[flang-commits] [flang] [Flang][MLIR][OpenMP] Create a deferred declare target marking process for Bridge.cpp (PR #78502)

[via flang-commits]

================
@@ -4471,6 +4518,47 @@ bool Fortran::lower::isOpenMPDeviceDeclareTarget(
       ompDecl.u);
 }
 
+// In certain cases such as subroutine or function interfaces which declare
+// but do not define or directly call the subroutine or function in the same
+// module, their lowering is delayed until after the declare target construct
+// itself is processed, so there symbol is not within the table.
+//
+// This function will also return true if we encounter any device declare
+// target cases, to satisfy checking if we require the requires attributes
+// on the module.
+bool Fortran::lower::markOpenMPDeferredDeclareTargetFunctions(
+    mlir::Operation *mod,
+    llvm::SmallVectorImpl<OMPDeferredDeclTarInfo> &deferredDeclareTargets,
+    AbstractConverter &converter) {
+  bool deviceCodeFound = false;
+  if (auto modOp = llvm::dyn_cast<mlir::ModuleOp>(mod)) {
+    for (auto sym : deferredDeclareTargets) {
+      mlir::Operation *op =
+          modOp.lookupSymbol(converter.mangleName(std::get<2>(sym)));
+
+      // Due to interfaces being optionally emitted on usage in a module,
+      // not finding an operation at this point cannot be a hard error, we
+      // simply ignore it for now.
----------------
agozillon wrote:

It's a TODO, it's something that will likely have to be handled at the semantic checking level, there isn't a way for us to hard-error out for not finding an operation even at the end of the module (and this is likely not the place to do so, I would imagine most of these checks should be done via semantic checks). Flang does not generate an operation for interfaces not in use by the module, so whenever that's the case we'd have emitted a TODO error message which would break perfectly valid code.

https://github.com/llvm/llvm-project/pull/78502