Branches performance
Pull requests
last run
3 days ago
chore(deps): lock file maintenance
3 days ago
44aafa0
renovate/lock-file-maintenance
-1%
last run
5 days ago
chore(deps): update actions/upload-artifact action to v7
5 days ago
17912e4
renovate/major-github-artifact-actions
0%
last run
11 days ago
perf(neon): combine 4 escape masks in SIMD domain before GPR extraction
The chunk loop (64 bytes/iteration) previously extracted each vector's
escape mask to a GPR individually, creating 4 serial shrn→fmov→cbz
chains on the critical path (~32 cycles on Apple Silicon).
Now compute all 4 mask vectors independently (can pipeline), combine
them with 3 vorrq instructions in the vector domain, and extract a
single combined bitmask. Individual bitmasks are only extracted lazily
in the slow path (when escapes exist).
Fast path savings per 64-byte chunk:
- 3 fewer shrn instructions (6 cycles)
- 3 fewer fmov d→x transfers (15 cycles)
- 3 fewer cbz branches (3 cycles)
- Stores optimized from 3×str to 2×stp by compiler
Benchmark (5.5MB source fixtures): ~3.3% faster (10.47ms → 10.12ms)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
11 days ago
7649a17
perf/neon-combined-mask-check
0%
last run
15 days ago
chore(deps): update release-plz/action action to v0.5.127
15 days ago
1f4e028
renovate/release-plz-action-0.x
-1%