Typing conformance results

No changes detected ✅

Memory usage report

Memory usage unchanged ✅

ecosystem-analyzer results

No diagnostic changes detected ✅

Full report with detailed diff (timing results)

I also considered using a SmallVec, but cache can grow large

Wow, 1423 is larger than I would have guessed. I wonder if we could (separately) experiment with using a fancier cache that has some kind of LRU cache eviction. My guess would be that some small portion of those 1423 entries are repeatedly hit, while most are hit rarely, if at all

Wow, 1423 is larger than I would have guessed. I wonder if we could (separately) experiment with using a fancier cache that has some kind of LRU cache eviction. My guess would be that some small portion of those 1423 entries are repeatedly hit, while most are hit rarely, if at all

I'm not sure this is worth it. This adds another write to the cached paths and the cost is relatively low, because we don't store CycleDetectors.

I still think experimenting could be interesting, but it's definitely out of scope here

Sure. We just see in Salsa that LRU cache has a high cost because it adds a write to every read path. But, the scales are probably different here.

Merging this PR will improve performance by 5.25%

⚡ 1 improved benchmark
✅ 70 untouched benchmarks
⏩ 64 skipped benchmarks¹

Performance Changes

_{Comparing micha/cycle-detector-inline-cache (bb6d6d0) with main (ae306a0)}