[OpenMP 6.0 ]Codegen for Reduction over private variables with reduction clause

clang/docs/OpenMPSupport.rst

@@ -406,7 +406,8 @@ implementation.
 +-------------------------------------------------------------+---------------------------+---------------------------+--------------------------------------------------------------------------+
 | Extensions to atomic construct                              | :none:`unclaimed`         | :none:`unclaimed`         |                                                                          |
 +-------------------------------------------------------------+---------------------------+---------------------------+--------------------------------------------------------------------------+ 
-| Private reductions                                          | :part:`partial`           | :none:`unclaimed`         | Parse/Sema:https://github.com/llvm/llvm-project/pull/129938              |
+| Private reductions                                          | :good:`mostly`            | :none:`unclaimed`         | Parse/Sema:https://github.com/llvm/llvm-project/pull/129938              |
+|                                                             |                           |                           | Codegen: https://github.com/llvm/llvm-project/pull/134709                |
 +-------------------------------------------------------------+---------------------------+---------------------------+--------------------------------------------------------------------------+
 | Self maps                                                   | :part:`partial`           | :none:`unclaimed`         | parsing/sema done: https://github.com/llvm/llvm-project/pull/129888      |
 +-------------------------------------------------------------+---------------------------+---------------------------+--------------------------------------------------------------------------+

clang/docs/ReleaseNotes.rst

@@ -937,15 +937,7 @@ OpenMP Support
 - Added support 'no_openmp_constructs' assumption clause.
 - Added support for 'self_maps' in map and requirement clause.
 - Added support for 'omp stripe' directive.
-- Fixed a crashing bug with ``omp unroll partial`` if the argument to
-  ``partial`` was an invalid expression. (#GH139267)
-- Fixed a crashing bug with ``omp tile sizes`` if the argument to ``sizes`` was
-  an invalid expression. (#GH139073)
-- Fixed a crashing bug with ``omp simd collapse`` if the argument to
-  ``collapse`` was an invalid expression. (#GH138493)
-- Fixed a crashing bug with a malformed ``cancel`` directive. (#GH139360)
-- Fixed a crashing bug with ``omp distribute dist_schedule`` if the argument to
-  ``dist_schedule`` was not strictly positive. (#GH139266)
+- Added support for private variable reduction.
 
 Improvements
 ^^^^^^^^^^^^

clang/lib/CodeGen/CGOpenMPRuntime.cpp

@@ -4898,6 +4898,274 @@ void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF,
   }
 }
 
+void CGOpenMPRuntime::emitPrivateReduction(
+    CodeGenFunction &CGF, SourceLocation Loc, const Expr *Privates,
+    const Expr *LHSExprs, const Expr *RHSExprs, const Expr *ReductionOps) {
+
+  //  Create a shared global variable (__shared_reduction_var) to accumulate the
+  //  final result.
+  //
+  //  Call __kmpc_barrier to synchronize threads before initialization.
+  //
+  //  The master thread (thread_id == 0) initializes __shared_reduction_var
+  //    with the identity value or initializer.
+  //
+  //  Call __kmpc_barrier to synchronize before combining.
+  //  For each i:
+  //    - Thread enters critical section.
+  //    - Reads its private value from LHSExprs[i].
+  //    - Updates __shared_reduction_var[i] = RedOp_i(__shared_reduction_var[i],
+  //    LHSExprs[i]).
+  //    - Exits critical section.
+  //
+  //  Call __kmpc_barrier after combining.
+  //
+  //  Each thread copies __shared_reduction_var[i] back to LHSExprs[i].
+  //
+  //  Final __kmpc_barrier to synchronize after broadcasting
+  QualType PrivateType = Privates->getType();
+  llvm::Type *LLVMType = CGF.ConvertTypeForMem(PrivateType);
+
+  llvm::Constant *InitVal = nullptr;
+  const OMPDeclareReductionDecl *UDR = getReductionInit(ReductionOps);
+  // Determine the initial value for the shared reduction variable
+  if (!UDR) {
+    InitVal = llvm::Constant::getNullValue(LLVMType);
+    if (const auto *DRE = dyn_cast<DeclRefExpr>(Privates)) {
+      if (const auto *VD = dyn_cast<VarDecl>(DRE->getDecl())) {
+        const Expr *InitExpr = VD->getInit();
+        if (InitExpr) {
+          Expr::EvalResult Result;
+          if (InitExpr->EvaluateAsRValue(Result, CGF.getContext())) {
+            APValue &InitValue = Result.Val;
+            if (InitValue.isInt())
+              InitVal = llvm::ConstantInt::get(LLVMType, InitValue.getInt());
+            else if (InitValue.isFloat())
+              InitVal = llvm::ConstantFP::get(LLVMType, InitValue.getFloat());
+            else if (InitValue.isComplexInt()) {
+              // For complex int: create struct { real, imag }
+              llvm::Constant *Real = llvm::ConstantInt::get(
+                  cast<llvm::StructType>(LLVMType)->getElementType(0),
+                  InitValue.getComplexIntReal());
+              llvm::Constant *Imag = llvm::ConstantInt::get(
+                  cast<llvm::StructType>(LLVMType)->getElementType(1),
+                  InitValue.getComplexIntImag());
+              InitVal = llvm::ConstantStruct::get(
+                  cast<llvm::StructType>(LLVMType), {Real, Imag});
+            } else if (InitValue.isComplexFloat()) {
+              llvm::Constant *Real = llvm::ConstantFP::get(
+                  cast<llvm::StructType>(LLVMType)->getElementType(0),
+                  InitValue.getComplexFloatReal());
+              llvm::Constant *Imag = llvm::ConstantFP::get(
+                  cast<llvm::StructType>(LLVMType)->getElementType(1),
+                  InitValue.getComplexFloatImag());
+              InitVal = llvm::ConstantStruct::get(
+                  cast<llvm::StructType>(LLVMType), {Real, Imag});
+            }
+          }
+        }
+      }
+    }
+  } else {
+    InitVal = llvm::Constant::getNullValue(LLVMType);
+  }
+  std::string ReductionVarNameStr;
+  if (const auto *DRE = dyn_cast<DeclRefExpr>(Privates->IgnoreParenCasts()))
+    ReductionVarNameStr = DRE->getDecl()->getNameAsString();
+  else
+    ReductionVarNameStr = "unnamed_priv_var";
+
+  // Create an internal shared variable
+  std::string SharedName =
+      CGM.getOpenMPRuntime().getName({"internal_pivate_", ReductionVarNameStr});
+  llvm::GlobalVariable *SharedVar = new llvm::GlobalVariable(
+      CGM.getModule(), LLVMType, false, llvm::GlobalValue::InternalLinkage,
+      InitVal, ".omp.reduction." + SharedName, nullptr,
+      llvm::GlobalVariable::NotThreadLocal);
+
+  SharedVar->setAlignment(
+      llvm::MaybeAlign(CGF.getContext().getTypeAlign(PrivateType) / 8));
+
+  Address SharedResult(SharedVar, SharedVar->getValueType(),
+                       CGF.getContext().getTypeAlignInChars(PrivateType));
+
+  llvm::Value *ThreadId = getThreadID(CGF, Loc);
+  llvm::Value *BarrierLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE);
+  llvm::Value *BarrierArgs[] = {BarrierLoc, ThreadId};
+
+  llvm::BasicBlock *InitBB = CGF.createBasicBlock("init");
+  llvm::BasicBlock *InitEndBB = CGF.createBasicBlock("init.end");
+
+  llvm::Value *IsWorker = CGF.Builder.CreateICmpEQ(
+      ThreadId, llvm::ConstantInt::get(ThreadId->getType(), 0));
+  CGF.Builder.CreateCondBr(IsWorker, InitBB, InitEndBB);
+
+  CGF.EmitBlock(InitBB);
+
+  auto EmitSharedInit = [&]() {
+    if (UDR) { // Check if it's a User-Defined Reduction
+      if (const Expr *UDRInitExpr = UDR->getInitializer()) {
+        std::pair<llvm::Function *, llvm::Function *> FnPair =
+            getUserDefinedReduction(UDR);
+        llvm::Function *InitializerFn = FnPair.second;
+        if (InitializerFn) {
+          if (const auto *CE =
+                  dyn_cast<CallExpr>(UDRInitExpr->IgnoreParenImpCasts())) {
+            const auto *OutDRE = cast<DeclRefExpr>(
+                cast<UnaryOperator>(CE->getArg(0)->IgnoreParenImpCasts())
+                    ->getSubExpr());
+            const VarDecl *OutVD = cast<VarDecl>(OutDRE->getDecl());
+
+            CodeGenFunction::OMPPrivateScope LocalScope(CGF);
+            LocalScope.addPrivate(OutVD, SharedResult);
+
+            (void)LocalScope.Privatize();
+            if (const auto *OVE = dyn_cast<OpaqueValueExpr>(
+                    CE->getCallee()->IgnoreParenImpCasts())) {
+              CodeGenFunction::OpaqueValueMapping OpaqueMap(
+                  CGF, OVE, RValue::get(InitializerFn));
+              CGF.EmitIgnoredExpr(CE);
+            } else {
+              CGF.EmitAnyExprToMem(UDRInitExpr, SharedResult,
+                                   PrivateType.getQualifiers(), true);
+            }
+          } else {
+            CGF.EmitAnyExprToMem(UDRInitExpr, SharedResult,
+                                 PrivateType.getQualifiers(), true);
+          }
+        } else {
+          CGF.EmitAnyExprToMem(UDRInitExpr, SharedResult,
+                               PrivateType.getQualifiers(), true);
+        }
+      } else {
+        // EmitNullInitialization handles default construction for C++ classes
+        // and zeroing for scalars, which is a reasonable default.
+        CGF.EmitNullInitialization(SharedResult, PrivateType);
+      }
+      return; // UDR initialization handled
+    }
+    if (const auto *DRE = dyn_cast<DeclRefExpr>(Privates)) {
+      if (const auto *VD = dyn_cast<VarDecl>(DRE->getDecl())) {
+        if (const Expr *InitExpr = VD->getInit()) {
+          CGF.EmitAnyExprToMem(InitExpr, SharedResult,
+                               PrivateType.getQualifiers(), true);
+          return;
+        }
+      }
+    }
+    CGF.EmitNullInitialization(SharedResult, PrivateType);
+  };
+  EmitSharedInit();
+  CGF.Builder.CreateBr(InitEndBB);
+  CGF.EmitBlock(InitEndBB);
+
+  CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
+                          CGM.getModule(), OMPRTL___kmpc_barrier),
+                      BarrierArgs);
+
+  const Expr *ReductionOp = ReductionOps;
+  const OMPDeclareReductionDecl *CurrentUDR = getReductionInit(ReductionOp);
+  LValue SharedLV = CGF.MakeAddrLValue(SharedResult, PrivateType);
+  LValue LHSLV = CGF.EmitLValue(LHSExprs);
+
+  auto EmitCriticalReduction = [&](auto ReductionGen) {
+    std::string CriticalName = getName({"reduction_critical"});
+    emitCriticalRegion(CGF, CriticalName, ReductionGen, Loc);
+  };
+
+  if (CurrentUDR) {
+    // Handle user-defined reduction.
+    auto ReductionGen = [&](CodeGenFunction &CGF, PrePostActionTy &Action) {
+      Action.Enter(CGF);
+      std::pair<llvm::Function *, llvm::Function *> FnPair =
+          getUserDefinedReduction(CurrentUDR);
+      if (FnPair.first) {
+        if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) {
+          const auto *OutDRE = cast<DeclRefExpr>(
+              cast<UnaryOperator>(CE->getArg(0)->IgnoreParenImpCasts())
+                  ->getSubExpr());
+          const auto *InDRE = cast<DeclRefExpr>(
+              cast<UnaryOperator>(CE->getArg(1)->IgnoreParenImpCasts())
+                  ->getSubExpr());
+          CodeGenFunction::OMPPrivateScope LocalScope(CGF);
+          LocalScope.addPrivate(cast<VarDecl>(OutDRE->getDecl()),
+                                SharedLV.getAddress());
+          LocalScope.addPrivate(cast<VarDecl>(InDRE->getDecl()),
+                                LHSLV.getAddress());
+          (void)LocalScope.Privatize();
+          emitReductionCombiner(CGF, ReductionOp);
+        }
+      }
+    };
+    EmitCriticalReduction(ReductionGen);
+  } else {
+    // Handle built-in reduction operations.
+    const Expr *ReductionClauseExpr = ReductionOp->IgnoreParenCasts();
+    if (const auto *Cleanup = dyn_cast<ExprWithCleanups>(ReductionClauseExpr))
+      ReductionClauseExpr = Cleanup->getSubExpr()->IgnoreParenCasts();
+
+    const Expr *AssignRHS = nullptr;
+    if (const auto *BinOp = dyn_cast<BinaryOperator>(ReductionClauseExpr)) {
+      if (BinOp->getOpcode() == BO_Assign)
+        AssignRHS = BinOp->getRHS();
+    } else if (const auto *OpCall =
+                   dyn_cast<CXXOperatorCallExpr>(ReductionClauseExpr)) {
+      if (OpCall->getOperator() == OO_Equal)
+        AssignRHS = OpCall->getArg(1);
+    }
+
+    if (!AssignRHS)
+      return;
+
+    auto ReductionGen = [&](CodeGenFunction &CGF, PrePostActionTy &Action) {
+      Action.Enter(CGF);
+      const auto *OmpOutDRE =
+          dyn_cast<DeclRefExpr>(LHSExprs->IgnoreParenImpCasts());
+      const auto *OmpInDRE =
+          dyn_cast<DeclRefExpr>(RHSExprs->IgnoreParenImpCasts());
+      if (!OmpOutDRE || !OmpInDRE)
+        return;
+      const VarDecl *OmpOutVD = cast<VarDecl>(OmpOutDRE->getDecl());
+      const VarDecl *OmpInVD = cast<VarDecl>(OmpInDRE->getDecl());
+      CodeGenFunction::OMPPrivateScope LocalScope(CGF);
+      LocalScope.addPrivate(OmpOutVD, SharedLV.getAddress());
+      LocalScope.addPrivate(OmpInVD, LHSLV.getAddress());
+      (void)LocalScope.Privatize();
+      // Emit the actual reduction operation
+      CGF.EmitIgnoredExpr(ReductionOp);
+    };
+    EmitCriticalReduction(ReductionGen);
+  }
+
+  CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
+                          CGM.getModule(), OMPRTL___kmpc_barrier),
+                      BarrierArgs);
+
+  // Broadcast final result
+  bool IsAggregate = PrivateType->isAggregateType();
+  LValue SharedLV1 = CGF.MakeAddrLValue(SharedResult, PrivateType);
+  llvm::Value *FinalResultVal = nullptr;
+  Address FinalResultAddr = Address::invalid();
+
+  if (IsAggregate)
+    FinalResultAddr = SharedResult;
+  else
+    FinalResultVal = CGF.EmitLoadOfScalar(SharedLV1, Loc);
+
+  LValue TargetLHSLV = CGF.EmitLValue(LHSExprs);
+  if (IsAggregate) {
+    CGF.EmitAggregateCopy(TargetLHSLV,
+                          CGF.MakeAddrLValue(FinalResultAddr, PrivateType),
+                          PrivateType, AggValueSlot::DoesNotOverlap, false);
+  } else {
+    CGF.EmitStoreOfScalar(FinalResultVal, TargetLHSLV);
+  }
+  // Final synchronization barrier
+  CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction(
+                          CGM.getModule(), OMPRTL___kmpc_barrier),
+                      BarrierArgs);
+}
+
 void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
                                     ArrayRef<const Expr *> Privates,
                                     ArrayRef<const Expr *> LHSExprs,
@@ -5153,7 +5421,7 @@ void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
       } else {
         // Emit as a critical region.
         auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *,
-                                           const Expr *, const Expr *) {
+                                     const Expr *, const Expr *) {
           CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime();
           std::string Name = RT.getName({"atomic_reduction"});
           RT.emitCriticalRegion(
@@ -5200,6 +5468,26 @@ void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc,
 
   CGF.EmitBranch(DefaultBB);
   CGF.EmitBlock(DefaultBB, /*IsFinished=*/true);
+  if (Options.IsPrivateVarReduction) {
+    if (LHSExprs.empty() || Privates.empty() || ReductionOps.empty())
+      return;
+    if (LHSExprs.size() != Privates.size() ||
+        LHSExprs.size() != ReductionOps.size())
+      return;
+    assert(!LHSExprs.empty() && "PrivateVarReduction: LHSExprs is empty");
+    assert(!Privates.empty() && "PrivateVarReduction: Privates is empty");
+    assert(!ReductionOps.empty() &&
+           "PrivateVarReduction: ReductionOps is empty");
+    assert(LHSExprs.size() == Privates.size() &&
+           "PrivateVarReduction: Privates size mismatch");
+    assert(LHSExprs.size() == ReductionOps.size() &&
+           "PrivateVarReduction: ReductionOps size mismatch");
+    for (unsigned I :
+         llvm::seq<unsigned>(std::min(ReductionOps.size(), LHSExprs.size()))) {
+      emitPrivateReduction(CGF, Loc, Privates[I], LHSExprs[I], RHSExprs[I],
+                           ReductionOps[I]);
+    }
+  }
 }
 
 /// Generates unique name for artificial threadprivate variables.

clang/lib/CodeGen/CGOpenMPRuntime.h

@@ -1201,8 +1201,20 @@ class CGOpenMPRuntime {
   struct ReductionOptionsTy {
     bool WithNowait;
     bool SimpleReduction;
+    bool IsPrivateVarReduction;
     OpenMPDirectiveKind ReductionKind;
   };
+
+  /// Emits code for private variable reduction
+  /// \param Privates List of private copies for original reduction arguments.
+  /// \param LHSExprs List of LHS in \a ReductionOps reduction operations.
+  /// \param RHSExprs List of RHS in \a ReductionOps reduction operations.
+  /// \param ReductionOps List of reduction operations in form 'LHS binop RHS'
+  /// or 'operator binop(LHS, RHS)'.
+  void emitPrivateReduction(CodeGenFunction &CGF, SourceLocation Loc,
+                            const Expr *Privates, const Expr *LHSExprs,
+                            const Expr *RHSExprs, const Expr *ReductionOps);
+
   /// Emit a code for reduction clause. Next code should be emitted for
   /// reduction:
   /// \code