[llvm] [LV] Use ExtractLane(LastActiveLane, V) live outs when tail-folding. (PR #149042)

[Florian Hahn via llvm-commits]

================
@@ -3342,6 +3370,34 @@ void VPlanTransforms::convertToConcreteRecipes(VPlan &Plan) {
         ToRemove.push_back(Expr);
       }
 
+      // Expand LastActiveLane into Not + FirstActiveLane + Sub.
+      auto *LastActiveL = dyn_cast<VPInstruction>(&R);
+      if (LastActiveL &&
+          LastActiveL->getOpcode() == VPInstruction::LastActiveLane) {
+        // Create Not(Mask) for all operands.
+        SmallVector<VPValue *, 2> NotMasks;
+        for (VPValue *Op : LastActiveL->operands()) {
+          VPValue *NotMask = Builder.createNot(Op, LastActiveL->getDebugLoc());
+          NotMasks.push_back(NotMask);
+        }
+
+        // Create FirstActiveLane on the inverted masks.
+        VPValue *FirstInactiveLane = Builder.createNaryOp(
----------------
fhahn wrote:

Having the LastActiveLane check separate has the potential advantage that it could expose more CSE/simplification opportuntiies. The restriction for prefix-masks at the moment is only there due to how it gets lowered using FirstActiveLane.

I think we can just remove the restriction once we have a use-case for non-prefix-masks and add dedicated lowering that supports them.

https://github.com/llvm/llvm-project/pull/149042