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

[Sairudra More via flang-commits]

================
@@ -8605,6 +8702,42 @@ convertOmpTarget(Operation &opInst, llvm::IRBuilderBase &builder,
             targetOp.getPrivateNeedsBarrier(), &mappedPrivateVars)))
       return llvm::make_error<PreviouslyReportedError>();
 
+    // The target body accesses each in_reduction variable through its
+    // map_entries block argument. Redirect that block argument to the per-task
+    // private storage returned by __kmpc_task_reduction_get_th_data so the body
+    // accumulates into the reduction-private copy rather than the mapped
+    // original. The lookup must run inside the target task body so the gtid
+    // corresponds to the executing thread. The descriptor argument is NULL: the
+    // runtime walks enclosing taskgroups to locate the matching task_reduction
+    // registration for `origPtr`. Mirrors the in_reduction handling on
+    // omp.taskloop.context.
+    if (!inRedOrigPtrs.empty()) {
+      // OpenMPIRBuilder owns the runtime-generation path: it computes the
+      // target body gtid once (so several in_reduction items do not each emit a
+      // __kmpc_global_thread_num) and emits the per-item
+      // __kmpc_task_reduction_get_th_data lookup with the required
+      // address-space casts. The MLIR-specific work stays here: collect, per
+      // item, the type the private pointer must have (the map block argument's
+      // type), and, through the callback, rebind the map block argument that
+      // stands in for each in_reduction list item to the per-task
+      // reduction-private storage the runtime returns.
----------------
Saieiei wrote:

Done!

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