gh-154401: Skip fetching the thread state for non-GC types in _Py_Dealloc#154430
Open
johng wants to merge 1 commit into
Open
gh-154401: Skip fetching the thread state for non-GC types in _Py_Dealloc#154430johng wants to merge 1 commit into
_Py_Dealloc#154430johng wants to merge 1 commit into
Conversation
johng
force-pushed
the
optimize-dealloc-nongc
branch
from
July 22, 2026 06:47
c4ef5f6 to
e91e3d1
Compare
johng
changed the title
_Py_Dealloc_Py_Dealloc
_Py_Dealloc() fetched the current thread state and computed the C recursion margin on every deallocation, but both are only used by the trashcan mechanism, which applies only to GC-tracked types. Gate them behind the Py_TPFLAGS_HAVE_GC check so the common non-GC case (int, float, str and similar atomic types) skips them. On platforms where _PyThreadState_GET() reads a _Thread_local via a function call (e.g. macOS TLV), this removes a call from every non-GC deallocation. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
johng
force-pushed
the
optimize-dealloc-nongc
branch
from
July 22, 2026 06:51
e91e3d1 to
8a4463e
Compare
Contributor
Author
|
Raised via AI, have been verifying findings as best I can |
Contributor
|
You have to show full comparison report between main and this PR from pyperformance suite. |
Contributor
|
Changes look good to me, waiting to see results from pyperformance. |
Contributor
Author
Benchmarks: this change vs
|
| Benchmark | main | A2 (change) | Change | Significance |
|---|---|---|---|---|
| chaos | 29.9 ms | 29.1 ms | 1.03x faster | Significant (t=6.90) |
| comprehensions | 7.50 us | 7.34 us | 1.02x faster | Significant (t=4.56) |
| coroutines | 11.0 ms | 10.7 ms | 1.02x faster | Significant (t=5.20) |
| crypto_pyaes | 42.0 ms | 41.8 ms | 1.01x faster | Not significant |
| deepcopy | 118 us | 114 us | 1.03x faster | Significant (t=4.36) |
| deepcopy_memo | 11.6 us | 11.3 us | 1.02x faster | Significant (t=6.83) |
| deepcopy_reduce | 1.27 us | 1.24 us | 1.02x faster | Significant (t=4.06) |
| deltablue | 1.68 ms | 1.62 ms | 1.04x faster | Significant (t=7.85) |
| fannkuch | 176 ms | 165 ms | 1.07x faster | Significant (t=6.25) |
| float | 33.8 ms | 32.2 ms | 1.05x faster | Significant (t=13.25) |
| generators | 18.5 ms | 18.2 ms | 1.02x faster | Not significant |
| go | 52.8 ms | 52.3 ms | 1.01x faster | Not significant |
| hexiom | 2.94 ms | 2.95 ms | 1.00x slower | Not significant |
| json_dumps | 4.09 ms | 4.11 ms | 1.01x slower | Not significant |
| json_loads | 14.0 us | 14.1 us | 1.00x slower | Not significant |
| logging_format | 2.66 us | 2.65 us | 1.00x faster | Not significant |
| logging_silent | 42.2 ns | 41.9 ns | 1.01x faster | Not significant |
| logging_simple | 2.41 us | 2.42 us | 1.00x slower | Not significant |
| meteor_contest | 48.4 ms | 48.9 ms | 1.01x slower | Not significant |
| nbody | 55.5 ms | 54.6 ms | 1.02x faster | Not significant |
| nqueens | 36.2 ms | 36.8 ms | 1.02x slower | Not significant |
| pickle | 5.82 us | 5.84 us | 1.00x slower | Not significant |
| pickle_dict | 13.2 us | 13.1 us | 1.01x faster | Not significant |
| pickle_list | 2.39 us | 2.40 us | 1.00x slower | Not significant |
| pprint_pformat | 729 ms | 742 ms | 1.02x slower | Not significant |
| pprint_safe_repr | 356 ms | 359 ms | 1.01x slower | Not significant |
| pyflate | 183 ms | 184 ms | 1.01x slower | Not significant |
| raytrace | 134 ms | 132 ms | 1.01x faster | Not significant |
| regex_compile | 57.5 ms | 57.7 ms | 1.00x slower | Not significant |
| regex_dna | 86.2 ms | 85.1 ms | 1.01x faster | Not significant |
| regex_effbot | 1.28 ms | 1.25 ms | 1.03x faster | Significant (t=7.63) |
| regex_v8 | 9.35 ms | 9.09 ms | 1.03x faster | Significant (t=12.39) |
| richards | 25.0 ms | 24.4 ms | 1.03x faster | Significant (t=8.46) |
| scimark_fft | 150 ms | 144 ms | 1.04x faster | Significant (t=11.74) |
| scimark_lu | 55.8 ms | 54.8 ms | 1.02x faster | Not significant |
| scimark_monte_carlo | 31.6 ms | 31.4 ms | 1.00x faster | Not significant |
| scimark_sor | 52.5 ms | 52.1 ms | 1.01x faster | Not significant |
| scimark_sparse_mat_mult | 2.50 ms | 2.40 ms | 1.04x faster | Significant (t=8.17) |
| spectral_norm | 47.5 ms | 45.9 ms | 1.04x faster | Significant (t=14.45) |
| telco | 3.14 ms | 3.22 ms | 1.03x slower | Significant (t=-8.14) |
| unpack_sequence | 16.5 ns | 17.1 ns | 1.03x slower | Significant (t=-14.47) |
| unpickle | 7.38 us | 7.52 us | 1.02x slower | Not significant |
x86-64 / Linux (2 vCPU) — geometric mean 1.02x faster (16 significant faster, 3 significant slower)
| Benchmark | main | A2 (change) | Change | Significance |
|---|---|---|---|---|
| chaos | 102 ms | 101 ms | 1.01x faster | Not significant |
| comprehensions | 27.5 us | 27.9 us | 1.02x slower | Not significant |
| coroutines | 36.9 ms | 35.0 ms | 1.05x faster | Not significant |
| crypto_pyaes | 144 ms | 140 ms | 1.03x faster | Not significant |
| deepcopy | 455 us | 467 us | 1.03x slower | Not significant |
| deepcopy_memo | 48.0 us | 46.5 us | 1.03x faster | Not significant |
| deepcopy_reduce | 5.81 us | 5.40 us | 1.08x faster | Significant (t=2.88) |
| deltablue | 5.83 ms | 5.80 ms | 1.01x faster | Not significant |
| fannkuch | 630 ms | 643 ms | 1.02x slower | Not significant |
| float | 135 ms | 129 ms | 1.04x faster | Significant (t=2.79) |
| generators | 51.6 ms | 54.5 ms | 1.06x slower | Significant (t=-2.27) |
| go | 187 ms | 176 ms | 1.07x faster | Significant (t=2.29) |
| hexiom | 9.86 ms | 9.88 ms | 1.00x slower | Not significant |
| json_dumps | 17.2 ms | 17.3 ms | 1.00x slower | Not significant |
| json_loads | 41.3 us | 39.4 us | 1.05x faster | Significant (t=3.34) |
| logging_format | 12.4 us | 11.3 us | 1.10x faster | Significant (t=3.14) |
| logging_silent | 150 ns | 148 ns | 1.01x faster | Not significant |
| logging_simple | 10.0 us | 10.2 us | 1.01x slower | Not significant |
| meteor_contest | 173 ms | 170 ms | 1.02x faster | Not significant |
| nbody | 195 ms | 166 ms | 1.18x faster | Significant (t=5.98) |
| nqueens | 150 ms | 148 ms | 1.01x faster | Not significant |
| pickle | 19.0 us | 18.9 us | 1.01x faster | Not significant |
| pickle_dict | 45.4 us | 45.4 us | 1.00x faster | Not significant |
| pickle_list | 9.04 us | 8.79 us | 1.03x faster | Not significant |
| pprint_pformat | 3.01 sec | 3.05 sec | 1.01x slower | Not significant |
| pprint_safe_repr | 1.43 sec | 1.47 sec | 1.03x slower | Not significant |
| pyflate | 673 ms | 677 ms | 1.01x slower | Not significant |
| raytrace | 489 ms | 472 ms | 1.04x faster | Significant (t=3.15) |
| regex_compile | 267 ms | 255 ms | 1.04x faster | Significant (t=2.21) |
| regex_dna | 253 ms | 249 ms | 1.01x faster | Not significant |
| regex_effbot | 3.72 ms | 3.82 ms | 1.03x slower | Significant (t=-2.02) |
| regex_v8 | 30.8 ms | 29.0 ms | 1.06x faster | Significant (t=3.14) |
| richards | 80.3 ms | 85.9 ms | 1.07x slower | Significant (t=-3.03) |
| scimark_fft | 534 ms | 499 ms | 1.07x faster | Significant (t=3.07) |
| scimark_lu | 219 ms | 230 ms | 1.05x slower | Not significant |
| scimark_monte_carlo | 132 ms | 124 ms | 1.06x faster | Significant (t=2.48) |
| scimark_sor | 206 ms | 193 ms | 1.07x faster | Significant (t=4.03) |
| scimark_sparse_mat_mult | 8.88 ms | 8.23 ms | 1.08x faster | Significant (t=4.84) |
| spectral_norm | 163 ms | 156 ms | 1.04x faster | Significant (t=5.07) |
| telco | 310 ms | 292 ms | 1.06x faster | Significant (t=2.68) |
| unpack_sequence | 60.4 ns | 60.1 ns | 1.00x faster | Not significant |
| unpickle | 25.9 us | 25.3 us | 1.02x faster | Significant (t=2.07) |
Contributor
Author
|
@sergey-miryanov please see above for the table of results, and also link to gist of raw json too |
Contributor
|
It would be good to see results with enabled PGO+LTO, and for GIL-enabled build. |
ZeroIntensity
approved these changes
Jul 22, 2026
ZeroIntensity
left a comment
ZeroIntensity
left a comment
Member
There was a problem hiding this comment.
The 7% slowdown on x86-64 richards is a little concerning, but I suspect that's just noise. I can't imagine how this would possibly slow things down, because we're not adding any additional checks beyond what was already there.
Contributor
Author
|
Thanks! Added the three way comparison for both arch's https://gist.github.com/johng/388e14f9a40c88f8a2031f2383f6938d Likely need to find a better x86 host in the future, there's likely a bit too much noise but no trends either way for that arch as expected. Arm I think is showing a consistent small speedup across the tests. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Demonstration for
_Py_Deallocfor MacOS ARM #154401