More optimization

• Some fixes and optimization.

juno_samples/rodinia/srad/src/cpu.sch

@@ -8,6 +8,7 @@ macro simpl!(X) {
   infer-schedules(X);
 }
 
+no-memset(srad@scratch);
 phi-elim(*);
 let loop1 = outline(srad@loop1);
 let loop2 = outline(srad@loop2);
@@ -31,8 +32,18 @@ simpl!(*);
 fork-interchange[0, 1](loop1);
 reduce-slf(*);
 simpl!(*);
+slf(*);
+simpl!(*);
+
+fork-tile[32, 0, false, false](loop2);
+let split = fork-split(loop2);
+let loop2_body = outline(split.srad_1.fj1);
+simpl!(loop2, loop2_body);
+
+inline(srad@loop2);
+delete-uncalled(*);
 
-fork-split(*);
-unforkify(*);
+fork-split(extract, compress, loop1, loop2_body, loop3);
+unforkify(extract, compress, loop1, loop2_body, loop3);
 
 gcm(*);