[llvm] [RISCV] Keep same SEW/LMUL ratio if possible in forward transfer (PR #69788)

[Philip Reames via llvm-commits]

================
@@ -342,14 +342,14 @@ define half @vpreduce_ord_fadd_nxv64f16(half %s, <vscale x 64 x half> %v, <vscal
 ; ZVFHMIN-NEXT:    vsetvli a4, zero, e16, m4, ta, ma
 ; ZVFHMIN-NEXT:    vfwcvt.f.f.v v24, v8
 ; ZVFHMIN-NEXT:    fcvt.s.h fa5, fa0
-; ZVFHMIN-NEXT:    vsetivli zero, 1, e32, m1, ta, ma
+; ZVFHMIN-NEXT:    vsetvli zero, zero, e32, m8, ta, ma
 ; ZVFHMIN-NEXT:    vfmv.s.f v8, fa5
 ; ZVFHMIN-NEXT:    vsetvli zero, a0, e32, m8, ta, ma
 ; ZVFHMIN-NEXT:    vfredosum.vs v8, v24, v8, v0.t
 ; ZVFHMIN-NEXT:    vfmv.f.s fa5, v8
 ; ZVFHMIN-NEXT:    fcvt.h.s fa5, fa5
 ; ZVFHMIN-NEXT:    fcvt.s.h fa5, fa5
-; ZVFHMIN-NEXT:    vsetivli zero, 1, e32, m1, ta, ma
+; ZVFHMIN-NEXT:    vsetivli zero, 1, e32, m8, ta, ma
----------------
preames wrote:

Just to flag this - this points out that we're changing LMUL without considering whether having the same ratio is enough to actually emit the vl preserving form.  I'm noting this for the record, but I do not think this is a practical problem.  

https://github.com/llvm/llvm-project/pull/69788