Merge pull request #22 from arrdem/patch-1

Implement a better memoize for parallel use cases

Author: amalloy

src/flatland/useful/parallel.clj

@@ -17,3 +17,34 @@
                       (let [body (wrap-fn #(doall (map f slice)))]
                         (future-call body)))
                     (slice *pcollect-thread-num* coll))))))
+
+
+(defn- assoc-noclobber 
+  "An assoc wrapper which ensures that existing keys will not be
+  clobbered by subsequent assoc invocations. 
+
+  Used as a helper for locking-memoize to ensure that (delay) refs
+  cannot be lost by swap! retry behavior."
+  
+  [m k v]
+  (if (contains? m k) m
+      (assoc m k v)))
+
+(defn pmemoize 
+  "Memoizes the function f, using the same approach as
+  clojure.core/memoize. The practical difference is that this function
+  provides the gurantee that in spite of parallel invocations of the
+  memoized function each input to f will only ever be memoized
+  once. This resolves an implementation detail in clojure.core/memoize
+  which allows f to be applied to args without locking the cache to
+  prevent other threads duplicating the work."
+
+  [f]
+  (let [mem (atom {})]
+    (fn [ & args ]
+      (if-let [e (find @mem args)]
+        (deref (val e))
+        (-> (swap! mem assoc-noclobber
+                   args (delay (apply f args)))
+            (get args)
+            (deref))))))