[flang-commits] [flang] [llvm] [mlir] [flang][OpenMP] Lower target in_reduction (PR #199967)

[Sairudra More via flang-commits]

================
@@ -1218,6 +1218,57 @@ Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) {
       "omp_global_thread_num");
 }
 
+OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetInReduction(
+    const LocationDescription &Loc, ArrayRef<Value *> OrigPtrs,
+    ArrayRef<Type *> ResultPtrTys,
+    function_ref<void(unsigned, Value *)> MapPrivate) {
+  assert(OrigPtrs.size() == ResultPtrTys.size() &&
+         "expected one result pointer type per in_reduction item");
+  if (!updateToLocation(Loc))
+    return Loc.IP;
+  if (OrigPtrs.empty())
+    return Builder.saveIP();
+
+  // Compute the executing thread's gtid once for the whole target body and
+  // reuse it for every in_reduction lookup, so a target with several
+  // in_reduction items does not emit a redundant __kmpc_global_thread_num per
+  // item. Owning the gtid setup and the lookup loop here keeps the runtime
+  // generation inside OpenMPIRBuilder rather than the MLIR translation layer.
----------------
Saieiei wrote:

Done!

https://github.com/llvm/llvm-project/pull/199967