[WIP][OPENMP] Fix assertion error for using alignas with OpenMP directive

[cchen]

Summary

For the below case, Clang seems to regress (asserts) after https://github.com/llvm/llvm-project/commit/715f7a1bd058c64a39cc4773114dfb46ae8cc8a3

struct FOO
{
  static const int vec_align_bytes = 32;

  void foo()
  {
    alignas(vec_align_bytes) double a;
#pragma omp parallel // I've tried some other directives like master and target and all having same assertion failure
    a++;
  }

};

We can avoid the assertion failure by either removing alignas or
OpenMP directive in the above code. It is apparent the usage of alignas
is resonable so the issue might be stem from OpenMP code.
In addition, replacing OpenMP directive with some arbitrary Clang directive
can avoid the assertion failure.

[cchen]

Haven't added test yet since I'm not sure in which file should I add the test.

[jdoerfert]

We can avoid the assertion failure by either removing alignas or OpenMP directive in the above code.

What OpenMP directive?

Haven't added test yet since I'm not sure in which file should I add the test.

Let's go with something like clang/test/OpenMP/sema_alignas.cpp so we can start discussing the test at least.

[ABataev]

In D80590#2055937, @cchen wrote:

Haven't added test yet since I'm not sure in which file should I add the test.

Where is the directive?

[cchen]

In D80590#2056009, @ABataev wrote:

In D80590#2055937, @cchen wrote:

Haven't added test yet since I'm not sure in which file should I add the test.

Where is the directive?

I just added it, sorry for my carelessness.

[cchen]

Add test

[ABataev]

Investigated the bug. Seems to me, the bug is not related to OpenMP. The following code crashes the compiler too:

struct FOO
{
  static const int vec_align_bytes = 32;

  void foo()
  {
    alignas(vec_align_bytes) double a;
    ;
  }

};

Seems to me, the bug is in Actions.ClassifyName which does not clean up the list of the possibly odr-used decls. Or in the code, which tries to parse the align argument Parser::ParseAlignArgument.

[ABataev]

Must be fixed in 2f7269b6773de2750f9cd1417ef5f21cd6cf7a91

clang/lib/Sema/SemaOpenMP.cpp

if (const auto *D = dyn_cast_or_null<OMPExecutableDirective>(CurDirective)) {
     if (!CurContext->isDependentContext())
       checkAllocateClauses(*this, DSAStack, D->clauses());
     checkReductionClauses(*this, DSAStack, D->clauses());
   }
 
   DSAStack->pop();
   DiscardCleanupsInEvaluationContext();
   PopExpressionEvaluationContext();
+  CleanupVarDeclMarking();
 }
 
 static bool FinishOpenMPLinearClause(OMPLinearClause &Clause, DeclRefExpr *IV,
                                      Expr *NumIterations, Sema &SemaRef,
                                      Scope *S, DSAStackTy *Stack);
 
 namespace {
 

clang/test/OpenMP/sema_alignas.cpp

+// RUN: %clang_cc1 -verify=expected,omp45 -fopenmp -fopenmp-version=45 %s
+// RUN: %clang_cc1 -verify=expected,omp50 -fopenmp -fopenmp-version=50 %s
+
+// RUN: %clang_cc1 -verify=expected,omp45 -fopenmp-simd -fopenmp-version=45 %s
+// RUN: %clang_cc1 -verify=expected,omp50 -fopenmp-simd -fopenmp-version=50 %s
+
+// expected-no-diagnostics
+
+#ifndef HEADER
+#define HEADER
+
+struct FOO {
+  static const int vec_align_bytes = 32;
+
+  void foo() {
+    alignas(vec_align_bytes) double a;
+#pragma omp parallel
+    a++;
+
+#pragma omp master
+    a++;
+
+#pragma omp parallel private(a)
+    a++;
+  }
+};
+
+namespace {
+class FOO {
+public:
+  static const int vec_align_bytes = 32;
+
+  void foo() {
+    alignas(vec_align_bytes) double a;
+#pragma omp parallel
+    a++;
+
+#pragma omp master
+    a++;
+
+#pragma omp parallel private(a)
+    a++;
+  }
+};
+} // namespace
+
+template <typename T>
+struct FOO : public T {
+  static const T vec_align_bytes = 32;
+
+  void foo() {
+    alignas(vec_align_bytes) double a;
+#pragma omp parallel
+    a++;
+
+#pragma omp master
+    a++;
+
+#pragma omp parallel private(a)
+    a++;
+  }
+};
+
+namespace {
+namespace {
+template <typename T>
+class FOO : public T {
+public:
+  static const T vec_align_bytes = 32;
+
+  void foo() {
+    alignas(vec_align_bytes) double a;
+#pragma omp parallel
+    a++;
+
+#pragma omp master
+    a++;
+
+#pragma omp parallel private(a)
+    a++;
+  }
+};
+} // namespace
+} // namespace
+
+#endif