ben-manes / caffeine / #5544

100%

master: 100%

push

github

Count correlated references as hits in IRR dedup policies

Lirs2, ClockPro, and ClockProPlus collapse a consecutive duplicate
access (a correlated reference) because their per-access transition
is not idempotent. The guard returned before recordOperation() and
recordHit(), which dropped the duplicate from the hit-rate
denominator.

The reference implementations -- Song Jiang's lirs.c / clock-pro.c
and Chen Zhong's C++ -- increment the denominator before their
`continue`, so a correlated reference stays counted as a non-miss
(a hit). The early return diverged ~0.3-1pp on dup-heavy traces;
e.g. ClockPro on cs reported 27.1% vs the reference's 28.1% at
size 500.

Score the duplicate as a hit before returning so it stays in the
denominator. Miss counts are unchanged -- already bit-for-bit with
the references -- only the denominator is corrected. Verified on cs:
ClockPro now matches clock-pro.c (28.1/57.2/79.2% at 500/1k/2k) and
Lirs2 matches Zhong.

Add CorrelatedAccessTest covering the LIRS / CLOCK-Pro family, and
correct comments that over-generalized the dedup as an "original
LIRS convention" -- LirsPolicy is idempotent and needs no guard.

4049 of 4054 branches covered (99.88%)

8242 of 8242 relevant lines covered (100.0%)

1.0 hits per line