[clang] [OpenMP][CodeGen] Improved codegen for combined loop directives (PR #87278)

[Eli Friedman via cfe-commits]

================
@@ -6135,6 +6137,79 @@ processImplicitMapsWithDefaultMappers(Sema &S, DSAStackTy *Stack,
   }
 }
 
+namespace {
+/// A 'teams loop' with a nested 'loop bind(parallel)' or generic function
+/// call in the associated loop-nest cannot be a 'parallel for'.
+class TeamsLoopChecker final : public ConstStmtVisitor<TeamsLoopChecker> {
+  Sema &SemaRef;
+
+public:
+  bool teamsLoopCanBeParallelFor() const { return TeamsLoopCanBeParallelFor; }
+
+  // Is there a nested OpenMP loop bind(parallel)
+  void VisitOMPExecutableDirective(const OMPExecutableDirective *D) {
+    if (D->getDirectiveKind() == llvm::omp::Directive::OMPD_loop) {
+      if (const auto *C = D->getSingleClause<OMPBindClause>())
+        if (C->getBindKind() == OMPC_BIND_parallel) {
+          TeamsLoopCanBeParallelFor = false;
+          // No need to continue visiting any more
+          return;
+        }
+    }
+    for (const Stmt *Child : D->children())
+      if (Child)
+        Visit(Child);
+  }
+
+  void VisitCallExpr(const CallExpr *C) {
----------------
efriedma-quic wrote:

I'm concerned detecting CallExprs like this is going to have unexpected effects on the generated code: there are a lot of constructs that are written as "calls", but don't actually call anything external.  So minor changes to the user's code could have an unexpectedly large impact on code generation.

Is there some way we can do this transform in an LLVM IR optimization pass?

https://github.com/llvm/llvm-project/pull/87278