gh-130415: Narrowing to constants in branches involving is comparisons with a constant #143895

reidenong · 2026-01-15T22:58:48Z

This PR does constant narrowing in the JIT optimizer for branches with _IS_OP comparisons with the constant singletons True, False, None.

It introduces a new optimizer symbol, predicate, which describes a relationship between two objects. We create a new optimizer symbol upon encountering a comparison with a constant, and when we have definitive information on a object after a branch is taken (or not taken) we narrow it to a constant. Ideally, this new symbol will be easily extendable for other ops like (==, !=, contains etc.).

Unit tests intend to show constant narrowing by performing narrowing, then constant folding with other ops and showing the proof of the latter.

All feedback is welcome, with emphasis on the overall approach, how this new symbol fits into the lattice, and naming.

Thanks in advance.

Issue: Better constant narrowing in the JIT optimizer #130415

Fidget-Spinner · 2026-01-15T23:11:25Z

This is pretty cool! Thanks for doing this, I'll review it soon.

markshannon

I like the idea, it provides some more reasoning about values and fits nicely into the existing framework.

Thanks for doing this.

I would recommend changing the construction of predicates to be more relaxed and only check for constants when the predicate is resolved. It would be more powerful, easier to reason about and maybe faster.

As an example:

p0 = a is b
p1 = b is None
exit_if(not p1)
exit_if(not p0)
# a must be None

By restricting p0 to cases where a is a const or b is a const, we cannot infer that a is None after the second exit.

markshannon · 2026-01-16T09:58:48Z

Include/internal/pycore_optimizer.h

 extern JitOptRef _Py_uop_sym_new_compact_int(JitOptContext *ctx);
 extern void _Py_uop_sym_set_compact_int(JitOptContext *ctx,  JitOptRef sym);
+extern JitOptRef _Py_uop_sym_new_predicate(JitOptContext *ctx, JitOptRef subject_ref, JitOptRef constant_ref, JitOptPredicateKind kind, bool invert);
+extern bool _Py_uop_sym_is_known_singleton(JitOptContext *ctx, JitOptRef sym);


It doesn't matters whether something is a singleton or not, so you can drop this.

markshannon · 2026-01-16T09:59:41Z

Include/internal/pycore_optimizer_types.h


+typedef enum {
+    JIT_PRED_IS,
+    // JIT_PRED_EQ,


Either remove this, or add a TODO note that we'll be adding EQ, NE, etc.

markshannon · 2026-01-16T10:02:03Z

Python/optimizer_bytecodes.c


    op(_IS_OP, (left, right -- b, l, r)) {
-        b = sym_new_type(ctx, &PyBool_Type);
+        bool invert = (oparg != 0);


It would be both more powerful and result in cleaner code, if we left all analysis until the predicate is resolved.
We will often have more information then.

markshannon · 2026-01-16T10:06:59Z

Python/optimizer_symbols.c

            return;
+        case JIT_SYM_PREDICATE_TAG:
+            if (!PyBool_Check(const_val) ||
+                (_Py_uop_sym_is_const(ctx, ref) &&


Drop these extra checks. I'm not sure they are necessary and may be creating a contradiction where none exists.

If const_val is a boolean, then we can resolve the predicate, sym_apply_predicate_narrowing(ctx, flag, as_bool(const_val))

markshannon · 2026-01-16T10:09:59Z

Python/optimizer_symbols.c

 }

+JitOptRef
+_Py_uop_sym_new_predicate(JitOptContext *ctx, JitOptRef subject_ref, JitOptRef constant_ref, JitOptPredicateKind kind, bool invert)


We don't need to restrict this to the case where one of the values is a constant. If a is b is true, then a and b refer to the same object regardless of whether one is constant.

bedevere-app · 2026-01-16T10:21:50Z

A Python core developer has requested some changes be made to your pull request before we can consider merging it. If you could please address their requests along with any other requests in other reviews from core developers that would be appreciated.

Once you have made the requested changes, please leave a comment on this pull request containing the phrase I have made the requested changes; please review again. I will then notify any core developers who have left a review that you're ready for them to take another look at this pull request.

- Delay analysis until predicate resolves - Generalize predicate to lhs, rhs - Expand to accept all safe constants (not just singletons)

reidenong · 2026-01-16T21:04:25Z

Hi @markshannon I think I've addressed the stated problems, a summary of changes is that

the predicate symbol now takes general lhs and rhs references
analysis is done when predicate resolves
all constants are accepted (not just singletons)

Additionally, you mentioned that for sym_set_const, we can resolve the predicate with sym_apply_predicate_narrowing(ctx, flag, as_bool(const_val)). I may be mistaken, but I think that it is not necessary to place this within sym_set_const, because for each _GUARD_IS_*_POP we will apply predicate narrowing first before setting the predicate symbol to const. (Is there a case where we set the predicate symbol to const not from _GUARD_IS_*_POP?)

Another implementation I initially considered would be to move the sym_apply_predicate_narrowing call from _GUARD_IS_*_POP to within sym_set_const, but I thought that might introduce an additional layer of indirection and make the code harder to reason about. Let me know if this would be a preferred approach.

Thanks!

Fidget-Spinner · 2026-01-16T21:33:48Z

We need a test in the optimizer_symbols.c file (at the bottom, the last function).

Also CI is failing, can you please investigate? Thanks!

python-cla-bot · 2026-01-16T22:31:12Z

All commit authors signed the Contributor License Agreement.

Fidget-Spinner · 2026-01-16T22:36:18Z

CI failing due to #142911

reidenong · 2026-01-16T22:46:10Z

CI failing due to #142911

ok, will take a look and get back asap

Fidget-Spinner · 2026-01-17T00:35:12Z

@reidenong you should be able to pull in main now and it will build.

reidenong added 3 commits January 13, 2026 22:04

Add failing regression tests

181d096

Refine failing regression tests

e30a46b

Add new predicate symbol and IS_OP constant narrowing

eba715e

reidenong requested review from Fidget-Spinner, markshannon, savannahostrowski and tomasr8 as code owners January 15, 2026 22:58

bedevere-app bot mentioned this pull request Jan 15, 2026

Better constant narrowing in the JIT optimizer #130415

Open

17 tasks

bedevere-app bot added the awaiting review label Jan 15, 2026

Fidget-Spinner added the skip news label Jan 15, 2026

Fidget-Spinner self-assigned this Jan 15, 2026

markshannon requested changes Jan 16, 2026

View reviewed changes

bedevere-app bot added awaiting changes and removed awaiting review labels Jan 16, 2026

reidenong added 3 commits January 16, 2026 16:14

Remove comment

8834e12

Refactor predicate optimizer symbol

d828f8c

- Delay analysis until predicate resolves - Generalize predicate to lhs, rhs - Expand to accept all safe constants (not just singletons)

Remove unnecessary checks in sym_set_const

24ac326

reidenong requested a review from markshannon January 16, 2026 20:52

Merge branch 'main' into narrowing-to-constants-in-branches-with-is

8c95c5f

Uh oh!

gh-130415: Narrowing to constants in branches involving is comparisons with a constant #143895

Are you sure you want to change the base?

gh-130415: Narrowing to constants in branches involving is comparisons with a constant #143895

Conversation

reidenong commented Jan 15, 2026 • edited by bedevere-app bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Fidget-Spinner commented Jan 15, 2026

Uh oh!

markshannon left a comment

Choose a reason for hiding this comment

Uh oh!

markshannon Jan 16, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Jan 16, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Jan 16, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Jan 16, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Jan 16, 2026

Choose a reason for hiding this comment

Uh oh!

bedevere-app bot commented Jan 16, 2026

Uh oh!

reidenong commented Jan 16, 2026

Uh oh!

Fidget-Spinner commented Jan 16, 2026

Uh oh!

python-cla-bot bot commented Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Fidget-Spinner commented Jan 16, 2026

Uh oh!

reidenong commented Jan 16, 2026

Uh oh!

Fidget-Spinner commented Jan 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

reidenong commented Jan 15, 2026 •

edited by bedevere-app bot

Loading

python-cla-bot bot commented Jan 16, 2026 •

edited

Loading