llvm-project
0de93f30 - [NVPTX] Mark SMUL_LOHI and UMUL_LOHI as Expand for v2i16 (#225808)

Commit
13 days ago
[NVPTX] Mark SMUL_LOHI and UMUL_LOHI as Expand for v2i16 (#225808) `SMUL_LOHI` and `UMUL_LOHI` were inadvertently left as `Legal` for `v2i16` in NVPTX, despite the backend having no instruction patterns to select them. As a result, the vector legalizer expands `v2i16` high-multiplies (`mulhs`/`mulhu`) and constant divisions into `v2i16` `smul_lohi`/`umul_lohi` nodes instead of scalarizing them into `i16` operations, causing an instruction selection crash. Marking both operations as `Expand` for `v2i16` ensures they are properly scalarized. Assisted-by: Gemini --------- Co-authored-by: Yonah Goldberg <ygoldberg@nvidia.com>
Parents
Loading