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

[David Sherwood via llvm-commits]

================
@@ -1835,6 +1835,27 @@ void VPlanTransforms::simplifyRecipes(VPlan &Plan) {
   }
 }
 
+void VPlanTransforms::simplifyReverses(VPlan &Plan) {
+  auto m_ReverseReverse = [](auto X) {
+    VPValue *EVL;
+    return m_CombineOr(
+        m_Reverse(m_Reverse(X)),
+        m_vp_Reverse(m_vp_Reverse(X, m_VPValue(), m_VPValue(EVL)), m_VPValue(),
+                     m_Deferred(EVL)));
+  };
+
+  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
+           vp_depth_first_deep(Plan.getEntry()))) {
+    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+      VPValue *X;
+      if (match(&R, m_ReverseReverse(m_VPValue(X)))) {
----------------
david-arm wrote:

In general I have also seen IR generated the vectoriser like this:

%rev1 = reverse %a
%fadd = fadd fast %rev1, splat (float 1.0e+00)
%rev2 = reverse %fadd

where there is only one use of %rev1 and one use of %fadd. I think these pair-reverse simplifications could be used in more places to reduce the cost, since instcombine would fold away these reverses.

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