[Mlir-commits] [mlir] [OpenMP][mlir] Add Groupprivate op in omp dialect. (PR #162704)

[Kareem Ergawy]

================
@@ -6128,6 +6143,75 @@ convertTargetFreeMemOp(Operation &opInst, llvm::IRBuilderBase &builder,
   return success();
 }
 
+/// Converts an OpenMP Groupprivate operation into LLVM IR.
+static LogicalResult
+convertOmpGroupprivate(Operation &opInst, llvm::IRBuilderBase &builder,
+                       LLVM::ModuleTranslation &moduleTranslation) {
+  llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
+  auto groupprivateOp = cast<omp::GroupprivateOp>(opInst);
+
+  if (failed(checkImplementationStatus(opInst)))
+    return failure();
+
+  bool isTargetDevice = ompBuilder->Config.isTargetDevice();
+  auto deviceType = groupprivateOp.getDeviceType();
+
+  // skip allocation based on device_type
+  bool shouldAllocate = true;
+  if (deviceType.has_value()) {
+    switch (*deviceType) {
+    case mlir::omp::DeclareTargetDeviceType::host:
+      // Only allocate on host
+      shouldAllocate = !isTargetDevice;
+      break;
+    case mlir::omp::DeclareTargetDeviceType::nohost:
+      // Only allocate on device
+      shouldAllocate = isTargetDevice;
+      break;
+    case mlir::omp::DeclareTargetDeviceType::any:
+      // Allocate on both
+      shouldAllocate = true;
+      break;
+    }
+  }
+
+  Value symAddr = groupprivateOp.getSymAddr();
+  Operation *symOp = symAddr.getDefiningOp();
+
+  LLVM::GlobalOp global =
+      getGlobalFromSymbol(symOp, moduleTranslation, &opInst);
+  if (!global)
+    return failure();
+
+  llvm::GlobalValue *globalValue = moduleTranslation.lookupGlobal(global);
+  llvm::Value *resultPtr;
+
+  if (shouldAllocate) {
+    // Get the size of the variable
+    llvm::Type *varType = globalValue->getValueType();
+    llvm::Module *llvmModule = moduleTranslation.getLLVMModule();
+    llvm::DataLayout DL = llvmModule->getDataLayout();
+    uint64_t typeSize = DL.getTypeAllocSize(varType);
+    // Call omp_alloc_shared to allocate memory for groupprivate variable.
+    llvm::FunctionCallee allocSharedFn = ompBuilder->getOrCreateRuntimeFunction(
+        *llvmModule, llvm::omp::OMPRTL___kmpc_alloc_shared);
+    // Call runtime to allocate shared memory for this group
+    resultPtr = builder.CreateCall(allocSharedFn, {builder.getInt64(typeSize)});
+  } else {
+    // Use original global address when not allocating group-private storage
+    resultPtr = moduleTranslation.lookupValue(symAddr);
+    if (!resultPtr)
+      resultPtr = globalValue;
+  }
+
+  llvm::Type *ptrTy = builder.getPtrTy();
+  if (resultPtr->getType() != ptrTy)
+    resultPtr = builder.CreateBitCast(resultPtr, ptrTy);
----------------
ergawy wrote:

Just for my understanding, why do we need the bitcast here?

https://github.com/llvm/llvm-project/pull/162704