[llvm] [MachinePipeliner] Add an abstract layer to manipulate Data Dependenc… (PR #109918)

[Ryotaro Kasuga via llvm-commits]

================
@@ -1675,24 +1653,33 @@ void SwingSchedulerDAG::Circuits::createAdjacencyStructure(
         OutputDeps[N] = BackEdge;
       }
       // Do not process a boundary node, an artificial node.
-      // A back-edge is processed only if it goes to a Phi.
-      if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() ||
-          (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI()))
+      if (OE.getDst()->isBoundaryNode() || OE.isArtificial())
+        continue;
+
+      // This code is retained o preserve previous behavior and prevent
+      // regression. This condition means that anti-dependnecies within an
+      // iteration are ignored when searching circuits. Therefore it's natural
+      // to consider this dependence as well.
+      // FIXME: Remove this code if it doesn't have significant impact on
+      // performance.
+      if (OE.isAntiDep())
----------------
kasuga-fj wrote:

I don't think it is necessary. Dependencies marked as Anti and going into PHI are converted to Data in the `SwingSchedulerDDGEdge` ctor.

https://github.com/llvm/llvm-project/pull/109918