Merge branch 'maint'

diff --cc test_inputs/codegen/unroll4.c
index 0000000,9bf1884..f461e6c
mode 000000,100644..100644
--- /dev/null
--- 2/test_inputs/codegen/unroll4.c
+++ b/test_inputs/codegen/unroll4.c
@@@ -1,0 -1,22 +1,22 @@@
 -    if (c3 == t2 + 32 || (c3 == t2 && t1 + 4 * floord(-((t2 + 1) % 2) - t1 + t2 + 3, 4) >= 5 && t1 + 4 * floord(-((t2 + 1) % 2) - t1 + t2 + 3, 4) <= 8))
+ {
+   write_shared_A(3, ((t1 + 3) % 4) + 1, ((t2 + 31) % 32) + 1);
+   if (t1 % 3 == 0 && t2 >= 1 && t2 <= 2)
+     write_shared_A(3, (-t1 + 12) / 3, t2 + 32);
+   {
+     int c3 = t2 <= 1 || (t2 >= 2 && ((3 * t1 + t2 + 3) % 4) + 1 >= t2) ? t2 + 32 : t2;
 -    if (c3 == t2 + 32 || (c3 == t2 && t1 + 4 * floord(-((t2 + 1) % 2) - t1 + t2 + 3, 4) >= 5 && t1 + 4 * floord(-((t2 + 1) % 2) - t1 + t2 + 3, 4) <= 8))
++    if (c3 == t2 + 32 || (c3 == t2 && ((-t1 + 8) % 4) + t2 >= ((t2 + 1) % 2) + 5))
+       write_shared_A(3, ((t1 + 3) % 4) + 5, c3);
+   }
+   if (t2 >= t1 + 1 && t2 <= 4 && t1 >= 1)
+     write_shared_A(3, t1 + 4, t2 + 32);
+   write_shared_A(4, ((t1 + 3) % 4) + 1, ((t2 + 31) % 32) + 1);
+   if (t1 % 3 == 0 && t2 >= 1 && t2 <= 2)
+     write_shared_A(4, (-t1 + 12) / 3, t2 + 32);
+   {
+     int c3 = t2 <= 1 || (t2 >= 2 && ((3 * t1 + t2 + 3) % 4) + 1 >= t2) ? t2 + 32 : t2;
++    if (c3 == t2 + 32 || (c3 == t2 && ((-t1 + 8) % 4) + t2 >= ((t2 + 1) % 2) + 5))
+       write_shared_A(4, ((t1 + 3) % 4) + 5, c3);
+   }
+   if (t2 >= t1 + 1 && t2 <= 4 && t1 >= 1)
+     write_shared_A(4, t1 + 4, t2 + 32);
+ }