[llvm] [AArch64] Add patterns for selecting dec from sub (PR #209467)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

> To me this feels like it should be a canonicalisation to remove the need for what's effectively a duplicate set of patterns. Looking at `DAGCombiner::visitSUB()` I can see the combine exists but is currently restricted to single uses. I don't see a reason for that and think "vscale(imm)" should be treated like any other constant.

It looks like removing the single use restriction makes things worse in cases where vscale*-C can't be folded into all its uses, e.g. in multiple_uses_sub_vscale_i64 in llvm/test/CodeGen/AArch64/sve-vscale-combine.ll where currently we get
```
multiple_uses_sub_vscale_i64:           // @multiple_uses_sub_vscale_i64
// %bb.0:
        rdvl    x8, #1
        lsr     x8, x8, #4
        sub     x9, x0, x8
        add     x8, x1, x8
        mul     x0, x9, x8
        ret
```
but with doing that we get
```
multiple_uses_sub_vscale_i64:           // @multiple_uses_sub_vscale_i64
// %bb.0:                                                                                           
        rdvl    x8, #1   
        rdvl    x9, #-1                                                                             
        lsr     x8, x8, #4                                                                          
        asr     x9, x9, #4
        add     x9, x0, x9
        add     x8, x1, x8
        mul     x0, x9, x8
        ret    
```
There's no incq/decq instructions, so we need to do rdvl+shift to get vscale, and because we have both +vscale and -vscale we need to do it twice.

It also makes things worse for load/store addressing, e.g. in dech_inch_scalar_i64 in llvm/test/CodeGen/AArch64/sve-vl-arith.ll where we have ``store val (add x, vscale * -8)``. This gets selected to ``STR val, [x, vscale*-8]``, but there's no cnth with negative immediate so we end up with cnth+neg:
```
dech_inch_scalar_i64:                   // @dech_inch_scalar_i64
        .cfi_startproc
// %bb.0:
        cnth    x8
        inch    x0
        neg     x8, x8
        str     x0, [x1, x8]
        ret
```

https://github.com/llvm/llvm-project/pull/209467