[llvm] [RISCV] Recognize VLA shift pairs from shuffle masks (PR #127710)

[Philip Reames via llvm-commits]

================
@@ -822,7 +820,7 @@ define void @vnsrl_32_i32_two_source(ptr %in0, ptr %in1, ptr %out) {
 ; V-NEXT:    vle32.v v8, (a0)
 ; V-NEXT:    vle32.v v9, (a1)
 ; V-NEXT:    vmv.v.i v0, 1
-; V-NEXT:    vrgather.vi v9, v8, 1, v0.t
+; V-NEXT:    vslidedown.vi v9, v8, 1, v0.t
----------------
preames wrote:

I think this is a BP3 oddity, and not something we should tune for.  Having vrgather.vi, vslidedown.vi and vslideup.vi all be the same cost at mf2, but for there to be a constant 1 cycle difference at all higher LMUL seems quite odd.  I'm honestly a bit suspicious of the numbers at all.  

https://github.com/llvm/llvm-project/pull/127710