Vectorization: Skip scalar epilogue when it's unnecessary.
The vectorizer now avoids generating a scalar epilogue for loops when it's not needed, improving performance.
This patch optimizes vectorized loops by preventing the generation of unnecessary scalar epilogues. Specifically, when a loop has no live-out values, only one exit point, and no side effects, the scalar epilogue is skipped. This avoids redundant scalar code execution, leading to improved performance. For example, loops that perform comparisons and return based on the results can now avoid the scalar epilogue.
In Details
The patch introduces LOOP_VINFO_EARLY_BRK_NEEDS_EPILOG and adds early_break_needs_epilogue to _loop_vec_info. It modifies tree-vect-data-refs.cc to determine when the scalar epilogue is not needed. This optimization is particularly relevant for loops with early exits where no further computation is required.
For Context
Vectorization aims to improve performance by processing multiple data elements in parallel using SIMD instructions. After a vectorized loop, a scalar epilogue might be needed to handle remaining elements or perform final computations. However, in some cases, the scalar epilogue is redundant and can be skipped. This optimization identifies such cases and prevents the generation of the unnecessary scalar epilogue, leading to improved performance by reducing redundant code execution.