SLP subgraph merging handles overlapping vector parts
SLP vectorization now merges subgraphs that overlap in low/high parts or use VEC_PERM, fixing SSA verification failures.
This commit enhances SLP (Superword Level Parallelism) vectorization by enabling the merging of subgraphs that overlap in their low or high parts, or through full vector copies using VEC_PERM nodes. It introduces mechanisms to build reverse mappings for scalar statements to SLP nodes and checks for containment within candidate subgraphs. This resolves issues in testcases like PR126053 and PR126099, preventing SSA verification failures by correctly handling overlapping vector data. New testcases are added to verify these improvements.
In Details
This patch enhances SLP vectorization by improving subgraph merging. It introduces vect_cse_gather_part_starts and vect_cse_slp_node_parts to handle merging of SLP nodes that overlap in lowpart or highpart extractions, or through VEC_PERM nodes. The logic correctly identifies overlapping parts of SLP nodes without requiring full vector equality, addressing issues like those in PR126099 where schedule verification would elide graph entries due to these overlaps. It specifically targets cases where vectorizable_slp_permutation restrictions would prevent merging.
- SLP
- Superword Level Parallelism, a compiler optimization technique that finds instructions operating on scalars that can be vectorized to operate on multiple data elements simultaneously.
- VEC_PERM
- A GIMPLE operation representing a vector permutation, used to rearrange elements within a vector.
- SSA
- Static Single Assignment form, an intermediate representation where every variable is assigned exactly once, simplifying many compiler analyses and optimizations.
- lowpart/highpart
- Refers to the lower or higher-order bits/elements of a larger data type or vector, often relevant in type promotions or vector element extraction.