[llvm] [VPlan] Extract reverse operation for reverse accesses (PR #146525)

[Luke Lau via llvm-commits]

================
@@ -3341,3 +3363,34 @@ void VPlanTransforms::addBranchWeightToMiddleTerminator(
       MDB.createBranchWeights({1, VectorStep - 1}, /*IsExpected=*/false);
   MiddleTerm->addMetadata(LLVMContext::MD_prof, BranchWeights);
 }
+
+void VPlanTransforms::adjustRecipesForReverseAccesses(VPlan &Plan) {
----------------
lukel97 wrote:

What I guess I would eventually like to see is that our optimisations to remove the header masks just become plain old peepholes, written pattern match style like in similarRecipes. 

E.g. for a regular load we would try and match:

```
(load ptr, header-mask) -> (vp.load ptr, all-true, evl)
```

And if we were to split out the reverses for both the data and mask into separate recipes, and add a VF operand like what we currently do for the end pointer, we would also have:

```   
(reverse (load (end-ptr p, vf), (reverse header-mask, vf)), vf)
   ->
(reverse (vp.load (end-ptr p, evl), all-true, evl), evl)
OR
(vp.strided-load p, all-true, stride=-1, evl)
```

I think these patterns seem simple enough, and we could probably write them with the VPlanPatternMatch. Most importantly, these transformations don't change the semantics and are just optimisations. So if somehow we miss one of these transforms it will still be correct.

For permutation elimination we would just need to have:

```
(reverse (reverse x N) N) -> x
```


https://github.com/llvm/llvm-project/pull/146525