[clang] [OpenMP 6.0 ]Codegen for Reduction over private variables with reduction clause (PR #134709)

[CHANDRA GHALE via cfe-commits]

================
@@ -4899,6 +4899,234 @@ void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF,
   }
 }
 
+void CGOpenMPRuntime::emitPrivateReduction(
+    CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates,
+    ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs,
+    ArrayRef<const Expr *> ReductionOps) {
+  if (LHSExprs.empty() || Privates.empty() || ReductionOps.empty())
+    return;
+
+  if (LHSExprs.size() != Privates.size() ||
+      LHSExprs.size() != ReductionOps.size())
+    return;
+
+  //  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.
----------------
chandraghale wrote:

Agreed. Handling simple cases with atomics for performance is a good idea.I would like to address this in a follow-up patch. Currently, synchronization generated here is independent of the `nowait` clause.


https://github.com/llvm/llvm-project/pull/134709