================
@@ -4898,6 +4898,234 @@ 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);
+
+ const OMPDeclareReductionDecl *UDR = getReductionInit(ReductionOps);
+ 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,
+ llvm::Constant::getNullValue(LLVMType), ".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);
----------------
alexey-bataev wrote:
```suggestion
PrivateType.getQualifiers(),
/*IsInitializer=*/true);
```
https://github.com/llvm/llvm-project/pull/134709
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
