8000 Provide an "aggressive" specialization/JIT mode for debugging · Issue #129386 · python/cpython · GitHub
[go: up one dir, main page]
Skip to content

Provide an "aggressive" specialization/JIT mode for debugging #129386

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
brandtbucher opened this issue Jan 28, 2025 · 4 comments
Open

Provide an "aggressive" specialization/JIT mode for debugging #129386

brandtbucher opened this issue Jan 28, 2025 · 4 comments
Assignees
@brandtbucher
Labels
interpreter-core (Objects, Python, Grammar, and Parser dirs) performance Performance or resource usage topic-JIT type-feature A feature request or enhancement

Comments

@brandtbucher
Copy link
Member
brandtbucher commented Jan 28, 2025 •
edited by bedevere-app bot
Loading

@markshannon recently pointed out that our thresholds for JIT stuff are high enough that running the test suite isn't quite the stress test that it used to be. While these thresholds will probably come down with time, it might make sense to provide a mode where specialization and the JIT happen much more aggressively, for development purposes.

Our current values are:

  • ADAPTIVE_WARMUP_VALUE: 1 (specialize the 2nd time an instruction is run)
  • ADAPTIVE_COOLDOWN_VALUE: 52 (re-specialize the 53rd time a guard fails)
  • JUMP_BACKWARD_INITIAL_VALUE: 4095 (JIT a new "root" trace the 4096th time we jump backwards)
  • SIDE_EXIT_INITIAL_VALUE: 4095 (JIT a new side-exit trace the 4096th time a guard fails)

I propose that this new mode use the following values:

  • ADAPTIVE_WARMUP_VALUE: 1 (specialize the 2nd time an instruction is run)
  • ADAPTIVE_COOLDOWN_VALUE: 1 (re-specialize the 2nd time a guard fails)
  • JUMP_BACKWARD_INITIAL_VALUE: 15 (JIT a new "root" trace the 16th time we jump backwards)
  • SIDE_EXIT_INITIAL_VALUE: 15 (JIT a new side-exit trace the 16th time a guard fails)

A few notes:

  • Specializing the first time an instruction runs leads to weird situations (for example, STORE_ATTR can fail to specialize if storage for the new attribute doesn't exist yet), so it doesn't make sense to reduce ADAPTIVE_WARMUP_VALUE further.
  • We fall into infinite loops if ADAPTIVE_COOLDOWN_VALUE is reduced to zero. This probably indicates a bug where we are specializing, then immediately deopting (because the specialization is incorrect somehow).
  • 16 was chosen for the JIT thresholds so that branch counters have time to stabilize.
  • We should probably make this a configure option, not a runtime-togglable value, since these counters are set and reset constantly using static inline functions and constant values, and we want to avoid introducing branching and indirection. Maybe it makes sense to just turn this on if configured --with-pydebug to keep things simple?

Before moving forward with this, we first need to update all tests that hardcode assumptions about these values (I already have a branch for this). This is worth doing anyways so the values can be tweaked in the future with minimal disruption.

Linked PRs
@brandtbucher brandtbucher added 3.14 bugs and security fixes interpreter-core (Objects, Python, Grammar, and Parser dirs) performance Performance or resource usage topic-JIT labels Jan 28, 2025
@brandtbucher brandtbucher self-assigned this Jan 28, 2025
@bedevere-app bedevere-app bot mentioned this issue Jan 28, 2025
Merged
brandtbucher added a commit that referenced this issue Jan 29, 2025
@brandtbucher
GH-129386: Use symbolic constants for specialization tests (GH-129415)
002c4e2
@encukou
Copy link
Member
encukou commented Jan 30, 2025

The PR broke a lot of Refleaks buildbots. @brandtbucher,can you fix it topay, or should we revert it for now?

@brandtbucher
Copy link
Member Author

I can look at it today. Didn't see any failures posted to the PR though (and I'm surprised it's causing leaks).

@bedevere-app bedevere-app bot mentioned this issue Jan 30, 2025
Merged
@brandtbucher
Copy link
Member Author

Thanks for the heads-up! I've opened a PR with a fix.

@brandtbucher
Copy link
Member Author

After talking with @markshannon, what we probably want for this issue is to put all of these counters in one place to make them easier to tweak. A new mode isn't very flexible, and is quite a bit of work. So we'll start with reorganization and helpful comments.

encukou pushed a commit that referenced this issue Jan 31, 2025
@brandtbucher
GH-129386: Add test.support.reset_code (GH-129486)
674befb
@picnixz picnixz added type-feature A feature request or enhancement and removed 3.14 bugs and security fixes labels Jan 31, 2025
srinivasreddy pushed a commit to srinivasreddy/cpython that referenced this issue Feb 7, 2025
@brandtbucher @srinivasreddy
pythonGH-129386: Use symbolic constants for specialization tests (pyt…
282755b 
…honGH-129415)
srinivasreddy pushed a commit to srinivasreddy/cpython that referenced this issue Feb 7, 2025
@brandtbucher @srinivasreddy
pythonGH-129386: Add test.support.reset_code (pythonGH-129486)
030c800
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@brandtbucher brandtbucher

Labels
interpreter-core (Objects, Python, Grammar, and Parser dirs) performance Performance or resource usage topic-JIT type-feature A feature request or enhancement
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@encukou @picnixz @brandtbucher
0