Merge pull request #12547 from bradcray/reduce-atomics-future

Add future requesting the ability to + reduce atomics

[trivial, not reviewed]

It seems that we should be able to simply do reductions across arrays of atomics, yet this seems not to work without providing some additional helper functions yourself... This may relate somewhat to issue #8847.

test/reductions/reduceAtomics.bad

@@ -0,0 +1 @@
+reduceAtomics.chpl:12: error: Cannot directly add atomic variables

test/reductions/reduceAtomics.chpl

@@ -0,0 +1,35 @@
+var A: [1..10] atomic int;
+
+forall i in 1..10 do
+  A[i].write(i);
+
+var AofA: [1..3] [1..10] atomic int;
+
+forall i in 1..3 do
+  forall j in 1..10 do
+    AofA[i][j].write(i*10 + j);
+
+var B = + reduce A;
+var C = + reduce AofA;
+
+writeln(B);
+writeln(C);
+
+
+// It's annoying that I have to define these in order to run a +
+// reduction over an array of atomic ints, but that seems to be
+// the case at present...
+/*
+proc +(x: atomic int, y: atomic int) {
+  return x.read() + y.read();
+}
+
+proc +=(ref x: int, y: atomic int) {
+  x += y.read();
+}
+
+proc +=(X: [?D] int, Y: [D] atomic int) {
+  forall i in D do 
+    X[i] += Y[i].read();
+}
+*/

test/reductions/reduceAtomics.future

@@ -0,0 +1,8 @@
+feature request: Support reductions on arrays of atomics
+
+It seems that we should support reductions on arrays of atomics, yet
+today we don't.  Here, I focus on + reductions and atomic ints, but I
+think we should have a general solution for all cases.  The commented
+out code demonstrates overloads that will make this program work as
+intended if I uncomment it.  But I don't think I should have to write
+it and wouldn't expect most users to figure out how to do it.

test/reductions/reduceAtomics.good

@@ -0,0 +1,2 @@
+55
+63 66 69 72 75 78 81 84 87 90