Improving application performance is an important practical task for supercomputer calculations. Along with parallelization of calculations between cluster nodes (for example, using MPI tools), as well as multithreaded programming (for example, using OpenMP), program code vectorization is used, which provides parallelism at the level of individual instructions. The AVX-512 vector instruction set of the Intel microprocessor architecture has a number of unique features that can be used to vectorize program code from a very wide range of applications. The use of special mask registers makes it possible to effectively vectorize code containing conditional statements, and the use of profile information about the probability of performing operations in the source code allows performing code transformations that lead to even more efficient use of automatic vectorization. The paper considers the splitting transformation of a flat loop by a condition when we have knowledge about of a high probability of this condition. Practical problems in which this context is present are considered. For these problems, the conditions for splitting cycles to achieve efficient vectorization are highlighted. The described transformations were performed, and the effectiveness of the resulting code was verified by runs on microprocessors of the Intel Xeon Cascade Lake and Intel Xeon Phi Knights Landing families, the results of the runs are presented.