[llvm] [AArch64][SVE] Select immediate offsets for multi-vector instructions (PR #201637)

[Benjamin Maxwell via llvm-commits]

================
@@ -4564,6 +4573,16 @@ defm : store_pn_x2<nxv2f64, int_aarch64_sve_stnt1_pn_x2, STNT1D_2Z_IMM>;
 
 multiclass store_pn_x4<ValueType Ty, SDPatternOperator Store,
                         Instruction RegImmInst> {
+  let AddedComplexity = 1 in {
+    // scalar + immediate (mul vl)
+    def : Pat<(Store Ty:$vec0, Ty:$vec1, Ty:$vec2, Ty:$vec3, aarch64svcount:$PNg,
+                (am_sve_indexed_s4 GPR64sp:$base, simm4s1:$offset)),
+              (RegImmInst (REG_SEQUENCE ZPR4Mul4, Ty:$vec0, zsub0, Ty:$vec1, zsub1,
+                                                  Ty:$vec2, zsub2, Ty:$vec3, zsub3),
+                          PNR:$PNg, GPR64:$base, simm4s1:$offset)>;
----------------
MacDue wrote:

That does not really work with `am_sve_indexed_s4`, particularly if placed on the result pattern as it'll divide the result of `am_sve_indexed_s4` by the scale, but `am_sve_indexed` has already applied the scale. 



https://github.com/llvm/llvm-project/pull/201637