More optimization

• Some fixes and optimization.

juno_samples/rodinia/srad/src/cpu.sch

@@ -40,6 +40,9 @@ let split = fork-split(loop2);
 let loop2_body = outline(split.srad_1.fj1);
 simpl!(loop2, loop2_body);
 
+inline(srad@loop2);
+delete-uncalled(*);
+
 fork-split(extract, compress, loop1, loop2_body, loop3);
 unforkify(extract, compress, loop1, loop2_body, loop3);