[llvm] [AMDGPU] Register allocation anti-hints to reduce MFMA hazard NOPs (PR #156943)

[Diana Picus via llvm-commits]

================
@@ -44,8 +45,68 @@ AllocationOrder AllocationOrder::create(Register VirtReg, const VirtRegMap &VRM,
       dbgs() << '\n';
     }
   });
-  assert(all_of(Hints,
-                [&](MCPhysReg Hint) { return is_contained(Order, Hint); }) &&
+
+  // Get anti-hints
+  SmallVector<MCPhysReg, 16> AntiHintedPhysRegs;
+  MRI.getPhysRegAntiHints(VirtReg, AntiHintedPhysRegs, &VRM);
+
+  LLVM_DEBUG({
+    if (!AntiHintedPhysRegs.empty()) {
+      dbgs() << "anti-hints:";
+      for (MCPhysReg AntiHint : AntiHintedPhysRegs)
+        dbgs() << ' ' << printReg(AntiHint, TRI);
+      dbgs() << '\n';
+    }
+  });
+
+  // Create allocation order object
+  AllocationOrder AO(std::move(Hints), Order, HardHints);
+
+  // Apply anti-hints filtering if needed
+  if (!AntiHintedPhysRegs.empty()) {
+    AO.applyAntiHints(AntiHintedPhysRegs, TRI);
+
+    LLVM_DEBUG({
+      if (!AO.Hints.empty()) {
+        dbgs() << "filtered hints:";
+        for (MCPhysReg Hint : AO.Hints)
+          dbgs() << ' ' << printReg(Hint, TRI);
+        dbgs() << '\n';
+      }
+    });
+  }
+
+  assert(all_of(AO.Hints,
+                [&](MCPhysReg Hint) { return is_contained(AO.Order, Hint); }) &&
          "Target hint is outside allocation order.");
-  return AllocationOrder(std::move(Hints), Order, HardHints);
+  return AO;
+}
+
+void AllocationOrder::applyAntiHints(ArrayRef<MCPhysReg> AntiHintedPhysRegs,
+                                     const TargetRegisterInfo *TRI) {
+  // Create filtered order
+  FilteredOrderStorage.clear();
+  FilteredOrderStorage.reserve(Order.size());
----------------
rovka wrote:

Can `FilteredOrderStorage` be local to this method? Also, would it work to use `std::stable_partition` or something instead of looping twice?

https://github.com/llvm/llvm-project/pull/156943