[llvm] Fold SVE mul and mul_u to neg during isel (PR #160828)

[via llvm-commits]

================
@@ -730,6 +730,21 @@ let Predicates = [HasSVE_or_SME] in {
   defm ABS_ZPmZ  : sve_int_un_pred_arit<  0b110, "abs",  AArch64abs_mt>;
   defm NEG_ZPmZ  : sve_int_un_pred_arit<  0b111, "neg",  AArch64neg_mt>;
 
+  // mul x (splat -1) -> neg x
+  let Predicates = [HasSVE_or_SME] in {
+    def : SVE_2_Op_Neg_One_Passthru_Pat<nxv16i8, AArch64mul_m1, nxv16i1, NEG_ZPmZ_B , i32>;
+    def : SVE_2_Op_Neg_One_Passthru_Pat<nxv8i16, AArch64mul_m1, nxv8i1, NEG_ZPmZ_H , i32>;
+    def : SVE_2_Op_Neg_One_Passthru_Pat<nxv4i32, AArch64mul_m1, nxv4i1, NEG_ZPmZ_S , i32>;
+    def : SVE_2_Op_Neg_One_Passthru_Pat<nxv2i64, AArch64mul_m1, nxv2i1, NEG_ZPmZ_D , i64>;
+
+    let AddedComplexity = 5 in {
+    defm : SVE_2_Op_Neg_One_Passthru_Pat_Comm<nxv16i8, AArch64mul_p, nxv16i1, NEG_ZPmZ_B , i32>;
----------------
Lukacma wrote:

Even though the generated assembly doesn't seem affected, these should lower to _UNDEF variant of NEG.

https://github.com/llvm/llvm-project/pull/160828