Up-to-date TODO list for per-interpreter GIL.

[ericsnowcurrently]

I'm moving the TODO, etc. to https://github.com/ericsnowcurrently/multi-core-python/wiki/0-The-Plan

---

Here's a rough TODO list, sort of in order (a lot can be done in parallel).

Note that most of these items fall out of Tools/c-analyzer/cpython/globals-to-fix.tsv. (Run ./python Tools/c-analyzer/table-file.py Tools/c-analyzer/cpython/globals-to-fix.tsv to get counts.)

(python/cpython#84692)

Primary Tasks

• analysis (summary)
• proposal: "A Per-Interpreter GIL"
  • initial python-dev discussion
  • clarify next steps with the steering council
  • write the proposal (PEP 684)
  • accepted
• tooling (to identify global variables)
  • analyze requirements & make design
  • implement the tool (Tools/c-analyzer)
  • re-enable CI check to prevent addition of more globals (may need fixes)
• resolve main blockers
  • objects exposed in public C-API (solve with immortal objects)
  • shared vs. per-interpreter allocators (solve with mimalloc)
  • impact on extension maintainers (solve with docs, work with numpy)
• implementation
  • prerequisites
    • simplify runtime init/fini code (PEP 432, PEP 587)
    • introduce _PyRuntimeState
  • consolidate globals (#81057)
    • ... (move a bunch of stuff into _PyRuntimeState)
    • ... (statically init a bunch of global objects)
    • (maybe?) move some state from PyInterpreterState to _PyRuntimeState
    • drop _Py_IDENTIFIER() in extensions (or modify it to use a statically initialized str?)
      • Modules/_asynciomodule.c
      • Modules/_ctypes/callbacks.c
      • Modules/_ctypes/callproc.c
      • Modules/_ctypes/_ctypes.c
      • Modules/_ctypes/stgdict.c
      • Modules/_cursesmodule.c
      • Modules/_datetimemodule.c
      • Modules/_elementtree.c
      • Modules/_json.c
      • Modules/ossaudiodev.c
      • Modules/_testcapimodule.c
      • Programs/_testembed.c
    • statically init latin-1 chars
    • statically init currently cached str objects
    • deal with PyArg_Parser (#90928, #82322, faster-cpython#281)
    • pull remaining objects into _PyRuntimeState
    • pull remaining state into _PyRuntimeState
  • prep for per-interpreter state
    • solve per-interpreter state for core/builtin static types (e.g. dynamic lookup for some attrs)
    • solve for extensions
    • make some vars atomic?
    • convert stdlib extensions to PEP 630 (see bpo-40077, PEP 687 & the SC discussion)
  • add new C-API and implement runtime machinery
    • ...
  • move state (incl. GIL) to PyInterpreterState
    • enforce restrictions on importing extensions...
    • finally: move some state/objects to PyInterpreterState
    • 💥 move the GIL to PyInterpreterState 💥

Solutions for Blockers

Immortal Objects:

• analysis, design, and initial discussion (bpo-40255)
• implement basic solution (gh-19474)
• run benchmarks (incl. with all objects immortalized at end of runtime init)
• (maybe?) prepare a contingency plan for if we don't end up with immortal objects
• write a proposal (PEP 683)
• present at the language summit
• enumerate other strategies for getting back to performance neutral
• implement them
• PEP accepted
• merge implementation

mimalloc (thread-safe, performant allocator):

• get mimalloc working as CPython's allocator (Christian: ...)
• add static analyzer support (ASAN, valgrind) (Daan)
• resolve problem with leaking large blocks (Daan/Christian)
• resolve PGO issues on Windows (Eric: contact MSVC team)
• merge the implementation
• sort out GIL guarantees for allocator API

Reduce Impact on Extension Maintainers:

• add docs (i.e. copy PEP 630)
• (probably) add some new C-API to solve pain points
• work with numpy, specifically, to add support for multiple interpreters
• work with cython

Secondary Tasks

• finish "c-analyzer.py data count"
• fix "c-analyzer.py data check" & add to CI
• statically init other immutables
• ...

Other Related Items

• PEP 554 (stdlib module to expose multiple interpreters)
• help numpy, cython, etc.
• PEP: add Public C-API as replacement for _Py_IDENTIFIER() (@zooba's staticstr?)
• "subinterpreters" project: https://github.com/orgs/python/projects/3/views/1

[ericsnowcurrently]

FYI, here's a recent breakdown of the remaining globals:

• core/builtins (651 total)
  • 298 static types (incl. 130 exception types)
  • 179 cached objects (e.g. str instances)
  • 57 other objects
  • 117 non-object (incl. 60 state)
• extension modules (605 total)
  • 78 static types
  • 317 cached objects (e.g. str instances)
  • 145 other objects
  • 65 non-object (incl. 29 state)

(the detailed breakdown)

./python Tools/c-analyzer/table-file.py Tools/c-analyzer/cpython/globals-to-fix.tsv
 --------------------------------------------------
  (365)       global objects to fix in core code
         119      static types
         130      builtin exception types
           9      singletons
   (98)           cached - initialized once
          39          manually cached PyUnicodeObject
          54          _PyArg_Parser (holds tuple of strings)
           1          holds strings
           4          other
    (9)           other
           7          initialized once
           2          state
 --------------------------------------------------
   (94)       global non-objects to fix in core code
   (29)           initialized/set once
           1          pre-allocated buffer
          13          during init
          15          lazy
   (20)           unlikely to change after init (or main thread)
          11          through C-API
           4          REPL
           5          handling C argv
   (45)           state
          15          allocator
           2          pre-allocated memory
           2          local buffer
           2          linked list
          24          other
 --------------------------------------------------
  (169)       global objects to fix in builtin modules
          22      modules
          49      static types
    (5)           non-static types - initialized once
           3          structseq types
           2          exception types
   (81)           cached - initialized once
          29          manually cached PyUnicodeOjbect
          52          _PyArg_Parser
   (12)           other
           6          initialized once
           6          state
 --------------------------------------------------
   (23)       global non-objects to fix in builtin modules
           8      initialized once
          15      state
 --------------------------------------------------
  (540)       global objects to fix in extension modules
          65      modules
          78      static types
   (34)           non-static types - initialized once
           3          structseq types
           6          heap types
          25          exception types
  (317)           cached - initialized once
          18          _Py_IDENTIFIER (global)
         149          _Py_IDENTIFIER (local)
           1          _Py_static_string
          12          manually cached PyUnicodeOjbect
         124          _PyArg_Parser
          13          other - during module init
   (46)           other
          31          initialized once
          15          state
 --------------------------------------------------
   (65)       global non-objects to fix in extension modules
   (36)           initialized once
           2          pre-allocated buffer
          34          other
          29      state
 --------------------------------------------------
(total: 1256)

[kumaraditya303]

statically init latin-1 chars is done https://bugs.python.org/issue46881