[llvm] [AArch64][SME] Propagate desired ZA states in the MachineSMEABIPass (PR #149510)

[Sander de Smalen via llvm-commits]

================
@@ -918,6 +987,43 @@ bool MachineSMEABI::runOnMachineFunction(MachineFunction &MF) {
       getAnalysis<EdgeBundlesWrapperLegacy>().getEdgeBundles();
 
   FunctionInfo FnInfo = collectNeededZAStates(SMEFnAttrs);
+
+  if (OptLevel != CodeGenOptLevel::None) {
+    // Propagate desired states forwards then backwards. We propagate forwards
+    // first as this propagates desired states from inner to outer loops.
+    // Backwards propagation is then used to fill in any gaps. Note: Doing both
+    // in one step can give poor results. For example:
+    //
+    //    ┌─────┐
+    //  ┌─┤ BB0 ◄───┐
+    //  │ └─┬───┘   │
+    //  │ ┌─▼───◄──┐│
+    //  │ │ BB1 │  ││
+    //  │ └─┬┬──┘  ││
+    //  │   │└─────┘│
+    //  │ ┌─▼───┐   │
+    //  │ │ BB2 ├───┘
+    //  │ └─┬───┘
+    //  │ ┌─▼───┐
+    //  └─► BB3 │
+    //    └─────┘
+    //
+    // If:
+    // - "BB0" and "BB2" (outer loop) has no state preference
----------------
sdesmalen-arm wrote:

I thought the entry block (BB0) would always have a state preference?

https://github.com/llvm/llvm-project/pull/149510