[llvm] [AMDGPU] Expand constant named barrier matching in SDag (PR #191506)

[Krzysztof Drewniak via llvm-commits]

================
@@ -11818,6 +11818,29 @@ static void initializeM0ToZeroForClusterLoad(SDValue Op, SelectionDAG &DAG,
   DAG.UpdateNodeOperands(N, Ops);
 }
 
+// Try to resolve a named barrier object to a constant barrier ID.
+// Named barrier IDs are encoded at bits [5:0] of a BARRIER address space
+// address.
+static std::optional<uint32_t>
+getConstantNamedBarrierNumber(SDValue BarOp, SelectionDAG &DAG) {
+  std::optional<uint64_t> BarVal;
+  if (auto *C = dyn_cast<ConstantSDNode>(BarOp)) {
+    BarVal = C->getZExtValue();
+  } else if (auto *GA = dyn_cast<GlobalAddressSDNode>(BarOp)) {
+    const auto *GV = cast<GlobalVariable>(GA->getGlobal());
+    if (AMDGPU::isNamedBarrier(*GV) &&
+        AMDGPUMachineFunctionInfo::get32BitAbsoluteAddress(*GV,
+                                                           AMDGPUAS::BARRIER)) {
+      auto *MFI = DAG.getMachineFunction().getInfo<SIMachineFunctionInfo>();
+      BarVal = MFI->allocateBarrierGlobal(DAG.getDataLayout(), *GV) +
+               GA->getOffset();
+    }
+  }
+  if (BarVal)
----------------
krzysz00 wrote:

What about poison? Should we just pick something arbitrarly here?

https://github.com/llvm/llvm-project/pull/191506