Regate the binary search pivots as it was just polluting cache

Author: tibvdm

sa-index/src/sa_searcher.rs

@@ -597,8 +597,15 @@ impl Searcher {
     /// call. One fetch will be wasted; both are free (single non-blocking CPU instruction).
     /// From iteration 2 onward the needed SA entry is already in L1/L2 cache.
     #[inline]
+    #[cfg_attr(not(feature = "mmap"), allow(unused_variables))]
     fn prefetch_binary_search_pivots(&self, lo: usize, center: usize, hi: usize) {
+        // Pivot prefetching pays off only for mmap where cold pages are the bottleneck.
+        // For preloaded (in-memory) data, hot SA positions stay in L3 from temporal
+        // locality, so PREFETCHT0 adds overhead without benefit and pollutes L1 for
+        // the text comparison that follows immediately in the same iteration.
+        #[cfg(feature = "mmap")]
         self.sa.prefetch_sa_index((lo + center) / 2);
+        #[cfg(feature = "mmap")]
         self.sa.prefetch_sa_index((center + hi) / 2);
     }