llvm-project

dee03e7d - [mlir][xegpu] Skip scalable gather/scatter lowering (#234725)

[mlir][xegpu] Skip scalable gather/scatter lowering (#234725) Leave scalable `vector.gather` and `vector.scatter` operations unconverted because XeGPU load/store offsets and masks require fixed-size vectors. Check this before creating offset arithmetic to avoid mixing scalable and fixed-size vector types and triggering a folding assertion. Add regression coverage for retaining scalable gathers and scatters. Fixes #202531 Assisted-by: Codex / GPT-6