[llvm] ISel/AArch64/SVE: custom lower vector ISD::[L]LRINT (PR #89035)

[David Green via llvm-commits]

davemgreen wrote:

Hi - Can we expand LRINT/LLRINT to just do `frintx + [fcvtl] + fcvtzs`? It seems that we scalarize the vector operations at the moment, and the scalar operations work that way.

Could we just be lowering LRINT/LLRINT to ISD::FRINT+ISD::FP_TO_SINT and let the rest of legalization deal with the results? It should then hopefully be able to handle any vector types without much extra work.

https://github.com/llvm/llvm-project/pull/89035