RMK012: The vectorization cost model states the loop is not a SIMD opportunity because conditional execution renders vectorization inefficient
The loop appears to be vectorizable, but the statements in the conditions of the loop render vectorization inefficient.
If the conditions in the loop do not depend on the data, remove them using loop unswitching or loop fission.
The main assumption of vectorization is that the same operation is applied to a vector of data instead of to a single piece of data. Loops with conditional statements do the opposite: they apply different operations to the data depending on some condition.
This doesn't mean that vectorization is impossible because conditional statements can be simulated in a vectorization framework: what was earlier a conditional statement is now executed unconditionally, but in the final steps the results are filtered out if the condition is not true. Because the processor is executing instructions whose result will possibly be thrown away, this decreases the efficiency of vectorization.
Occasionally, the loops with conditional statements can be vectorized efficiently:
-
If most statements in the loop are executed unconditionally, but there are a few of them that are executed conditionally, the loop can benefit from vectorization.
-
If the condition in the loop is always evaluated to a loop-invariant value (i.e. its value is either true or false across the execution of the loop), this condition can be moved outside of the loop (see loop unswitching).
-
If the condition in the loop depends on iterator variables only, the conditions can be removed by splitting the loop into several loops using loop fission.