[llvm] [RISCV] Fold vector shift of sext/zext to widening multiply (PR #121563)

[Piotr Fusik via llvm-commits]

================
@@ -4943,16 +4943,17 @@ define <8 x i64> @mgather_baseidx_v8i16_v8i64(ptr %base, <8 x i16> %idxs, <8 x i
 ; RV32ZVE32F-LABEL: mgather_baseidx_v8i16_v8i64:
 ; RV32ZVE32F:       # %bb.0:
 ; RV32ZVE32F-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
-; RV32ZVE32F-NEXT:    vsext.vf2 v10, v8
+; RV32ZVE32F-NEXT:    vmv.v.x v10, a1
+; RV32ZVE32F-NEXT:    li a1, 8
 ; RV32ZVE32F-NEXT:    vsetvli zero, zero, e8, mf2, ta, ma
 ; RV32ZVE32F-NEXT:    vmv.x.s t0, v0
-; RV32ZVE32F-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
-; RV32ZVE32F-NEXT:    vsll.vi v8, v10, 3
 ; RV32ZVE32F-NEXT:    andi a3, t0, 1
-; RV32ZVE32F-NEXT:    vadd.vx v8, v8, a1
+; RV32ZVE32F-NEXT:    vsetvli zero, zero, e16, m1, ta, ma
+; RV32ZVE32F-NEXT:    vwmaccus.vx v10, a1, v8
----------------
pfusik wrote:

`vsext`+`vsll`+`vadd` folded into `vwmaccus`.

https://github.com/llvm/llvm-project/pull/121563