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

[Philip Reames via llvm-commits]

================
@@ -1259,15 +1260,16 @@ define <4 x i16> @mgather_truemask_v4i16(<4 x ptr> %ptrs, <4 x i16> %passthru) {
 ; RV64ZVE32F-NEXT:  .LBB20_5: # %cond.load
 ; RV64ZVE32F-NEXT:    ld a2, 0(a0)
 ; RV64ZVE32F-NEXT:    lh a2, 0(a2)
-; RV64ZVE32F-NEXT:    vsetivli zero, 4, e16, mf2, tu, ma
+; RV64ZVE32F-NEXT:    vsetvli zero, zero, e16, m1, tu, ma
 ; RV64ZVE32F-NEXT:    vmv.s.x v8, a2
 ; RV64ZVE32F-NEXT:    andi a2, a1, 2
 ; RV64ZVE32F-NEXT:    beqz a2, .LBB20_2
 ; RV64ZVE32F-NEXT:  .LBB20_6: # %cond.load1
 ; RV64ZVE32F-NEXT:    ld a2, 8(a0)
 ; RV64ZVE32F-NEXT:    lh a2, 0(a2)
-; RV64ZVE32F-NEXT:    vsetivli zero, 2, e16, mf2, tu, ma
+; RV64ZVE32F-NEXT:    vsetvli zero, zero, e16, m1, ta, ma
 ; RV64ZVE32F-NEXT:    vmv.s.x v9, a2
+; RV64ZVE32F-NEXT:    vsetivli zero, 2, e16, mf2, tu, ma
----------------
preames wrote:

To flag - we're actually adding a vsetvli which didn't exist.  This is arguably a regression, but a pretty minor one since the "added" vsetvli is effectively a vtype toggle (i.e. the line I commented on "moved").  I think this also goes away once the patch on review for doing this in the backwards pass lands.  

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