[Mlir-commits] [mlir] [mlir] Translating task_reduction clause for pass-by-value vars to LLVMIR (PR #125218)

[Tom Eccles]

================
@@ -2469,6 +2473,228 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
   return success();
 }
 
+template <typename OP>
+static llvm::Value *createTaskReductionFunction(
+    llvm::IRBuilderBase &builder, const std::string &name, llvm::Type *redTy,
+    LLVM::ModuleTranslation &moduleTranslation,
+    SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls, Region &region,
+    OP &op, unsigned cnt,
+    SmallVectorImpl<llvm::Value *> &privateReductionVariables,
+    DenseMap<Value, llvm::Value *> &reductionVariableMap) {
+
+  llvm::LLVMContext &Context = builder.getContext();
+  // TODO: by-ref reduction variables are yet to be handled.
+  llvm::Type *OpaquePtrTy = llvm::PointerType::get(Context, 0);
+  if (region.empty() && name == "red_fini")
+    // Finalization is optional for reductions.
+    return llvm::Constant::getNullValue(OpaquePtrTy);
+
+  // Prepare a general structure of the function to be emitted
+  llvm::FunctionType *funcType =
+      llvm::FunctionType::get(OpaquePtrTy, {OpaquePtrTy, OpaquePtrTy}, false);
+  llvm::Function *function =
+      llvm::Function::Create(funcType, llvm::Function::ExternalLinkage, name,
+                             builder.GetInsertBlock()->getModule());
+  function->setDoesNotRecurse();
+  llvm::BasicBlock *entry =
+      llvm::BasicBlock::Create(Context, "entry", function);
+  llvm::IRBuilder<> bbBuilder(entry);
+
+  // Prepare the function arguments
+  llvm::Value *arg0 = function->getArg(0);
+  llvm::Value *arg1 = function->getArg(1);
+
+  if (name == "red_init") {
+    // For the initialization, map the reduction variables
+    // to the arguments of the function
+    function->addParamAttr(0, llvm::Attribute::NoAlias);
+    function->addParamAttr(1, llvm::Attribute::NoAlias);
+    mlir::omp::DeclareReductionOp &reduction = reductionDecls[cnt];
+    Region &initializerRegion = reduction.getInitializerRegion();
+    Block &entry = initializerRegion.front();
+
+    mlir::Value mlirSource = op.getTaskReductionVars()[cnt];
+    llvm::Value *llvmSource = moduleTranslation.lookupValue(mlirSource);
+    llvm::Value *origVal = llvmSource;
+
+    moduleTranslation.mapValue(reduction.getInitializerMoldArg(), origVal);
+
+    if (entry.getNumArguments() > 1) {
+      llvm::Value *allocation =
+          reductionVariableMap.lookup(op.getReductionVars()[cnt]);
+      moduleTranslation.mapValue(reduction.getInitializerAllocArg(),
+                                 allocation);
+    }
+
+  } else if (name == "red_comb") {
+    // For the combiner, perform a load for each argument
+    // and map it to the combiner region.
+    llvm::Value *arg0L = bbBuilder.CreateLoad(redTy, arg0);
+    llvm::Value *arg1L = bbBuilder.CreateLoad(redTy, arg1);
+    moduleTranslation.mapValue(region.front().getArgument(0), arg0L);
+    moduleTranslation.mapValue(region.front().getArgument(1), arg1L);
+  }
+
+  // Emit an empty function body in case of empty region
+  if (region.empty()) {
+    bbBuilder.CreateRet(arg0); // Return from the function
+    return function;
+  }
+
+  SmallVector<llvm::Value *, 1> phis;
+  if (failed(inlineConvertOmpRegions(region, "", bbBuilder, moduleTranslation,
+                                     &phis)))
+    return nullptr;
+  assert(
+      phis.size() == 1 &&
+      "expected one value to be yielded from the reduction declaration region");
+  bbBuilder.CreateStore(phis[0], arg0);
+  bbBuilder.CreateRet(arg0); // Return from the function
+  return function;
+}
+
+void emitTaskRedInitCall(
+    llvm::IRBuilderBase &builder, LLVM::ModuleTranslation &moduleTranslation,
+    const llvm::OpenMPIRBuilder::LocationDescription &ompLoc, int arraySize,
+    llvm::Value *ArrayAlloca) {
+
+  llvm::LLVMContext &Context = builder.getContext();
+  uint32_t SrcLocStrSize;
+  llvm::Constant *SrcLocStr =
+      moduleTranslation.getOpenMPBuilder()->getOrCreateSrcLocStr(ompLoc,
+                                                                 SrcLocStrSize);
+  llvm::Value *Ident = moduleTranslation.getOpenMPBuilder()->getOrCreateIdent(
+      SrcLocStr, SrcLocStrSize);
+  llvm::Value *ThreadID =
+      moduleTranslation.getOpenMPBuilder()->getOrCreateThreadID(Ident);
+  llvm::Constant *ConstInt =
+      llvm::ConstantInt::get(llvm::Type::getInt32Ty(Context), arraySize);
+
+  llvm::Function *TaskRedInitFn =
+      moduleTranslation.getOpenMPBuilder()->getOrCreateRuntimeFunctionPtr(
+          llvm::omp::OMPRTL___kmpc_taskred_init);
+  builder.CreateCall(TaskRedInitFn, {ThreadID, ConstInt, ArrayAlloca});
+}
+
+template <typename OP>
+static LogicalResult allocAndInitializeTaskReductionVars(
+    OP op, ArrayRef<BlockArgument> reductionArgs, llvm::IRBuilderBase &builder,
+    LLVM::ModuleTranslation &moduleTranslation,
+    llvm::OpenMPIRBuilder::InsertPointTy &allocaIP,
+    SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls,
+    SmallVectorImpl<llvm::Value *> &privateReductionVariables,
+    DenseMap<Value, llvm::Value *> &reductionVariableMap) {
+
+  if (op.getNumReductionVars() == 0)
+    return success();
+
+  llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
+  llvm::LLVMContext &Context = builder.getContext();
+  SmallVector<DeferredStore> deferredStores;
+
+  // Save the current insertion point
+  auto oldIP = builder.saveIP();
+
+  // Set insertion point after the allocations
+  builder.SetInsertPoint(allocaIP.getBlock()->getTerminator());
+
+  // Define the kmp_taskred_input_t structure
+  llvm::StructType *kmp_taskred_input_t =
+      llvm::StructType::create(Context, "kmp_taskred_input_t");
+  llvm::Type *OpaquePtrTy = llvm::PointerType::get(Context, 0); // void*
+  llvm::Type *SizeTy = builder.getInt64Ty(); // size_t (assumed to be i64)
----------------
tblah wrote:

You could use llvm::DataLayout::getIntPtrType for a portable size_t equivalent.

https://github.com/llvm/llvm-project/pull/125218