[llvm] [VPlan] Simplify reverse(reverse(x)) -> x (PR #199057)

[Luke Lau via llvm-commits]

================
@@ -3138,11 +3148,22 @@ void VPlanTransforms::optimizeEVLMasks(VPlan &Plan) {
     }
   }
 
-  // Fold the following splice patterns into vp.reverse for reverse accesses:
+  // Fold the following splice patterns:
+  //   splice.right(splice.left(poison, x, evl), poison, evl) -> x
   //   vector.reverse(splice.left(poison, x, evl))  -> vp.reverse(x, true, evl)
   //   splice.right(vector.reverse(x), poison, evl) -> vp.reverse(x, true, evl)
   for (VPUser *U : collectUsersRecursively(EVL)) {
+    auto *Def = cast<VPRecipeBase>(U);
     VPValue *X;
+    if (match(U, m_Intrinsic<Intrinsic::vector_splice_right>(
----------------
lukel97 wrote:

The regression isn't with the cost model but with the codegen.  InstCombine currently cancels out the reverses when VPlan doesn't, but if we cancel out the reverses in VPlan but leave behind the splices, InstCombine won't remove them. 

E.g. this loop will regress:

```c
void f(int *x, int *y, int n) {
  for (int i = n - 1; i >= 0; i--)
    x[i] = y[i];
}
```

With splice elimination:

```asm
.LBB0_5:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
	vsetvli	a7, a5, e8, mf2, ta, ma
	not	a3, a4
	add	t1, a3, a2
	sh2add	t0, t1, a1
	sub	t2, a6, a7
	sh2add	a3, t2, t0
	vle32.v	v8, (a3)
	sh2add	a3, t1, a0
	sh2add	a3, t2, a3
	sub	a5, a5, a7
	vse32.v	v8, (a3)
	add	a4, a4, a7
	bnez	a5, .LBB0_5
```

Without

```asm
	vsetvli	t0, a4, e8, mf2, ta, ma
	not	a3, a5
	add	t2, a3, a2
	sh2add	t1, t2, a1
	sub	t3, a6, t0
	sh2add	a3, t3, t1
	vle32.v	v8, (a3)
	sub	t1, a7, t0
	vsetvli	a3, zero, e32, m2, ta, ma
	vslideup.vx	v10, v8, t1
	vsetvli	zero, a4, e32, m2, ta, ma
	vslidedown.vx	v8, v10, t1
	sh2add	a3, t2, a0
	sh2add	a3, t3, a3
	sub	a4, a4, t0
	vse32.v	v8, (a3)
	add	a5, a5, t0
	bnez	a4, .LBB0_5
```


https://github.com/llvm/llvm-project/pull/199057