cpython/Objects/frame_layout.md

6.4 KiB

The Frame Stack

Each call to a Python function has an activation record, commonly known as a "frame". Python semantics allows frames to outlive the activation, so they have (before 3.11) been allocated on the heap. This is expensive as it requires many allocations and results in poor locality of reference.

In 3.11, rather than have these frames scattered about memory, as happens for heap-allocated objects, frames are allocated contiguously in a per-thread stack. This improves performance significantly for two reasons:

  • It reduces allocation overhead to a pointer comparison and increment.
  • Stack allocated data has the best possible locality and will always be in CPU cache.

Generator and coroutines still need heap allocated activation records, but can be linked into the per-thread stack so as to not impact performance too much.

Layout

Each activation record consists of four conceptual sections:

  • Local variables (including arguments, cells and free variables)
  • Evaluation stack
  • Specials: The per-frame object references needed by the VM: globals dict, code object, etc.
  • Linkage: Pointer to the previous activation record, stack depth, etc.

Layout

The specials and linkage sections are a fixed size, so are grouped together.

Each activation record is laid out as:

  • Specials and linkage
  • Locals
  • Stack

This seems to provide the best performance without excessive complexity. It needs the interpreter to hold two pointers, a frame pointer and a stack pointer.

Alternative layout

An alternative layout that was used for part of 3.11 alpha was:

  • Locals
  • Specials and linkage
  • Stack

This has the advantage that no copying is required when making a call, as the arguments on the stack are (usually) already in the correct location for the parameters. However, it requires the VM to maintain an extra pointer for the locals, which can hurt performance.

A variant that only needs the need two pointers is to reverse the numbering of the locals, so that the last one is numbered 0, and the first in memory is numbered N-1. This allows the locals, specials and linkage to accessed from the frame pointer. We may implement this in the future.

Note:

In a contiguous stack, we would need to save one fewer registers, as the top of the caller's activation record would be the same at the base of the callee's. However, since some activation records are kept on the heap we cannot do this.

Generators and Coroutines

Generators and coroutines contain a _PyInterpreterFrame The specials sections contains the following pointers:

  • Globals dict
  • Builtins dict
  • Locals dict (not the "fast" locals, but the locals for eval and class creation)
  • Code object
  • Heap allocated PyFrameObject for this activation record, if any.
  • The function.

The pointer to the function is not strictly required, but it is cheaper to store a strong reference to the function and borrowed references to the globals and builtins, than strong references to both globals and builtins.

Frame objects

When creating a backtrace or when calling sys._getframe() the frame becomes visible to Python code. When this happens a new PyFrameObject is created and a strong reference to it placed in the frame_obj field of the specials section. The frame_obj field is initially NULL.

The PyFrameObject may outlive a stack-allocated _PyInterpreterFrame. If it does then _PyInterpreterFrame is copied into the PyFrameObject, except the evaluation stack which must be empty at this point. The linkage section is updated to reflect the new location of the frame.

This mechanism provides the appearance of persistent, heap-allocated frames for each activation, but with low runtime overhead.

Generators and Coroutines

Generator objects have a _PyInterpreterFrame embedded in them. This means that creating a generator requires only a single allocation, reducing allocation overhead and improving locality of reference. The embedded frame is linked into the per-thread frame when iterated or awaited.

If a frame object associated with a generator outlives the generator, then the embedded _PyInterpreterFrame is copied into the frame object.

All the above applies to coroutines and async generators as well.

Field names

Many of the fields in _PyInterpreterFrame were copied from the 3.10 PyFrameObject. Thus, some of the field names may be a bit misleading.

For example the f_globals field has a f_ prefix implying it belongs to the PyFrameObject struct, although it belongs to the _PyInterpreterFrame struct. We may rationalize this naming scheme for 3.12.

Shim frames

On entry to _PyEval_EvalFrameDefault() a shim _PyInterpreterFrame is pushed. This frame is stored on the C stack, and popped when _PyEval_EvalFrameDefault() returns. This extra frame is inserted so that RETURN_VALUE, YIELD_VALUE, and RETURN_GENERATOR do not need to check whether the current frame is the entry frame. The shim frame points to a special code object containing the INTERPRETER_EXIT instruction which cleans up the shim frame and returns.

The Instruction Pointer

_PyInterpreterFrame has two fields which are used to maintain the instruction pointer: instr_ptr and return_offset.

When a frame is executing, instr_ptr points to the instruction currently being executed. In a suspended frame, it points to the instruction that would execute if the frame were to resume. After frame.f_lineno is set, instr_ptr points to the next instruction to be executed. During a call to a python function, instr_ptr points to the call instruction, because this is what we would expect to see in an exception traceback.

The return_offset field determines where a RETURN should go in the caller, relative to instr_ptr. It is only meaningful to the callee, so it needs to be set in any instruction that implements a call (to a Python function), including CALL, SEND and BINARY_SUBSCR_GETITEM, among others. If there is no callee, then return_offset is meaningless. It is necessary to have a separate field for the return offset because (1) if we apply this offset to instr_ptr while executing the RETURN, this is too early and would lose us information about the previous instruction which we could need for introspecting and debugging. (2) SEND needs to pass two offsets to the generator: one for RETURN and one for YIELD. It uses the oparg for one, and the return_offset for the other.