The UNROLL pragma specifies to the compiler how many times a loop should be unrolled. The UNROLL pragma is useful for helping the compiler utilize SIMD instructions. It is also useful in cases where better utilization of software pipeline resources are needed over a non-unrolled loop.
The optimizer must be invoked (use --opt_level=[1|2|3] or -O1, -O2, or -O3) in order for pragma-specified loop unrolling to take place. The compiler has the option of ignoring this pragma.
No statements are allowed between the UNROLL pragma and the for, while, or do-while loop to which it applies. However, other pragmas, such as MUST_ITERATE and PROB_ITERATE, can appear between the UNROLL pragma and the loop.
The syntax of the pragma for C and C++ is:
| #pragma UNROLL(n) |
If possible, the compiler unrolls the loop so there are n copies of the original loop. The compiler only unrolls if it can determine that unrolling by a factor of n is safe. In order to increase the chances the loop is unrolled, the compiler needs to know certain properties:
The compiler can sometimes obtain this information itself by analyzing the code. However, sometimes the compiler can be overly conservative in its assumptions and therefore generates more code than is necessary when unrolling. This can also lead to not unrolling at all. Furthermore, if the mechanism that determines when the loop should exit is complex, the compiler may not be able to determine these properties of the loop. In these cases, you must tell the compiler the properties of the loop by using the MUST_ITERATE pragma.
Specifying #pragma UNROLL(1) asks that the loop not be unrolled. Automatic loop unrolling also is not performed in this case.
If multiple UNROLL pragmas are specified for the same loop, it is undefined which pragma is used, if any.