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 +10,15 @@ macro simpl!(X) {
 
 phi-elim(bfs);
 no-memset(bfs@cost);
-outline(bfs@cost_init);
+let init = outline(bfs@cost_init);
 let loop1 = outline(bfs@loop1);
 let loop2 = outline(bfs@loop2);
+parallel-reduce(loop1);
 
 simpl!(*);
 predication(*);
 const-inline(*);
+loop-bound-canon(*);
 simpl!(*);
 fixpoint {
   forkify(*);
@@ -26,5 +28,12 @@ simpl!(*);
 predication(*);
 simpl!(*);
 
-unforkify(*);
+fork-tile[32, 0, false, false](loop1);
+let split = fork-split(loop1);
+let out = outline(split.bfs_1.fj1);
+unforkify(out);
+inline(bfs@loop1);
+delete-uncalled(*);
+
+unforkify(init);
 gcm(*);