* Transform opcodes into opnames
* Print the whole stack at each opcode, and eliminate prtrace output at each (push/pop/stackadj)
* Display info about the function at each resume_frame
Remove the Include/code.h header file. C extensions should only
include the main <Python.h> header file.
Python.h includes directly Include/cpython/code.h instead.
* Moves the bytecode to the end of the corresponding PyCodeObject, and quickens it in-place.
* Removes the almost-always-unused co_varnames, co_freevars, and co_cellvars member caches
* _PyOpcode_Deopt is a new mapping from all opcodes to their un-quickened forms.
* _PyOpcode_InlineCacheEntries is renamed to _PyOpcode_Caches
* _Py_IncrementCountAndMaybeQuicken is renamed to _PyCode_Warmup
* _Py_Quicken is renamed to _PyCode_Quicken
* _co_quickened is renamed to _co_code_adaptive (and is now a read-only memoryview).
* Do not emit unused nonzero opargs anymore in the compiler.
Remove the private undocumented function
_PyEval_GetCoroutineOriginTrackingDepth() from the C API. Call the
public sys.get_coroutine_origin_tracking_depth() function instead.
Change the internal function
_PyEval_SetCoroutineOriginTrackingDepth():
* Remove the 'tstate' parameter;
* Add return value and raises an exception if depth is negative;
* No longer export the function: call the public
sys.set_coroutine_origin_tracking_depth() function instead.
Uniformize also function declarations in pycore_ceval.h.
When an exception is created in a nested call to PyObject_GetAttr, any
external calls will override the context information of the
AttributeError that we have already placed in the most internal call.
This will cause the suggestions we create to nor work properly as the
attribute name and object that we will be using are the incorrect ones.
To avoid this, we need to check first if these attributes are already
set and bail out if that's the case.
Rename also struct _cframe to struct _PyCFrame.
Add a comment suggesting using public functions rather than using
directly the private _PyCFrame structure.
We're no longer using _Py_IDENTIFIER() (or _Py_static_string()) in any core CPython code. It is still used in a number of non-builtin stdlib modules.
The replacement is: PyUnicodeObject (not pointer) fields under _PyRuntimeState, statically initialized as part of _PyRuntime. A new _Py_GET_GLOBAL_IDENTIFIER() macro facilitates lookup of the fields (along with _Py_GET_GLOBAL_STRING() for non-identifier strings).
https://bugs.python.org/issue46541#msg411799 explains the rationale for this change.
The core of the change is in:
* (new) Include/internal/pycore_global_strings.h - the declarations for the global strings, along with the macros
* Include/internal/pycore_runtime_init.h - added the static initializers for the global strings
* Include/internal/pycore_global_objects.h - where the struct in pycore_global_strings.h is hooked into _PyRuntimeState
* Tools/scripts/generate_global_objects.py - added generation of the global string declarations and static initializers
I've also added a --check flag to generate_global_objects.py (along with make check-global-objects) to check for unused global strings. That check is added to the PR CI config.
The remainder of this change updates the core code to use _Py_GET_GLOBAL_IDENTIFIER() instead of _Py_IDENTIFIER() and the related _Py*Id functions (likewise for _Py_GET_GLOBAL_STRING() instead of _Py_static_string()). This includes adding a few functions where there wasn't already an alternative to _Py*Id(), replacing the _Py_Identifier * parameter with PyObject *.
The following are not changed (yet):
* stop using _Py_IDENTIFIER() in the stdlib modules
* (maybe) get rid of _Py_IDENTIFIER(), etc. entirely -- this may not be doable as at least one package on PyPI using this (private) API
* (maybe) intern the strings during runtime init
https://bugs.python.org/issue46541
* Add PRECALL_FUNCTION opcode.
* Move 'call shape' varaibles into struct.
* Replace CALL_NO_KW and CALL_KW with KW_NAMES and CALL instructions.
* Specialize for builtin methods taking using the METH_FASTCALL | METH_KEYWORDS protocol.
* Allow kwnames for specialized calls to builtin types.
* Specialize calls to tuple(arg) and str(arg).
* Add RETURN_GENERATOR and JUMP_NO_INTERRUPT opcodes.
* Trim frame and generator by word each.
* Minor refactor of frame.c
* Update test.test_sys to account for smaller frames.
* Treat generator functions as normal functions when evaluating and specializing.
Previously, the main interpreter was allocated on the heap during runtime initialization. Here we instead embed it into _PyRuntimeState, which means it is statically allocated as part of the _PyRuntime global. The same goes for the initial thread state (of each interpreter, including the main one). Consequently there are fewer allocations during runtime/interpreter init, fewer possible failures, and better memory locality.
FYI, this also helps efforts to consolidate globals, which in turns helps work on subinterpreter isolation.
https://bugs.python.org/issue45953
* Do not PUSH/POP traceback or type to the stack as part of exc_info
* Remove exc_traceback and exc_type from _PyErr_StackItem
* Add to what's new, because this change breaks things like Cython
This parallels _PyRuntimeState.interpreters. Doing this helps make it more clear what part of PyInterpreterState relates to its threads.
https://bugs.python.org/issue46008
This falls into the category of keep-allocation-and-initialization separate. It also allows us to use _PyEval_InitState() safely in functions that return void.
https://bugs.python.org/issue46008
* Make generator, coroutine and async gen structs all the same size.
* Store interpreter frame in generator (and coroutine). Reduces the number of allocations neeeded for a generator from two to one.
* Split exit paths into exceptional and non-exceptional.
* Move exit tracing code to individual bytecodes.
* Wrap all trace entry and exit events in macros to make them clearer and easier to enhance.
* Move return sequence into RETURN_VALUE, YIELD_VALUE and YIELD_FROM. Distinguish between normal trace events and dtrace events.
* Make internal APIs that take PyFrameConstructor take a PyFunctionObject instead.
* Add reference to function to frame, borrow references to builtins and globals.
* Add COPY_FREE_VARS instruction to allow specialization of calls to inner functions.
* Avoid making C calls for most calls to Python functions.
* Change initialize_locals(steal=true) and _PyTuple_FromArraySteal to consume the argument references regardless of whether they succeed or fail.
Add PyThreadState_EnterTracing() and PyThreadState_LeaveTracing()
functions to the limited C API to suspend and resume tracing and
profiling.
Add an unit test on the PyThreadState C API to _testcapi.
Add also internal _PyThreadState_DisableTracing() and
_PyThreadState_ResetTracing().
* Never change types' cached keys. It could invalidate inline attribute objects.
* Lazily create object dictionaries.
* Update specialization of LOAD/STORE_ATTR.
* Don't update shared keys version for deletion of value.
* Update gdb support to handle instance values.
* Rename SPLIT_KEYS opcodes to INSTANCE_VALUE.
Redefining the PyThreadState_GET() macro in pycore_pystate.h is
useless since it doesn't affect files not including it. Either use
_PyThreadState_GET() directly, or don't use pycore_pystate.h internal
C API. For example, the _testcapi extension don't use the internal C
API, but use the public PyThreadState_Get() function instead.
Replace PyThreadState_Get() with _PyThreadState_GET(). The
_PyThreadState_GET() macro is more efficient than PyThreadState_Get()
and PyThreadState_GET() function calls which call fail with a fatal
Python error.
posixmodule.c and _ctypes extension now include <windows.h> before
pycore header files (like pycore_call.h).
_PyTraceback_Add() now uses _PyErr_Fetch()/_PyErr_Restore() instead
of PyErr_Fetch()/PyErr_Restore().
The _decimal and _xxsubinterpreters extensions are now built with the
Py_BUILD_CORE_MODULE macro defined to get access to the internal C
API.
* Move _PyObject_CallNoArgs() to pycore_call.h (internal C API).
* _ssl, _sqlite and _testcapi extensions now call the public
PyObject_CallNoArgs() function, rather than _PyObject_CallNoArgs().
* _lsprof extension is now built with Py_BUILD_CORE_MODULE macro
defined to get access to internal _PyObject_CallNoArgs().
Fix typo in the private _PyObject_CallNoArg() function name: rename
it to _PyObject_CallNoArgs() to be consistent with the public
function PyObject_CallNoArgs().
Places the locals between the specials and stack. This is the more "natural" layout for a C struct, makes the code simpler and gives a slight speedup (~1%)
* Refactor dispatch logic to make flow of control clearer. Moves lltrace and dxprofile instrumentation into DISPATCH macro.
* Remove switch in interpreter loop when using computed gotos. There is no need for two nearly-duplicate dispatch tables.
* Generalize cache names for LOAD_ATTR to allow store and delete specializations.
* Factor out specialization of attribute dictionary access.
* Specialize STORE_ATTR.
* Convert "specials" array to InterpreterFrame struct, adding f_lasti, f_state and other non-debug FrameObject fields to it.
* Refactor, calls pushing the call to the interpreter upward toward _PyEval_Vector.
* Compute f_back when on thread stack, only filling in value when frame object outlives stack invocation.
* Move ownership of InterpreterFrame in generator from frame object to generator object.
* Do not create frame objects for Python calls.
* Do not create frame objects for generators.
A TypeError is now raised instead of an AttributeError in
"with" and "async with" statements for objects which do not
support the context manager or asynchronous context manager
protocols correspondingly.
* Specialize obj.__class__ with LOAD_ATTR_SLOT
* Specialize instance attribute lookup with attribute on class, provided attribute on class is not an overriding descriptor.
* Add stat for how many times the unquickened instruction has executed.
Currently, if an arg value escapes (into the closure for an inner function) we end up allocating two indices in the fast locals even though only one gets used. Additionally, using the lower index would be better in some cases, such as with no-arg `super()`. To address this, we update the compiler to fix the offsets so each variable only gets one "fast local". As a consequence, now some cell offsets are interspersed with the locals (only when an arg escapes to an inner function).
https://bugs.python.org/issue43693
* Specialize LOAD_ATTR with LOAD_ATTR_SLOT and LOAD_ATTR_SPLIT_KEYS
* Move dict-common.h to internal/pycore_dict.h
* Add LOAD_ATTR_WITH_HINT specialized opcode.
* Quicken in function if loopy
* Specialize LOAD_ATTR for module attributes.
* Add specialization stats
This was reverted in GH-26596 (commit 6d518bb) due to some bad memory accesses.
* Add the MAKE_CELL opcode. (gh-26396)
The memory accesses have been fixed.
https://bugs.python.org/issue43693
This moves logic out of the frame initialization code and into the compiler and eval loop. Doing so simplifies the runtime code and allows us to optimize it better.
https://bugs.python.org/issue43693
These were reverted in gh-26530 (commit 17c4edc) due to refleaks.
* 2c1e258 - Compute deref offsets in compiler (gh-25152)
* b2bf2bc - Add new internal code objects fields: co_fastlocalnames and co_fastlocalkinds. (gh-26388)
This change fixes the refleaks.
https://bugs.python.org/issue43693
* Add co_firstinstr field to code object.
* Implement barebones quickening.
* Use non-quickened bytecode when tracing.
* Add NEWS item
* Add new file to Windows build.
* Don't specialize instructions with EXTENDED_ARG.
* Revert "bpo-43693: Compute deref offsets in compiler (gh-25152)"
This reverts commit b2bf2bc1ec.
* Revert "bpo-43693: Add new internal code objects fields: co_fastlocalnames and co_fastlocalkinds. (gh-26388)"
This reverts commit 2c1e2583fd.
These two commits are breaking the refleak buildbots.
Merges locals and cells into a single array.
Saves a pointer in the interpreter and means that we don't need the LOAD_CLOSURE opcode any more
https://bugs.python.org/issue43693
A number of places in the code base (notably ceval.c and frameobject.c) rely on mapping variable names to indices in the frame "locals plus" array (AKA fast locals), and thus opargs. Currently the compiler indirectly encodes that information on the code object as the tuples co_varnames, co_cellvars, and co_freevars. At runtime the dependent code must calculate the proper mapping from those, which isn't ideal and impacts performance-sensitive sections. This is something we can easily address in the compiler instead.
This change addresses the situation by replacing internal use of co_varnames, etc. with a single combined tuple of names in locals-plus order, along with a minimal array mapping each to its kind (local vs. cell vs. free). These two new PyCodeObject fields, co_fastlocalnames and co_fastllocalkinds, are not exposed to Python code for now, but co_varnames, etc. are still available with the same values as before (though computed lazily).
Aside from the (mild) performance impact, there are a number of other benefits:
* there's now a clear, direct relationship between locals-plus and variables
* code that relies on the locals-plus-to-name mapping is simpler
* marshaled code objects are smaller and serialize/de-serialize faster
Also note that we can take this approach further by expanding the possible values in co_fastlocalkinds to include specific argument types (e.g. positional-only, kwargs). Doing so would allow further speed-ups in _PyEval_MakeFrameVector(), which is where args get unpacked into the locals-plus array. It would also allow us to shrink marshaled code objects even further.
https://bugs.python.org/issue43693
* Move up the comment about fields using in hashing/comparision.
* Group the fields more clearly.
* Add co_ncellvars and co_nfreevars.
* Raise ValueError if nlocals != len(varnames), rather than aborting.
* Remove 'zombie' frames. We won't need them once we are allocating fixed-size frames.
* Add co_nlocalplus field to code object to avoid recomputing size of locals + frees + cells.
* Move locals, cells and freevars out of frame object into separate memory buffer.
* Use per-threadstate allocated memory chunks for local variables.
* Move globals and builtins from frame object to per-thread stack.
* Move (slow) locals frame object to per-thread stack.
* Move internal frame functions to internal header.
"Zero cost" exception handling.
* Uses a lookup table to determine how to handle exceptions.
* Removes SETUP_FINALLY and POP_TOP block instructions, eliminating (most of) the runtime overhead of try statements.
* Reduces the size of the frame object by about 60%.
* Add Py_TPFLAGS_SEQUENCE and Py_TPFLAGS_MAPPING, add to all relevant standard builtin classes.
* Set relevant flags on collections.abc.Sequence and Mapping.
* Use flags in MATCH_SEQUENCE and MATCH_MAPPING opcodes.
* Inherit Py_TPFLAGS_SEQUENCE and Py_TPFLAGS_MAPPING.
* Add NEWS
* Remove interpreter-state map_abc and seq_abc fields.
Dennis Sweeney
Irit Katriel
jonasdlindner
Brandt Bucher
Victor Stinner
Christian Heimes
Hood Chatham
Mark Shannon
Kumar Aditya
neonene
Oleg Iarygin
Pablo Galindo Salgado
Inada Naoki
Nikita Sobolev
Eric Snow
Ken Jin
Brandt Bucher
Yury Selivanov
penguin_wwy
Dong-hee Na
Serhiy Storchaka
Gabriele N. Tornetta