More optimizations

• A lot of misc. optimization work, notably parallelization of Rodinia CPU schedules.
• Support >32 tiles in GPU backend.
• Tweak some transformations.

juno_samples/rodinia/bfs/src/cpu.sch

@@ -10,13 +10,14 @@ macro simpl!(X) {
 
 phi-elim(bfs);
 no-memset(bfs@cost);
-outline(bfs@cost_init);
-let loop1 = outline(bfs@loop1);
-let loop2 = outline(bfs@loop2);
+let init = outline(bfs@cost_init);
+let traverse = outline(bfs@loop1);
+let collect = outline(bfs@loop2);
 
 simpl!(*);
 predication(*);
 const-inline(*);
+loop-bound-canon(*);
 simpl!(*);
 fixpoint {
   forkify(*);
@@ -25,6 +26,37 @@ fixpoint {
 simpl!(*);
 predication(*);
 simpl!(*);
+reduce-slf(*);
+simpl!(*);
+slf(*);
+simpl!(*);
+
+fixpoint {
+  forkify(collect);
+  fork-guard-elim(collect);
+}
+simpl!(collect);
+
+parallel-fork(traverse, collect);
+parallel-reduce(traverse, collect);
 
-unforkify(*);
+fork-tile[32, 0, false, true](traverse, collect);
+let (outer, inner) = fork-reshape[[1], [0]](traverse);
+let traverse_body = outline(inner);
+let (outer, inner) = fork-reshape[[1], [0]](collect);
+let collect_body = outline(inner);
+
+let init_body = init;
+// Following code seems to generate breaking RT code
+//fork-tile[32, 0, false, true](init);
+//let (outer, inner) = fork-reshape[[1], [0]](init);
+//let init_body = outline(inner);
+//inline(bfs@cost_init);
+
+inline(bfs@loop1, bfs@loop2);
+delete-uncalled(*);
+const-inline(*);
+
+unforkify(init_body, traverse_body, collect_body);
+simpl!(*);
 gcm(*);