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

[Michael Marjieh via llvm-commits]

================
@@ -114,10 +115,123 @@ class MachinePipeliner : public MachineFunctionPass {
   bool useWindowScheduler(bool Changed);
 };
 
+/// Represents a dependence between two instruction.
+class SwingSchedulerDDGEdge {
+  SUnit *Dst = nullptr;
+  SDep Pred;
+  unsigned Distance = 0;
+
+public:
+  /// Creates an edge corresponding to an edge represented by \p PredOrSucc and
+  /// \p Dep in the original DAG. This pair has no information about the
+  /// direction of the edge, so we need to pass an additional argument \p
+  /// IsSucc.
+  SwingSchedulerDDGEdge(SUnit *PredOrSucc, const SDep &Dep, bool IsSucc)
+      : Dst(PredOrSucc), Pred(Dep), Distance(0u) {
+    SUnit *Src = Dep.getSUnit();
+
+    if (IsSucc) {
+      std::swap(Src, Dst);
+      Pred.setSUnit(Src);
+    }
+
+    // An anti-dependence to PHI means loop-carried dependence.
+    if (Pred.getKind() == SDep::Anti && Src->getInstr()->isPHI()) {
+      Distance = 1;
+      std::swap(Src, Dst);
+      auto Reg = Pred.getReg();
+      Pred = SDep(Src, SDep::Kind::Data, Reg);
----------------
mmarjieh wrote:

Let's track what happens when sending this edge to your DDGEdge constructor:
Input:
some_instruction -> phi; anti dep

When evaluation the phi's successor to initialize the edge, DDGEdge will be called with the following arguments:
![image](https://github.com/user-attachments/assets/a8cdfac4-fb90-4699-baa8-d7caa3208444)

PredOrSucc: phi
IsSucc = true,
Dep: some_instruction; anti dep


First you swap Src and Dst:
![image](https://github.com/user-attachments/assets/80b2af70-8268-4b48-bbcf-9885cc53de98)

Meaning now we have:
Src = phi
Dst = some_instruction
Pred's SUnit will be: phi

Then this code will be executed:
![image](https://github.com/user-attachments/assets/44139e80-fbb3-4d88-9f31-91644e2ba25f)
Meaning the edge lost its "anti" kind and now we model it as data.


My problem is that now we don't consider that
some_instruction is a successor of the phi in some parts of the pipeliner's code:
computeNodeOrder
succ_L

I believe that, in addition to checking the anti-deps on the input edges, we need to check edges that have a non-zero distance.





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