14 KiB
A higher level definition of the bytecode interpreter
Abstract
The CPython interpreter is defined in C, meaning that the semantics of the bytecode instructions, the dispatching mechanism, error handling, and tracing and instrumentation are all intermixed.
This document proposes defining a custom C-like DSL for defining the instruction semantics and tools for generating the code deriving from the instruction definitions.
These tools would be used to:
- Generate the main interpreter (done)
- Generate the tier 2 interpreter
- Generate documentation for instructions
- Generate metadata about instructions, such as stack use (done).
- Generate the tier 2 optimizer's abstract interpreter.
Having a single definition file ensures that there is a single source of truth for bytecode semantics.
Other tools that operate on bytecodes, like frame.setlineno
and the dis module, will be derived from the common semantic
definition, reducing errors.
Motivation
The bytecode interpreter of CPython has traditionally been defined as standard C code, but with a lot of macros. The presence of these macros and the nature of bytecode interpreters means that the interpreter is effectively defined in a domain specific language (DSL).
Rather than using an ad-hoc DSL embedded in the C code for the interpreter, a custom DSL should be defined and the semantics of the bytecode instructions, and the instructions defined in that DSL.
Generating the interpreter decouples low-level details of dispatching and error handling from the semantics of the instructions, resulting in more maintainable code and a potentially faster interpreter.
It also provides the ability to create and check optimizers and optimization passes from the semantic definition, reducing errors.
Rationale
As we improve the performance of CPython, we need to optimize larger regions of code, use more complex optimizations and, ultimately, translate to machine code.
All of these steps introduce the possibility of more bugs, and require more code to be written. One way to mitigate this is through the use of code generators. Code generators decouple the debugging of the code (the generator) from checking the correctness (the DSL input).
For example, we are likely to want a new interpreter for the tier 2 optimizer to be added in 3.12. That interpreter will have a different API, a different set of instructions and potentially different dispatching mechanism. But the instructions it will interpret will be built from the same building blocks as the instructions for the tier 1 (PEP 659) interpreter.
Rewriting all the instructions is tedious and error-prone, and changing the instructions is a maintenance headache as both versions need to be kept in sync.
By using a code generator and using a common source for the instructions, or parts of instructions, we can reduce the potential for errors considerably.
Specification
This specification is a work in progress. We update it as the need arises.
Syntax
Each op definition has a kind, a name, a stack and instruction stream effect, and a piece of C code describing its semantics::
file:
(definition | family | pseudo)+
definition:
"inst" "(" NAME ["," stack_effect] ")" "{" C-code "}"
|
"op" "(" NAME "," stack_effect ")" "{" C-code "}"
|
"macro" "(" NAME ")" "=" uop ("+" uop)* ";"
stack_effect:
"(" [inputs] "--" [outputs] ")"
inputs:
input ("," input)*
outputs:
output ("," output)*
input:
object | stream | array
output:
object | array
uop:
NAME | stream
object:
NAME [":" type] [ "if" "(" C-expression ")" ]
type:
NAME ["*"]
stream:
NAME "/" size
size:
INTEGER
array:
object "[" C-expression "]"
family:
"family" "(" NAME ")" = "{" NAME ("," NAME)+ [","] "}" ";"
pseudo:
"pseudo" "(" NAME ")" = "{" NAME ("," NAME)+ [","] "}" ";"
The following definitions may occur:
inst: A normal instruction, as previously defined byTARGET(NAME)inceval.c.op: A part instruction from which macros can be constructed.macro: A bytecode instruction constructed from ops and cache effects.
NAME can be any ASCII identifier that is a C identifier and not a C or Python keyword.
foo_1 is legal. $ is not legal, nor is struct or class.
The optional type in an object is the C type. It defaults to PyObject *.
The objects before the "--" are the objects on top of the stack at the start of
the instruction. Those after the "--" are the objects on top of the stack at the
end of the instruction.
An inst without stack_effect is a transitional form to allow the original C code
definitions to be copied. It lacks information to generate anything other than the
interpreter, but is useful for initial porting of code.
Stack effect names may be unused, indicating the space is to be reserved
but no use of it will be made in the instruction definition.
This is useful to ensure that all instructions in a family have the same
stack effect.
The number in a stream define how many codeunits are consumed from the
instruction stream. It returns a 16, 32 or 64 bit value.
If the name is unused the size can be any value and that many codeunits
will be skipped in the instruction stream.
By convention cache effects (stream) must precede the input effects.
The name oparg is pre-defined as a 32 bit value fetched from the instruction stream.
Special instruction annotations
Instruction headers may be prefixed by one or more annotations. The non-exhaustive list of annotations and their meanings are as follows:
override. For external use by other interpreter definitions to override the current instruction definition.pure. This instruction has no side effects.
Special functions/macros
The C code may include special functions that are understood by the tools as part of the DSL.
Those functions include:
DEOPT_IF(cond, instruction). Deoptimize ifcondis met.ERROR_IF(cond, label). Jump to error handler atlabelifcondis true.DECREF_INPUTS(). GeneratePy_DECREF()calls for the input stack effects.SYNC_SP(). Synchronizes the physical stack pointer with the stack effects.
Note that the use of DECREF_INPUTS() is optional -- manual calls
to Py_DECREF() or other approaches are also acceptable
(e.g. calling an API that "steals" a reference).
Variables can either be defined in the input, output, or in the C code.
Variables defined in the input may not be assigned in the C code.
If an ERROR_IF occurs, all values will be removed from the stack;
they must already be DECREF'ed by the code block.
If a DEOPT_IF occurs, no values will be removed from the stack or
the instruction stream; no values must have been DECREF'ed or created.
These requirements result in the following constraints on the use of
DEOPT_IF and ERROR_IF in any instruction's code block:
- Until the last
DEOPT_IF, no objects may be allocated,INCREFed, orDECREFed. - Before the first
ERROR_IF, all input values must beDECREFed, and no objects may be allocated orINCREFed, with the exception of attempting to create an object and checking for success usingERROR_IF(result == NULL, label). (TODO: Unclear what to do with intermediate results.) - No
DEOPT_IFmay follow anERROR_IFin the same block.
(There is some wiggle room: these rules apply to dynamic code paths, not to static occurrences in the source code.)
If code detects an error condition before the first DECREF of an input,
two idioms are valid:
- Use
goto error. - Use a block containing the appropriate
DECREFcalls ending inERROR_IF(true, error).
An example of the latter would be:
res = PyObject_Add(left, right);
if (res == NULL) {
DECREF_INPUTS();
ERROR_IF(true, error);
}
Semantics
The underlying execution model is a stack machine. Operations pop values from the stack, and push values to the stack. They also can look at, and consume, values from the instruction stream.
All members of a family (which represents a specializable instruction and its specializations) must have the same stack and instruction stream effect.
The same is true for all members of a pseudo instruction (which is mapped by the bytecode compiler to one of its members).
Examples
(Another source of examples can be found in the tests.)
Some examples:
Output stack effect
inst ( LOAD_FAST, (-- value) ) {
value = frame->f_localsplus[oparg];
Py_INCREF(value);
}
This would generate:
TARGET(LOAD_FAST) {
PyObject *value;
value = frame->f_localsplus[oparg];
Py_INCREF(value);
PUSH(value);
DISPATCH();
}
Input stack effect
inst ( STORE_FAST, (value --) ) {
SETLOCAL(oparg, value);
}
This would generate:
TARGET(STORE_FAST) {
PyObject *value = PEEK(1);
SETLOCAL(oparg, value);
STACK_SHRINK(1);
DISPATCH();
}
Input stack effect and cache effect
op ( CHECK_OBJECT_TYPE, (owner, type_version/2 -- owner) ) {
PyTypeObject *tp = Py_TYPE(owner);
assert(type_version != 0);
DEOPT_IF(tp->tp_version_tag != type_version);
}
This might become (if it was an instruction):
TARGET(CHECK_OBJECT_TYPE) {
PyObject *owner = PEEK(1);
uint32 type_version = read32(next_instr);
PyTypeObject *tp = Py_TYPE(owner);
assert(type_version != 0);
DEOPT_IF(tp->tp_version_tag != type_version);
next_instr += 2;
DISPATCH();
}
More examples
For explanations see "Generating the interpreter" below.)
op ( CHECK_HAS_INSTANCE_VALUES, (owner -- owner) ) {
PyDictOrValues dorv = *_PyObject_DictOrValuesPointer(owner);
DEOPT_IF(!_PyDictOrValues_IsValues(dorv));
}
op ( LOAD_INSTANCE_VALUE, (owner, index/1 -- null if (oparg & 1), res) ) {
res = _PyDictOrValues_GetValues(dorv)->values[index];
DEOPT_IF(res == NULL);
Py_INCREF(res);
null = NULL;
Py_DECREF(owner);
}
macro ( LOAD_ATTR_INSTANCE_VALUE ) =
counter/1 + CHECK_OBJECT_TYPE + CHECK_HAS_INSTANCE_VALUES +
LOAD_INSTANCE_VALUE + unused/4 ;
op ( LOAD_SLOT, (owner, index/1 -- null if (oparg & 1), res) ) {
char *addr = (char *)owner + index;
res = *(PyObject **)addr;
DEOPT_IF(res == NULL);
Py_INCREF(res);
null = NULL;
Py_DECREF(owner);
}
macro ( LOAD_ATTR_SLOT ) = counter/1 + CHECK_OBJECT_TYPE + LOAD_SLOT + unused/4;
inst ( BUILD_TUPLE, (items[oparg] -- tuple) ) {
tuple = _PyTuple_FromArraySteal(items, oparg);
ERROR_IF(tuple == NULL, error);
}
inst ( PRINT_EXPR ) {
PyObject *value = POP();
PyObject *hook = _PySys_GetAttr(tstate, &_Py_ID(displayhook));
PyObject *res;
if (hook == NULL) {
_PyErr_SetString(tstate, PyExc_RuntimeError,
"lost sys.displayhook");
Py_DECREF(value);
goto error;
}
res = PyObject_CallOneArg(hook, value);
Py_DECREF(value);
ERROR_IF(res == NULL);
Py_DECREF(res);
}
Defining an instruction family
A family maps a specializable instruction to its specializations.
Example: These opcodes all share the same instruction format):
family(load_attr) = { LOAD_ATTR, LOAD_ATTR_INSTANCE_VALUE, LOAD_SLOT };
Defining a pseudo instruction
A pseudo instruction is used by the bytecode compiler to represent a set of possible concrete instructions.
Example: JUMP may expand to JUMP_FORWARD or JUMP_BACKWARD:
pseudo(JUMP) = { JUMP_FORWARD, JUMP_BACKWARD };
Generating the interpreter
The generated C code for a single instruction includes a preamble and dispatch at the end which can be easily inserted. What is more complex is ensuring the correct stack effects and not generating excess pops and pushes.
For example, in CHECK_HAS_INSTANCE_VALUES, owner occurs in the input, so it cannot be
redefined. Thus it doesn't need to written and can be read without adjusting the stack pointer.
The C code generated for CHECK_HAS_INSTANCE_VALUES would look something like:
{
PyObject *owner = stack_pointer[-1];
PyDictOrValues dorv = *_PyObject_DictOrValuesPointer(owner);
DEOPT_IF(!_PyDictOrValues_IsValues(dorv));
}
When combining ops together to form instructions, temporary values should be used,
rather than popping and pushing, such that LOAD_ATTR_SLOT would look something like:
case LOAD_ATTR_SLOT: {
PyObject *s1 = stack_pointer[-1];
/* CHECK_OBJECT_TYPE */
{
PyObject *owner = s1;
uint32_t type_version = read32(next_instr + 1);
PyTypeObject *tp = Py_TYPE(owner);
assert(type_version != 0);
if (tp->tp_version_tag != type_version) goto deopt;
}
/* LOAD_SLOT */
{
PyObject *owner = s1;
uint16_t index = *(next_instr + 1 + 2);
char *addr = (char *)owner + index;
PyObject *null;
PyObject *res = *(PyObject **)addr;
if (res == NULL) goto deopt;
Py_INCREF(res);
null = NULL;
Py_DECREF(owner);
if (oparg & 1) {
stack_pointer[0] = null;
stack_pointer += 1;
}
s1 = res;
}
next_instr += (1 + 1 + 2 + 1 + 4);
stack_pointer[-1] = s1;
DISPATCH();
}
Other tools
From the instruction definitions we can generate the stack marking code used in frame.set_lineno(),
and the tables for use by disassemblers.