gh-141504: Factor out tracing and optimization heuristics into a single object by corona10 · Pull Request #143381 · python/cpython

corona10 · 2026-01-03T07:02:03Z

Issue: Factor out tracing and optimization heuristics into a single "policy" object. #141504

corona10 · 2026-01-03T07:42:09Z

Python/pystate.c

+    _tstate->jit_metrics.side_exit_initial_value = SIDE_EXIT_INITIAL_VALUE;
+    _tstate->jit_metrics.side_exit_initial_backoff = SIDE_EXIT_INITIAL_BACKOFF;
+
+    char *env = Py_GETENV("PYTHON_JIT_JUMP_BACKWARD_INITIAL_VALUE");


We may need to extract utility function to handle those things.

Fidget-Spinner · 2026-01-03T11:22:39Z

Include/internal/pycore_tstate.h

+typedef struct _PyJitMetrics {
+    uint16_t jump_backward_initial_value;
+    uint16_t jump_backward_initial_backoff;
+    uint16_t side_exit_initial_value;
+    uint16_t side_exit_initial_backoff;
+} _PyJitMetrics;
+


Can you rename this to _PyJitPolicy pleases?

Also please then wrap this into an overall struct called _PyPolicies. The reason is that we might want to add more structs later, like for the interpreter.

So it should be

Suggested change

typedef struct _PyJitMetrics {

uint16_t jump_backward_initial_value;

uint16_t jump_backward_initial_backoff;

uint16_t side_exit_initial_value;

uint16_t side_exit_initial_backoff;

} _PyJitMetrics;

typedef struct _PyJitPolicy {

uint16_t jump_backward_initial_value;

uint16_t jump_backward_initial_backoff;

uint16_t side_exit_initial_value;

uint16_t side_exit_initial_backoff;

} _PyJitPolicy;

typedef struct _PyPolicy {

_PyJitPolicy jit;

} _PyPolicy;

corona10 · 2026-01-03T12:11:22Z

Python/specialize.c

+        PyThreadState *tstate = _PyThreadState_GET();
+        _PyThreadStateImpl *tstate_impl = (_PyThreadStateImpl *)tstate;
+        jump_counter = initial_jump_backoff_counter(
+            tstate_impl->policy.jit.jump_backward_initial_value,


@Fidget-Spinner
The crash is directly related to this line.
But ENABLE_SPECIALIZATION_FT does not mean TIER2 is enabled.

Do we need to move jump_backward_initial_value and jump_backward_initial_backoff to policy.interpreter?

JUMP_BACKWARD_INITIAL_VALUE is only used by the JIT. On normal build/FT build, the CACHE of JUMP_BACKWARD is unused. So we can just keep it for the JIT.

Hmm I realised the CI is now failing because FT doesn't build with JIT.

In that case, yes please move this to an interpreter policy.

Fidget-Spinner · 2026-01-03T13:53:52Z

Include/internal/pycore_tstate.h

 #endif
 #if _Py_TIER2
    _PyJitTracerState jit_tracer_state;
+    _PyPolicy policy;


This needs to be moved out of the ifdef. CI is failing.

Fidget-Spinner

Some comments to cleanup the code. After this I will approve. Thanks!

Fidget-Spinner · 2026-01-03T14:26:55Z

Include/internal/pycore_backoff.h

 #define JUMP_BACKWARD_INITIAL_BACKOFF 6
 static inline _Py_BackoffCounter
-initial_jump_backoff_counter(void)
+initial_jump_backoff_counter(uint16_t initial_value, uint16_t initial_backoff)


This should take a single value _tstate/tstate and grab the jump backward values from it.

Fidget-Spinner · 2026-01-03T14:26:59Z

Include/internal/pycore_backoff.h


 static inline _Py_BackoffCounter
-initial_temperature_backoff_counter(void)
+initial_temperature_backoff_counter(uint16_t initial_value, uint16_t initial_backoff)


Fidget-Spinner · 2026-01-03T14:27:25Z

Python/pystate.c

 }

+static inline void
+init_jit_metric(uint16_t *target, const char *env_name, uint16_t default_value,


Rename to init_policy

Fidget-Spinner · 2026-01-03T14:27:35Z

Python/pystate.c

    _tstate->asyncio_running_task = NULL;
-
+    // Initialize interpreter policy from environment variables
+    init_jit_metric(&_tstate->policy.interp.jump_backward_initial_value,


rename here as well

Fidget-Spinner

Great. Thanks!

Fidget-Spinner · 2026-01-03T15:25:42Z

Note: the names are a little lengthy. But considering this is an internal CPython only env var used for testing, we can determine proper names later when/if we decide to expose them.

markshannon · 2026-01-05T14:07:41Z

I have some questions.

Why set thresholds per thread, rather than per interpreter. Why would we want per-thread values?
Why set and store the threshold and backoff separately rather than as a single backoff counter?

I've added a new comment on the issue, that will hopefully clarify what I was after.

corona10 added 3 commits January 3, 2026 15:30

pythongh-141504: Configurable JIT optimization thresholds

0ebb67a

Add NEWS.d

2b1cbbc

nit

7fce731

corona10 requested review from Fidget-Spinner, ZeroIntensity, ericsnowcurrently and markshannon as code owners January 3, 2026 07:02

bedevere-app bot added the awaiting core review label Jan 3, 2026

bedevere-app bot mentioned this pull request Jan 3, 2026

Factor out tracing and optimization heuristics into a single "policy" object. #141504

Open

corona10 changed the title ~~gh-141504: Factor out tracing and optimization heuristics into a single object~~ [WIP] gh-141504: Factor out tracing and optimization heuristics into a single object Jan 3, 2026

corona10 marked this pull request as draft January 3, 2026 07:23

bedevere-app bot removed the awaiting core review label Jan 3, 2026

corona10 commented Jan 3, 2026

View reviewed changes

fix

d01f572

corona10 changed the title ~~[WIP] gh-141504: Factor out tracing and optimization heuristics into a single object~~ gh-141504: Factor out tracing and optimization heuristics into a single object Jan 3, 2026

nit

1942f72

corona10 marked this pull request as ready for review January 3, 2026 07:59

bedevere-app bot added the awaiting core review label Jan 3, 2026

Fidget-Spinner reviewed Jan 3, 2026

View reviewed changes

corona10 added 2 commits January 3, 2026 20:55

Address code review

a0e6965

nit

196d204

corona10 commented Jan 3, 2026

View reviewed changes

Address code review

f0ccb28

corona10 requested a review from Fidget-Spinner January 3, 2026 13:38

Fidget-Spinner reviewed Jan 3, 2026

View reviewed changes

fix

b3b608b

corona10 requested a review from Fidget-Spinner January 3, 2026 14:22

Fidget-Spinner reviewed Jan 3, 2026

View reviewed changes

corona10 added 2 commits January 3, 2026 23:49

Address code review

6a32b9a

nit

f55ce78

Fidget-Spinner approved these changes Jan 3, 2026

View reviewed changes

bedevere-app bot removed the awaiting core review label Jan 3, 2026

fix

03fe066

bedevere-app bot added the awaiting merge label Jan 3, 2026

corona10 added 2 commits January 3, 2026 23:54

nit

fe0d38a

nit

304c70c

corona10 enabled auto-merge (squash) January 3, 2026 14:56

nit

e6e3202

corona10 disabled auto-merge January 3, 2026 14:58

corona10 enabled auto-merge (squash) January 3, 2026 14:58

corona10 merged commit 12283f6 into python:main Jan 3, 2026
69 of 70 checks passed

bedevere-app bot removed the awaiting merge label Jan 3, 2026

Uh oh!

Conversation

corona10 commented Jan 3, 2026 • edited by bedevere-app bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Fidget-Spinner left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Fidget-Spinner left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Fidget-Spinner commented Jan 3, 2026

Uh oh!

markshannon commented Jan 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

corona10 commented Jan 3, 2026 •

edited by bedevere-app bot

Loading

markshannon commented Jan 5, 2026 •

edited

Loading