[OpenMP][OMPIRBuilder] Add delayed privatization support for wsloop

Extend MLIR to LLVM lowering by adding support for `omp.wsloop` for
delayed privatization. This also refactors a few bit of code to isolate
the logic needed for `firstprivate` initialization in a shared util that
can be used across constructs that need it.

Author: ergawy

mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp

@@ -268,7 +268,6 @@ static LogicalResult checkImplementationStatus(Operation &op) {
         checkAllocate(op, result);
         checkLinear(op, result);
         checkOrder(op, result);
-        checkPrivate(op, result);
       })
       .Case([&](omp::ParallelOp op) { checkAllocate(op, result); })
       .Case([&](omp::SimdOp op) {
@@ -1302,6 +1301,7 @@ allocatePrivateVars(llvm::IRBuilderBase &builder,
                     MutableArrayRef<mlir::Value> mlirPrivateVars,
                     llvm::SmallVectorImpl<llvm::Value *> &llvmPrivateVars,
                     const llvm::OpenMPIRBuilder::InsertPointTy &allocaIP) {
+  llvm::IRBuilderBase::InsertPointGuard guard(builder);
   // Allocate private vars
   llvm::BranchInst *allocaTerminator =
       llvm::cast<llvm::BranchInst>(allocaIP.getBlock()->getTerminator());
@@ -1363,6 +1363,86 @@ allocatePrivateVars(llvm::IRBuilderBase &builder,
   return afterAllocas;
 }
 
+static LogicalResult
+initFirstPrivateVars(llvm::IRBuilderBase &builder,
+                     LLVM::ModuleTranslation &moduleTranslation,
+                     SmallVectorImpl<mlir::Value> &mlirPrivateVars,
+                     SmallVectorImpl<llvm::Value *> &llvmPrivateVars,
+                     SmallVectorImpl<omp::PrivateClauseOp> &privateDecls,
+                     llvm::BasicBlock *afterAllocas) {
+  llvm::IRBuilderBase::InsertPointGuard guard(builder);
+  // Apply copy region for firstprivate.
+  bool needsFirstprivate =
+      llvm::any_of(privateDecls, [](omp::PrivateClauseOp &privOp) {
+        return privOp.getDataSharingType() ==
+               omp::DataSharingClauseType::FirstPrivate;
+      });
+
+  if (!needsFirstprivate)
+    return success();
+
+  assert(afterAllocas->getSinglePredecessor());
+
+  // Find the end of the allocation blocks
+  builder.SetInsertPoint(afterAllocas->getSinglePredecessor()->getTerminator());
+  llvm::BasicBlock *copyBlock =
+      splitBB(builder, /*CreateBranch=*/true, "omp.private.copy");
+  builder.SetInsertPoint(copyBlock->getFirstNonPHIOrDbgOrAlloca());
+
+  for (auto [decl, mlirVar, llvmVar] :
+       llvm::zip_equal(privateDecls, mlirPrivateVars, llvmPrivateVars)) {
+    if (decl.getDataSharingType() != omp::DataSharingClauseType::FirstPrivate)
+      continue;
+
+    // copyRegion implements `lhs = rhs`
+    Region &copyRegion = decl.getCopyRegion();
+
+    // map copyRegion rhs arg
+    llvm::Value *nonPrivateVar = moduleTranslation.lookupValue(mlirVar);
+    assert(nonPrivateVar);
+    moduleTranslation.mapValue(decl.getCopyMoldArg(), nonPrivateVar);
+
+    // map copyRegion lhs arg
+    moduleTranslation.mapValue(decl.getCopyPrivateArg(), llvmVar);
+
+    // in-place convert copy region
+    builder.SetInsertPoint(builder.GetInsertBlock()->getTerminator());
+    if (failed(inlineConvertOmpRegions(copyRegion, "omp.private.copy", builder,
+                                       moduleTranslation)))
+      return decl.emitError("failed to inline `copy` region of `omp.private`");
+
+    // ignore unused value yielded from copy region
+
+    // clear copy region block argument mapping in case it needs to be
+    // re-created with different sources for reuse of the same reduction
+    // decl
+    moduleTranslation.forgetMapping(copyRegion);
+  }
+
+  return success();
+}
+
+static LogicalResult
+cleanupPrivateVars(llvm::IRBuilderBase &builder,
+                   LLVM::ModuleTranslation &moduleTranslation, Location loc,
+                   SmallVectorImpl<llvm::Value *> &llvmPrivateVars,
+                   SmallVectorImpl<omp::PrivateClauseOp> &privateDecls) {
+  // private variable deallocation
+  SmallVector<Region *> privateCleanupRegions;
+  llvm::transform(privateDecls, std::back_inserter(privateCleanupRegions),
+                  [](omp::PrivateClauseOp privatizer) {
+                    return &privatizer.getDeallocRegion();
+                  });
+
+  if (failed(inlineOmpRegionCleanup(
+          privateCleanupRegions, llvmPrivateVars, moduleTranslation, builder,
+          "omp.private.dealloc", /*shouldLoadCleanupRegionArg=*/false)))
+    return mlir::emitError(loc, "failed to inline `dealloc` region of an "
+                                "`omp.private` op in");
+
+  return success();
+}
+
 static LogicalResult
 convertOmpSections(Operation &opInst, llvm::IRBuilderBase &builder,
                    LLVM::ModuleTranslation &moduleTranslation) {
@@ -1622,50 +1702,10 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
     if (handleError(afterAllocas, *taskOp).failed())
       return llvm::make_error<PreviouslyReportedError>();
 
-    // Apply copy region for firstprivate
-    bool needsFirstPrivate =
-        llvm::any_of(privateDecls, [](omp::PrivateClauseOp &privOp) {
-          return privOp.getDataSharingType() ==
-                 omp::DataSharingClauseType::FirstPrivate;
-        });
-    if (needsFirstPrivate) {
-      // Find the end of the allocation blocks
-      assert(afterAllocas.get()->getSinglePredecessor());
-      builder.SetInsertPoint(
-          afterAllocas.get()->getSinglePredecessor()->getTerminator());
-      llvm::BasicBlock *copyBlock =
-          splitBB(builder, /*CreateBranch=*/true, "omp.private.copy");
-      builder.SetInsertPoint(copyBlock->getFirstNonPHIOrDbgOrAlloca());
-    }
-    for (auto [decl, mlirVar, llvmVar] :
-         llvm::zip_equal(privateDecls, mlirPrivateVars, llvmPrivateVars)) {
-      if (decl.getDataSharingType() != omp::DataSharingClauseType::FirstPrivate)
-        continue;
-
-      // copyRegion implements `lhs = rhs`
-      Region &copyRegion = decl.getCopyRegion();
-
-      // map copyRegion rhs arg
-      llvm::Value *nonPrivateVar = moduleTranslation.lookupValue(mlirVar);
-      assert(nonPrivateVar);
-      moduleTranslation.mapValue(decl.getCopyMoldArg(), nonPrivateVar);
-
-      // map copyRegion lhs arg
-      moduleTranslation.mapValue(decl.getCopyPrivateArg(), llvmVar);
-
-      // in-place convert copy region
-      builder.SetInsertPoint(builder.GetInsertBlock()->getTerminator());
-      if (failed(inlineConvertOmpRegions(copyRegion, "omp.private.copy",
-                                         builder, moduleTranslation)))
-        return llvm::createStringError(
-            "failed to inline `copy` region of an `omp.private` op in taskOp");
-
-      // ignore unused value yielded from copy region
-
-      // clear copy region block argument mapping in case it needs to be
-      // re-created with different source for reuse of the same reduction decl
-      moduleTranslation.forgetMapping(copyRegion);
-    }
+    if (failed(initFirstPrivateVars(builder, moduleTranslation, mlirPrivateVars,
+                                    llvmPrivateVars, privateDecls,
+                                    afterAllocas.get())))
+      return llvm::make_error<PreviouslyReportedError>();
 
     // translate the body of the task:
     builder.restoreIP(codegenIP);
@@ -1674,19 +1714,11 @@ convertOmpTaskOp(omp::TaskOp taskOp, llvm::IRBuilderBase &builder,
     if (failed(handleError(continuationBlockOrError, *taskOp)))
       return llvm::make_error<PreviouslyReportedError>();
 
-    // private variable deallocation
-    SmallVector<Region *> privateCleanupRegions;
-    llvm::transform(privateDecls, std::back_inserter(privateCleanupRegions),
-                    [](omp::PrivateClauseOp privatizer) {
-                      return &privatizer.getDeallocRegion();
-                    });
-
     builder.SetInsertPoint(continuationBlockOrError.get()->getTerminator());
-    if (failed(inlineOmpRegionCleanup(
-            privateCleanupRegions, llvmPrivateVars, moduleTranslation, builder,
-            "omp.private.dealloc", /*shouldLoadCleanupRegionArg=*/false)))
-      return llvm::createStringError("failed to inline `dealloc` region of an "
-                                     "`omp.private` op in an omp.task");
+
+    if (failed(cleanupPrivateVars(builder, moduleTranslation, taskOp.getLoc(),
+                                  llvmPrivateVars, privateDecls)))
+      return llvm::make_error<PreviouslyReportedError>();
 
     return llvm::Error::success();
   };
@@ -1777,22 +1809,56 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
     chunk = builder.CreateSExtOrTrunc(chunkVar, ivType);
   }
 
+  MutableArrayRef<BlockArgument> privateBlockArgs =
+      cast<omp::BlockArgOpenMPOpInterface>(*wsloopOp).getPrivateBlockArgs();
+  SmallVector<mlir::Value> mlirPrivateVars;
+  SmallVector<llvm::Value *> llvmPrivateVars;
+  SmallVector<omp::PrivateClauseOp> privateDecls;
+  mlirPrivateVars.reserve(privateBlockArgs.size());
+  llvmPrivateVars.reserve(privateBlockArgs.size());
+  collectPrivatizationDecls(wsloopOp, privateDecls);
+
+  for (mlir::Value privateVar : wsloopOp.getPrivateVars())
+    mlirPrivateVars.push_back(privateVar);
+
   SmallVector<omp::DeclareReductionOp> reductionDecls;
   collectReductionDecls(wsloopOp, reductionDecls);
   llvm::OpenMPIRBuilder::InsertPointTy allocaIP =
       findAllocaInsertPoint(builder, moduleTranslation);
 
   SmallVector<llvm::Value *> privateReductionVariables(
       wsloopOp.getNumReductionVars());
+
+  llvm::Expected<llvm::BasicBlock *> afterAllocas = allocatePrivateVars(
+      builder, moduleTranslation, privateBlockArgs, privateDecls,
+      mlirPrivateVars, llvmPrivateVars, allocaIP);
+  if (handleError(afterAllocas, opInst).failed())
+    return failure();
+
   DenseMap<Value, llvm::Value *> reductionVariableMap;
 
   MutableArrayRef<BlockArgument> reductionArgs =
       cast<omp::BlockArgOpenMPOpInterface>(opInst).getReductionBlockArgs();
 
-  if (failed(allocAndInitializeReductionVars(
-          wsloopOp, reductionArgs, builder, moduleTranslation, allocaIP,
-          reductionDecls, privateReductionVariables, reductionVariableMap,
-          isByRef)))
+  SmallVector<DeferredStore> deferredStores;
+
+  if (failed(allocReductionVars(wsloopOp, reductionArgs, builder,
+                                moduleTranslation, allocaIP, reductionDecls,
+                                privateReductionVariables, reductionVariableMap,
+                                deferredStores, isByRef)))
+    return failure();
+
+  if (failed(initFirstPrivateVars(builder, moduleTranslation, mlirPrivateVars,
+                                  llvmPrivateVars, privateDecls,
+                                  afterAllocas.get())))
+    return failure();
+
+  assert(afterAllocas.get()->getSinglePredecessor());
+  if (failed(initReductionVars(wsloopOp, reductionArgs, builder,
+                               moduleTranslation,
+                               afterAllocas.get()->getSinglePredecessor(),
+                               reductionDecls, privateReductionVariables,
+                               reductionVariableMap, isByRef, deferredStores)))
     return failure();
 
   // TODO: Replace this with proper composite translation support.
@@ -1899,9 +1965,13 @@ convertOmpWsloop(Operation &opInst, llvm::IRBuilderBase &builder,
   builder.restoreIP(afterIP);
 
   // Process the reductions if required.
-  return createReductionsAndCleanup(wsloopOp, builder, moduleTranslation,
-                                    allocaIP, reductionDecls,
-                                    privateReductionVariables, isByRef);
+  if (failed(createReductionsAndCleanup(wsloopOp, builder, moduleTranslation,
+                                        allocaIP, reductionDecls,
+                                        privateReductionVariables, isByRef)))
+    return failure();
+
+  return cleanupPrivateVars(builder, moduleTranslation, wsloopOp.getLoc(),
+                            llvmPrivateVars, privateDecls);
 }
 
 /// Converts the OpenMP parallel operation to LLVM IR.
@@ -1959,52 +2029,12 @@ convertOmpParallel(omp::ParallelOp opInst, llvm::IRBuilderBase &builder,
             deferredStores, isByRef)))
       return llvm::make_error<PreviouslyReportedError>();
 
-    // Apply copy region for firstprivate.
-    bool needsFirstprivate =
-        llvm::any_of(privateDecls, [](omp::PrivateClauseOp &privOp) {
-          return privOp.getDataSharingType() ==
-                 omp::DataSharingClauseType::FirstPrivate;
-        });
-    if (needsFirstprivate) {
-      // Find the end of the allocation blocks
-      assert(afterAllocas.get()->getSinglePredecessor());
-      builder.SetInsertPoint(
-          afterAllocas.get()->getSinglePredecessor()->getTerminator());
-      llvm::BasicBlock *copyBlock =
-          splitBB(builder, /*CreateBranch=*/true, "omp.private.copy");
-      builder.SetInsertPoint(copyBlock->getFirstNonPHIOrDbgOrAlloca());
-    }
-    for (auto [decl, mlirVar, llvmVar] :
-         llvm::zip_equal(privateDecls, mlirPrivateVars, llvmPrivateVars)) {
-      if (decl.getDataSharingType() != omp::DataSharingClauseType::FirstPrivate)
-        continue;
-
-      // copyRegion implements `lhs = rhs`
-      Region &copyRegion = decl.getCopyRegion();
-
-      // map copyRegion rhs arg
-      llvm::Value *nonPrivateVar = moduleTranslation.lookupValue(mlirVar);
-      assert(nonPrivateVar);
-      moduleTranslation.mapValue(decl.getCopyMoldArg(), nonPrivateVar);
-
-      // map copyRegion lhs arg
-      moduleTranslation.mapValue(decl.getCopyPrivateArg(), llvmVar);
-
-      // in-place convert copy region
-      builder.SetInsertPoint(builder.GetInsertBlock()->getTerminator());
-      if (failed(inlineConvertOmpRegions(copyRegion, "omp.private.copy",
-                                         builder, moduleTranslation)))
-        return llvm::createStringError(
-            "failed to inline `copy` region of `omp.private`");
-
-      // ignore unused value yielded from copy region
-
-      // clear copy region block argument mapping in case it needs to be
-      // re-created with different sources for reuse of the same reduction
-      // decl
-      moduleTranslation.forgetMapping(copyRegion);
-    }
+    if (failed(initFirstPrivateVars(builder, moduleTranslation, mlirPrivateVars,
+                                    llvmPrivateVars, privateDecls,
+                                    afterAllocas.get())))
+      return llvm::make_error<PreviouslyReportedError>();
 
+    assert(afterAllocas.get()->getSinglePredecessor());
     if (failed(
             initReductionVars(opInst, reductionArgs, builder, moduleTranslation,
                               afterAllocas.get()->getSinglePredecessor(),
@@ -2089,17 +2119,9 @@ convertOmpParallel(omp::ParallelOp opInst, llvm::IRBuilderBase &builder,
       return llvm::createStringError(
           "failed to inline `cleanup` region of `omp.declare_reduction`");
 
-    SmallVector<Region *> privateCleanupRegions;
-    llvm::transform(privateDecls, std::back_inserter(privateCleanupRegions),
-                    [](omp::PrivateClauseOp privatizer) {
-                      return &privatizer.getDeallocRegion();
-                    });
-
-    if (failed(inlineOmpRegionCleanup(
-            privateCleanupRegions, llvmPrivateVars, moduleTranslation, builder,
-            "omp.private.dealloc", /*shouldLoadCleanupRegionArg=*/false)))
-      return llvm::createStringError(
-          "failed to inline `dealloc` region of `omp.private`");
+    if (failed(cleanupPrivateVars(builder, moduleTranslation, opInst.getLoc(),
+                                  llvmPrivateVars, privateDecls)))
+      return llvm::make_error<PreviouslyReportedError>();
 
     builder.restoreIP(oldIP);
     return llvm::Error::success();

mlir/test/Target/LLVMIR/openmp-todo.mlir

@@ -635,22 +635,3 @@ llvm.func @wsloop_order(%lb : i32, %ub : i32, %step : i32) {
   }
   llvm.return
 }
-
-// -----
-
-omp.private {type = private} @x.privatizer : !llvm.ptr alloc {
-^bb0(%arg0: !llvm.ptr):
-  %0 = llvm.mlir.constant(1 : i32) : i32
-  %1 = llvm.alloca %0 x i32 : (i32) -> !llvm.ptr
-  omp.yield(%1 : !llvm.ptr)
-}
-llvm.func @wsloop_private(%lb : i32, %ub : i32, %step : i32, %x : !llvm.ptr) {
-  // expected-error@below {{not yet implemented: Unhandled clause privatization in omp.wsloop operation}}
-  // expected-error@below {{LLVM Translation failed for operation: omp.wsloop}}
-  omp.wsloop private(@x.privatizer %x -> %arg0 : !llvm.ptr) {
-    omp.loop_nest (%iv) : i32 = (%lb) to (%ub) step (%step) {
-      omp.yield
-    }
-  }
-  llvm.return
-}