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/gpu.sch

@@ -10,14 +10,17 @@ macro simpl!(X) {
 
 phi-elim(bfs);
 no-memset(bfs@cost);
-let cost_init = outline(bfs@cost_init);
-let loop1 = outline(bfs@loop1);
-let loop2 = outline(bfs@loop2);
-gpu(loop1, loop2);
+let init = outline(bfs@cost_init);
+let traverse = outline(bfs@loop1);
+let collect = outline(bfs@loop2);
+parallel-reduce(traverse, collect);
+no-memset(make_stop_prod);
+gpu(traverse, make_stop_prod, collect);
 
 simpl!(*);
 predication(*);
 const-inline(*);
+loop-bound-canon(*);
 simpl!(*);
 fixpoint {
   forkify(*);
@@ -26,14 +29,17 @@ fixpoint {
 simpl!(*);
 predication(*);
 simpl!(*);
-
-unforkify(cost_init);
-parallel-reduce(loop1);
-forkify(*);
-fork-guard-elim(*);
-simpl!(*);
-predication(*);
 reduce-slf(*);
 simpl!(*);
 
+fixpoint {
+  forkify(collect);
+  fork-guard-elim(collect);
+}
+simpl!(collect);
+
+fork-tile[1024, 0, false, true](traverse, collect);
+fork-split(traverse, collect);
+
+unforkify(init);
 gcm(*);