perf: separate cached and cold fetch paths (#1184)

Author: MichaReiser

benches/compare.rs

@@ -165,6 +165,17 @@ fn inputs(c: &mut Criterion) {
         )
     });
 
+    group.bench_function(BenchmarkId::new("cached", "Input"), |b| {
+        let db = salsa::DatabaseImpl::default();
+        let input = Input::new(&db, "hello, world!".to_owned());
+
+        // Prewarm the memo. Every measured call reads a value that has already been verified in
+        // the current revision.
+        assert_eq!(length(&db, input), 13);
+
+        b.iter(|| length(black_box(&db), black_box(input)))
+    });
+
     group.bench_function(BenchmarkId::new("new", "SupertypeInput"), |b| {
         b.iter_batched_ref(
             || {

src/function/fetch.rs

@@ -59,10 +59,14 @@ where
         let memo_ingredient_index = self.memo_ingredient_index(zalsa, id);
 
         loop {
-            if let Some(memo) = self
-                .fetch_hot(zalsa, id, memo_ingredient_index)
-                .or_else(|| self.fetch_cold(zalsa, zalsa_local, db, id, memo_ingredient_index))
-            {
+            // Keep the hot and cold probes in distinct control-flow blocks. Using `or_else`
+            // here can outline both into one function, making hot hits pay for the cold path's
+            // stack frame.
+            if let Some(memo) = self.fetch_hot(zalsa, id, memo_ingredient_index) {
+                return memo;
+            }
+
+            if let Some(memo) = self.fetch_cold(zalsa, zalsa_local, db, id, memo_ingredient_index) {
                 return memo;
             }
         }