llvm-project

1301afd8 - [LLVM][CodeGen][SVE] Improve lowering for v1f32/f64 when NEON is not available. (#229724)

[LLVM][CodeGen][SVE] Improve lowering for v1f32/f64 when NEON is not available. (#229724) When NEON is not available it is better to scalarise single element floating-point vectors than widening them to use Streaming-SVE. Explicitly make v1f64 scalar_to_vector operations always legal, because we can use scalar instructions and add a combine to avoid "nop" casts.