[llvm] [clang-tools-extra] [compiler-rt] [clang] [flang] [Flang][OpenMP] Restructure recursive lowering in `createBodyOfOp` (PR #77761)

[Krzysztof Parzyszek via cfe-commits]

================
@@ -2223,37 +2247,64 @@ static void createBodyOfOp(
                                             mlir::omp::YieldOp>(
         firOpBuilder, eval.getNestedEvaluations());
 
-  // Insert the terminator.
-  Fortran::lower::genOpenMPTerminator(firOpBuilder, op.getOperation(), loc);
-  // Reset the insert point to before the terminator.
-  resetBeforeTerminator(firOpBuilder, storeOp, block);
+  // Start with privatization, so that the lowering of the nested
+  // code will use the right symbols.
+  constexpr bool isLoop = std::is_same_v<Op, mlir::omp::WsLoopOp> ||
+                          std::is_same_v<Op, mlir::omp::SimdLoopOp>;
+  bool privatize = clauses && !outerCombined;
 
-  // Handle privatization. Do not privatize if this is the outer operation.
-  if (clauses && !outerCombined) {
-    constexpr bool isLoop = std::is_same_v<Op, mlir::omp::WsLoopOp> ||
-                            std::is_same_v<Op, mlir::omp::SimdLoopOp>;
+  firOpBuilder.setInsertionPoint(marker);
+  std::optional<DataSharingProcessor> tempDsp;
+  if (privatize) {
     if (!dsp) {
-      DataSharingProcessor proc(converter, *clauses, eval);
-      proc.processStep1();
-      proc.processStep2(op, isLoop);
-    } else {
-      if (isLoop && args.size() > 0)
-        dsp->setLoopIV(converter.getSymbolAddress(*args[0]));
-      dsp->processStep2(op, isLoop);
+      tempDsp.emplace(converter, *clauses, eval);
+      tempDsp->processStep1();
     }
-
-    if (storeOp)
-      firOpBuilder.setInsertionPointAfter(storeOp);
   }
 
   if constexpr (std::is_same_v<Op, mlir::omp::ParallelOp>) {
     threadPrivatizeVars(converter, eval);
-    if (clauses)
+    if (clauses) {
+      firOpBuilder.setInsertionPoint(marker);
       ClauseProcessor(converter, *clauses).processCopyin();
+    }
   }
 
-  if (genNested)
+  if (genNested) {
+    // genFIR(Evaluation&) tries to patch up unterminated blocks, causing
+    // a lot of trouble if the terminator generation is delayed past this
+    // point. Insert a temporary terminator here, then delete it.
+    firOpBuilder.setInsertionPointToEnd(&op.getRegion().back());
+    auto *temp = Fortran::lower::genOpenMPTerminator(firOpBuilder,
+                                                     op.getOperation(), loc);
+    firOpBuilder.setInsertionPointAfter(marker);
     genNestedEvaluations(converter, eval);
+    temp->erase();
+  }
+
+  if (auto *exitBlock = findExitBlock(op.getRegion())) {
+    firOpBuilder.setInsertionPointToEnd(exitBlock);
+    auto *term = Fortran::lower::genOpenMPTerminator(firOpBuilder,
+                                                     op.getOperation(), loc);
+    // Only insert lastprivate code when there actually is an exit block.
+    // Such a block may not exist if the nested code produced an infinite
+    // loop (this may not make sense in production code, but a user could
+    // write that and we should handle it).
----------------
kparzysz wrote:

What I'm going to do is to create a unique exiting block:  create a new block, and for each yet-unterminated block, will insert a branch to the new one.  Then insert the OMP terminator into it, and process the lastprivate in there.

This will get rid of the assertion.

https://github.com/llvm/llvm-project/pull/77761