Sashiko AI Introducing Linux Performance Specialist Review Stage

The Google-developed Sashiko for agentic AI code review initially focused on the Linux kernel continues helping to uncover new bugs and issues as new patches hit the kernel mailing list as well as those manually running this open-source AI software. In addition to recently introducing an experimental patch series design assessment, they are introducing another new skill for helping with performance/benchmark evaluation of kernel changes. Beyond pointing out coding issues and other straight-forward issues by Sashiko, an interesting and important avenue now being explored too is on the performance side. As we all know, Linux kernel performance regressions can be all too common, unfortunately. A new review stage added in Friday's Sashiko 0.9.5 release is around performance and benchmarking: "Linux Performance Specialist Review Stage and Benchmarking: Add an optional performance specialist analysis stage to linux_patch_review (available in daemon and local sashiko review, including --stages performance and --cheap) with second-order side-effect tracing and anti-noise guardrails Add expect_clean neutral benchmark support, per-stage false-positive tracking, and the balanced 40-commit benchmark_performance.json suite (20 upstream regressions + 20 soak-verified clean commits)" Interesting and any efforts to help prevent Linux kernel performance regressions is certainly very welcome to see. This commit further elaborates on the initial implementation: Integrate an optional performance specialist stage into ANALYSIS_STAGES in linux_patch_review.rs for both local (sashiko review, including --stages performance and --cheap) and daemon review modes, backed by projects/linux/prompts/performance.md and aligned rules in false-positive-guide.md and severity.md. Because kernel performance regressions are usually unexpected side effects of ordinary bug fixes, cleanups, and struct layout changes: - Schedule performance on any non-trivial C, Rust, or struct/header change while skipping pure docs, comments, or ID tables. - Trace callers and outer embedding structs 1-2 hops outward using repository tools, forbidding self-dismissals based on short critical sections, cold teardown rationalizations, or treating a stated correctness mechanism as an accepted throughput/latency cliff. - Exclude speculative micro-optimizations, zero-delta guard(mutex) cleanups, and bounded one-shot control paths without fan-out. Controlled A/B benchmark on benchmark_performance.json (40 commits: 20 upstream regressions + 20 neutral expect_clean commits) using gemini-4-argon at high effort: - Positive detection: 3/20 (15.0%) baseline -> 14/20 (70.0%) (+55.0 pp, 9 of 11 net-new wins originated solely from performance). - Neutral clean rate: 11/20 (55.0%) baseline -> 12/20 (60.0%) (0 new false-positive neutral commits; 52 -> 51 total unmatched findings). - Resource overhead: +11.0% median input tokens (91.4% cached) and -3.1% median per-commit duration. Sashiko 0.9.5 also brings a variety of other improvements too.

Original Source

Read the full article at Phoronix →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.