[flang-commits] [flang] [llvm] [Proof-of-Concept][flang][OpenMP] Implicitely map allocatable record fields (PR #117867)

[via flang-commits]

================
@@ -485,6 +492,153 @@ class MapInfoFinalizationPass
       // clear all local allocations we made for any boxes in any prior
       // iterations from previous function scopes.
       localBoxAllocas.clear();
+      func->walk([&](mlir::omp::MapInfoOp op) {
+        mlir::Type underlyingType =
+            fir::unwrapRefType(op.getVarPtr().getType());
+
+        if (!fir::isRecordWithAllocatableMember(underlyingType))
+          return mlir::WalkResult::advance();
+
+        mlir::omp::TargetOp target =
+            mlir::dyn_cast_if_present<mlir::omp::TargetOp>(
+                getFirstTargetUser(op));
+
+        if (!target)
+          return mlir::WalkResult::advance();
+
+        auto mapClauseOwner =
+            llvm::dyn_cast<mlir::omp::MapClauseOwningOpInterface>(*target);
+
+        // TODO Add as a method to MapClauseOwningOpInterface.
+        unsigned mapVarIdx = 0;
+        for (auto [idx, mapOp] : llvm::enumerate(mapClauseOwner.getMapVars())) {
+          if (mapOp == op) {
+            mapVarIdx = idx;
+            break;
+          }
+        }
+
+        auto argIface =
+            llvm::dyn_cast<mlir::omp::BlockArgOpenMPOpInterface>(*target);
+        mlir::BlockArgument opBlockArg = argIface.getMapBlockArgs()[mapVarIdx];
----------------
agozillon wrote:

Might need to be a tad wary of how this works with use_device_addr/ptr, as they're both map info holders, and soon with Pranav's and your work I believe private may also be, and I don't think they have the same implicit connotations regular map does. We may also eventually need to be careful of this with declare mappers as well, as they're explicit user defined mappings that shouldn't have the implicit behavior but that's a problem for another day/PR when that work is further along :-) 

https://github.com/llvm/llvm-project/pull/117867