[Mlir-commits] [mlir] [mlir] Translating task_reduction clause for pass-by-value vars to LLVMIR (PR #125218)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Fri Jan 2 05:15:21 PST 2026
https://github.com/NimishMishra updated https://github.com/llvm/llvm-project/pull/125218
>From 55ecb7f882cd3deb7e58132280a79b5b023fae42 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Thu, 1 Jan 2026 23:24:28 +0530
Subject: [PATCH 1/7] Rebase and Address review comments
---
.../mlir/Dialect/OpenMP/OpenMPClauses.td | 8 +
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 247 +++++++++++++++++-
2 files changed, 245 insertions(+), 10 deletions(-)
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
index 05e2ee4e5632b..a1f987553e050 100644
--- a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
+++ b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
@@ -1388,6 +1388,14 @@ class OpenMP_TaskReductionClauseSkip<
OptionalAttr<SymbolRefArrayAttr>:$task_reduction_syms
);
+ let extraClassDeclaration = [{
+ /// Returns the number of reduction variables.
+ unsigned getNumReductionVars() { return getTaskReductionVars().size(); }
+
+ /// Returns the reduction symbols
+ auto getReductionSyms() { return getTaskReductionSyms(); }
+ }];
+
let description = [{
The `task_reduction` clause specifies a reduction among tasks. For each list
item, the number of copies is unspecified. Any copies associated with the
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 03d67a52853f6..2db77c628876f 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -361,11 +361,6 @@ static LogicalResult checkImplementationStatus(Operation &op) {
op.getReductionMod().value() != omp::ReductionModifier::defaultmod)
result = todo("reduction with modifier");
};
- auto checkTaskReduction = [&todo](auto op, LogicalResult &result) {
- if (!op.getTaskReductionVars().empty() || op.getTaskReductionByref() ||
- op.getTaskReductionSyms())
- result = todo("task_reduction");
- };
auto checkUntied = [&todo](auto op, LogicalResult &result) {
if (op.getUntied())
result = todo("untied");
@@ -399,10 +394,7 @@ static LogicalResult checkImplementationStatus(Operation &op) {
checkAllocate(op, result);
checkInReduction(op, result);
})
- .Case([&](omp::TaskgroupOp op) {
- checkAllocate(op, result);
- checkTaskReduction(op, result);
- })
+ .Case([&](omp::TaskgroupOp op) { checkAllocate(op, result); })
.Case([&](omp::TaskwaitOp op) {
checkDepend(op, result);
checkNowait(op, result);
@@ -2469,6 +2461,220 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
return success();
}
+template <typename OP>
+llvm::Value *createTaskReductionFunction(
+ llvm::IRBuilderBase &builder, const std::string &name, llvm::Type *redTy,
+ LLVM::ModuleTranslation &moduleTranslation,
+ SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls, Region ®ion,
+ 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);
+ 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);
+
+ llvm::Value *arg0 = function->getArg(0);
+ llvm::Value *arg1 = function->getArg(1);
+
+ if (name == "red_init") {
+ 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") {
+ 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);
+ }
+ if (region.empty()) {
+ // Emit an empty function body in case of empty region
+ 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)
+ llvm::Type *FlagsTy = llvm::Type::getInt32Ty(Context); // flags (i32)
+
+ // Structure members
+ std::vector<llvm::Type *> structMembers = {
+ OpaquePtrTy, // reduce_shar (void*)
+ OpaquePtrTy, // reduce_orig (void*)
+ SizeTy, // reduce_size (size_t)
+ OpaquePtrTy, // reduce_init (void*)
+ OpaquePtrTy, // reduce_fini (void*)
+ OpaquePtrTy, // reduce_comb (void*)
+ FlagsTy // flags (i32)
+ };
+
+ kmp_taskred_input_t->setBody(structMembers);
+ int arraySize = op.getTaskReductionVars().size();
+ llvm::ArrayType *ArrayTy =
+ llvm::ArrayType::get(kmp_taskred_input_t, arraySize);
+
+ // Allocate the array for kmp_taskred_input_t
+ llvm::AllocaInst *ArrayAlloca =
+ builder.CreateAlloca(ArrayTy, nullptr, "kmp_taskred_array");
+
+ // Restore the insertion point
+ builder.restoreIP(oldIP);
+ llvm::DataLayout DL = builder.GetInsertBlock()->getModule()->getDataLayout();
+
+ for (int Cnt = 0; Cnt < arraySize; ++Cnt) {
+ llvm::Value *shared =
+ moduleTranslation.lookupValue(op.getTaskReductionVars()[Cnt]);
+
+ // Create a GEP to access the reduction element
+ llvm::Value *StructPtr = builder.CreateGEP(
+ ArrayTy, ArrayAlloca, {builder.getInt32(0), builder.getInt32(Cnt)},
+ "red_element");
+
+ llvm::Value *FieldPtrReduceShar = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 0, "reduce_shar");
+ builder.CreateStore(shared, FieldPtrReduceShar);
+
+ llvm::Value *FieldPtrReduceOrig = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 1, "reduce_orig");
+ builder.CreateStore(shared, FieldPtrReduceOrig);
+
+ // Store size of the reduction variable
+ llvm::Value *FieldPtrReduceSize = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 2, "reduce_size");
+ llvm::Type *redTy =
+ moduleTranslation.convertType(reductionDecls[Cnt].getType());
+ uint64_t sizeInBytes = DL.getTypeAllocSize(redTy);
+ llvm::ConstantInt *sizeConst =
+ llvm::ConstantInt::get(llvm::Type::getInt64Ty(Context), sizeInBytes);
+ builder.CreateStore(sizeConst, FieldPtrReduceSize);
+
+ // Initialize reduction variable
+ llvm::Value *FieldPtrReduceInit = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 3, "reduce_init");
+ llvm::Value *initFunction = createTaskReductionFunction(
+ builder, "red_init", redTy, moduleTranslation, reductionDecls,
+ reductionDecls[Cnt].getInitializerRegion(), op, Cnt,
+ privateReductionVariables, reductionVariableMap);
+ builder.CreateStore(initFunction, FieldPtrReduceInit);
+
+ // Create finish and combine functions
+ llvm::Value *FieldPtrReduceFini = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 4, "reduce_fini");
+ llvm::Value *finiFunction = createTaskReductionFunction(
+ builder, "red_fini", redTy, moduleTranslation, reductionDecls,
+ reductionDecls[Cnt].getCleanupRegion(), op, Cnt,
+ privateReductionVariables, reductionVariableMap);
+ builder.CreateStore(finiFunction, FieldPtrReduceFini);
+
+ llvm::Value *FieldPtrReduceComb = builder.CreateStructGEP(
+ kmp_taskred_input_t, StructPtr, 5, "reduce_comb");
+ llvm::Value *combFunction = createTaskReductionFunction(
+ builder, "red_comb", redTy, moduleTranslation, reductionDecls,
+ reductionDecls[Cnt].getReductionRegion(), op, Cnt,
+ privateReductionVariables, reductionVariableMap);
+ builder.CreateStore(combFunction, FieldPtrReduceComb);
+
+ llvm::Value *FieldPtrFlags =
+ builder.CreateStructGEP(kmp_taskred_input_t, StructPtr, 6, "flags");
+ llvm::ConstantInt *flagVal =
+ llvm::ConstantInt::get(llvm::Type::getInt64Ty(Context), 0);
+ builder.CreateStore(flagVal, FieldPtrFlags);
+ }
+
+ // Emit the runtime call
+ emitTaskRedInitCall(builder, moduleTranslation, ompLoc, arraySize,
+ ArrayAlloca);
+ return success();
+}
+
/// Converts an OpenMP taskgroup construct into LLVM IR using OpenMPIRBuilder.
static LogicalResult
convertOmpTaskgroupOp(omp::TaskgroupOp tgOp, llvm::IRBuilderBase &builder,
@@ -2477,8 +2683,29 @@ convertOmpTaskgroupOp(omp::TaskgroupOp tgOp, llvm::IRBuilderBase &builder,
if (failed(checkImplementationStatus(*tgOp)))
return failure();
+ SmallVector<omp::DeclareReductionOp> reductionDecls;
+ SmallVector<llvm::Value *> privateReductionVariables(
+ tgOp.getNumReductionVars());
+ DenseMap<Value, llvm::Value *> reductionVariableMap;
+ MutableArrayRef<BlockArgument> reductionArgs =
+ tgOp.getRegion().getArguments();
+ LogicalResult bodyGenStatus = success();
+
+ std::optional<ArrayAttr> attr = tgOp.getTaskReductionSyms();
+ assert(attr && "Missing task reduction symbols");
+ reductionDecls.reserve(reductionDecls.size() + tgOp.getNumReductionVars());
+ for (auto symbolRef : attr->getAsRange<SymbolRefAttr>()) {
+ reductionDecls.push_back(
+ SymbolTable::lookupNearestSymbolFrom<omp::DeclareReductionOp>(
+ tgOp, symbolRef));
+ }
auto bodyCB = [&](InsertPointTy allocaIP, InsertPointTy codegenIP) {
builder.restoreIP(codegenIP);
+
+ if (failed(allocAndInitializeTaskReductionVars(
+ tgOp, reductionArgs, builder, moduleTranslation, allocaIP,
+ reductionDecls, privateReductionVariables, reductionVariableMap)))
+ bodyGenStatus = failure();
return convertOmpOpRegions(tgOp.getRegion(), "omp.taskgroup.region",
builder, moduleTranslation)
.takeError();
@@ -2494,7 +2721,7 @@ convertOmpTaskgroupOp(omp::TaskgroupOp tgOp, llvm::IRBuilderBase &builder,
return failure();
builder.restoreIP(*afterIP);
- return success();
+ return bodyGenStatus;
}
static LogicalResult
>From 3e1e9f50b76d2b8a09bd486dc01699a5d3120586 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 12:36:02 +0530
Subject: [PATCH 2/7] Add tests and address review comments
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 67 +++++--
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 189 ++++++++++++++++++
mlir/test/Target/LLVMIR/openmp-todo.mlir | 52 ++---
3 files changed, 262 insertions(+), 46 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 2db77c628876f..565c9e9601de3 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -361,6 +361,15 @@ static LogicalResult checkImplementationStatus(Operation &op) {
op.getReductionMod().value() != omp::ReductionModifier::defaultmod)
result = todo("reduction with modifier");
};
+ auto checkTaskReduction = [&todo](auto op, LogicalResult &result) {
+ if (op.getTaskReductionByref()) {
+ llvm::ArrayRef<bool> ByrefAttrs = op.getTaskReductionByref().value();
+ for (bool ByrefAttr : ByrefAttrs) {
+ if (ByrefAttr)
+ result = todo("task_reduction with pass by reference argument");
+ }
+ }
+ };
auto checkUntied = [&todo](auto op, LogicalResult &result) {
if (op.getUntied())
result = todo("untied");
@@ -394,7 +403,10 @@ static LogicalResult checkImplementationStatus(Operation &op) {
checkAllocate(op, result);
checkInReduction(op, result);
})
- .Case([&](omp::TaskgroupOp op) { checkAllocate(op, result); })
+ .Case([&](omp::TaskgroupOp op) {
+ checkAllocate(op, result);
+ checkTaskReduction(op, result);
+ })
.Case([&](omp::TaskwaitOp op) {
checkDepend(op, result);
checkNowait(op, result);
@@ -2462,11 +2474,11 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
}
template <typename OP>
-llvm::Value *createTaskReductionFunction(
+static llvm::Value *createTaskReductionFunction(
llvm::IRBuilderBase &builder, const std::string &name, llvm::Type *redTy,
LLVM::ModuleTranslation &moduleTranslation,
SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls, Region ®ion,
- OP &op, unsigned Cnt,
+ OP &op, unsigned cnt,
SmallVectorImpl<llvm::Value *> &privateReductionVariables,
DenseMap<Value, llvm::Value *> &reductionVariableMap) {
@@ -2476,6 +2488,8 @@ llvm::Value *createTaskReductionFunction(
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 =
@@ -2486,17 +2500,20 @@ llvm::Value *createTaskReductionFunction(
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];
+ mlir::omp::DeclareReductionOp &reduction = reductionDecls[cnt];
Region &initializerRegion = reduction.getInitializerRegion();
Block &entry = initializerRegion.front();
- mlir::Value mlirSource = op.getTaskReductionVars()[Cnt];
+ mlir::Value mlirSource = op.getTaskReductionVars()[cnt];
llvm::Value *llvmSource = moduleTranslation.lookupValue(mlirSource);
llvm::Value *origVal = llvmSource;
@@ -2504,22 +2521,25 @@ llvm::Value *createTaskReductionFunction(
if (entry.getNumArguments() > 1) {
llvm::Value *allocation =
- reductionVariableMap.lookup(op.getReductionVars()[Cnt]);
+ 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);
}
- if (region.empty()) {
- // Emit an empty function body in case of empty region
+
+ // 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,
@@ -2609,13 +2629,13 @@ static LogicalResult allocAndInitializeTaskReductionVars(
builder.restoreIP(oldIP);
llvm::DataLayout DL = builder.GetInsertBlock()->getModule()->getDataLayout();
- for (int Cnt = 0; Cnt < arraySize; ++Cnt) {
+ for (int cnt = 0; cnt < arraySize; ++cnt) {
llvm::Value *shared =
- moduleTranslation.lookupValue(op.getTaskReductionVars()[Cnt]);
+ moduleTranslation.lookupValue(op.getTaskReductionVars()[cnt]);
// Create a GEP to access the reduction element
llvm::Value *StructPtr = builder.CreateGEP(
- ArrayTy, ArrayAlloca, {builder.getInt32(0), builder.getInt32(Cnt)},
+ ArrayTy, ArrayAlloca, {builder.getInt32(0), builder.getInt32(cnt)},
"red_element");
llvm::Value *FieldPtrReduceShar = builder.CreateStructGEP(
@@ -2630,7 +2650,7 @@ static LogicalResult allocAndInitializeTaskReductionVars(
llvm::Value *FieldPtrReduceSize = builder.CreateStructGEP(
kmp_taskred_input_t, StructPtr, 2, "reduce_size");
llvm::Type *redTy =
- moduleTranslation.convertType(reductionDecls[Cnt].getType());
+ moduleTranslation.convertType(reductionDecls[cnt].getType());
uint64_t sizeInBytes = DL.getTypeAllocSize(redTy);
llvm::ConstantInt *sizeConst =
llvm::ConstantInt::get(llvm::Type::getInt64Ty(Context), sizeInBytes);
@@ -2641,7 +2661,7 @@ static LogicalResult allocAndInitializeTaskReductionVars(
kmp_taskred_input_t, StructPtr, 3, "reduce_init");
llvm::Value *initFunction = createTaskReductionFunction(
builder, "red_init", redTy, moduleTranslation, reductionDecls,
- reductionDecls[Cnt].getInitializerRegion(), op, Cnt,
+ reductionDecls[cnt].getInitializerRegion(), op, cnt,
privateReductionVariables, reductionVariableMap);
builder.CreateStore(initFunction, FieldPtrReduceInit);
@@ -2650,7 +2670,7 @@ static LogicalResult allocAndInitializeTaskReductionVars(
kmp_taskred_input_t, StructPtr, 4, "reduce_fini");
llvm::Value *finiFunction = createTaskReductionFunction(
builder, "red_fini", redTy, moduleTranslation, reductionDecls,
- reductionDecls[Cnt].getCleanupRegion(), op, Cnt,
+ reductionDecls[cnt].getCleanupRegion(), op, cnt,
privateReductionVariables, reductionVariableMap);
builder.CreateStore(finiFunction, FieldPtrReduceFini);
@@ -2658,7 +2678,7 @@ static LogicalResult allocAndInitializeTaskReductionVars(
kmp_taskred_input_t, StructPtr, 5, "reduce_comb");
llvm::Value *combFunction = createTaskReductionFunction(
builder, "red_comb", redTy, moduleTranslation, reductionDecls,
- reductionDecls[Cnt].getReductionRegion(), op, Cnt,
+ reductionDecls[cnt].getReductionRegion(), op, cnt,
privateReductionVariables, reductionVariableMap);
builder.CreateStore(combFunction, FieldPtrReduceComb);
@@ -2692,12 +2712,15 @@ convertOmpTaskgroupOp(omp::TaskgroupOp tgOp, llvm::IRBuilderBase &builder,
LogicalResult bodyGenStatus = success();
std::optional<ArrayAttr> attr = tgOp.getTaskReductionSyms();
- assert(attr && "Missing task reduction symbols");
- reductionDecls.reserve(reductionDecls.size() + tgOp.getNumReductionVars());
- for (auto symbolRef : attr->getAsRange<SymbolRefAttr>()) {
- reductionDecls.push_back(
- SymbolTable::lookupNearestSymbolFrom<omp::DeclareReductionOp>(
- tgOp, symbolRef));
+ if (attr) {
+ reductionDecls.reserve(reductionDecls.size() + tgOp.getNumReductionVars());
+ for (auto symbolRef : attr->getAsRange<SymbolRefAttr>()) {
+ reductionDecls.push_back(
+ SymbolTable::lookupNearestSymbolFrom<omp::DeclareReductionOp>(
+ tgOp, symbolRef));
+ }
+ assert(reductionDecls.size() == tgOp.getNumReductionVars() &&
+ "Missing reduction declaration");
}
auto bodyCB = [&](InsertPointTy allocaIP, InsertPointTy codegenIP) {
builder.restoreIP(codegenIP);
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index 1eb501ca02703..a4165837d9df7 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -3074,6 +3074,195 @@ llvm.func @omp_taskgroup_task(%x: i32, %y: i32, %zaddr: !llvm.ptr) {
// -----
+// CHECK-LABEL: define void @_QPtaskred_integer_arg() {
+// CHECK: %[[VAR:.*]] = alloca i32, i64 1, align 4
+// CHECK: %kmp_taskred_array = alloca [1 x %kmp_taskred_input_t], align 8
+// CHECK: %[[TID:.*]] = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: call void @__kmpc_taskgroup(ptr @1, i32 %[[TID]])
+// CHECK: %[[RED_ELEMENT:.*]] = getelementptr [1 x %kmp_taskred_input_t], ptr %kmp_taskred_array, i32 0, i32 0
+// CHECK: %[[REDUCE_SHAR:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 0
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_SHAR]], align 8
+// CHECK: %[[REDUCE_ORIG:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 1
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_ORIG]], align 8
+// CHECK: %[[REDUCE_SIZE:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 2
+// CHECK: store i64 4, ptr %[[REDUCE_SIZE]], align 4
+// CHECK: %[[REDUCE_INIT:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 3
+// CHECK: store ptr @red_init, ptr %[[REDUCE_INIT]], align 8
+// CHECK: %[[REDUCE_FINI:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 4
+// CHECK: store ptr null, ptr %reduce_fini, align 8
+// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
+// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
+// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
+// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
+
+// CHECK: define ptr @red_init(ptr noalias %0, ptr noalias %1) #2 {
+// CHECK: entry:
+// CHECK: store i32 0, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+// CHECK: ; Function Attrs: norecurse
+// CHECK: define ptr @red_comb(ptr %0, ptr %1) #2 {
+// CHECK: entry:
+// CHECK: %2 = load i32, ptr %0, align 4
+// CHECK: %3 = load i32, ptr %1, align 4
+// CHECK: %4 = add i32 %2, %3
+// CHECK: store i32 %4, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+omp.declare_reduction @add_reduction_i32 : i32 init {
+^bb0(%arg0: i32):
+ %0 = llvm.mlir.constant(0 : i32) : i32
+ omp.yield(%0 : i32)
+} combiner {
+^bb0(%arg0: i32, %arg1: i32):
+ %0 = llvm.add %arg0, %arg1 : i32
+ omp.yield(%0 : i32)
+}
+llvm.func @_QPtaskred_integer_arg() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 {bindc_name = "s"} : (i64) -> !llvm.ptr
+ omp.taskgroup task_reduction(@add_reduction_i32 %1 -> %arg0 : !llvm.ptr) {
+ omp.terminator
+ }
+ llvm.return
+}
+
+// -----
+
+// CHECK-LABEL: define void @_QPfloat_arg() {
+// CHECK: %[[VAR:.*]] = alloca float, i64 1, align 4
+// CHECK: %kmp_taskred_array = alloca [1 x %kmp_taskred_input_t], align 8
+// CHECK: %[[TID:.*]] = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: call void @__kmpc_taskgroup(ptr @1, i32 %[[TID]])
+// CHECK: %[[RED_ELEMENT:.*]] = getelementptr [1 x %kmp_taskred_input_t], ptr %kmp_taskred_array, i32 0, i32 0
+// CHECK: %[[REDUCE_SHAR:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 0
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_SHAR]], align 8
+// CHECK: %[[REDUCE_ORIG:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 1
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_ORIG]], align 8
+// CHECK: %[[REDUCE_SIZE:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 2
+// CHECK: store i64 4, ptr %[[REDUCE_SIZE]], align 4
+// CHECK: %[[REDUCE_INIT:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 3
+// CHECK: store ptr @red_init, ptr %[[REDUCE_INIT]], align 8
+// CHECK: %[[REDUCE_FINI:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 4
+// CHECK: store ptr null, ptr %reduce_fini, align 8
+// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
+// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
+// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
+// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
+
+
+// CHECK: ; Function Attrs: norecurse
+// CHECK: define ptr @red_init(ptr noalias %0, ptr noalias %1) #2 {
+// CHECK: entry:
+// CHECK: store float 1.000000e+00, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+// CHECK: ; Function Attrs: norecurse
+// CHECK: define ptr @red_comb(ptr %0, ptr %1) #2 {
+// CHECK: entry:
+// CHECK: %2 = load float, ptr %0, align 4
+// CHECK: %3 = load float, ptr %1, align 4
+// CHECK: %4 = fmul contract float %2, %3
+// CHECK: store float %4, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+
+omp.declare_reduction @multiply_reduction_f32 : f32 init {
+^bb0(%arg0: f32):
+ %0 = llvm.mlir.constant(1.000000e+00 : f32) : f32
+ omp.yield(%0 : f32)
+} combiner {
+^bb0(%arg0: f32, %arg1: f32):
+ %0 = llvm.fmul %arg0, %arg1 {fastmathFlags = #llvm.fastmath<contract>} : f32
+ omp.yield(%0 : f32)
+}
+llvm.func @_QPfloat_arg() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x f32 {bindc_name = "s"} : (i64) -> !llvm.ptr
+ omp.taskgroup task_reduction(@multiply_reduction_f32 %1 -> %arg0 : !llvm.ptr) {
+ omp.terminator
+ }
+ llvm.return
+}
+
+// -----
+
+// CHECK: define void @_QPlogical_reduction() {
+// CHECK: %[[VAR:.*]] = alloca i32, i64 1, align 4
+// CHECK: %kmp_taskred_array = alloca [1 x %kmp_taskred_input_t], align 8
+// CHECK: %[[TID:.*]] = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: call void @__kmpc_taskgroup(ptr @1, i32 %[[TID]])
+// CHECK: %[[RED_ELEMENT:.*]] = getelementptr [1 x %kmp_taskred_input_t], ptr %kmp_taskred_array, i32 0, i32 0
+// CHECK: %[[REDUCE_SHAR:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 0
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_SHAR]], align 8
+// CHECK: %[[REDUCE_ORIG:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 1
+// CHECK: store ptr %[[VAR]], ptr %[[REDUCE_ORIG]], align 8
+// CHECK: %[[REDUCE_SIZE:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 2
+// CHECK: store i64 4, ptr %[[REDUCE_SIZE]], align 4
+// CHECK: %[[REDUCE_INIT:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 3
+// CHECK: store ptr @red_init, ptr %[[REDUCE_INIT]], align 8
+// CHECK: %[[REDUCE_FINI:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 4
+// CHECK: store ptr null, ptr %reduce_fini, align 8
+// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
+// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
+// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
+// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
+// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
+
+
+// CHECK: define ptr @red_init(ptr noalias %0, ptr noalias %1) #2 {
+// CHECK: entry:
+// CHECK: store i32 1, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+// CHECK: ; Function Attrs: norecurse
+// CHECK: define ptr @red_comb(ptr %0, ptr %1) #2 {
+// CHECK: entry:
+// CHECK: %2 = load i32, ptr %0, align 4
+// CHECK: %3 = load i32, ptr %1, align 4
+// CHECK: %4 = icmp ne i32 %2, 0
+// CHECK: %5 = icmp ne i32 %3, 0
+// CHECK: %6 = and i1 %4, %5
+// CHECK: %7 = zext i1 %6 to i32
+// CHECK: store i32 %7, ptr %0, align 4
+// CHECK: ret ptr %0
+// CHECK: }
+
+
+omp.declare_reduction @and_reduction_l32 : i32 init {
+^bb0(%arg0: i32):
+ %0 = llvm.mlir.constant(1 : i64) : i32
+ omp.yield(%0 : i32)
+} combiner {
+^bb0(%arg0: i32, %arg1: i32):
+ %0 = llvm.mlir.constant(0 : i64) : i32
+ %1 = llvm.icmp "ne" %arg0, %0 : i32
+ %2 = llvm.icmp "ne" %arg1, %0 : i32
+ %3 = llvm.and %1, %2 : i1
+ %4 = llvm.zext %3 : i1 to i32
+ omp.yield(%4 : i32)
+}
+llvm.func @_QPlogical_reduction() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 {bindc_name = "s"} : (i64) -> !llvm.ptr
+ omp.taskgroup task_reduction(@and_reduction_l32 %1 -> %arg0 : !llvm.ptr) {
+ omp.terminator
+ }
+ llvm.return
+}
+
+// -----
+
llvm.func @test_01() attributes {sym_visibility = "private"}
llvm.func @test_02() attributes {sym_visibility = "private"}
// CHECK-LABEL: define void @_QPomp_task_priority() {
diff --git a/mlir/test/Target/LLVMIR/openmp-todo.mlir b/mlir/test/Target/LLVMIR/openmp-todo.mlir
index 396c57af81c44..f240ab6d39227 100644
--- a/mlir/test/Target/LLVMIR/openmp-todo.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-todo.mlir
@@ -295,31 +295,35 @@ llvm.func @taskgroup_allocate(%x : !llvm.ptr) {
// -----
-omp.declare_reduction @add_f32 : f32
-init {
-^bb0(%arg: f32):
- %0 = llvm.mlir.constant(0.0 : f32) : f32
- omp.yield (%0 : f32)
-}
-combiner {
-^bb1(%arg0: f32, %arg1: f32):
- %1 = llvm.fadd %arg0, %arg1 : f32
- omp.yield (%1 : f32)
-}
-atomic {
-^bb2(%arg2: !llvm.ptr, %arg3: !llvm.ptr):
- %2 = llvm.load %arg3 : !llvm.ptr -> f32
- llvm.atomicrmw fadd %arg2, %2 monotonic : !llvm.ptr, f32
- omp.yield
-}
-llvm.func @taskgroup_task_reduction(%x : !llvm.ptr) {
- // expected-error at below {{not yet implemented: Unhandled clause task_reduction in omp.taskgroup operation}}
- // expected-error at below {{LLVM Translation failed for operation: omp.taskgroup}}
- omp.taskgroup task_reduction(@add_f32 %x -> %prv : !llvm.ptr) {
- omp.terminator
- }
+omp.declare_reduction @add_reduction_byref_i32 : !llvm.ptr alloc {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 : (i64) -> !llvm.ptr
+ %2 = llvm.mlir.constant(1 : i64) : i64
+ omp.yield(%1 : !llvm.ptr)
+} init {
+ ^bb0(%arg0: !llvm.ptr, %arg1: !llvm.ptr):
+ %0 = llvm.mlir.constant(0 : i32) : i32
+ llvm.store %0, %arg1 : i32, !llvm.ptr
+ omp.yield(%arg1 : !llvm.ptr)
+} combiner {
+ ^bb0(%arg0: !llvm.ptr, %arg1: !llvm.ptr):
+ %0 = llvm.load %arg0 : !llvm.ptr -> i32
+ %1 = llvm.load %arg1 : !llvm.ptr -> i32
+ %2 = llvm.add %0, %1 : i32
+ llvm.store %2, %arg0 : i32, !llvm.ptr
+ omp.yield(%arg0 : !llvm.ptr)
+}
+llvm.func @_QPtask_reduction_byref() {
+ %0 = llvm.mlir.constant(1 : i64) : i64
+ %1 = llvm.alloca %0 x i32 {bindc_name = "x"} : (i64) -> !llvm.ptr
+ %2 = llvm.mlir.constant(1 : i64) : i64
+ // expected-error at below {{not yet implemented: Unhandled clause task_reduction with pass by reference argument in omp.taskgroup operation}}
+ // expected-error at below {{LLVM Translation failed for operation: omp.taskgroup}}
+ omp.taskgroup task_reduction(byref @add_reduction_byref_i32 %1 -> %arg0 : !llvm.ptr) {
+ omp.terminator
+ }
llvm.return
-}
+}
// -----
>From c17469d2067a683a6f76591ebe9e0f92100980f9 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 12:42:07 +0530
Subject: [PATCH 3/7] Format
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 22 +++++++++----------
1 file changed, 11 insertions(+), 11 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 565c9e9601de3..c33d5bf3fa083 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2538,19 +2538,19 @@ static llvm::Value *createTaskReductionFunction(
// Emit an empty function body in case of empty region
if (region.empty())
bbBuilder.CreateRet(arg0); // Return from the function
- return 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;
+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(
>From 743d4cd6d0b303052f5e4f85f120d86b593163da Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 12:50:23 +0530
Subject: [PATCH 4/7] Format
---
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 21 +++++++++----------
1 file changed, 10 insertions(+), 11 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index c33d5bf3fa083..5cc8093bc806e 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2539,18 +2539,17 @@ static llvm::Value *createTaskReductionFunction(
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;
+ 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(
>From dd24d8d809b7b13b6c2b274aec8b4ab3d1cffcb0 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 13:04:30 +0530
Subject: [PATCH 5/7] Fix test
---
.../LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 5cc8093bc806e..c7bbfeda85439 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2536,9 +2536,10 @@ static llvm::Value *createTaskReductionFunction(
}
// Emit an empty function body in case of empty region
- if (region.empty())
+ if (region.empty()) {
bbBuilder.CreateRet(arg0); // Return from the function
- return function;
+ return function;
+ }
SmallVector<llvm::Value *, 1> phis;
if (failed(inlineConvertOmpRegions(region, "", bbBuilder, moduleTranslation,
>From 31d9c847bd088522860d6d8901e8396f51227a59 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 18:33:50 +0530
Subject: [PATCH 6/7] Address review comments
---
.../mlir/Dialect/OpenMP/OpenMPClauses.td | 8 +-
.../OpenMP/OpenMPToLLVMIRTranslation.cpp | 78 +++++++++++--------
mlir/test/Target/LLVMIR/openmp-llvm.mlir | 6 +-
3 files changed, 52 insertions(+), 40 deletions(-)
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
index a1f987553e050..95e1f607d7d51 100644
--- a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
+++ b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
@@ -1389,11 +1389,11 @@ class OpenMP_TaskReductionClauseSkip<
);
let extraClassDeclaration = [{
- /// Returns the number of reduction variables.
- unsigned getNumReductionVars() { return getTaskReductionVars().size(); }
+ /// Returns the number of reduction variables.
+ unsigned getNumReductionVars() { return getTaskReductionVars().size(); }
- /// Returns the reduction symbols
- auto getReductionSyms() { return getTaskReductionSyms(); }
+ /// Returns the reduction symbols
+ auto getReductionSyms() { return getTaskReductionSyms(); }
}];
let description = [{
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index c7bbfeda85439..f76cb433bf20c 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2473,18 +2473,29 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
return success();
}
-template <typename OP>
+/*
+ * Utility function for translating `red_init`, `red_comb`, and `red_fini` to
+ * LLVMIR. The ulitity first (commonly) generates a skeleton for any of the
+ * three functions, and then generates the function body based on the
+ * specific operations involved in `red_init` (codegen related to initialization
+ * of task reduction variables) and `red_comb` (codegen related to combination).
+ * Currently, codegen for `red_fini` is skipped since finalization is optional
+ * for `task_reduction` clause, but this ulitity has the capability of defining
+ * finalization if needed. Finally, the returned `llvm::Function` is used to
+ * populate the relevant entries in the task reduction specific data structure.
+ */
static llvm::Value *createTaskReductionFunction(
- llvm::IRBuilderBase &builder, const std::string &name, llvm::Type *redTy,
- LLVM::ModuleTranslation &moduleTranslation,
- SmallVectorImpl<omp::DeclareReductionOp> &reductionDecls, Region ®ion,
- OP &op, unsigned cnt,
+ omp::TaskgroupOp &op, llvm::IRBuilderBase &builder, const std::string &name,
+ llvm::Type *redTy, LLVM::ModuleTranslation &moduleTranslation,
+ omp::DeclareReductionOp &reductionDecl, Region ®ion, 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);
+ llvm::DataLayout DL = builder.GetInsertBlock()->getModule()->getDataLayout();
+ llvm::Type *OpaquePtrTy =
+ llvm::PointerType::get(Context, DL.getProgramAddressSpace());
if (region.empty() && name == "red_fini")
// Finalization is optional for reductions.
return llvm::Constant::getNullValue(OpaquePtrTy);
@@ -2509,20 +2520,19 @@ static llvm::Value *createTaskReductionFunction(
// 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();
+ Region &initializerRegion = reductionDecl.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);
+ moduleTranslation.mapValue(reductionDecl.getInitializerMoldArg(), origVal);
if (entry.getNumArguments() > 1) {
llvm::Value *allocation =
- reductionVariableMap.lookup(op.getReductionVars()[cnt]);
- moduleTranslation.mapValue(reduction.getInitializerAllocArg(),
+ reductionVariableMap.lookup(op.getTaskReductionVars()[cnt]);
+ moduleTranslation.mapValue(reductionDecl.getInitializerAllocArg(),
allocation);
}
@@ -2553,10 +2563,11 @@ static llvm::Value *createTaskReductionFunction(
return function;
}
-void emitTaskRedInitCall(
- llvm::IRBuilderBase &builder, LLVM::ModuleTranslation &moduleTranslation,
- const llvm::OpenMPIRBuilder::LocationDescription &ompLoc, int arraySize,
- llvm::Value *ArrayAlloca) {
+static 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;
@@ -2590,19 +2601,15 @@ static LogicalResult allocAndInitializeTaskReductionVars(
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());
+ llvm::DataLayout DL = builder.GetInsertBlock()->getModule()->getDataLayout();
// 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)
+ llvm::Type *OpaquePtrTy =
+ llvm::PointerType::get(Context,
+ DL.getProgramAddressSpace()); // void*
+ llvm::Type *SizeTy = DL.getIntPtrType(Context); // size_t
llvm::Type *FlagsTy = llvm::Type::getInt32Ty(Context); // flags (i32)
// Structure members
@@ -2621,13 +2628,18 @@ static LogicalResult allocAndInitializeTaskReductionVars(
llvm::ArrayType *ArrayTy =
llvm::ArrayType::get(kmp_taskred_input_t, arraySize);
+ // Save the current insertion point
+ auto oldIP = builder.saveIP();
+
+ // Set insertion point after the allocations
+ builder.SetInsertPoint(allocaIP.getBlock()->getTerminator());
+
// Allocate the array for kmp_taskred_input_t
llvm::AllocaInst *ArrayAlloca =
builder.CreateAlloca(ArrayTy, nullptr, "kmp_taskred_array");
// Restore the insertion point
builder.restoreIP(oldIP);
- llvm::DataLayout DL = builder.GetInsertBlock()->getModule()->getDataLayout();
for (int cnt = 0; cnt < arraySize; ++cnt) {
llvm::Value *shared =
@@ -2660,8 +2672,8 @@ static LogicalResult allocAndInitializeTaskReductionVars(
llvm::Value *FieldPtrReduceInit = builder.CreateStructGEP(
kmp_taskred_input_t, StructPtr, 3, "reduce_init");
llvm::Value *initFunction = createTaskReductionFunction(
- builder, "red_init", redTy, moduleTranslation, reductionDecls,
- reductionDecls[cnt].getInitializerRegion(), op, cnt,
+ op, builder, "red_init", redTy, moduleTranslation, reductionDecls[cnt],
+ reductionDecls[cnt].getInitializerRegion(), cnt,
privateReductionVariables, reductionVariableMap);
builder.CreateStore(initFunction, FieldPtrReduceInit);
@@ -2669,23 +2681,23 @@ static LogicalResult allocAndInitializeTaskReductionVars(
llvm::Value *FieldPtrReduceFini = builder.CreateStructGEP(
kmp_taskred_input_t, StructPtr, 4, "reduce_fini");
llvm::Value *finiFunction = createTaskReductionFunction(
- builder, "red_fini", redTy, moduleTranslation, reductionDecls,
- reductionDecls[cnt].getCleanupRegion(), op, cnt,
- privateReductionVariables, reductionVariableMap);
+ op, builder, "red_fini", redTy, moduleTranslation, reductionDecls[cnt],
+ reductionDecls[cnt].getCleanupRegion(), cnt, privateReductionVariables,
+ reductionVariableMap);
builder.CreateStore(finiFunction, FieldPtrReduceFini);
llvm::Value *FieldPtrReduceComb = builder.CreateStructGEP(
kmp_taskred_input_t, StructPtr, 5, "reduce_comb");
llvm::Value *combFunction = createTaskReductionFunction(
- builder, "red_comb", redTy, moduleTranslation, reductionDecls,
- reductionDecls[cnt].getReductionRegion(), op, cnt,
+ op, builder, "red_comb", redTy, moduleTranslation, reductionDecls[cnt],
+ reductionDecls[cnt].getReductionRegion(), cnt,
privateReductionVariables, reductionVariableMap);
builder.CreateStore(combFunction, FieldPtrReduceComb);
llvm::Value *FieldPtrFlags =
builder.CreateStructGEP(kmp_taskred_input_t, StructPtr, 6, "flags");
llvm::ConstantInt *flagVal =
- llvm::ConstantInt::get(llvm::Type::getInt64Ty(Context), 0);
+ llvm::ConstantInt::get(llvm::Type::getInt32Ty(Context), 0);
builder.CreateStore(flagVal, FieldPtrFlags);
}
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index a4165837d9df7..dd0236b3cdb52 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -3093,7 +3093,7 @@ llvm.func @omp_taskgroup_task(%x: i32, %y: i32, %zaddr: !llvm.ptr) {
// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
-// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: store i32 0, ptr %[[FLAGS]], align 4
// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
@@ -3152,7 +3152,7 @@ llvm.func @_QPtaskred_integer_arg() {
// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
-// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: store i32 0, ptr %[[FLAGS]], align 4
// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
@@ -3214,7 +3214,7 @@ llvm.func @_QPfloat_arg() {
// CHECK: %[[REDUCE_COMB:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 5
// CHECK: store ptr @red_comb, ptr %[[REDUCE_COMB]], align 8
// CHECK: %[[FLAGS:.*]] = getelementptr inbounds nuw %kmp_taskred_input_t, ptr %[[RED_ELEMENT]], i32 0, i32 6
-// CHECK: store i64 0, ptr %[[FLAGS]], align 4
+// CHECK: store i32 0, ptr %[[FLAGS]], align 4
// CHECK: %omp_global_thread_num1 = call i32 @__kmpc_global_thread_num(ptr @1)
// CHECK: %[[INIT:.*]] = call ptr @__kmpc_taskred_init(i32 %omp_global_thread_num1, i32 1, ptr %kmp_taskred_array)
>From ddb8c9f79e77d97b38611f8cd97cb5fddb885ed6 Mon Sep 17 00:00:00 2001
From: NimishMishra <neelam.nimish at gmail.com>
Date: Fri, 2 Jan 2026 18:45:06 +0530
Subject: [PATCH 7/7] Fix format
---
.../Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index f76cb433bf20c..08a6695166eb8 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2610,7 +2610,7 @@ static LogicalResult allocAndInitializeTaskReductionVars(
llvm::PointerType::get(Context,
DL.getProgramAddressSpace()); // void*
llvm::Type *SizeTy = DL.getIntPtrType(Context); // size_t
- llvm::Type *FlagsTy = llvm::Type::getInt32Ty(Context); // flags (i32)
+ llvm::Type *FlagsTy = llvm::Type::getInt32Ty(Context); // flags (i32)
// Structure members
std::vector<llvm::Type *> structMembers = {
More information about the Mlir-commits
mailing list