Xu 0.8.11
Xu 0.8.11 is the stable release for this optimization cycle. Download the SDK, installers, checksums and release notes.
What changed
- x64 integer reductions use verified SSE2 prefixes. Eligible System V reductions split the induction dependency across two vectors.
- Eligible leaf integer kernels can use an additional saved register. SHA-256 software hashing avoids per-block state copies and expands its circular schedule in groups of 16 rounds.
- ARM64 base-case returns can avoid building a stack frame when their parameter and effect checks permit it.
- Windows bootstrap stack sizing, entry alignment and argument handling are repaired. SDK executable selection and artifact download paths are corrected.
Validation and performance results
The release passed compiler fixed-point checks, relevant local correctness regressions, native bootstrap and SDK checks, and the six-platform benchmark output checks. The benchmark page links to its exact Actions run and compiler identities.
The current quick matrix measures 13 workloads on six platforms. C and Rust comparisons use the lowest median among O1, O2 and O3 on the same platform. Missing measurements are shown as a dash. SHA-256 uses hardware instructions where supported on ARM64; x64 currently uses software. Whole-program timing includes process startup, and timings from different machines are not optimization speedup evidence.
Memory measurement correction
2026-10-16: legacy Linux RSS included the Python launcher's memory peak and has been withdrawn. A corrected sampler completed a six-platform rerun of the same stable commit, now reflected on the benchmark page. Linux RSS comes from a separate post-exec run; tracing is excluded from timing. Resource ranges across workloads do not establish an overall winner.
Next development cycle
Development continues at 0.8.12-dev.1. The website and default downloads remain on the published stable SDK.