[mlir] Add conversion from SCF parallel loops to OpenMP

[ftynse]

Summary

Introduce a conversion pass from SCF parallel loops to OpenMP dialect
constructs - parallel region and workshare loop. Loops with reductions are not
supported because the OpenMP dialect cannot model them yet.

The conversion currently targets only one level of parallelism, i.e. only
one top-level omp.parallel operation is produced even if there are nested
scf.parallel operations that could be mapped to omp.wsloop. Nested
parallelism support is left for future work.

[ftynse]

Add cmake

[ftynse]

Sprinkle around a bit more documentation

[bondhugula]

Nice. Looks like you missed adding test cases.

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

25

Brief doc comment please.

58

Missing doc comment.

78

Comment here or somewhere in the block.

88

func.getContext() and make this method static?

[ftynse]

Forgotten git add

[ftynse]

Address (most of) the review

[ftynse]

Address review

[kiranchandramohan]

Thanks for this patch.

We should be able to do something similar for the Fortran "Do concurrent" loop. @schweitz

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

24

Nit: OpenMP paralle loop or OpenMP parallel + workshare loop.

68

Are you not interested in nested parallelism?

[SouraVX]

Please ignore if this is out of context.
So Flang lower all loops into fir.do_loop(which is essentially a superset of scf.for) then it may lower to scf.for(depending on feasibility). So the question I have is what's the intended lowering targeted here for openmp worksharing loop ?

FORTRAN -> fir.loop -> scf.for -> ws.loop
OR 
FORTRAN -> fir.loop -> ws.loop

OR something else ?

is it already discussed/finalized in RFC?
If this patch intend to implements some part from the RFC discussion ? Would you mind adding the link here.
Thanks!

[ftynse]

In D91982#2413912, @SouraVX wrote:

Please ignore if this is out of context.
So Flang lower all loops into fir.loop(which is essentially a superset of scf.for) then it may lower to scf.for(depending on feasibility). So the question I have is what's the intended lowering targeted here for openmp worksharing loop ?

FORTRAN -> fir.loop -> scf.for -> ws.loop
OR 
FORTRAN -> fir.loop -> ws.loop

OR something else ?

is it already discussed/finalized in RFC?
If this patch intend to implements some part from the RFC discussion ? Would you mind adding the link here.
Thanks!

This is fully orthogonal to Flang. I have code that comes from higher-level abstractions in core MLIR (Linalg) to SCF, and I need this code to run in parallel on a platform that supports OpenMP. Flang can choose to target the SCF dialect or the OpenMP dialect regardless of is done here. If you think this needs an RFC within core MLIR context, I can send one, but to me this looks like relatively straightforward plumbing work.

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

24

This pattern does not create the OpenMP parallel operation, only the workshare loop, so I maintain that the original comment is correct and the suggested replacement will be misleading.

68

I don't have a use case for it as of yet. We can add more features when we need them.

[kiranchandramohan]

LGTM.

Do you need the lowering to LLVM IR also sometime soon?

Previously, did you mention something about a parallelisation strategy and SCF ops carrying an attribute to determine whether they should be lowered to OpenMP parallel loops.

@SouraVX we can discuss the lowering separately for Flang. Since we decided to represent the work-sharing loop as a loop like operation (omp.wsloop), lowering directly to the work-sharing loop from the parse tree will be the straight forward method. If there are difficulties with that we can think of lowering to fir.do_loop and then converting to omp.wsloop but I think this might need modification of fir.do_loop or addition of a directive like operation in FIR.

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

68

OK. Can the commit message/description carry this info?

[ftynse]

In D91982#2414182, @kiranchandramohan wrote:

LGTM.

Do you need the lowering to LLVM IR also sometime soon?

Previously, did you mention something about a parallelisation strategy and SCF ops carrying an attribute to determine whether they should be lowered to OpenMP parallel loops.

@SouraVX we can discuss the lowering separately for Flang. Since we decided to represent the work-sharing loop as a loop like operation (omp.wsloop), lowering directly to the work-sharing loop from the parse tree will be the straight forward method. If there are difficulties with that we can think of lowering to fir.do_loop and then converting to omp.wsloop but I think this might need modification of fir.do_loop or addition of a directive like operation in FIR.

Thanks for the quick review!

I am in the process of writing a translation to LLVM IR, happy to use anything that is already available.

We discussed parallelization strategies in several ODMs, but there was no firm decision. Most of the discussion was in the GPU context, I believe, where we need not only to say which loops remain parallel but also how they are mapped to the device. We can add annotations that this pass would consume. Alternatively, it is possible to first transform any SCF parallel that should not be converted to SCF for, and then just run a blanket conversion of all remaining parallel operations. We already have the former conversion implemented, can just as well add a way to control it at a finer grain.

We can expose any part of this transformation somewhere in Utils.h if you find them useful for Flang/

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

68

Sure, will put this into the commit message and add a TODO here.

[kiranchandramohan]

In D91982#2414233, @ftynse wrote:

In D91982#2414182, @kiranchandramohan wrote:

LGTM.

Do you need the lowering to LLVM IR also sometime soon?

Previously, did you mention something about a parallelisation strategy and SCF ops carrying an attribute to determine whether they should be lowered to OpenMP parallel loops.

@SouraVX we can discuss the lowering separately for Flang. Since we decided to represent the work-sharing loop as a loop like operation (omp.wsloop), lowering directly to the work-sharing loop from the parse tree will be the straight forward method. If there are difficulties with that we can think of lowering to fir.do_loop and then converting to omp.wsloop but I think this might need modification of fir.do_loop or addition of a directive like operation in FIR.

Thanks for the quick review!

I am in the process of writing a translation to LLVM IR, happy to use anything that is already available.

@Meinersbur added a canonical loop in the OpenMPIRBuilder (https://reviews.llvm.org/D90830). The idea was to create the worksharing loop using this canonical loop.
FYI, we have a call ( 4pm BST/11am ET) for the OpenMP for Flang work where the OpenMP dialect work is also discussed. Link to the call and minutes can be found below.
https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjhiMzZiOWMtNTU5Ni00YzljLWEwNDctYjZkNTZhODg2ZTIx%40thread.v2/0?context=%7b%22Tid%22%3a%223dd8961f-e488-4e60-8e11-a82d994e183d%22%2c%22Oid%22%3a%22ce0d0f4a-6fbf-4ba7-b0c7-20184e3bcb53%22%7d
https://docs.google.com/document/d/1yA-MeJf6RYY-ZXpdol0t7YoDoqtwAyBhFLr5thu5pFI/edit#

We discussed parallelization strategies in several ODMs, but there was no firm decision. Most of the discussion was in the GPU context, I believe, where we need not only to say which loops remain parallel but also how they are mapped to the device. We can add annotations that this pass would consume. Alternatively, it is possible to first transform any SCF parallel that should not be converted to SCF for, and then just run a blanket conversion of all remaining parallel operations. We already have the former conversion implemented, can just as well add a way to control it at a finer grain.

We can expose any part of this transformation somewhere in Utils.h if you find them useful for Flang/

OK, thanks for the explanation. We might need it (not sure though).

[ftynse]

In D91982#2414309, @kiranchandramohan wrote:

In D91982#2414233, @ftynse wrote:

In D91982#2414182, @kiranchandramohan wrote:

LGTM.

Do you need the lowering to LLVM IR also sometime soon?

I have something that works here https://reviews.llvm.org/D92055 but it needs some discussions.

Previously, did you mention something about a parallelisation strategy and SCF ops carrying an attribute to determine whether they should be lowered to OpenMP parallel loops.

@SouraVX we can discuss the lowering separately for Flang. Since we decided to represent the work-sharing loop as a loop like operation (omp.wsloop), lowering directly to the work-sharing loop from the parse tree will be the straight forward method. If there are difficulties with that we can think of lowering to fir.do_loop and then converting to omp.wsloop but I think this might need modification of fir.do_loop or addition of a directive like operation in FIR.

Thanks for the quick review!

I am in the process of writing a translation to LLVM IR, happy to use anything that is already available.

@Meinersbur added a canonical loop in the OpenMPIRBuilder (https://reviews.llvm.org/D90830). The idea was to create the worksharing loop using this canonical loop.

Familiar names :) I indeed intended to use the canonical loop creation functionality.

FYI, we have a call ( 4pm BST/11am ET) for the OpenMP for Flang work where the OpenMP dialect work is also discussed. Link to the call and minutes can be found below.
https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZjhiMzZiOWMtNTU5Ni00YzljLWEwNDctYjZkNTZhODg2ZTIx%40thread.v2/0?context=%7b%22Tid%22%3a%223dd8961f-e488-4e60-8e11-a82d994e183d%22%2c%22Oid%22%3a%22ce0d0f4a-6fbf-4ba7-b0c7-20184e3bcb53%22%7d
https://docs.google.com/document/d/1yA-MeJf6RYY-ZXpdol0t7YoDoqtwAyBhFLr5thu5pFI/edit#

Good to know, thanks! I may drop by some time.

We discussed parallelization strategies in several ODMs, but there was no firm decision. Most of the discussion was in the GPU context, I believe, where we need not only to say which loops remain parallel but also how they are mapped to the device. We can add annotations that this pass would consume. Alternatively, it is possible to first transform any SCF parallel that should not be converted to SCF for, and then just run a blanket conversion of all remaining parallel operations. We already have the former conversion implemented, can just as well add a way to control it at a finer grain.

We can expose any part of this transformation somewhere in Utils.h if you find them useful for Flang/

OK, thanks for the explanation. We might need it (not sure though).

Don't hesitate to ping me.

mlir/include/mlir/Conversion/Passes.h

 #include "mlir/Conversion/GPUToSPIRV/ConvertGPUToSPIRVPass.h"
 #include "mlir/Conversion/GPUToVulkan/ConvertGPUToVulkanPass.h"
 #include "mlir/Conversion/LinalgToLLVM/LinalgToLLVM.h"
 #include "mlir/Conversion/LinalgToSPIRV/LinalgToSPIRVPass.h"
 #include "mlir/Conversion/LinalgToStandard/LinalgToStandard.h"
 #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h"
 #include "mlir/Conversion/PDLToPDLInterp/PDLToPDLInterp.h"
 #include "mlir/Conversion/SCFToGPU/SCFToGPUPass.h"
+#include "mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h"
 #include "mlir/Conversion/SCFToStandard/SCFToStandard.h"
 #include "mlir/Conversion/SPIRVToLLVM/ConvertSPIRVToLLVMPass.h"
 #include "mlir/Conversion/ShapeToStandard/ShapeToStandard.h"
 #include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVMPass.h"
 #include "mlir/Conversion/StandardToSPIRV/ConvertStandardToSPIRVPass.h"
 #include "mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h"
 #include "mlir/Conversion/VectorToROCDL/VectorToROCDL.h"
 #include "mlir/Conversion/VectorToSCF/VectorToSCF.h"

mlir/include/mlir/Conversion/Passes.td

 
 def ConvertPDLToPDLInterp : Pass<"convert-pdl-to-pdl-interp", "ModuleOp"> {
   let summary = "Convert PDL ops to PDL interpreter ops";
   let constructor = "mlir::createPDLToPDLInterpPass()";
   let dependentDialects = ["pdl_interp::PDLInterpDialect"];
 }
 
 //===----------------------------------------------------------------------===//
+// SCFToOpenMP
+//===----------------------------------------------------------------------===//
+
+def ConvertSCFToOpenMP : FunctionPass<"convert-scf-to-openmp"> {
+  let summary = "Convert SCF parallel loop to OpenMP parallel + workshare "
+                "constructs.";
+  let constructor = "mlir::createConvertSCFToOpenMPPass()";
+  let dependentDialects = ["omp::OpenMPDialect"];
+}
+
+//===----------------------------------------------------------------------===//
 // SCFToStandard
 //===----------------------------------------------------------------------===//
 
 def SCFToStandard : Pass<"convert-scf-to-std"> {
   let summary = "Convert SCF dialect to Standard dialect, replacing structured"
                 " control flow with a CFG";
   let constructor = "mlir::createLowerToCFGPass()";
   let dependentDialects = ["StandardOpsDialect"];

mlir/include/mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h

+//===- ConvertSCFToOpenMP.h - SCF to OpenMP pass entrypoint -----*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_CONVERSION_SCFTOOPENMP_SCFTOOPENMP_H
+#define MLIR_CONVERSION_SCFTOOPENMP_SCFTOOPENMP_H
+
+#include <memory>
+
+namespace mlir {
+class FuncOp;
+template <typename T>
+class OperationPass;
+
+std::unique_ptr<OperationPass<FuncOp>> createConvertSCFToOpenMPPass();
+
+} // namespace mlir
+
+#endif // MLIR_CONVERSION_SCFTOOPENMP_SCFTOOPENMP_H

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

let arguments = (ins Optional<AnyType>:$if_expr_var,
              Variadic<AnyType>:$shared_vars,
              Variadic<AnyType>:$copyin_vars,
              Variadic<AnyType>:$allocate_vars,
              Variadic<AnyType>:$allocators_vars,
              OptionalAttr<ProcBindKind>:$proc_bind_val);
 
   let regions = (region AnyRegion:$region);
 
+  let builders = [
+    OpBuilderDAG<(ins CArg<"ArrayRef<NamedAttribute>", "{}">:$attributes)>
+  ];
   let parser = [{ return parseParallelOp(parser, result); }];
   let printer = [{ return printParallelOp(p, *this); }];
   let verifier = [{ return ::verifyParallelOp(*this); }];
 }
 
 def TerminatorOp : OpenMP_Op<"terminator", [Terminator]> {
   let summary = "terminator for OpenMP regions";
   let description = [{
let arguments = (ins Variadic<IntLikeType>:$lowerBound,
              Variadic<AnyType>:$linear_step_vars,
              OptionalAttr<ScheduleKind>:$schedule_val,
              Optional<AnyType>:$schedule_chunk_var,
              Confined<OptionalAttr<I64Attr>, [IntMinValue<0>]>:$collapse_val,
              OptionalAttr<UnitAttr>:$nowait,
              Confined<OptionalAttr<I64Attr>, [IntMinValue<0>]>:$ordered_val,
              OptionalAttr<OrderKind>:$order_val);
 
+  let builders = [
+    OpBuilderDAG<(ins "ValueRange":$lowerBound, "ValueRange":$upperBound,
+                  "ValueRange":$step,
+                  CArg<"ArrayRef<NamedAttribute>", "{}">:$attributes)>
+  ];
+
   let regions = (region AnyRegion:$region);
 }
 
 def YieldOp : OpenMP_Op<"yield", [NoSideEffect, ReturnLike, Terminator,
                                   HasParent<"WsLoopOp">]> {
   let summary = "loop yield and termination operation";
   let description = [{
     "omp.yield" yields SSA values from the OpenMP dialect op region and

mlir/lib/Conversion/CMakeLists.txt

 add_subdirectory(AffineToStandard)
 add_subdirectory(AsyncToLLVM)
 add_subdirectory(AVX512ToLLVM)
 add_subdirectory(GPUCommon)
 add_subdirectory(GPUToNVVM)
 add_subdirectory(GPUToROCDL)
 add_subdirectory(GPUToSPIRV)
 add_subdirectory(GPUToVulkan)
 add_subdirectory(LinalgToLLVM)
 add_subdirectory(LinalgToSPIRV)
 add_subdirectory(LinalgToStandard)
 add_subdirectory(OpenMPToLLVM)
 add_subdirectory(PDLToPDLInterp)
 add_subdirectory(SCFToGPU)
+add_subdirectory(SCFToOpenMP)
 add_subdirectory(SCFToSPIRV)
 add_subdirectory(SCFToStandard)
 add_subdirectory(ShapeToStandard)
 add_subdirectory(SPIRVToLLVM)
 add_subdirectory(StandardToLLVM)
 add_subdirectory(StandardToSPIRV)
 add_subdirectory(VectorToROCDL)
 add_subdirectory(VectorToLLVM)
 add_subdirectory(VectorToSCF)
 add_subdirectory(VectorToSPIRV)

mlir/lib/Conversion/PassDetail.h

 class LLVMDialect;
 class LLVMAVX512Dialect;
 } // end namespace LLVM
 
 namespace NVVM {
 class NVVMDialect;
 } // end namespace NVVM
 
+namespace omp {
+class OpenMPDialect;
+} // end namespace omp
+
 namespace pdl_interp {
 class PDLInterpDialect;
 } // end namespace pdl_interp
 
 namespace ROCDL {
 class ROCDLDialect;
 } // end namespace ROCDL
 

mlir/lib/Conversion/SCFToOpenMP/CMakeLists.txt

+add_mlir_conversion_library(MLIRSCFToOpenMP
+  SCFToOpenMP.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Conversion/SCFToStandard
+
+  DEPENDS
+  MLIRConversionPassIncGen
+
+  LINK_COMPONENTS
+  Core
+
+  LINK_LIBS PUBLIC
+  MLIROpenMP
+  MLIRSCF
+  MLIRTransforms
+  )

mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp

+//===- SCFToOpenMP.cpp - Structured Control Flow to OpenMP conversion -----===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass to convert scf.parallel operations into OpenMP
+// parallel loops.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Conversion/SCFToOpenMP/SCFToOpenMP.h"
+#include "../PassDetail.h"
+#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
+#include "mlir/Dialect/SCF/SCF.h"
+#include "mlir/Transforms/DialectConversion.h"
+
+using namespace mlir;
+
+namespace {
+
+/// Converts SCF parallel operation into an OpenMP workshare loop construct.
+struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> {
+  using OpRewritePattern<scf::ParallelOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(scf::ParallelOp parallelOp,
+                                PatternRewriter &rewriter) const override {
+    // TODO: add support for reductions when OpenMP loops have them.
+    if (parallelOp.getNumResults() != 0)
+      return rewriter.notifyMatchFailure(
+          parallelOp,
+          "OpenMP dialect does not yet support loops with reductions");
+
+    // Replace SCF yield with OpenMP yield.
+    {
+      OpBuilder::InsertionGuard guard(rewriter);
+      rewriter.setInsertionPointToEnd(parallelOp.getBody());
+      assert(llvm::hasSingleElement(parallelOp.region()) &&
+             "expected scf.parallel to have one block");
+      rewriter.replaceOpWithNewOp<omp::YieldOp>(
+          parallelOp.getBody()->getTerminator(), ValueRange());
+    }
+
+    // Replace the loop.
+    auto loop = rewriter.create<omp::WsLoopOp>(
+        parallelOp.getLoc(), parallelOp.lowerBound(), parallelOp.upperBound(),
+        parallelOp.step());
+    rewriter.inlineRegionBefore(parallelOp.region(), loop.region(),
+                                loop.region().begin());
+    rewriter.eraseOp(parallelOp);
+    return success();
+  }
+};
+
+/// A pass converting SCF operations to OpenMP operations.
+struct SCFToOpenMPPass : public ConvertSCFToOpenMPBase<SCFToOpenMPPass> {
+  /// Inserts OpenMP "parallel" operations around top-level SCF "parallel"
+  /// operations in the given function. This is implemented as a direct IR
+  /// modification rather than as a conversion pattern because it does not
+  /// modify the top-level operation it matches, which is a requirement for
+  /// rewrite patterns.
+  void insertOpenMPParallel(FuncOp func) {
+    // Collect top-level SCF "parallel" ops.
+    SmallVector<scf::ParallelOp, 4> topLevelParallelOps;
+    func.walk([&topLevelParallelOps](scf::ParallelOp parallelOp) {
+      // Ignore ops that are already within OpenMP parallel construct.
+      if (!parallelOp.getParentOfType<scf::ParallelOp>())
+        topLevelParallelOps.push_back(parallelOp);
+    });
+
+    // Wrap SCF ops into OpenMP "parallel" ops.
+    for (scf::ParallelOp parallelOp : topLevelParallelOps) {
+      OpBuilder builder(parallelOp);
+      auto omp = builder.create<omp::ParallelOp>(parallelOp.getLoc());
+      Block *block = builder.createBlock(&omp.getRegion());
+      builder.create<omp::TerminatorOp>(parallelOp.getLoc());
+      block->getOperations().splice(
+          block->begin(),
+          parallelOp.getOperation()->getBlock()->getOperations(),
+          parallelOp.getOperation());
+    }
+  }
+
+  /// Applies the conversion patterns in the given function.
+  LogicalResult applyPatterns(FuncOp func) {
+    ConversionTarget target(getContext());
+    target.addIllegalOp<scf::ParallelOp>();
+    target.addDynamicallyLegalOp<scf::YieldOp>([](scf::YieldOp op) {
+      return !isa<scf::ParallelOp>(op.getParentOp());
+    });
+    target.addLegalDialect<omp::OpenMPDialect>();
+
+    OwningRewritePatternList patterns;
+    patterns.insert<ParallelOpLowering>(&getContext());
+    FrozenRewritePatternList frozen(std::move(patterns));
+    return applyPartialConversion(getFunction(), target, frozen);
+  }
+
+  /// Pass entry point.
+  void runOnFunction() override {
+    insertOpenMPParallel(getFunction());
+    if (failed(applyPatterns(getFunction())))
+      signalPassFailure();
+  }
+};
+
+} // end namespace
+
+std::unique_ptr<OperationPass<FuncOp>> mlir::createConvertSCFToOpenMPPass() {
+  return std::make_unique<SCFToOpenMPPass>();
+}

mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp

 #include "mlir/Dialect/OpenMP/OpenMPOps.cpp.inc"
       >();
 }
 
 //===----------------------------------------------------------------------===//
 // ParallelOp
 //===----------------------------------------------------------------------===//
 
+void ParallelOp::build(OpBuilder &builder, OperationState &state,
+                       ArrayRef<NamedAttribute> attributes) {
+  ParallelOp::build(
+      builder, state, /*if_expr_var=*/nullptr, /*num_threads_var=*/nullptr,
+      /*default_val=*/nullptr, /*private_vars=*/ValueRange(),
+      /*firstprivate_vars=*/ValueRange(), /*shared_vars=*/ValueRange(),
+      /*copyin_vars=*/ValueRange(), /*allocate_vars=*/ValueRange(),
+      /*allocators_vars=*/ValueRange(), /*proc_bind_val=*/nullptr);
+  state.addAttributes(attributes);
+}
+
 /// Parse a list of operands with types.
 ///
 /// operand-and-type-list ::= `(` ssa-id-and-type-list `)`
 /// ssa-id-and-type-list ::= ssa-id-and-type |
 ///                          ssa-id-and-type `,` ssa-id-and-type-list
 /// ssa-id-and-type ::= ssa-id `:` type
 static ParseResult
 parseOperandAndTypeList(OpAsmParser &parser,
static ParseResult parseParallelOp(OpAsmParser &parser,
   Region *body = result.addRegion();
   SmallVector<OpAsmParser::OperandType, 4> regionArgs;
   SmallVector<Type, 4> regionArgTypes;
   if (parser.parseRegion(*body, regionArgs, regionArgTypes))
     return failure();
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// WsLoopOp
+//===----------------------------------------------------------------------===//
+
+void WsLoopOp::build(OpBuilder &builder, OperationState &state,
+                     ValueRange lowerBound, ValueRange upperBound,
+                     ValueRange step, ArrayRef<NamedAttribute> attributes) {
+  build(builder, state, TypeRange(), lowerBound, upperBound, step,
+        /*private_vars=*/ValueRange(),
+        /*firstprivate_vars=*/ValueRange(), /*lastprivate_vars=*/ValueRange(),
+        /*linear_vars=*/ValueRange(), /*linear_step_vars=*/ValueRange(),
+        /*schedule_val=*/nullptr, /*schedule_chunk_var=*/nullptr,
+        /*collapse_val=*/nullptr,
+        /*nowait=*/nullptr, /*ordered_val=*/nullptr, /*order_val=*/nullptr);
+  state.addAttributes(attributes);
+}
+
 #define GET_OP_CLASSES
 #include "mlir/Dialect/OpenMP/OpenMPOps.cpp.inc"

mlir/test/Conversion/SCFToOpenMP/scf-to-openmp.mlir

+// RUN: mlir-opt -convert-scf-to-openmp %s | FileCheck %s
+
+// CHECK-LABEL: @parallel
+func @parallel(%arg0: index, %arg1: index, %arg2: index,
+          %arg3: index, %arg4: index, %arg5: index) {
+  // CHECK: omp.parallel {
+  // CHECK:  "omp.wsloop"({{.*}}) ( {
+  scf.parallel (%i, %j) = (%arg0, %arg1) to (%arg2, %arg3) step (%arg4, %arg5) {
+    // CHECK:   test.payload
+    "test.payload"(%i, %j) : (index, index) -> ()
+    // CHECK:   omp.yield
+    // CHECK: }
+  }
+  // CHECK:   omp.terminator
+  // CHECK: }
+  return
+}
+
+// CHECK-LABEL: @nested_loops
+func @nested_loops(%arg0: index, %arg1: index, %arg2: index,
+                   %arg3: index, %arg4: index, %arg5: index) {
+  // CHECK: omp.parallel {
+  // CHECK:  "omp.wsloop"({{.*}}) ( {
+  // CHECK-NOT: omp.parallel
+  scf.parallel (%i) = (%arg0) to (%arg2) step (%arg4) {
+    // CHECK:   "omp.wsloop"({{.*}}) ( {
+    scf.parallel (%j) = (%arg1) to (%arg3) step (%arg5) {
+      // CHECK:   test.payload
+      "test.payload"(%i, %j) : (index, index) -> ()
+      // CHECK:   omp.yield
+      // CHECK: }
+    }
+    // CHECK:   omp.yield
+    // CHECK: }
+  }
+  // CHECK:   omp.terminator
+  // CHECK: }
+  return
+}
+
+func @adjacent_loops(%arg0: index, %arg1: index, %arg2: index,
+                     %arg3: index, %arg4: index, %arg5: index) {
+  // CHECK: omp.parallel {
+  // CHECK:  "omp.wsloop"({{.*}}) ( {
+  scf.parallel (%i) = (%arg0) to (%arg2) step (%arg4) {
+    // CHECK:   test.payload1
+    "test.payload1"(%i) : (index) -> ()
+    // CHECK:   omp.yield
+    // CHECK: }
+  }
+  // CHECK:   omp.terminator
+  // CHECK: }
+
+  // CHECK: omp.parallel {
+  // CHECK:  "omp.wsloop"({{.*}}) ( {
+  scf.parallel (%j) = (%arg1) to (%arg3) step (%arg5) {
+    // CHECK:   test.payload2
+    "test.payload2"(%j) : (index) -> ()
+    // CHECK:   omp.yield
+    // CHECK: }
+  }
+  // CHECK:   omp.terminator
+  // CHECK: }
+  return
+}