[llvm] [CodeGen][NewPM] Port RegAllocEvictionAdvisor analysis to NPM (PR #117309)

[Akshat Oke via llvm-commits]

================
@@ -191,13 +193,66 @@ class RegAllocEvictionAdvisorAnalysis : public ImmutablePass {
   const AdvisorMode Mode;
 };
 
+/// Common provider for legacy and new pass managers.
+/// This keeps the state for logging, and sets up and holds the provider.
+/// The legacy pass itself used to keep the logging state and provider,
+/// so this extraction helps the NPM analysis to reuse the logic.
+class RegAllocEvictionAdvisorProvider {
+public:
+  enum class AdvisorMode : int { Default, Release, Development };
+  RegAllocEvictionAdvisorProvider(AdvisorMode Mode) : Mode(Mode) {}
+
+  virtual ~RegAllocEvictionAdvisorProvider() = default;
+
+  virtual bool doInitialization(Module &M) { return false; }
+
+  virtual void logRewardIfNeeded(const MachineFunction &MF,
+                                 llvm::function_ref<float()> GetReward) {}
+
+  virtual std::unique_ptr<RegAllocEvictionAdvisor>
+  getAdvisor(const MachineFunction &MF, const RAGreedy &RA) = 0;
+
+  /// Set the analyses that the advisor needs to use as they might not be
+  /// available before the advisor is created.
+  virtual void setAnalyses(std::initializer_list<llvm::Any> AnalysisP) {}
+
+  AdvisorMode getAdvisorMode() const { return Mode; }
+
+private:
+  const AdvisorMode Mode;
+};
+
+RegAllocEvictionAdvisorProvider *createReleaseModeAdvisorProvider();
----------------
optimisan wrote:

They do for the legacy analyses, but these are only used for the new (and only one) analysis.
Without the methods, the new PM analysis has to be in `MLRegAlloc..cpp` to keep the ML advisors local there, and also have to move the Default Provider out to `llvm::`.
Or - just separate out the ML advisor initialization in the new analysis to the `MLRegAlloc..cpp`?

https://github.com/llvm/llvm-project/pull/117309