SLP vectorizer now skips reduction subgroups of size one.
GCC’s SLP vectorizer now explicitly avoids creating size-one reduction subgroups, improving code generation quality.
GCC’s SLP vectorizer now requires reduction subgroups to have a size greater than one. Previously, size-one groups were analyzed, leading to inefficient code generation. The vect_analyze_slp_reduction_group function now returns false for groups of size one or less, ensuring these cases correctly fall back to single-lane reductions and avoiding suboptimal output.
In Details
Within GCC's tree-optimization pass, specifically in tree-vect-slp.cc, the SLP (Superword Level Parallelism) vectorizer's reduction subgroup analysis is refined. The vect_analyze_slp_reduction_group function now explicitly checks group_size and returns false if it is less than or equal to one. This prevents the vectorizer from attempting SLP on trivial, single-element reductions. While such groups might technically 'succeed' in the analysis, their vectorization typically leads to worse code than scalar execution, often introducing unnecessary shuffle or extract operations. By making r…
For Context
Compilers like GCC try to make your code run faster by using special processor features, one of which is called vectorization. This means performing the same operation on multiple pieces of data simultaneously, often by grouping them into 'vectors.' A particular technique called SLP (Superword Level Parallelism) looks for patterns in your code that can be vectorized, especially for operations that combine many values into a single result, known as reductions (like summing an array). This update in GCC improves how its SLP vectorizer handles these reductions. Previously, if the compiler found a 'group' of just one item to reduce, it would still try to vectorize it. This often made the code slower because the overhead of setting up the vector operation outweighed any benefit. Now, the compiler is smarter: if a reduction group only has one element, it skips the vectorization attempt and reverts to a simpler, more efficient single-item operation. This leads to better performance for such…