Home - Speed - Summaries: (main - py2.7) : (py3.10) : (py3.11) : Everything : RPython - Nightly builds - Waterfall - Builders - About

Console View


Categories: aarch64 benchmark-run linux32 linux64 macos-arm64 macos-x86_64 win64
Legend:   Passed Failed Warnings Failed Again Running Exception Offline No data

aarch64 benchmark-run linux32 linux64 macos-arm64 macos-x86_64 win64
Matti Picus
remove debug printing
Matti Picus
Merge pull request #5473 from pypy/is_cpyext_builtin
tweak _is_cpyext* check for inpect
mattip
add _Py_NO_RETURN to _Py_FatalErrorFunc like CPython
Matti Picus
Merge pull request #5474 from Julian/darwin-perflevel0-l2cache
Estimate GC nursery size via performance-core L2 cache (not efficiency) on Apple Silicon
CF Bolz-Tereick
same for add and sub
Matti Picus
don't crash if ansiramp is missing
mattip
remove unicode
mattip
Merge branch 'main' into py3.11
Matti Picus
Merge pull request #5468 from pypy/int-ovf-to-long-without-fromint
Int ovf to long without fromint
Matti Picus
tweak _is_cpyext* check for inpect
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
mattip
speed up hypothesis test from 103 to 10 secs
CF Bolz-Tereick
speed up int << int -> long shifts
CF Bolz-Tereick
fix test_pow_ovf_without_fromint
to achieve this, reinstate (or implement, I am not sure) an optimization
we had on pypy2 that leaves the exponent of long ** int as an int
mattip
Revert "fix w_char for windows64"
This reverts commit 2783f700f22c6925a3afb7602cc4821e80637a23.
Julian Berman
gc: detect performance-cluster L2 cache on Apple Silicon
On Apple Silicon, `hw.l2cachesize` returns the *efficiency* cluster's L2
(e.g. 4 MB on a typical M-series), not the performance cluster's (16 MB
and up). The result was that `get_L2cache_darwin()` returned 4 MB, which
in `best_nursery_size_for_L2cache()` falls into the "no L3 present"
branch (L2 <= 8 MB) and gives the 4 MB `NURSERY_SIZE_UNKNOWN_CACHE`
fallback rather than half of the real performance-cluster L2.

Apple documents per-cluster topology under `hw.perflevelN.*` and states
that "lower values of N indicate higher-performance core types" — so
`hw.perflevel0.l2cachesize` is the L2 for the cluster the work actually
runs against:

  https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_system_capabilities#3901385

Read it first, falling back to the legacy `hw.l2cachesize` when it's
absent (Intel Macs). On an M-series Mac this changes the returned size
from 4 MB to 16 MB, so the heuristic returns an 8 MB nursery (half of
L2) — closer to what the Intel-Mac path gives.

`hw.l3cachesize` (and Apple's documented `hw.perflevel0.l3cachesize`)
both appear unpopulated on at least some M-series chips, so this commit
doesn't try to recover an "L3" contribution there; heavily-allocating
workloads may still want to set `PYPY_GC_NURSERY` higher, but the
default is no longer pathological on Apple Silicon.

Refs: 39b141959b (raise default gc nursery size to 4MB)
Refs: #4939

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
mattip
fix w_char for windows64
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
CF Bolz-Tereick
don't call fromint for pow
mattip
Merge branch 'main' into py3.11
CF Bolz-Tereick
preserve RPython-level traceback when reraising an OperationError untranslated
this helps debugging with --pdb to find out where an OperationError
comes from, when one occurs in an untranslated test