[llvm-branch-commits] [clang] [lld] [llvm] [mlir] [RFC][AMDGPU] Add BARRIER address space (PR #195613)

[Nicolai Hähnle via llvm-branch-commits]

================
@@ -3322,10 +3341,32 @@ bool AMDGPULegalizerInfo::legalizeGlobalValue(
   MachineFunction &MF = B.getMF();
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
 
+  const auto TrapAndPoison = [&] {
+    B.buildTrap();
+    B.buildUndef(DstReg);
+    MI.eraseFromParent();
+    return true;
+  };
+
+  if (AS == AMDGPUAS::BARRIER) {
+    const GlobalVariable *GVar = cast<GlobalVariable>(GV);
+    if (!AMDGPU::isNamedBarrier(*GVar)) {
+      const Function &Fn = MF.getFunction();
+      Fn.getContext().diagnose(DiagnosticInfoUnsupported(
+          Fn, "Unsupported use of BARRIER address space!", MI.getDebugLoc(),
+          DS_Error));
+      return TrapAndPoison();
+    }
+
+    B.buildConstant(DstReg,
+                    MFI->allocateBarrierGlobal(B.getDataLayout(), *GVar));
+    MI.eraseFromParent();
+    return true;
+  }
----------------
nhaehnle wrote:

This isn't super crucial, but I'd make it an IR validity rule that global variables in this address space have to be of the right type, so the duplicate code for inserting traps here could be removed (and optionally replaced with an `assert`).

https://github.com/llvm/llvm-project/pull/195613