3.15 Vectorization on loops containing pointers

When accessing arrays, the compiler can often prove that memory accesses do not overlap. When using pointers, this is less likely to be possible, and either requires a runtime test, or requires you to use the restrict keyword.

The compiler is able to vectorize loops containing pointers if it can determine that the loop is safe. Both array references and pointer references in loops are analyzed to see if there is any vector access to memory. In some cases, the compiler creates a run-time test, and executes a vector version or scalar version of the loop depending on the result of the test.

Often, function arguments are passed as pointers. If several pointer variables are passed to a function, it is possible that pointing to overlapping sections of memory can occur. Often, at runtime, this is not the case but the compiler always follows the safe method and avoids optimizing loops that involve pointers appearing on both the left and right sides of an assignment operator. For example, consider the following function.

void func (int *pa, int *pb, int x)
    int i;
    for (i = 0; i < 100; i++)
       *(pa + i) = *(pb + i) + x;

In this example, if pa and pb overlap in memory in a way that causes results from one loop pass to feed back to a subsequent loop pass, then vectorization of the loop can give incorrect results. If the function is called with the following arguments, vectorization might be ambiguous:

int *a;
func (a, a-1);

The compiler performs a runtime test to see if pointer aliasing occurs. If pointer aliasing does not occur, it executes a vectorized version of the code. If pointer aliasing occurs, the original nonvectorized code executes instead. This leads to a small cost in runtime efficiency and code size.

In practice, it is very rare for data dependence to exist because of function arguments. Programs that pass overlapping pointers are very hard to understand and debug, apart from any vectorization concerns.

In the example above, adding restrict to pa is sufficient to avoid the runtime test.

Non-ConfidentialPDF file icon PDF versionARM DUI0472M
Copyright © 2010-2016 ARM Limited or its affiliates. All rights reserved.