the number of code units to jump, typically given as a named constant.)

Serializing non-zero cache entries would present a problem because the serialization (:mod:marshal) format must be independent of the machine byte order.

More information about the use of inline caches can be found in PEP 659.

The evaluation stack

Most instructions read or write some data in the form of object references (PyObject *). The CPython bytecode interpreter is a stack machine, meaning that its instructions operate by pushing data onto and popping it off the stack. The stack is forms part of the frame for the code object. Its maximum depth is calculated by the compiler and stored in the co_stacksize field of the code object, so that the stack can be pre-allocated is a contiguous array of PyObject* pointers, when the frame is created.

The stack effects of each instruction are also exposed through the opcode metadata through two functions that report how many stack elements the instructions consumes, and how many it produces (_PyOpcode_num_popped and _PyOpcode_num_pushed). For example, the BINARY_OP instruction pops two objects from the stack and pushes the result back onto the stack.

The stack grows up in memory; the operation PUSH(x) is equivalent to *stack_pointer++ = x, whereas x = POP() means x = *--stack_pointer. Overflow and underflow checks are active in debug mode, but are otherwise optimized away.

At any point during execution, the stack level is knowable based on the instruction pointer alone, and some properties of each item on the stack are also known. In particular, only a few instructions may push a NULL onto the stack, and the positions that may be NULL are known. A few other instructions (GET_ITER, FOR_ITER) push or pop an object that is known to be an iterator.

Instruction sequences that do not allow statically knowing the stack depth are deemed illegal; the bytecode compiler never generates such sequences. For example, the following sequence is illegal, because it keeps pushing items on the stack:

LOAD_FAST 0
JUMP_BACKWARD 2

[!NOTE] Do not confuse the evaluation stack with the call stack, which is used to implement calling and returning from functions.

Error handling

When the implementation of an opcode raises an exception, it jumps to the exception_unwind label in Python/ceval.c. The exception is then handled as described in the exception handling documentation.

Python-to-Python calls

The _PyEval_EvalFrameDefault() function is recursive, because sometimes the interpreter calls some C function that calls back into the interpreter. In 3.10 and before, this was the case even when a Python function called another Python function: The CALL opcode would call the tp_call dispatch function of the callee, which would extract the code object, create a new frame for the call stack, and then call back into the interpreter. This approach is very general but consumes several C stack frames for each nested Python call, thereby increasing the risk of an (unrecoverable) C stack overflow.

Since 3.11, the CALL instruction special-cases function objects to "inline" the call. When a call gets inlined, a new frame gets pushed onto the call stack and the interpreter "jumps" to the start of the callee's bytecode. When an inlined callee executes a RETURN_VALUE instruction, the frame is popped off the call stack and the interpreter returns to its caller, by popping a frame off the call stack and "jumping" to the return address. There is a flag in the frame (frame->is_entry) that indicates whether the frame was inlined (set if it wasn't). If RETURN_VALUE finds this flag set, it performs the usual cleanup and returns from _PyEval_EvalFrameDefault() altogether, to a C caller.

A similar check is performed when an unhandled exception occurs.

The call stack

Up through 3.10, the call stack was implemented as a singly-linked list of frame objects. This was expensive because each call would require a heap allocation for the stack frame.

Since 3.11, frames are no longer fully-fledged objects. Instead, a leaner internal _PyInterpreterFrame structure is used, which is allocated using a custom allocator function (_PyThreadState_BumpFramePointer()), which allocates and initializes a frame structure. Usually a frame allocation is just a pointer bump, which improves memory locality.

Sometimes an actual PyFrameObject is needed, such as when Python code calls sys._getframe() or an extension module calls PyEval_GetFrame(). In this case we allocate a proper PyFrameObject and initialize it from the _PyInterpreterFrame.

Things get more complicated when generators are involved, since those do not follow the push/pop model. This includes async functions, which are based on the same mechanism. A generator object has space for a _PyInterpreterFrame structure, including the variable-size part (used for locals and the eval stack). When a generator (or async) function is first called, a special opcode RETURN_GENERATOR is executed, which is responsible for creating the generator object. The generator object's _PyInterpreterFrame is initialized with a copy of the current stack frame. The current stack frame is then popped off the frame stack and the generator object is returned. (Details differ depending on the is_entry flag.) When the generator is resumed, the interpreter pushes its _PyInterpreterFrame onto the frame stack and resumes execution. See also the generators section.

Introducing a new bytecode instruction

It is occasionally necessary to add a new opcode in order to implement a new feature or change the way that existing features are compiled. This section describes the changes required to do this.

First, you must choose a name for the bytecode, implement it in Python/bytecodes.c and add a documentation entry in Doc/library/dis.rst. Then run make regen-cases to assign a number for it (see Include/opcode_ids.h) and regenerate a number of files with the actual implementation of the bytecode in Python/generated_cases.c.h and metadata about it in additional files.

With a new bytecode you must also change what is called the "magic number" for .pyc files: bump the value of the variable MAGIC_NUMBER in Lib/importlib/_bootstrap_external.py. Changing this number will lead to all .pyc files with the old MAGIC_NUMBER to be recompiled by the interpreter on import. Whenever MAGIC_NUMBER is changed, the ranges in the magic_values array in PC/launcher.c may also need to be updated. Changes to Lib/importlib/_bootstrap_external.py will take effect only after running make regen-importlib.

[!NOTE] Running make regen-importlib before adding the new bytecode target to Python/bytecodes.c (followed by make regen-cases) will result in an error. You should only run make regen-importlib after the new bytecode target has been added.

[!NOTE] On Windows, running the ./build.bat script will automatically regenerate the required files without requiring additional arguments.

Finally, you need to introduce the use of the new bytecode. Update Python/codegen.c to emit code with this bytecode. Optimizations in Python/flowgraph.c may also need to be updated. If the new opcode affects a control flow or the block stack, you may have to update the frame_setlineno() function in Objects/frameobject.c. It may also be necessary to update Lib/dis.py if the new opcode interprets its argument in a special way (like FORMAT_VALUE or MAKE_FUNCTION).

If you make a change here that can affect the output of bytecode that is already in existence and you do not change the magic number, make sure to delete your old .py(c|o) files! Even though you will end up changing the magic number if you change the bytecode, while you are debugging your work you may be changing the bytecode output without constantly bumping up the magic number. This can leave you with stale .pyc files that will not be recreated. Running find . -name '*.py[co]' -exec rm -f '{}' + should delete all .pyc files you have, forcing new ones to be created and thus allow you test out your new bytecode properly. Run make regen-importlib for updating the bytecode of frozen importlib files. You have to run make again after this to recompile the generated C files.

Specialization

Bytecode specialization, which was introduced in PEP 659, speeds up program execution by rewriting instructions based on runtime information. This is done by replacing a generic instruction with a faster version that works for the case that this program encounters. Each specializable instruction is responsible for rewriting itself, using its inline caches for bookkeeping.

When an adaptive instruction executes, it may attempt to specialize itself, depending on the argument and the contents of its cache. This is done by calling one of the _Py_Specialize_XXX functions in Python/specialize.c.

The specialized instructions are responsible for checking that the special-case assumptions still apply, and de-optimizing back to the generic version if not.

Families of instructions

A family of instructions consists of an adaptive instruction along with the specialized instructions that it can be replaced by. It has the following fundamental properties:

• It corresponds to a single instruction in the code generated by the bytecode compiler.
• It has a single adaptive instruction that records an execution count and, at regular intervals, attempts to specialize itself. If not specializing, it executes the base implementation.
• It has at least one specialized form of the instruction that is tailored for a particular value or set of values at runtime.
• All members of the family must have the same number of inline cache entries, to ensure correct execution. Individual family members do not need to use all of the entries, but must skip over any unused entries when executing.

The current implementation also requires the following, although these are not fundamental and may change:

• All families use one or more inline cache entries, the first entry is always the counter.
• All instruction names should start with the name of the adaptive instruction.
• Specialized forms should have names describing their specialization.

Example family

The LOAD_GLOBAL instruction (in Python/bytecodes.c) already has an adaptive family that serves as a relatively simple example.

The LOAD_GLOBAL instruction performs adaptive specialization, calling _Py_Specialize_LoadGlobal() when the counter reaches zero.

There are two specialized instructions in the family, LOAD_GLOBAL_MODULE which is specialized for global variables in the module, and LOAD_GLOBAL_BUILTIN which is specialized for builtin variables.

Performance analysis

The benefit of a specialization can be assessed with the following formula: Tbase/Tadaptive.

Where Tbase is the mean time to execute the base instruction, and Tadaptive is the mean time to execute the specialized and adaptive forms.

Tadaptive = (sum(Ti*Ni) + Tmiss*Nmiss)/(sum(Ni)+Nmiss)

Ti is the time to execute the ith instruction in the family and Ni is the number of times that instruction is executed. Tmiss is the time to process a miss, including de-optimzation and the time to execute the base instruction.

The ideal situation is where misses are rare and the specialized forms are much faster than the base instruction. LOAD_GLOBAL is near ideal, Nmiss/sum(Ni) ≈ 0. In which case we have Tadaptive ≈ sum(Ti*Ni). Since we can expect the specialized forms LOAD_GLOBAL_MODULE and LOAD_GLOBAL_BUILTIN to be much faster than the adaptive base instruction, we would expect the specialization of LOAD_GLOBAL to be profitable.

Design considerations

While LOAD_GLOBAL may be ideal, instructions like LOAD_ATTR and CALL_FUNCTION are not. For maximum performance we want to keep Ti low for all specialized instructions and Nmiss as low as possible.

Keeping Nmiss low means that there should be specializations for almost all values seen by the base instruction. Keeping sum(Ti*Ni) low means keeping Ti low which means minimizing branches and dependent memory accesses (pointer chasing). These two objectives may be in conflict, requiring judgement and experimentation to design the family of instructions.

The size of the inline cache should as small as possible, without impairing performance, to reduce the number of EXTENDED_ARG jumps, and to reduce pressure on the CPU's data cache.

Gathering data

Before choosing how to specialize an instruction, it is important to gather some data. What are the patterns of usage of the base instruction? Data can best be gathered by instrumenting the interpreter. Since a specialization function and adaptive instruction are going to be required, instrumentation can most easily be added in the specialization function.

Choice of specializations

The performance of the specializing adaptive interpreter relies on the quality of specialization and keeping the overhead of specialization low.

Specialized instructions must be fast. In order to be fast, specialized instructions should be tailored for a particular set of values that allows them to:

1. Verify that incoming value is part of that set with low overhead.
2. Perform the operation quickly.

This requires that the set of values is chosen such that membership can be tested quickly and that membership is sufficient to allow the operation to performed quickly.

For example, LOAD_GLOBAL_MODULE is specialized for globals() dictionaries that have a keys with the expected version.

This can be tested quickly:

• globals->keys->dk_version == expected_version

and the operation can be performed quickly:

• value = entries[cache->index].me_value;.

Because it is impossible to measure the performance of an instruction without also measuring unrelated factors, the assessment of the quality of a specialization will require some judgement.

As a general rule, specialized instructions should be much faster than the base instruction.

Implementation of specialized instructions

In general, specialized instructions should be implemented in two parts:

1. A sequence of guards, each of the form DEOPT_IF(guard-condition-is-false, BASE_NAME).
2. The operation, which should ideally have no branches and a minimum number of dependent memory accesses.

In practice, the parts may overlap, as data required for guards can be re-used in the operation.

If there are branches in the operation, then consider further specialization to eliminate the branches.

Maintaining stats

Finally, take care that stats are gathered correctly. After the last DEOPT_IF has passed, a hit should be recorded with STAT_INC(BASE_INSTRUCTION, hit). After an optimization has been deferred in the adaptive instruction, that should be recorded with STAT_INC(BASE_INSTRUCTION, deferred).

Additional resources

• Brandt Bucher's talk about the specializing interpreter at PyCon US 2023. Slides Video