[meziantou]

BoundedCache used least-recently-used (LRU) eviction. That wrote a shared Interlocked counter and a per-entry timestamp on every hit, and every eviction copied all the entries to an array and sorted them. It now uses two generations of half the capacity:

• A hit in the current generation writes nothing.
• A hit in the previous generation copies the entry back into the current one, so entries that are used often stay in the cache.
• When the current generation is full, it becomes the previous one and the old previous one is dropped. No sorting is needed.

The cache still holds at most Capacity entries, plus a few that may be added during a rotation. Each entry no longer needs its own Entry object.

Why

This came from looking into whether the 8,192-entry cache of OverloadCandidateIndex is too small for large compilations. That cache is shared by every rule that uses OverloadFinder (MA0002, MA0011, MA0032, MA0042, MA0054, MA0074, MA0167, MA0209).

Measurements per run on meziantou.framework/src (6,747 files), with the 8 OverloadFinder rules and sequential analysis:

Cache	Hits	Misses (distinct keys: 68,508)	Time creating values	Time evicting
LRU, 8,192 (before)	1.418M	76.1k–76.9k	63–80 ms	36–73 ms
LRU, 32,768	1.425M	69.0k–69.5k	~80 ms	25–42 ms
LRU, 131,072	1.426M	68.5k	62–76 ms	0
Two generations, 8,192 (this PR)	1.417M	78.0k	88 ms	~0

The capacity is not the bottleneck. The hit rate is already about 95%, and evictions cause only about 8k recomputations at about 1 µs each. Most of the CreateParameterTypes time in profiles is Roslyn loading symbol data (AllInterfaces, GetMembers) the first time. That cost is the same whatever the cache size, so the capacity is left at 8,192. The only cost that could be removed was the eviction itself and the write on every hit, which matters most under concurrent analysis (1.4M hits per run). This PR removes both, at the cost of about 1.5k extra misses.

Notes for the reviewer

• BoundedCache is also used by RegexCache, RegexUsageAnalyzerBase, and DoNotUseBannedSyntaxAnalyzer (capacity 128). Their behavior is the same apart from which entries get evicted.
• I didn't finish an end-to-end wall-clock comparison: the machine was heavily loaded (load average around 19 on 15 cores), which made the runs too noisy. The numbers above come from counters added to a local build only for measuring.
• I replaced TheLeastRecentlyUsedEntriesAreEvicted with TheEntriesThatAreNotUsedAreEvicted and added TheEntriesThatAreUsedOftenAreKept.
• dotnet test passes on all Roslyn versions (4.8, 4.14, 5.0, 5.6, 5.9): 28,363 passed, 0 failed, 45 skipped.